Exercise for preventing falls in older people living in the community

Abstract

Background

At least one‐third of community‐dwelling people over 65 years of age fall each year. Exercises that target balance, gait and muscle strength have been found to prevent falls in these people. An up‐to‐date synthesis of the evidence is important given the major long‐term consequences associated with falls and fall‐related injuries

Objectives

To assess the effects (benefits and harms) of exercise interventions for preventing falls in older people living in the community.

Search methods

We searched CENTRAL, MEDLINE, Embase, three other databases and two trial registers up to 2 May 2018, together with reference checking and contact with study authors to identify additional studies.

Selection criteria

We included randomised controlled trials (RCTs) evaluating the effects of any form of exercise as a single intervention on falls in people aged 60+ years living in the community. We excluded trials focused on particular conditions, such as stroke.

Data collection and analysis

We used standard methodological procedures expected by Cochrane. Our primary outcome was rate of falls.

Main results

We included 108 RCTs with 23,407 participants living in the community in 25 countries. There were nine cluster‐RCTs. On average, participants were 76 years old and 77% were women. Most trials had unclear or high risk of bias for one or more items. Results from four trials focusing on people who had been recently discharged from hospital and from comparisons of different exercises are not described here.

Exercise (all types) versus control

Eighty‐one trials (19,684 participants) compared exercise (all types) with control intervention (one not thought to reduce falls). Exercise reduces the rate of falls by 23% (rate ratio (RaR) 0.77, 95% confidence interval (CI) 0.71 to 0.83; 12,981 participants, 59 studies; high‐certainty evidence). Based on an illustrative risk of 850 falls in 1000 people followed over one year (data based on control group risk data from the 59 studies), this equates to 195 (95% CI 144 to 246) fewer falls in the exercise group. Exercise also reduces the number of people experiencing one or more falls by 15% (risk ratio (RR) 0.85, 95% CI 0.81 to 0.89; 13,518 participants, 63 studies; high‐certainty evidence). Based on an illustrative risk of 480 fallers in 1000 people followed over one year (data based on control group risk data from the 63 studies), this equates to 72 (95% CI 52 to 91) fewer fallers in the exercise group. Subgroup analyses showed no evidence of a difference in effect on both falls outcomes according to whether trials selected participants at increased risk of falling or not.

The findings for other outcomes are less certain, reflecting in part the relatively low number of studies and participants. Exercise may reduce the number of people experiencing one or more fall‐related fractures (RR 0.73, 95% CI 0.56 to 0.95; 4047 participants, 10 studies; low‐certainty evidence) and the number of people experiencing one or more falls requiring medical attention (RR 0.61, 95% CI 0.47 to 0.79; 1019 participants, 5 studies; low‐certainty evidence). The effect of exercise on the number of people who experience one or more falls requiring hospital admission is unclear (RR 0.78, 95% CI 0.51 to 1.18; 1705 participants, 2 studies, very low‐certainty evidence). Exercise may make little important difference to health‐related quality of life: conversion of the pooled result (standardised mean difference (SMD) ‐0.03, 95% CI ‐0.10 to 0.04; 3172 participants, 15 studies; low‐certainty evidence) to the EQ‐5D and SF‐36 scores showed the respective 95% CIs were much smaller than minimally important differences for both scales.

Adverse events were reported to some degree in 27 trials (6019 participants) but were monitored closely in both exercise and control groups in only one trial. Fourteen trials reported no adverse events. Aside from two serious adverse events (one pelvic stress fracture and one inguinal hernia surgery) reported in one trial, the remainder were non‐serious adverse events, primarily of a musculoskeletal nature. There was a median of three events (range 1 to 26) in the exercise groups.

Different exercise types versus control

Different forms of exercise had different impacts on falls (test for subgroup differences, rate of falls: P = 0.004, I² = 71%). Compared with control, balance and functional exercises reduce the rate of falls by 24% (RaR 0.76, 95% CI 0.70 to 0.81; 7920 participants, 39 studies; high‐certainty evidence) and the number of people experiencing one or more falls by 13% (RR 0.87, 95% CI 0.82 to 0.91; 8288 participants, 37 studies; high‐certainty evidence). Multiple types of exercise (most commonly balance and functional exercises plus resistance exercises) probably reduce the rate of falls by 34% (RaR 0.66, 95% CI 0.50 to 0.88; 1374 participants, 11 studies; moderate‐certainty evidence) and the number of people experiencing one or more falls by 22% (RR 0.78, 95% CI 0.64 to 0.96; 1623 participants, 17 studies; moderate‐certainty evidence). Tai Chi may reduce the rate of falls by 19% (RaR 0.81, 95% CI 0.67 to 0.99; 2655 participants, 7 studies; low‐certainty evidence) as well as reducing the number of people who experience falls by 20% (RR 0.80, 95% CI 0.70 to 0.91; 2677 participants, 8 studies; high‐certainty evidence). We are uncertain of the effects of programmes that are primarily resistance training, or dance or walking programmes on the rate of falls and the number of people who experience falls. No trials compared flexibility or endurance exercise versus control.

Authors' conclusions

Exercise programmes reduce the rate of falls and the number of people experiencing falls in older people living in the community (high‐certainty evidence). The effects of such exercise programmes are uncertain for other non‐falls outcomes. Where reported, adverse events were predominantly non‐serious.

Exercise programmes that reduce falls primarily involve balance and functional exercises, while programmes that probably reduce falls include multiple exercise categories (typically balance and functional exercises plus resistance exercises). Tai Chi may also prevent falls but we are uncertain of the effect of resistance exercise (without balance and functional exercises), dance, or walking on the rate of falls.

Plain language summary

Exercise for preventing falls in older people living in the community

Background

At least one‐third of community‐dwelling people over 65 years of age fall each year. Exercises that target balance, gait and muscle strength have previously been found to prevent falls in these people.

Review aim

To assess the effects (benefits and harms) of exercise interventions for preventing falls in older people living in the community.

Search date

We searched the healthcare literature for reports of randomised controlled trials relevant to this review up to 2 May 2018. In such studies, people are allocated at random to receive one of two or more interventions being compared in the study. Leaving group allocation to chance helps ensure the participant populations are similar in the intervention groups.

Study characteristics

This review includes 108 randomised controlled trials with 23,407 participants. These were carried out in 25 countries. On average, participants were 76 years old and 77% were women.

Certainty of the evidence

The majority of trials had unclear or high risk of bias, mainly reflecting lack of blinding of trial participants and personnel to the interventions. This could have influenced how the trial was conducted and outcome assessment. The certainty of the evidence for the overall effect of exercise on falls was high. Risk of fracture, hospitalisation, medical attention and adverse events were not well reported and, where reported, the evidence was low‐ to very low‐certainty. This leads to uncertainty regarding drawing conclusions from the evidence for these outcomes.

Key results

Eighty‐one trials compared exercise (all types) versus a control intervention that is not thought to reduce falls in people living in the community (who also had not recently been discharged from hospital). Exercise reduces the number of falls over time by around one‐quarter (23% reduction). By way of an example, these data indicate that if there were 850 falls in 1000 people followed over one year, exercise would result in 195 fewer falls. Exercise also reduces the number of people experiencing one or more falls (number of fallers) by around one‐sixth (15%) compared with control. For example, if there were 480 fallers who fell in 1000 people followed over one year, exercise would result in 72 fewer fallers. The effects on falls were similar whether the trials selected people who were at an increased risk of falling or not.

We found exercise that mainly involved balance and functional training reduced falls compared with an inactive control group. Programmes involving multiple types of exercise (most commonly balance and functional exercises plus resistance exercises) probably reduced falls, and Tai Chi may also reduce falls. We did not find enough evidence to determine the effects of exercise programmes classified as being mainly resistance exercises, dance, or walking programmes. We found no evidence to determine the effects of programmes that were mainly flexibility or endurance exercise.

There was considerably less evidence for non‐fall outcomes. Exercise may reduce the number of people experiencing fractures by over one‐quarter (27%) compared with control. However, more studies are needed to confirm this. Exercise may also reduce the risk of a fall requiring medical attention. We did not find enough evidence to determine the effects of exercise on the risk of a fall requiring hospital admission. Exercise may make very little difference to health‐related quality of life. The evidence for adverse events related to exercise was also limited. Where reported, adverse events were usually non‐serious events of a musculoskeletal nature; exceptionally one trial reported a pelvic stress fracture and a hernia.

Background

Description of the condition

At least one‐third of community‐dwelling people over 65 years of age fall each year (Campbell 1990; Tinetti 1988), and the rate of fall‐related injuries increases with age (Peel 2002). Falls can have serious consequences, such as fractures and head injuries (Peel 2002). Around 10% of falls result in a fracture (Campbell 1990; Tinetti 1988); fall‐associated fractures in older people are a significant source of morbidity and mortality (Burns 2016). Although most fall‐related injuries, such as bruising, lacerations and sprains, are less serious, they can still lead to pain, reduced function and substantial healthcare costs (Burns 2016).

Falls are associated with reduced quality of life (Stenhagen 2014), and can have psychological consequences: fear of falling and loss of confidence that can result in self‐restricted activity levels leading to a reduction in physical function and social interactions (Yardley 2002). Paradoxically, this restriction of activities may increase the risk of further falls by contributing to deterioration in physical abilities. Both injurious and non‐injurious falls can have these psychological and subsequent physical impacts.

Despite early attempts to achieve a consensus definition of a 'fall' (Anonymous 1987), many definitions still exist in the literature. It is particularly important for studies to use a clear, simple definition of a fall. An international researchers' consensus statement defines a fall as "an unexpected event in which the participant comes to rest on the ground, floor, or lower level" (Lamb 2005). The wording recommended when asking study participants is: "In the past month, have you had any fall including a slip or trip in which you lost your balance and landed on the floor or ground or lower level?" (Lamb 2005). 'Lower level' refers to a surface lower than the person's starting position so, for example, falling from a standing position to unintentionally sitting on a bed would be considered a fall.

In addition to the physical and psychological consequences for individuals and their families, falls can have important financial impacts on individuals, families and health and community care systems (Burns 2016). For example, falling is an independent predictor of admission to residential aged care facilities (Tinetti 1997).

Description of the intervention

Exercise is a physical activity that is planned, structured and repetitive and aims to improve or maintain physical fitness (Caspersen 1985). There is a wide range of possible types of exercise, such as strengthening exercise, balance and co‐ordination exercise and aerobic exercise. Exercise programmes often include one or more types of exercise. The Prevention of Falls Network Europe (ProFaNE) developed a taxonomy that classifies exercise type as: i) gait, balance, and functional (task) training; ii) strength/resistance (including power); iii) flexibility; iv) three‐dimensional (3D) exercise (e.g. Tai Chi, Qigong, dance); v) general physical activity; vi) endurance; and vii) other kinds of exercises (Lamb 2011). The taxonomy allows for more than one type of exercise to be delivered within a programme.

Formal exercise programmes are delivered by a wide range of individuals ranging from health professionals (such as physiotherapists, also known as physical therapists) and exercise professionals (such as trained fitness leaders) to trained volunteers. Exercise programmes may be supervised, unsupervised or involve a mixture of both.

This review considers all types of exercise and all delivery methods.

Exercise can also be delivered as part of a multiple component intervention, where people also receive one or more other fall or fracture prevention interventions, such as home‐hazard modification and vitamin D supplementation. The effects of multiple component interventions that include exercise are assessed in Hopewell 2018.

How the intervention might work

Many aspects of physical functioning deteriorate with increased age and inactivity. Impairments in muscle strength, balance control and gait are particularly strong risk factors for falls (Tinetti 1988). For example, those with poor leg extensor strength were found to be 43% more likely to fall at home than their stronger counterparts (Menant 2017). Systematic reviews have found that those with gait problems have twice the odds of falling than those without (Deandrea 2010), and that measures of balance and mobility such as the Berg Balance Scale, Timed Up and Go Test, and Five Times Sit‐to‐Stand Test can identify individuals at greater risk of future falls (Lusardi 2017).

Exercises that address these impairments are therefore likely to reduce the risk of falling. As Cochrane Reviews have now found that exercise improves both strength (Liu 2009), and balance (Howe 2011) in older people, exercise is likely to have a fall prevention effect through its impact on these key fall risk factors. A Cochrane Review found that exercise reduces the fear of falling (Kendrick 2014), which is also a strong predictor of falls.

A previous Cochrane Review found exercise as a single intervention, prevents falls (Gillespie 2012), and to be the most commonly tested single fall prevention intervention. Economic evaluations accompanying randomised trials have found exercise to be a cost‐effective fall‐prevention strategy (Davis 2010).

Exercise interventions have been found to be effective when delivered in a group‐based setting or on an individual basis. The optimal features of successful fall prevention exercise programmes are not yet clear, but programmes that are multicomponent (e.g. target both strength and balance; Gillespie 2012), and programmes that include balance training, appear to be particularly effective (Sherrington 2017).

Different approaches to exercise will have advantages and disadvantages in terms of cost, 'enjoyability', accessibility and impacts on various body systems and outcomes. These advantages and disadvantages are likely to vary between individuals and in different settings.

Exercise has the potential to lead to adverse events such as cardiovascular episodes and musculoskeletal injuries if not carefully prescribed and undertaken (Thompson 2013). Exercise may also increase the risk of falls, particularly in higher risk individuals. For example, exercise interventions aiming to improve balance and ultimately lessen the risk of falling, often involve a 'challenge' to balance that simultaneously puts the person at greater risk of falling (Sherrington 2017). The risk may be increased if an exercise participant becomes fatigued (due to deconditioning or as a result of comorbidities or medications) or are not encouraged to use support when needed (Skelton 2001). Trials and reviews should therefore record and report adverse events.

As the majority of fractures in older people involve falls, exercise has the potential to prevent fractures. Systematic reviews have suggested that exercise may prevent fractures (Gillespie 2012), and fall‐related injuries (Robertson 2002).

Why it is important to do this review

An update of the effects of exercise interventions on falls is warranted given the number of new trials published, the increasing number of older people living in the community and the major long‐term consequences associated with falls and fall‐related injuries to both the individual and to society.

It is also important to understand to what extent interventions designed to prevent falls will also prevent fall‐associated fractures, the need for medical attention and improve quality of life. Different exercise programmes may have different effects on falls and so careful analysis of the impact of different programmes is crucial to optimise the prescription of exercise interventions and inform public health promotion initiatives for healthy ageing. Additionally, looking for adverse events associated with the different exercise programmes, such as exercise‐related falls and muscle strains, is also important.

Objectives

To assess the effects (benefits and harms) of exercise interventions for preventing falls in older people living in the community.

Methods

Criteria for considering studies for this review

Types of studies

We included randomised controlled trials (RCTs), either individual or cluster randomised, evaluating the effects of exercise interventions on falls or fall‐related fractures in older people living in the community. We excluded trials that explicitly used methods of quasi‐randomisation (e.g. allocation to groups by alternation or date of birth).

Types of participants

We included trials if they specified an inclusion criterion of 60 years of age or over. Trials that included younger participants were included if the mean age minus one standard deviation was more than 60 years. We included trials where the majority of participants were living in the community, either at home or in places of residence that, on the whole, do not provide residential health‐related care or rehabilitative services; for example, retirement villages, or sheltered housing. Trials with mixed populations (community and higher dependency places of residence) were eligible for inclusion if data were provided for subgroups based on setting or the numbers in higher dependency residences were very few and balanced in the comparison groups.

We excluded studies that only included participants affected by particular clinical conditions that increase the risk of falls, such as stroke, Parkinson's disease, multiple sclerosis, dementia, hip fracture and severe visual impairment. Several of these topic areas are covered by other Cochrane Reviews (Canning 2015; Verheyden 2013). We acknowledge that some individuals with these (and other) health conditions may be included in studies of the general community; these we included.

As in our protocol, we also included trials recruiting participants in hospital if the majority were discharged to the community, where the majority of the intervention was delivered and falls recorded. As we considered such trials, whose participants were recently discharged from hospital, to be a distinct category we reported them separately.

Types of interventions

This review included all exercise interventions tested in trials that measured falls in older people. The intention was to include trials where exercise was a single intervention as opposed to a component of a broader intervention. We included trials where an additional low‐contact intervention (e.g. information on fall prevention) was given to one or both groups if we judged that the main purpose of the study was to investigate the role of exercise.

We classified exercise programmes on the basis of the primary exercise category and noted the presence of additional, secondary, exercise categories. Based on the Prevention of Falls Network Europe (ProFaNE) taxonomy (Lamb 2011), as shown in Appendix 1, we classified exercise programmes in the included trials as primarily involving the following exercise categories: i) gait, balance, co‐ordination and functional task training (referred to as 'balance and functional exercises' for simplicity); ii) strength/resistance training (including power training, using resistance so referred to as 'resistance exercises'); iii) flexibility; iv) three‐dimensional (3D) exercise (with separate Tai Chi and dance subcategories); iv) general physical activity (walking programmes); v) endurance; and vi) other kinds of exercises. We also formed another category for exercise programmes that included more than one of the above categories as the primary exercise category, e.g. a programme with 15 minutes of gait, balance, co‐ordination and functional task training followed by 15 minutes of strength/resistance training. We examined the descriptions of interventions used in individual trials and categorised the intervention accordingly. For example, some forms of yoga may have been categorised as flexibility exercise and others as 3D exercise.

We compared each of these types of exercise with control, comprising either 'usual care' (i.e. no change in usual activities) or a control intervention (i.e. an intervention that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls).

We first undertook an 'umbrella' comparison of 'exercise (all types) versus control', explored the impact of the use of an increased risk of falls as a trial inclusion criterion and the impact of participant age on the overall impact of exercise on falls, then set out the following comparisons.

1. Balance and functional exercises versus control.

2. Resistance exercises versus control.

3. Flexibility training versus control.

4. 3D (including Tai Chi, Qigong) exercise versus control.

5. 3D (dance) exercise versus control.

6. Walking programme versus control.

7. Endurance training versus control.

8. Other kinds of exercise versus control.

9. Multiple categories of exercise versus control (i.e. exercise programmes including more than one of the above categories versus control).

We also planned to undertake the following secondary comparisons of different exercise programmes.

1. Different types of exercise, based on the above categories.

2. Different modes of delivery (e.g. group versus individual) of the same type of exercise.

3. Different doses (e.g. higher intensity versus lower intensity) of the same type of exercise.

Types of outcome measures

Primary outcomes

1. Rate of falls (falls per person‐year)

Secondary outcomes

1. Number of people who experienced one or more falls (risk of falling)

2. Number of people who experienced one or more fall‐related fractures

3. Number of people who experienced one or more falls that resulted in hospital admission (newly listed outcome April 2018)

4. Number of people who experienced one or more falls that required medical attention

5. Health‐related quality of life, measured using validated scale, e.g. EQ‐5D or similar (newly listed outcome April 2018)

6. Number of people who experienced one or more adverse events (see below)

We chose 'rate of falls' as the single primary outcome for ease of interpretation of the results of the review. Furthermore, the rate of falls is likely to be more sensitive to change than the proportion of fallers, especially in samples with high fall rates. As falls are count data, dichotomisation to falling versus not falling represents a loss of information. Therefore, many trials use the rate of falls as their primary outcome and use negative binomial regression to compare the rates between intervention and control groups, as recommended in Robertson 2005.

Adverse events needed to be monitored closely in all groups using the same methods over the entire study period to be included in the analysis.

Other outcomes

We recorded and reported mortality data, distinguishing where possible, between those who were lost to the trials because they had died and those whose death was explicitly linked to trial participation.

We recorded and reported data regarding intervention adherence, cost and cost‐effectiveness, where available.

Timing of outcome measurement

The primary outcome included one time point from each study. For studies with outcomes measured at multiple time points, we used the closest to 18 months in the primary analysis. We included a separate longer‐term outcome for studies with follow‐up at more than 18 months after randomisation. To maximise the use of available information, we also included studies with just one time point that was longer than 18 months in the primary analysis.

Search methods for identification of studies

Electronic searches

Our search extended the searches performed up to February 2012 in Gillespie 2012. We searched: the Cochrane Bone, Joint and Muscle Trauma Group Specialised Register (February 2012 to 2 May 2018); the Cochrane Central Register of Controlled Trials (CENTRAL) (Cochrane Register of Studies Online) (2012 Issue 2 to 2018 Issue 5); MEDLINE (including Epub Ahead of Print, In‐Process & Other Non‐Indexed Citations and MEDLINE Daily) (January 2012 to 30 April 2018); Embase (March 2012 to 2018 Week 18); the Cumulative Index to Nursing and Allied Health Literature (CINAHL) (February 2012 to 2 May 2018); and the Physiotherapy Evidence Database (PEDro) (2012 to 2 May 2018), using tailored search strategies. We did not apply any language restrictions.

In MEDLINE, we combined subject‐specific search terms with the sensitivity‐ and precision‐maximising version of the Cochrane Highly Sensitive Search Strategy for identifying randomised trials (Lefebvre 2011). The search strategies for CENTRAL, MEDLINE, Embase, CINAHL and PEDro are shown in Appendix 2.

We also searched the World Health Organisation International Clinical Trials Registry Platform (WHO ICTRP) and ClinicalTrials.gov for ongoing and recently completed trials (May 2018) (see Appendix 2).

Searching other resources

We checked reference lists of other systematic reviews as well as contacting researchers in the field to assist in the identification of ongoing and recently completed trials.

Data collection and analysis

The intended methodology for data collection and analysis was described in our published protocol (Sherrington 2016), which was based on the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011).

Selection of studies

Pairs of review authors (CS, AT, NJF, ZAM) screened the title, abstract and descriptors of identified studies for possible inclusion. From the full text, two review authors (CS, AT, NJF, ZAM) independently assessed potentially eligible trials for inclusion and resolved any disagreement through discussion. We contacted authors for additional information as necessary.

Data extraction and management

Pairs of review authors (CS, AT, NJF, ZAM, GW) independently extracted data using a pretested data extraction form (based on the one used in Gillespie 2012). We extracted data from both newly included trials and those included in Gillespie 2012. For the latter trials, however, we primarily extracted information and data for additional outcomes that were not collected previously for Gillespie 2012. Disagreement was resolved by consensus or third party adjudication. Review authors were not blinded to authors and sources. Review authors did not assess their own trials.

We used the standardised data extraction form to record the following items.

1. General information: review author's name; date of data extraction; study ID; first author of study; author's contact address (if available); citation of paper; and trial objectives.

2. Trial details: trial design; location; setting; sample size; inclusion and exclusion criteria (with particular note of whether there was exclusion for cognitive impairment); comparability of groups; length of follow‐up; stratification; stopping rules; and funding source.

3. 'Risk of bias' assessment and justification for this judgement: sequence generation; allocation concealment; blinding (participants, personnel, outcome assessors); incomplete outcome data; selective outcome reporting; and other bias (recall bias).

4. Characteristics of participants: age; gender; ethnicity; the number randomised, analysed and lost to follow‐up; and dropouts in each arm (with reasons).

5. Interventions: experimental and control interventions; details of exercise programme (duration, frequency, intensity and individual‐ or group‐based delivery, level of supervision); timing of intervention; uptake of intervention (acceptance of exercise intervention), whether studies assessed adherence (compliance) with interventions and associated data (e.g. number of sessions attended); and additional co‐interventions (such as motivational strategies, additional information or support given to participants).

6. Outcomes measured: rate of falls; number of people experiencing one or more falls; number of people who experienced one or more fall‐related fractures; number of people who experienced one or more falls requiring medical attention; and number of people who experienced adverse events.

7. Other details: cost and cost‐effectiveness information related to fall outcomes.

We retrieved data from both full‐text and abstract reports of studies. Where these sources did not provide sufficient information, we contacted study authors for additional details. We also used data sourced from personal communication reported by Gillespie 2012.

In response to feedback on an earlier draft of this review we extended our data extraction to extract data on the number of people who experienced one or more falls resulting in hospital admission, mortality and health‐related quality of life (Differences between protocol and review).

We recorded and reported data on fracture, hospitalisation, medical attention, and health‐related quality of life only where separate data were available by intervention group.

Assessment of risk of bias in included studies

Pairs of two review authors (CS, AT, NJF, ZAM, GW) independently assessed risk of bias using Cochrane's 'Risk of bias' tool as described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). Review authors were not blinded to authors and sources. Review authors did not assess their own trials. Disagreement was resolved by consensus or third party adjudication (CS).

As outlined in Appendix 3 we assessed the following domains: random sequence generation (selection bias); allocation concealment (selection bias); blinding of participants and personnel (performance bias); blinding of outcome assessment (detection bias); incomplete outcome data (attrition bias); and selective outcome reporting bias. We also assessed bias in the recall of falls due to less reliable methods of ascertainment (Hannan 2010). We rated risk of bias as either low, high or unclear for each domain.

Specifically for trials using cluster‐randomisation, we considered the risk of additional bias relating to recruitment, baseline imbalance, loss of clusters, incorrect analysis and comparability with individually‐randomised trials, as described in Chapter 16 of the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011).

Measures of treatment effect

We reported the treatment effects for rate of falls as rate ratios (RaRs) with 95% confidence intervals (CIs). For the number of fallers, number of participants experiencing fall‐related fractures, fall‐related hospital admission, falls that required medical attention and adverse events, we reported risk ratios (RRs) and 95% CIs.

The rate of falls is the total number of falls per unit of person‐time that falls were monitored (e.g. falls per person‐year). The RaR compares the rate of falls in any two groups during each trial. We used a RaR (for example, incidence RaR or hazard ratio (HR) for all falls) with 95% CI if these were reported in the paper. If both adjusted and unadjusted RaRs were reported, we used the unadjusted estimate unless the adjustment was for clustering. If a RaR was not reported, but appropriate raw data were available, we used Excel to calculate a RaR and 95% CI. We used the reported rate of falls (falls per person‐year) in each group and the total number of falls for participants contributing data, or we calculated the rate of falls in each group from the total number of falls and the actual total length of time falls were monitored (person‐years) for participants contributing data. In cases where data were only available for people who had completed the study, or where the trial authors had stated there were no losses to follow‐up, we assumed that these participants had been followed up for the maximum possible period.

The risk ratio (RR) compares the number of people who fell once or more (fallers) between groups. We used a reported estimate of the RR, HR for first fall, or odds ratio (OR)) and 95% CI if available. If both adjusted and unadjusted estimates were reported we used the unadjusted estimate, unless the adjustment was for clustering. If an OR was reported, or an effect estimate and 95% CI was not, and appropriate data were available, we calculated a RR and 95% CI using the 'csi' command in Stata. For the calculations, we used the number of participants contributing data in each group, if this was known; if not reported, we used the number randomised to each group. The same approach was used for the number of people experiencing fractures, falls requiring medical attention and adverse events. Data regarding the number of people in each group experiencing the additional variables of falls resulting in hospitalisation and death were entered into Review Manager 5 directly (Review Manager 2014).

For continuous outcomes (health‐related quality of life), we presented the mean difference (MD) with 95% CIs where the same outcome measure was used, or standardised mean difference (SMD) with 95% CIs for outcomes measured using different scales. Final values, which were used in preference to change scores, were always available where these outcomes were reported.

Unit of analysis issues

For trials which were cluster randomised, for example by medical practice, we performed adjustments for clustering, as described in Higgins 2011, if this was not done in the published report. We used an intraclass correlation coefficient (ICC) of 0.01 as reported in Smeeth 2002. We ignored the possibility of a clustering effect in trials that randomised by household. We anticipated that trials would be unlikely to report details of clustering by household and that the clustering effect by household would be very small (if any).

The pooled exercise versus control comparisons necessitated the inclusion of more than one pair‐wise comparison (intervention versus control) from the same trial in the same meta‐analysis. Where multiple comparisons from the same trial were included in the same meta‐analysis the standard errors were inflated by 25% and the number of control participants shown in the analyses was 'shared' between different comparisons by dividing by the number of intervention groups in the same analysis. For example, if a trial had 100 participants in a control group, 100 participants in a resistance training group, and 100 participants in a balance training group, the standard errors in the resistance versus control and balance versus control comparisons would be inflated by 25% and the number of control participants would be shown as 50 in both the resistance versus control and balance versus control comparisons.

We did not include outcomes collected at different time points in the same trial in the same analysis.

Dealing with missing data

Some missing data are inevitable in studies of older people, given the increased risk of ill health and death, and the length of delivery of the intervention in fall prevention trials. We attempted to contact study investigators for any key missing or unclear data or information in their trial; clarification on outcome data was only sought for number of falls and number of people who experienced falls. We undertook sensitivity analyses excluding trials with more than 20% loss to follow‐up or where the loss to follow‐up was unclear.

Assessment of heterogeneity

Where we considered study interventions to be sufficiently similar to be combined in meta‐analyses, we assessed heterogeneity of treatment effects by visual inspection of forest plots and by using the Chi² test (with a significance level at P < 0.10) and the I² statistic. We based our interpretation of the I² results on that suggested by Higgins 2011: 0% to 40% might not be important; 30% to 60% may represent moderate heterogeneity; 50% to 90% may represent substantial heterogeneity; and 75% to 100% may represent very substantial (‘considerable’) heterogeneity.

Assessment of reporting biases

We constructed and visually inspected funnel plots for outcomes that included more than 10 data points.

Data synthesis

For our primary comparison, we pooled data from all relevant trials without stratification. We originally planned to present the umbrella comparison of exercise versus control subgrouped by the main exercise categories (Sherrington 2016). This change was made in response to editorial input and the request for additional subgroup and sensitivity analyses in a commissioning brief relating to the National Institute for Health and Care Excellence (NICE) guideline CG161 (NICE 2013).

We presented separate analyses for studies that recruited people in hospitals and delivered interventions after discharge as we considered these were a distinct population compared with general community‐dwelling older adults.

We grouped similar exercise interventions using the fall prevention classification system (taxonomy) developed by the Prevention of Falls Network Europe (ProFaNE) (Lamb 2011). Full details are available in Appendix 1 and the ProFaNE Taxonomy Manual.

When considered appropriate, we pooled results of comparable studies using random‐effects models. We used 95% CIs throughout. We planned not to pool data where there was considerable heterogeneity (I² ≥ 75%) that could not be explained by the diversity of methodological or clinical features among trials.

When considered appropriate, we pooled data using the generic inverse variance method in Review Manager 5 (Review Manager 2014). This method enables pooling of the adjusted and unadjusted treatment effect estimates (rate ratios or risk ratios) reported in the individual studies or which can be calculated from data presented in the published article (see Measures of treatment effect). The generic inverse variance option in Review Manager 5 requires entering the natural logarithm of the rate ratio or risk ratio and its standard error for each trial; we calculated these in Excel. For continuous outcomes (health‐related quality of life), we presented MDs with 95% CIs where the same outcome measure was used, or SMDs with 95% CIs for outcomes measured using different scales.

Where it was inappropriate to pool data, we present trial‐level data in the analyses and tables for illustrative purposes.

The statistician was not blind to study or group.

Subgroup analysis and investigation of heterogeneity

We undertook subgroup analyses for the fall and fracture outcomes for the pooled (all‐exercise types) versus control analyses to compare the effect of exercise on falls and fractures in trials that did and did not use an increased risk of falls as an inclusion criterion. In response to a request (Differences between protocol and review) to explore the potential effects of stratification by age (based on a threshold of 75 years), we undertook subgroup analyses for the falls and fracture outcomes for the pooled (all‐exercise types) versus control analyses. We compared the effects on falls outcomes in trials with predominantly older populations (defined by inclusion criteria 75 years or above, lower range limit more than 75 years, or mean age minus one standard deviation more than 75 years) and those with predominantly younger populations.

Prompted by feedback at editorial review, we extended the following subgroup analyses (originally established for different exercise categories) to the all‐exercise types versus control for fall outcomes: a) individual versus group‐based exercise; and b) exercise delivered by people with different qualifications (e.g. health professionals versus trained fitness leaders).

We presented separate analyses stratified by the different ProFaNE exercise intervention categories outlined above, and performed subgroup analyses for the fall and fracture outcomes. We then used subgroup analyses to explore effects within the different ProFaNE exercise intervention categories. When there were at least 10 trials in a comparison, we carried out subgroup analyses to compare effects in trials of: a) higher versus lower falls risk at enrolment (i.e. trials with participants selected for inclusion based on history of falling or other specific risk factors for falling versus trials with unselected participants); b) individual versus group‐based exercise; and c) exercise delivered by people with different qualifications.

We used the test for subgroup differences available in Review Manager 2014 to determine whether there was evidence for a difference in treatment effect between subgroups.

Sensitivity analysis

We carried out 10 sensitivity analyses to explore the stability of the results.

Sensitivity analysis 1 (participant age)

In response to a specific request (Differences between protocol and review) to explore the potential effects of changing the age threshold from 60 to 65 years for inclusion into the review, we set out a series of sensitivity analyses to explore the effects of removing trials that would have been excluded from the review if a 65 year or older inclusion threshold had been applied.

Sensitivity analyses 2‐5 (risk of bias in included trials)

To assist with the GRADE rating we undertook sensitivity analyses for all outcomes in the 'Summary of findings' table by removing trials with a high risk of bias in any item.

To explore the possible impact of risk of bias on the primary pooled estimates of treatment effect, we examined the effects of the following.

1. Inclusion of trials at high or unclear risk of selection bias from inadequate concealment of allocation.

2. Inclusion of trials at high or unclear risk of detection bias from inadequate blinding of outcome assessors.

3. Inclusion of trials at high or unclear risk of attrition bias from incomplete outcome data.

Sensitivity analyses 6‐7 (meta‐analysis decisions)

We also examined the impact on the results of the removal of the cluster‐randomised trials and the use of fixed‐effect rather than random‐effects models for data pooling.

Sensitivity analysis 8 (multiple exercise category components)

In order to assist in the interpretation of the results of the type of exercise subgroup 'multiple categories of exercise' comparisons, we undertook a sensitivity analyses for both falls outcomes which only included trials that were coded as having the two primary components balance/functional exercises and resistance exercises.

Sensitivity analyses 9a and 9b (different exercise type coding)

To explore the possible impact of how we classified exercise interventions, we examined the effects of the following for both falls outcomes.

1. Classification of interventions based on the Otago Exercise Program as multiple categories of exercise.

2. Classification of any intervention that included balance and functional exercises plus strength exercises as multiple categories of exercise.

Assessing the certainty of evidence and 'Summary of findings' tables

We used the GRADE approach to assess the quality of evidence related to all outcomes listed in the Types of outcome measures (Schünemann 2017). Using GRADEpro GDT (GRADEPro GDT 2015), we assessed the certainty of the evidence as ‘high’, ‘moderate’, ‘low’ or ‘very low’ depending on the presence and extent of five factors: risk of bias; inconsistency of effect; indirectness; imprecision; and publication bias. We prepared 'Summary of finding' tables featuring the seven listed outcomes for the umbrella comparison (exercise (all types) versus control) and for the rate of falls, risk of falling, fall‐related fractures and adverse events for the primary exercise categories versus control comparisons, where data were available (Types of interventions). We used standardised qualitative statements to describe the different combinations of effect size and the certainty of evidence (Cochrane Norway 2017).

Results

Description of studies

Results of the search

A total of 8007 records were downloaded from the following databases: Cochrane Bone, Joint and Muscle Trauma Group Specialised Register (7), CENTRAL (1650), MEDLINE (1601), Embase (2998), CINAHL (1104), PEDro (139), the WHO ICTRP (317), and ClinicalTrials.gov (191). We identified 359 studies from a prior Cochrane Review (Gillespie 2012), and other systematic reviews. We also found one study after the search process in September 2018 (Li 2018b)

Removal of duplicates and spurious records resulted in 4006 references. Upon screening of these, we excluded 3541 records and we obtained copies of 465 papers for consideration. A screening of these led to the removal of a further 230 records. The final round of study selection based on 235 reports resulted in the inclusion of 108 studies (194 reports), the exclusion of 21 studies (23 reports) (see Characteristics of excluded studies) and identification of 16 ongoing studies (Ongoing studies). Two further studies await classification (Jagdhane 2016; Li 2018b).

We contacted authors of two studies to request additional details to assess eligibility, and received responses from both studies; we included Hamrick 2017 and excluded Hinrichs 2016.

A flow diagram summarising the study selection process is shown in Figure 1.

1.

Study flow diagram.

Included studies

This review includes 108 trials with 23,407 participants. Details are provided in the Characteristics of included studies and are briefly summarised below. Due to the size of the review, not all links to references have been inserted in the following text but can be viewed in Appendix 4. Characteristics of the included studies are summarised in Table 8 and Table 9.

1. Study design, length of follow‐up, setting and trial size.

Study IDa	Study design	No. arms (clusters)	Length of follow‐up (months)	Setting	No. randomised	No. analysedb	% lost to follow‐up
Gait, balance, and functional training
Almeida 2013	Parallel	3	4	Brazil	119	76	36%
Arantes 2015	Parallel	2	3	Brazil	30	28	7%
Arkkukangas 2015	Parallel	2	3	Sweden	45	40	11%
Barnett 2003	Parallel	2	12	Australia	163	150	8%
Boongrid 2017	Parallel	2	12	Thailand	439	437	0%
Campbell 1997	Parallel	2	24	New Zealand	233	233	0%
Clegg 2014	Parallel	2	3	United Kingdom	84	70	17%
Clemson 2010	Parallel	2	6	Australia	34	34	0%
Clemson 2012 (Life Program)	Parallel	3	12	Australia	317	317	0%
Cornillon 2002	Parallel	2	12	France	303	303	0%
Dadgari 2016	Cluster	2 (25)	6	Iran	551	317	42%
Dangour 2011	Cluster	2 (28)	24	Chile	984	619	37%
Day 2002	Parallel	2	18	Australia	272	272	0%
Duque 2013	Parallel	2	9	Australia	60	60	0%
El‐Khoury 2015	Parallel	2	24	France	706	706	0%
Gschwind 2015	Parallel	2	6	Germany, Spain, Australia	153	136	11%
Halvarsson 2013	Parallel	2	15	Sweden	59	48	19%
Halvarsson 2016 (balance group)	Parallel	3	3	Sweden	96	69	28%
Hamrick 2017	Parallel	2	6	USA	43	38	12%
Hirase 2015	Parallel	3	4	Japan	93	86	8%
Iliffe 2015 (FaME and OEP groups)	Cluster	3 (42)	18	United Kingdom	1254	709	43%
Iwamoto 2009	Parallel	2	5	Japan	68	67	1%
Karinkanta 2007 (balance group)	Parallel	4	12	Finland	149	144	3%
Kerse 2010	Parallel	2	12	New Zealand	193	193	0%
Korpelainen 2006	Parallel	2	30	Finland	160	160	0%
Kovacs 2013	Parallel	2	12	Hungary	76	72	5%
Lin 2007	Parallel	2	6	Taiwan	100	100	0%
Liu‐Ambrose 2008	Parallel	2	12	Canada	74	59	30%
Liu‐Ambrose 2004 (agility group)	Parallel	3	6	Canada	104	98	6%
Lord 1995	Parallel	2	12	Australia	197	169	14%
Lord 2003	Cluster	2 (20)	12	Australia	551	508	8%
Luukinen 2007	Parallel	2	16	Finland	486	437	10%
Madureira 2007	Parallel	2	12	Brazil	66	60	9%
McMurdo 1997	Parallel	2	24	Scotland	118	92	22%
Miko 2017	Parallel	2	12	Hungary	100	97	3%
Morgan 2004	Parallel	2	12	USA	294	229	22%
Nitz 2004	Parallel	2	6	Australia	73	45	38%
Reinsch 1992	Cluster	2 (16)	12	USA	230	230	0%
Robertson 2001a	Parallel	2	12	New Zealand	240	240	0%
Sakamoto 2013	Parallel	2	6	Japan	1365	865	37%
Sales 2017	Parallel	2	12	Australia	66	48	27%
Siegrist 2016	Cluster	2 (40)	12	Germany	378	378	0%
Skelton 2005	Parallel	2	9	United Kingdom	81	81	0%
Smulders 2010	Parallel	2	12	Netherlands	96	92	4%
Trombetti 2011	Parallel	2	6	Switzerland	134	134	0%
Weerdesteyn 2006	Parallel	2	7	Netherlands	58	58	0%
Wolf 1996 (balance group)	Parallel	3	8	USA	200	200	0%
Yang 2012	Parallel	2	6	Australia	165	121	27%
Strength/resistance (including power)
Ansai 2015 (resistance group)	Parallel	3	4	Brazil	69	68	1%
Carter 2002	Parallel	2	5	Canada	93	80	14%
Fiatarone 1997	Parallel	2	4	USA	34	0	N/A
Grahn Kronhed 2009	Parallel	2	12	Sweden	65	65	0%
Karinkanta 2007 (resistance group)	Parallel	4	12	Finland	149	144	3%
Latham 2003c	Parallel	2	6	New Zealand and Australia	243	222	9%
Liu‐Ambrose 2004 (resistance group)	Parallel	3	6	Canada	104	98	6%
Vogler 2009 (seated group)c	Parallel	3	12	Australia	180	171	5%
Woo 2007 (resistance group)	Parallel	3	12	China	180	176	33%
3D
Day 2015	Parallel	2	12	Australia	503	409	19%
Huang 2010	Cluster	2 (4)	5	Taiwan	115	78	32%
Li 2005	Parallel	2	6	USA	256	188	27%
Logghe 2009	Parallel	2	12	Netherlands	269	269	0%
Merom 2016	Cluster	2 (23)	12	Australia	530	522	2%
Taylor 2012	Parallel	2	17	New Zealand	684	684	0%
Voukelatos 2007	Parallel	2	6	Australia	702	684	3%
Wolf 2003	Cluster	2 (20)	11	USA	311	286	8%
Wolf 1996 (Tai Chi group)	Parallel	3	8	USA	200	200	0%
Woo 2007 (Tai Chi group)	Parallel	3	12	China	180	176	3%
Wu 2010 (com‐ex group)	Parallel	3	4	USA	64	64	0%
Wu 2010 (home‐ex group)	Parallel	3	4	USA	64	64	0%
Wu 2010 (tel‐ex group)	Parallel	3	4	USA	64	64	0%
General physical activity
Ebrahim 1997	Parallel	2	24	United Kingdom	165	102	38%
Resnick 2002	Parallel	2	6	USA	20	17	15%
Voukelatos 2015	Parallel	2	12	Australia	386	339	12%
Multiple primary exercise categories
Ansai 2015 (multicomponent group)d	Parallel	3	4	Brazil	69	68	1%
Beyer 2007d	Parallel	2	12	Denmark	65	53	18%
Brown 2002d	Parallel	2	14	Australia	99	71	28%
Buchner 1997	Parallel	2	25	USA	105	100	5%
Bunout 2005d	Parallel	2	12	Chile	298	241	19%
Cerny 1998d	Parallel	2	6	USA	28	28	0%
Clemson 2012 (structured group)d	Parallel	3	12	Australia	317	317	0%
Gill 2016d	Parallel	2	42	USA	1635	1635	0%
Haines 2009c,d	Parallel	2	6	Australia	53	53	0%
Halvarsson 2016 (balance and physical activity group)	Parallel	3	3	Sweden	96	69	28%
Hauer 2001d	Parallel	2	6	Germany	57	56	2%
Irez 2011d	Parallel	2	3	Turkey	60	60	0%
Kamide 2009d	Parallel	2	6	Japan	57	43	25%
Karinkanta 2007 (resistance and balance groups)d	Parallel	4	12	Finland	149	144	3%
Kim 2014d	Parallel	2	12	Japan	105	103	2%
Lehtola 2000	Parallel	2	10	Finland	131	131	0%
Means 2005d	Parallel	2	6	USA	338	238	30%
Ng 2015d	Parallel	2	12	Singapore	98	92	6%
Park 2008	Parallel	2	11	Korea	50	45	10%
Rubenstein 2000	Parallel	2	3	USA	59	59	0%
Sherrington 2014c,d	Parallel	2	12	Australia	340	340	0%
Suzuki 2004d	Parallel	2	20	Japan	52	44	15%
Uusi‐Rasi 2015d	Parallel	2	24	Finland	205	186	9%
Vogler 2009 (weightbearing group)c	Parallel	3	12	Australia	180	171	5%
Exercise versus exercise
Ballard 2004	Parallel	2	16	USA	40	39	3%
Barker 2016	Parallel	2	6	Australia	53	44	17%
Clemson 2012	Parallel	3	12	Australia	317	286	10%
Davis 2011	Parallel	3	9	Canada	155	155	0%
Freiberger 2007	Parallel	2	24	Germany	134	127	5%
Helbostad 2004	Parallel	2	12	Norway	77	68	12%
Hwang 2016	Parallel	2	18	Taiwan	456	334	27%
Iliffe 2015	Cluster	3 (42)	18	United Kingdom	1254	709	43%
Karinkanta 2007	Parallel	4	12	Finland	149	144	3%
Kemmler 2010	Parallel	2	18	Germany	246	227	8%
Kwok 2016	Parallel	2	12	Singapore	80	80	0%
Kyrdalen 2014	Parallel	2	3	Norway	125	94	25%
LaStayo 2017	Parallel	2	12	USA	134	112	16%
Liston 2014	Parallel	2	6	United Kingdom	21	15	29%
Liu‐Ambrose 2004	Parallel	3	6	Canada	104	98	6%
Lurie 2013	Parallel	2	3	USA	64	59	8%
Mirelman 2016	Parallel	2	N/A	Belgium, Israel, Italy, Netherlands, and United Kingdom	152	0	N/A
Morone 2016	Parallel	2	3	Italy	38	0	N/A
Morrison 2018	Parallel	2	3	USA	65	46	29%
Okubo 2016	Parallel	2	16	Japan	105	90	14%
Shigematsu 2008	Parallel	2	8	Japan	68	68	0%
Steadman 2003	Parallel	2	1	United Kingdom	199	133	33%
Taylor 2012	Parallel	2	17	New Zealand	684	684	0%
Verrusio 2017	Parallel	2	12	Italy	150	147	2%
Wolf 1996	Parallel	3	8	USA	200	200	0%
Yamada 2010	Parallel	2	12	Japan	60	58	3%
Yamada 2012	Parallel	2	12	Japan	157	145	8%
Yamada 2013	Parallel	2	12	Japan	264	230	13%

^a Categorised by primary exercise category.
 ^b Number analysed for fall data.
 ^c Post‐hospital discharge study.
 ^d Indicates the primary interventions include gait, balance, and functional training and strength/resistance training.

2. Key characteristics of participants and intervention approach.

Study IDa	Age (mean)	% Women	High risk of falls	Duration of intervention (weeks)	Intervention delivered by health professional	Group exercise	Intervention progressed
Gait, balance, and functional training
Almeida 2013	79	83%	Yes	16	Yes	Yes	NR
Arantes 2015	73	100%	Yes	12	Yes	Yes	Yes
Arkkukangas 2015	83	71%	No	12	Yes	No	Yes
Barnett 2003	75	67%	Yes	52	No	Yes	Yes
Boongrid 2017	74	83%	Yes	52	Yes	No	Yes
Campbell 1997	84	100%	Yes	52	Yes	No	Yes
Clegg 2014	79	71%	Yes	12	Yes	No	Yes
Clemson 2010	81	47%	Yes	26	Yes	No	Yes
Clemson 2012 (Life Program)	83	55%	Yes	52	Yes	No	Yes
Cornillon 2002	71	83%	No	52	No	No	No
Dadgari 2016	70	49%	Yes	24	No	No	Yes
Dangour 2011	66	68%	No	104	No	Yes	Yes
Day 2002	76	59%	No	18	No	Yes	No
Duque 2013	77	62%	Yes	6	Yes	No	Yes
El‐Khoury 2015	79	100%	Yes	104	No	Yes	Yes
Gschwind 2015	75	61%	No	16	No	No	Yes
Halvarsson 2013	77	71%	Yes	12	Yes	Yes	Yes
Halvarsson 2016 (balance group)	76	98%	Yes	12	Yes	Yes	Yes
Hamrick 2017	70	79%	No	8	No	Yes	Yes
Hirase 2015	82	70%	Yes	16	Yes	No	No
Iliffe 2015	73	62%	No	24	No	OEP: no; FaME: Yes	Yes
Iwamoto 2009	76	90%	No	20	No	Yes	No
Karinkanta 2007 (balance group)	73	100%	No	52	No	Yes	No
Kerse 2010	81	58%	No	26	No	No	Yes
Korpelainen 2006	73	100%	No	130	Yes	Yes	Yes
Kovacs 2013	69	100%	No	25	Yes	Yes	Yes
Lin 2007	77	51%	Yes	16	Yes	No	Yes
Liu‐Ambrose 2004 (agility group)	79	100%	No	25	No	Yes	No
Liu‐Ambrose 2008	83	71%	Yes	26	Yes	No	Yes
Lord 1995	71	100%	No	52	No	Yes	No
Lord 2003	80	86%	No	52	No	Yes	No
Luukinen 2007	88	79%	Yes	70	Yes	No	Yes
Madureira 2007	74	100%	No	40	Yes	Yes	No
McMurdo 1997	65	100%	No	60	No	Yes	No
Miko 2017	69	100%	No	52	Yes	Yes	Yes
Morgan 2004	81	71%	Yes	8	Yes	Yes	Yes
Nitz 2004	76	92%	Yes	10	Yes	Yes	No
Reinsch 1992	74	80%	No	52	No	Yes	No
Robertson 2001a	84	68%	No	52	Yes	No	Yes
Sakamoto 2013	80	82%	Yes	26	No	No	Yes
Sales 2017	73	69%	Yes	18	Both	Yes	Yes
Siegrist 2016	78	75%	Yes	16	Yes	Yes	Yes
Skelton 2005	72	100%	Yes	36	No	Yes	Yes
Smulders 2010	71	94%	Yes	5.5	Yes	Yes	Yes
Trombetti 2011	76	96%	Yes	26	No	Yes	Yes
Weerdesteyn 2006	74	77%	Yes	5	No	Yes	Yes
Wolf 1996 (balance group)	76	81%	No	15	No	No	Yes
Yang 2012	81	44%	Yes	26	Yes	No	No
Strength/resistance (including power)
Ansai 2015 (resistance group)	82	68%	Yes	16	No	Yes	Yes
Carter 2002	69	100%	No	20	No	Yes	No
Fiatarone 1997	82	94%	Yes	16	No	No	No
Grahn Kronhed 2009	71	100%	No	16	Yes	Yes	Yes
Karinkanta 2007 (resistance group)	73	100%	No	52	No	Yes	Yes
Latham 2003b	80	53%	Yes	10	Yes	No	Yes
Liu‐Ambrose 2004 (resistance group)	79	100%	No	25	No	Yes	Yes
Vogler 2009 (seated group)	80	79%	Yes	12	Yes	No	Yes
Woo 2007 (resistance group)	69	50%	No	52	No	Yes	No
3D
Day 2015	77	70%	Yes	48	No	Yes	Yes
Huang 2010	71	30%	No	22	No	Yes	No
Li 2005	77	70%	No	26	No	Yes	No
Logghe 2009	77	71%	Yes	13	No	Yes	No
Merom 2016		85%	No	52	No	Yes	Yes
Taylor 2012	75	73%	Yes	20	No	Yes	No
Voukelatos 2007	69	84%	No	16	No	Yes	No
Wolf 1996 (Tai Chi group)	76	81%	No	15	No	Yes	Yes
Wolf 2003	81	94%	Yes	48	No	Yes	Yes
Woo 2007 (Tai Chi group)	69	50%	No	52	No	Yes	No
Wu 2010 (com‐ex group)	75	84%	Yes	15	No	Yes	No
Wu 2010 (home‐ex group)	75	84%	Yes	15	No	No	No
Wu 2010 (tel‐ex group)	75	84%	Yes	15	No	No	No
General physical activity
Ebrahim 1997	67	100%	No	104	Yes	No	Yes
Resnick 2002	88	100%	No	26	No	Yes	Yes
Voukelatos 2015	73	74%	No	48	No	No	No
Multiple primary exercise categories
Ansai 2015 (multicomponent group)c	82	68%	Yes	16	No	Yes	Yes
Beyer 2007c	78	100%	Yes	26	Yes	Yes	Yes
Brown 2002c		79%	No	16	Yes	Yes	Yes
Buchner 1997	75	51%	Yes	25	No	Yes	Yes
Bunout 2005c	75	70%	No	52	No	Yes	Yes
Cerny 1998c	71		No	24	No	Yes	NR
Clemson 2012 (structured group)c	83	55%	Yes	52	Yes	No	Yes
Gill 2016c	79	67%	Yes	96	No	Yes	Yes
Haines 2009b,c	81	60%	Yes	8	Yes	No	Yes
Halvarsson 2016 (balance and physical activity group)	76	98%	Yes	12	Yes	Yes	Yes
Hauer 2001c	82	100%	Yes	12	Yes	Yes	Yes
Irez 2011c	75	100%	No	12	No	Yes	Yes
Kamide 2009c	71	100%	No	26	Yes	No	No
Karinkanta 2007 (resistance and balance groups)c	73	100%	No	52	No	Yes	Yes
Kim 2014c	78	100%	Yes	52	No	Yes	Yes
Lehtola 2000	74	80%	No	26	No	Yes	Yes
Means 2005c	74	57%	No	6	Yes	Yes	Yes
Ng 2015c	70	61%	Yes	12	No	Yes	Yes
Park 2008	68	100%	No	48	NR	Yes	No
Rubenstein 2000	75	0%	Yes	12	No	Yes	Yes
Sherrington 2014b,c	81	74%	Yes	52	Yes	No	Yes
Suzuki 2004c	78	100%	No	26	No	Yes	No
Uusi‐Rasi 2015c	74	100%	Yes	104	Yes	Yes	Yes
Vogler 2009b (weightbearing group)	80	79%	Yes	12	Yes	No	Yes
Exercise versus exercise
Ballard 2004	73	100%	Yes	15 (Low intensity = 2)	No	Yes	NR
Barker 2016	69	88%	Yes	12	Yes	Pilates group: Yes; HEP group: No	Yes
Clemson 2012	83	55%	Yes	52	Yes	No	Yes
Davis 2011	78	100%	No	52	No	Yes	No
Freiberger 2007	76	44%	Yes	16	No	No	Yes
Helbostad 2004	81	81%	Yes		Yes	Yes	Combined training:No; Home training: Yes.
Hirase 2015	82	70%	Yes	16	Yes	No	No
Hwang 2016	72	67%	Yes	24	Tai Chi: No; other group: Yes	No	Yes
Karinkanta 2007	73	100%	No	52	No	Yes	Yes
Kemmler 2010	69	100%	No	78	No	Yes	High intensity: Yes; low intensity: No
Kwok 2016	70	85%	Yes	12	Yes	Yes	Yes
Kyrdalen 2014	83	73%	Yes	12	Yes	Group: Yes; Home: No	Yes
LaStayo 2017	76	65%	Yes	12	Yes	Yes	Yes
Liston 2014	77	85%	Yes	8	Yes	Yes	OEP: Yes; Stretching: No.
Liu‐Ambrose 2004	79	100%	No	25	No	Yes	Yes
Lurie 2013	80	59%	Yes	Variable	Yes	No	Yes
Mirelman 2016	83	35%	Yes	6	No	No	Yes
Morone 2016	69	100%	Yes	8	Yes	Yes	No
Morrison 2018	67	48%	No	12	No	Balance: Yes; Wii: No	Balance: No; Wii: Yes
Okubo 2016	71	63%	No	64	No	Yes	Yes
Shigematsu 2008	69	63%	No	12	No	Yes	No
Steadman 2003	83	82%	Yes	6	Yes	Yes	Yes
Taylor 2012	75	73%	Yes	20	No	Yes	No
Verrusio 2017	65	53%	Yes	52	Yes	No	NR
Wolf 1996	76	81%	No	15	No	Yes	Yes
Yamada 2010	80	Unknown	No		Yes	Yes	Yes
Yamada 2012	86	81%	No	24	Yes	Yes	Yes
Yamada 2013	77	57%	No	24	No	No	Yes

^a Categorised by primary exercise category.
 ^b Post‐hospital discharge study.
 ^c Indicates the primary interventions include gait, balance, and functional training and strength/resistance training.

We contacted authors of 49 included studies to request additional details regarding study design and outcome data and received responses for 26 trials; this resulted in additional information that is used in the review for 10 studies (Arkkukangas 2015; Clegg 2014; Dadgari 2016; Hamrick 2017; Kerse 2010; Kovacs 2013; Lord 2003; Morrison 2018; Sales 2017; Siegrist 2016). Trialists of the other 16 studies either reported they had no data to supply or they supplied data that could not be used in the review (Ansai 2015; Beyer 2007; Cerny 1998; Dangour 2011; Davis 2011; Duque 2013; Gschwind 2015; Huang 2010; Kyrdalen 2014; LaStayo 2017; Lurie 2013; Morgan 2004; Morone 2016; Okubo 2016; Park 2008; Resnick 2002). This account does not include the studies for which further information or data were sought or supplied regarding trials included in Gillespie 2012.

Trial design

All included studies were randomised controlled trials (RCTs). The majority of trials were individually randomised and nine were cluster randomised; either by unit of residence (Huang 2010; Lord 2003; Merom 2016; Wolf 2003), health centre (Dadgari 2016; Dangour 2011; Iliffe 2015; Siegrist 2016), or senior centre (Reinsch 1992). The included trials had 230 groups. Most trials (n = 95) had two groups included in this review (usually intervention and control), 10 studies had three groups (two intervention and one control: Almeida 2013; Ansai 2015; Clemson 2012; Halvarsson 2016; Hirase 2015; Iliffe 2015; Liu‐Ambrose 2004; Vogler 2009; Wolf 1996; Woo 2007; all intervention: Davis 2011; Wu 2010), and one study had four groups (3 intervention, 1 control) (Karinkanta 2007).

Trial size

The median number of participants randomised per trial was 134 (interquartile range (IQR) 65 to 262). The trials ranged in sample size from 20 participants in Resnick 2002 to 1635 participants in Gill 2016.

Trial setting

The included trials were carried out in 25 countries, the most common being Australia (19 trials), USA (18 trials), Japan (11 trials), the UK (7 trials), Finland (5 trials), Brazil (4 trials), Canada (4 trials), Germany (4 trials), New Zealand (4 trials), Sweden (4 trials), the Netherlands (3 trials), and Taiwan (3 trials). The remaining trials were conducted in Chile (2 trials), France (2 trials), Hungary (2 trials), Italy (2 trials), Norway (2 trials), Singapore (2 trials), China (1 trial), Denmark (1 trial), Iran (1 trial), Korea (1 trial), Switzerland (1 trial), Thailand (1 trial) and Turkey (1 trial). Of the three multinational trials, Gschwind 2015 included participants in Germany, Spain and Australia; Mirelman 2016 recruited from Belgium, Israel, Italy, Netherlands and the UK and Latham 2003 from Australia and New Zealand. See Appendix 4.

Participants

There were 23,407 participants randomised and 20,007 with fall data at follow‐up. Overall, 77% of included participants were women. All participants were women in 28 trials (see Appendix 4), and men in one trial (Rubenstein 2000). The average participant age in the included trials was 76 years.
 
 The inclusion/exclusion criteria and other participant details are listed for each study in the Characteristics of included studies.

Sixteen trials (15%) would have been excluded if the review inclusion criteria had been set at 65+ years of age (see Appendix 4).

Sixty included studies (56%) specified a history of falling or evidence of one or more risk factors for falling in their inclusion criteria (see Appendix 4).

Seventy‐two trials (67%) excluded participants with cognitive impairment, either defined as an exclusion criterion or implied by the stated requirement to be able to give informed consent and/or to follow instructions (see Appendix 4).

Four trials (4%) only included people who had recently been discharged from hospital (Haines 2009; Latham 2003; Sherrington 2014; Vogler 2009). It is possible other trials also included some participants who had been recently discharged from hospital or the emergency department, however this was not quantified.

Interventions

Exercise was compared with a control intervention (one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise) in 81 trials (19,684 participants) in people not recently discharged from hospital, and four trials (816 participants) in people who were recently discharged from hospital (Haines 2009; Latham 2003; Sherrington 2014; Vogler 2009). Twenty‐three trials, with 3527 participants, compared the effect of different types of exercise in people not recently discharged from hospital, and one trial (180 participants) compared the effect of different types of exercise in the post‐hospital population (Vogler 2009). Four trials (1021 participants) compared group versus individual exercise (Barker 2016; Helbostad 2004; Iliffe 2015; Kyrdalen 2014), and three trials (879 participants) compared high‐ versus low‐dose exercise (Ballard 2004; Davis 2011; Taylor 2012); see Appendix 4).

When interventions are grouped by the type of intervention (descriptors), as described in Data synthesis, there were 230 groups; 146 intervention arms and 84 control arms. There were 13 multiarm studies included in the review; 12 trials had three arms (Almeida 2013; Ansai 2015; Clemson 2012; Davis 2011; Halvarsson 2016; Hirase 2015; Iliffe 2015; Liu‐Ambrose 2004; Vogler 2009; Wolf 1996; Woo 2007; Wu 2010), and one trial had four arms (Karinkanta 2007). Buchner 1997 had four arms; however, because fall data were not available for individual intervention groups we made an a priori decision to report fall outcomes for all three exercise groups combined compared with control group. In 76 (52%) intervention arms, the exercise intervention was delivered in a group setting; in 43 (29%) intervention arms, it was delivered individually; and 27 (18%) intervention arms involved a combination of group‐based and individual exercise (see Appendix 4). In 67 (46%) intervention arms, the intervention was delivered by a health professional; in the 77 (53%) intervention arms where the intervention was not delivered by a trained health professional, personnel included trained physical educators, trained exercise leaders and Tai Chi instructors; in one intervention arm, the intervention was delivered by both types of personnel (Sales 2017); and in one trial the personnel were not specified (Park 2008).

The intervention arms were grouped by their primary exercise modality into six categories (Appendix 5) using the ProFaNE taxonomy (Appendix 1).

1. Most intervention arms (n = 78; 53%) included balance and functional exercises as the primary intervention (ProFaNE taxonomy code gait/balance/co‐ordination/functional task training).

2. Strength/resistance training was the primary component of 9 (6%) intervention arms.

3. Flexibility training was the primary component of one (1%) intervention arms.

4. 3D training (constant repetitive movement through all three spatial planes) was the primary component of 15 (10%) intervention arms.

5. General physical activity (walking groups) was the primary component of 6 (4%) intervention arms.

6. Endurance training alone was the primary component of one (1%) intervention arm.

7. Multiple categories of ProFaNE taxonomy were the primary intervention in 37 (25%) intervention arms. The majority (n = 19, 51%) of these intervention arms included balance and functional exercise as well as resistance training.

The number of studies, and how many of these are cluster‐RCTs, for the main exercise versus control comparison for each primary exercise category is summarised below, with further details including numbers of participants presented in Table 10, and associated study IDs in Appendix 6 (all trials) and Appendix 7 (trials contributing data to the rate of falls analysis). Note that these do not include the four post‐hospital discharge RCTs.

3. Numbers of studies and participants included in the exercise versus control comparison for each primary exercise category.

Comparisona	Number of trials (cluster) b	Number of participants randomised	Number of participants analysed for any one outcome	Number of trials (cluster) with participants analysed for rate of falls outcome c,d	Number of participants analysed for rate of falls outcome d
Exercise (all types) versus control	81 (9)	19684	13518	59 (6)	12,981
Balance and functional exercises versus control	48 (6)	11860	8288	39 (4)	7920
Resistance exercises versus control	7	694	327	5	327
Flexibility versus control	0	0	0	0	0
3D exercise (Tai Chi) versus control	10 (2)	3284	2677	7 (1)	2655
3D exercise (dance) versus control	1 (1)	530	522	1 (1)	522
General physical activity (walking programme) versus control	3	571	441	2	441
Endurance training versus control	0	0	0	0	0
Other kinds of exercise versus control	0	0	0	0	0
Multiple categories of exercise versus control	21	4073	1623	11	1374

^aExercise (all types) combines all categories of exercise. Multiple categories of exercise include studies containing two or more primary categories of exercise, as categorised using the ProFaNE taxonomy. The remaining analyses include only one primary category of exercise, as categorised using the ProFaNE taxonomy.
 ^bStudy IDs are shown in Appendix 6.
 ^cStudy IDs are shown in Appendix 7.
 ^dThese data apply to the follow‐up (at the time point included in main analysis) for the primary outcome (rate of falls) for the individual trials.

1. Exercise (all types) versus control: 81 RCTs (9 cluster‐RCTs).

2. Balance and functional exercises versus control: 48 RCTs (6 cluster‐RCTs).

3. Resistance exercises versus control: 7 RCTs.

4. Flexibility versus control: 0 RCTs.

5. 3D exercise (Tai Chi) versus control: 10 RCTs (2 cluster‐RCTs).

6. 3D exercise (dance) versus control: 1 RCTs (1 cluster‐RCT).

7. General physical activity (walking programme) versus control: 3 RCTs.

8. Endurance training versus control: 0 RCTs.

9. Other kinds of exercise versus control: 0 RCTs.

10. Multiple categories of exercise versus control: 21 RCTs.

The duration of the exercise intervention in these 81 trials ranged from 5 to 130 weeks; it was one year or more in 24 trials (30%) and two years or more in five trials (6%) (Table 9).

Additional details of the number of studies and number of participants included in the primary analysis (exercise versus control on rate of falls) for each primary category of exercise are shown in Appendix 8.

Outcomes

The source of data used for calculating outcomes for each trial for generic inverse variance analysis is shown in Appendix 9. Rate of falls was reported in 34 trials, and could be calculated from a further 43 trials. Data on risk of falling (number of fallers) were available in 17 trials and could be calculated for a further 61. Raw data for rate of falls and number of fallers, when available, are shown in Appendix 10. Six trials met our inclusion criteria but did not include data that could be included in these analyses (Almeida 2013; Fiatarone 1997; Mirelman 2016; Morone 2016; Morrison 2018; Resnick 2002). Two of these trials contained inadequate data to include in an analysis (Fiatarone 1997; Resnick 2002), but reported no significant between‐group difference in number of falls, and two trials reported zero falls in each group (Almeida 2013; Morrison 2018). Morone 2016 did not present fall data, but found balance training using Wii‐fit may have a greater effect on balance outcomes compared with conventional balance training. Mirelman 2016 found treadmill plus virtual reality training may be more effective in preventing falls than treadmill alone, six months after the end of a six‐week training period. The raw data for non‐fall outcomes for these studies are shown in Appendix 11.

Eleven trials reported a fracture outcome, two trials reported number of falls requiring hospitalisation, and five trials reported the number of people experiencing a fall requiring medical attention. Death was recorded in 40 trials and was listed as a reason for loss to follow‐up in all of these trials except Wolf 2003, which also assessed death as an adverse event. Deaths were not reported by group in two trials (Day 2002; Lord 1995; Appendix 12). None of the deaths were explicitly linked to the trial participation.

Adverse events

Two trials, including one in the post‐hospital population, measured the number of people experiencing adverse events in both groups throughout the trial period (Iliffe 2015; Latham 2003). No other studies reported adverse events that were monitored closely in all groups over the entire study period. Adverse events reported to any degree are described in Appendix 13. Adverse events were reported to a degree in the intervention and control groups in 16 trials, in the intervention group only in 13 trials, in two intervention groups in seven trials, and in two intervention plus control group in five trials.

Adherence

Adherence was measured in 78 studies and adherence data were reported in 77 studies (Appendix 14). The measures used to quantify adherence varied: the majority of studies summarised proportion of classes attended (n = 53) or proportion of scheduled sessions completed (n = 20), three studies quantified the amount of exercise performed (Boongrid 2017; Okubo 2016; Sherrington 2014), and two studies described the proportion of participants who started the programme (El‐Khoury 2015; Skelton 2005).

Excluded studies

We eliminated 253 reports on full‐text review. We retained 21 studies (23 reports) as excluded studies as they initially appeared to meet the inclusion criteria but were subsequently excluded (see Excluded studies for links to references, and the Characteristics of excluded studies and Appendix 15 for details). Of the identified trials:

1. one trial did not meet the review's inclusion criterion for age (Pereira 1998);

2. one trial included participants with a particular clinical condition that increases the risk of falls (Hsu 2017);

3. one trial included participants who were not community‐dwelling (DeSure 2013);

4. 15 trials did not involve exercise as a single intervention;

5. one trial included an ineligible comparator (Ohtake 2013);

6. one trial did not measure falls (Hinrichs 2016);

7. one trial withdrew three of the six fallers from the study because the falls resulted in injuries (Morris 2008).

Studies awaiting classification

Two studies are awaiting classification. Li 2018b is a large study (n = 670) comparing the effect of Tai Ji Quan, multimodal exercise and stretching in older people at high risk of falls. The other is a small (n = 6) study (Jagdhane 2016).

Ongoing studies

We identified 16 ongoing trials (see the Characteristics of ongoing studies). Seven trials are currently open to recruitment CTRI/2018/01/011214; NCT02617303; NCT02926105; NCT03211429; NCT03320668; NCT03417531; NCT03462654), and nine are ongoing but no longer recruiting (ACTRN 12613001161718; ACTRN 12615000138583; ACTRN 12615000865516; ISRCTN71002650; NCT01029171; NCT02126488; NCT02287740; NCT03404830; NCT03455179).

The median target sample size is 402 (IQR 280‐670) and two of the ongoing trials are cluster randomised (ACTRN 12613001161718; ISRCTN71002650). Half of the trials (8/16, 50%) specify increased fall‐risk as an inclusion criterion. Eight studies are investigating the effect of a programme of multiple categories of exercise (ACTRN 12615000865516; CTRI/2018/01/011214; ISRCTN71002650; NCT01029171; NCT02287740; NCT02617303; NCT02926105; NCT03455179), including four using the Otago Exercise Program (ACTRN 12615000865516; NCT01029171; NCT02617303; NCT02926105). There are three trials on resistance training (ACTRN 12613001161718; NCT03404830; NCT03455179), one on Tai Chi (NCT03211429), one on balance training (ACTRN 12615000138583), and a study evaluating slip training on the treadmill (NCT02126488). Two studies compare group versus individual delivery, using the LiFE Program (NCT03462654) and Otago Exercise Program (NCT03320668). There are no studies investigating the effect of flexibility training, general physical activity or endurance training alone.

Risk of bias in included studies

Details of the 'Risk of bias' assessment across all included trials and for each individual item in the included trials are shown in Characteristics of included studies, Figure 2 and Figure 3.

2.

Risk of bias graph: review authors' judgements about each risk of bias item presented as percentages across all included studies.

3.

Risk of bias summary: review authors' judgements about each risk of bias item for each included study.

Allocation

We judged the risk of bias in generation of the allocation sequence as low in 67% (n = 72/108) of trials, unclear in 33% (n = 36/108) and high in zero trials. We assessed the methods of concealment of the allocation prior to group assignment as low risk of bias in 35% (n = 38/108), unclear in 60% (n = 65/108) and high in the remaining 5% (5/108) of trials (Cerny 1998; Dangour 2011; Huang 2010; Lord 2003; Reinsch 1992).

Blinding

Blinding of participants and personnel

In the majority of studies (90%, n = 97/108) it was not possible to blind the personnel and participants to group allocation. As the likelihood of awareness of group allocation introducing performance bias was not clear, we assessed the risk of bias for non‐blinding as unclear for these trials. We judged the impact of performance bias as low in 5% (n = 5/108) of trials, unclear in 89% (97/108) of trials and high in 6% (6/108) of trials.

Blinding of outcome assessment

We assessed the risk of bias for blinding of outcome assessment separately for the following outcomes.

1. Rate of falls and risk of falling


   1. We judged the risk of detection bias in relation to the methods of ascertainment of the rate and/or risk of falls to be low in 40% (n = 43/108), high in 21% (n = 23/108) and unclear in 39% (n = 42/108) of the included trials.

2. Risk of fractures

   1. In trials reporting on the risk of fracture, we assessed the risk of bias for blinding of outcome assessment for the rate of fractures. We judged the risk of detection bias in relation to the methods of ascertainment of fractures to be low in 20% (n = 4/20), high in 35% (n = 7/20) and unclear in 45% (n = 9/20) of the included trials that measured fractures.

3. Requiring hospital admission/medical attention, adverse events

   1. In trials reporting on the risk of hospital admission and/or requiring medical attention and/or adverse events, we judged the risk of detection bias in relation to the method of ascertainment of these outcomes to be low in 15% (5/33) of trials, unclear in 67% (22/33) and high in 18% (6/33) of trials.

4. Health‐related quality of life

   1. In trials reporting on health‐related quality of life we judged the risk of detection bias in relation to the method of ascertainment of health‐related quality of life to be high in all studies (23/23), due to participants in these studies being unblinded to their allocated group and health‐related quality of life being a self‐reported outcome.

Incomplete outcome data

We judged the risk of bias due to incomplete outcome data to be low in 53% (n = 57/108), unclear in 20% (n = 22/108) and high in the remaining 27% of trials (n = 29/108).

Selective reporting

We assessed the risk of bias due to selective reporting of falls outcomes as low in 12% (n = 13/108) of studies, unclear in 40% (n = 43/108) and high in 48% (52/108).

Other potential sources of bias

Bias in the recall of falls due to less reliable methods of ascertainment

We assessed 58% of included studies (n = 63/108) as being at low risk of bias in the recall of falls (i.e. falls were recorded concurrently using recommended methods of monthly diaries or postcards). We judged the risk of bias to be high in 27% of trials (n = 29/108), in that ascertainment of falling episodes was by participant recall, at intervals during the study or at its conclusion. In 15% of trials (n = 16/108) the risk of bias was unclear, as retrospective recall was for a short period only, or details of ascertainment were not described.

Bias due to cluster‐randomisation

We assessed the nine cluster‐randomised trials for risk of bias associated with recruitment methods, baseline imbalance, loss of clusters, incorrect analysis and comparability with individually‐randomised trials. We judged the risk of bias due to factors associated with cluster‐randomised trials to be low in one (11%) trial, unclear in seven trials (78%) and high in the remaining trial (11%, Dadgari 2016).

Effects of interventions

See: Table 1; Table 2; Table 3; Table 4; Table 5; Table 6; Table 7

Summary of findings for the main comparison. Summary of findings: exercise (all types) versus control (e.g. usual activities).

Exercise (all types) versus control (e.g. usual activities) for preventing falls in older people living in the community
Patient or population: Older people living in the community (trials focusing on people recently discharged from hospital were not included) Settings: Community, either at home or in places of residence that, on the whole, do not provide residential health‐related care Intervention: Exercise of all typesa Comparison: Usual care (no change in usual activities) or a control (non‐active) interventionb
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Control	Exercise (all types)
Rate of falls (falls per person‐years) Follow‐up: range 3 to 30 months	All studies population		Rate ratio 0.77 (0.71 to 0.83)d	12,981 (59 RCTs)	⊕⊕⊕⊕ highe	Overall, there is a reduction of 23% (95% CI 17% to 29%) in the number of falls Guide to the data: If 1000 people were followed over 1 year, the number of falls in the overall population would be 655 (95% CI 604 to 706) compared with 850 in the group receiving usual care or attention control. In the unselected population, the corresponding data are 466 (95% CI 430 to 503) compared with 605 in the group receiving usual care or attention control. In the selected higher‐risk population, the corresponding data are 924 (95% CI 852 to 996) compared with 1200 in the control group
	850 per 1000c	655 per 1000 (604 to 706)
	Not selected for high risk population
	605 per 1000c	466 per 1000 (430 to 503)
	Selected for high risk population
	1200 per 1000c	924 per 1000 (852 to 996)
Number of people who experienced one or more falls Follow‐up: range 3 to 25 months	All studies population		RR 0.85 (0.81 to 0.89)g	13,518 (63 RCTs)	⊕⊕⊕⊕ highe	Overall, there is a reduction of 15% (95% CI 11% to 19%) in the number of people who experienced one or more falls Guide to the data: If 1000 people were followed over 1 year, the number of people who experienced one or more falls in the unselected population would be 408 (95% CI 389 to 428) compared with 480 in the group receiving usual care or attention control. In the unselected population, the corresponding data are 323 (95% CI 308 to 339) compared with 380 in the group receiving usual care or attention control. In the selected higher‐risk population, the corresponding data are 425 (95% CI 405 to 445) compared with 500 in the control group.
	480 per 1000f	408 per 1000 (389 to 428)
	Not selected for high risk population
	380 per 1000f	323 per 1000 (308 to 339)
	Selected for high risk population
	500 per 1000f	425 per 1000 (405 to 445)
Number of people who experienced one or more fall‐related fractures Follow‐up: range 4 to 42 months	All studies populationh		RR 0.73 (0.56 to 0.95)	4047 (10 RCTs)	⊕⊕⊝⊝ lowi	Overall, there may be a reduction of 27% (95% CI 5% to 44%) in the number of people who experienced one or more fall‐related fractures Guide to the data: If 1000 people were followed over 1 year, the number of people who experienced one or more fall‐related fractures may be 47 (95% CI 36 to 61) compared with 64 in the control group
	64 per 1000	47 per 1000 (36 to 61)
Number of people who experienced one of more falls that resulted in hospital admission Follow‐up: range 3 to 42 months	All studies populationh		RR 0.78 (0.51 to 1.18)	1705 (2 RCTs)	⊕⊝⊝⊝ very lowj	The evidence is very low certainty, hence we are uncertain of the findings of a reduction of 22% (95% CI 49% reduction to 18% increase) in the number of people who experienced one or more falls that required hospital admission. Of note is that the 95% CI includes the possibility of both reduced and increased hospitalisation. Guide to the data: If 1000 people were followed over 1 year, the number of people who experience one or more falls that required hospital admission in the general risk population may be 45 (95% CI 30 to 68) compared with 57 in the group receiving usual care or attention control
	57 per 1000	45 per 1000 (29 to 68)
Number of people who experienced one or more falls that required medical attention. Follow‐up: range 6 to 24 months	All studies populationh		RR 0.61 (0.47 to 0.79)	1019 (5 RCTs)	⊕⊕⊝⊝ lowk	Overall, there may be a reduction of 39% (95% CI 21% to 53%) in the number of people who experienced one or more falls that required medical attention Guide to the data: If 1000 people were followed over 1 year, the number of people who experienced one or more falls that required medical attention may be 129 (95% CI 100 to 167) compared with 211 in the group receiving usual care or attention control
	211 per 1000	129 per 1000 (100 to 167)
Health‐related quality of life Follow‐up: range 3 to 24 months (A higher score indicates better quality of life)	‐	The mean health‐related quality of life score in the intervention groups was 0.03 standard deviations lower (0.10 lower to 0.04 higher)	‐	3172 (15 RCTs)	⊕⊕⊝⊝ lowl	SMD was calculated from 4 trials with EQ‐5D, 5 trials with SF‐36, 3 trials with SF12, 1 trial with QUALEFFO‐41, 1 trial with WHOQOL‐BREF, and 1 with Assessment of QOL EQ‐5D: Mean difference = −0.0026 (95% CI −0.0086 to 0.0034). SMD was converted back to MD using EQ‐5D scale (0 to 1), based on data for 4 trials (6 comparisons) reporting endpoint scores.m MID for the EQ‐5D is typically 0.074 (Walters 2005) SF36: Mean difference = −0.36 (95% CI −1.20 to 0.48). SMD was converted back to MD using SF‐36 scale, based on data for 5 trials.m MID for the SF‐36 is typically 3 to 5 ( Walters 2003)
Adverse events	See comment		Not estimable	6019 (27 RCTs)	⊕⊝⊝⊝n very low	Adverse events were reported to various degrees, but predominantly in the intervention groups, in the 27 RCTs, 14 of which reported no adverse events. Aside from 2 serious adverse events (1 pelvic stress fracture and 1 inguinal hernia surgery) reported in 1 trial, the rest were non‐serious adverse events, primarily of a musculoskeletal nature. There was a median of 3 events (range 1 to 26) in the exercise groups
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; MID: minimally important difference; RR: risk ratio; SMD: standardised mean difference
GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect

^aExercise is a physical activity that is planned, structured and repetitive and aims to improve or maintain physical fitness. There is a wide range of possible types of exercise, and exercise programmes often include one or more types of exercise. We categorised exercise based on the Prevention of Falls Network Europe (ProFaNE) taxonomy that classifies exercise type as: i) gait, balance, and functional [task] training; ii) strength/resistance (including power); iii) flexibility; iv) three‐dimensional (3D) exercise (e.g. Tai Chi, Qigong, dance); v) general physical activity; vi) endurance; and vii) other kind of exercises. The taxonomy allows for more than one type of exercise to be delivered within a programme.
 ^bA control intervention is one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls.
 ^cThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 59 RCTs. We calculated the risk in the control group using the median falls per person‐year for the subgroups of trials for which a) an increased risk of falls was not an inclusion criterion (29 RCTs, 6123 participants), or b) increased risk of falls was an inclusion criterion (30 RCTs, 6858 participants).
 ^dSubgroup analysis found no difference based on whether risk of falls was an inclusion criterion or not (test for subgroup differences: Chi² = 0.90, df = 1, P = 0.34, I² = 0%).
 ^eThere was no downgrading, including for risk of bias, as results were essentially unchanged with removal of the trials with a high risk of bias on one or more items.
 ^fThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 63 RCTs. We calculated the risk in the control group using the median proportion of fallers for the subgroups of trials for which a) an increased risk of falls was not an inclusion criterion (28 RCTs, 6347 participants), or b) increased risk of falls was an inclusion criterion (35 RCTs, 7171 participants).
 ^gSubgroup analysis found no difference based on whether risk of falls was an inclusion criterion or not (test for subgroup differences: Chi² = 0.94, df = 1, P = 0.33, I² = 0%).
 ^hWe calculated the risk in the control group based on the number of events and the number of participants in the control group for this outcome.
 ⁱ Downgraded by two levels due to imprecision (few events and wide CI due to small sample size), and risk of publication bias (likelihood of reporting fractures only if there was a treatment effect; with some indication on viewing the funnel plot).
 ^jDowngraded by two levels due to imprecision (low event rate and wide confidence intervals) and because most of the 81 studies included in the review for this comparison do not contribute to the outcome. We further downgraded the evidence by one level for risk of bias because the evidence was dominated by one trial that was at high risk of bias in one or more items.
 ^kDowngraded by two levels due to imprecision and the high probability of publication bias (only 5 of 89 RCTs included in the review reported the outcome). We did not downgrade for risk of bias as results were essentially unchanged with removal of the trials at a high risk of bias in one or more items.
 ^lDowngraded by two levels due to inconsistency (there was considerable heterogeneity (I² = 76%)) and risk of bias (removing studies with high risk of bias in one or more items had a marked impact on results).
 ^mIn order to express the MD in the unit‐specific measurement instruments (ED‐5D and SF‐36), we multiplied the SMD by a typical among‐person standard deviation for that scale, using the pooled standard deviation of baseline scores in the largest study in the analysis. For EQ‐5D, Iliffe 2015 has a combined SD of 0.086; for SF36, Dangour 2011 has combined SD of 12.04.
 ⁿDowngraded by three levels due to limitations in design of studies, suggesting a very serious risk of bias and incomplete data. Only one trial measured the number of people experiencing adverse events in both groups throughout the trial period (Iliffe 2015).

Summary of findings 2. Summary of findings: balance and functional exercises versus control (e.g. usual activities).

Balance, and functional exercises versus control (e.g. usual activities) for preventing falls in older people in the community
Patient or population: Older people living in the community (trials focusing on people recently discharged from hospital were not included) Settings: Community, either at home or in places of residence that, on the whole, do not provide residential health‐related care Intervention: Exercise, type = gait, balance, and functional (task) traininga Comparison: Usual care (no change in usual activities) or a control (non‐active) interventionb
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Control	Exercise (gait, balance, and functional [task] training)
Rate of falls (falls per person‐years) Follow‐up: range 3 to 30 months	All studies population		Rate ratio 0.76 (0.70 to 0.81)	7920 (39 RCTs)	⊕⊕⊕⊕d high	Overall, there is a reduction of 24% (95% CI 19% to 30%) in the number of falls Guide to the data based on the all‐studies estimate. If 1000 people were followed over 1 year, the number of falls would be 646 (95% CI 595 to 689) compared with 850 in the group receiving usual care or attention control
	850 per 1000c	646 per 1000 (595 to 689)
	Specific exercise population
	930 per 1000c	707 per 1000 (651 to 754)
Number of people who experienced one of more falls Follow‐up: range 3 to 24 months	All studies population		RR 0.87 (0.82 to 0.91)	8288 (37 RCTs)	⊕⊕⊕⊕d high	Overall, there is a reduction of 13% (95% CI 9% to 18%) in the number of people who experienced one or more falls. Guide to the data based on the all‐studies estimate. If 1000 people were followed over 1 year, the number of people who experienced one or more falls would be 418 (95% CI 394 to 437) compared with 480 in the group receiving usual care or attention control
	480 per 1000e	418 per 1000 (394 to 437)
	Specific exercise population
	549 per 1000e	478 per 1000 (451 to 500)
Number of people who experienced one or more fall‐related fractures. Follow‐up: range 6 to 30 months	All studies population		RR 0.44 (0.25 to 0.76)	2139 (7 RCTs)	⊕⊕⊝⊝g low	Overall, there may be a reduction of 56% (95% CI 24% to 75%) in the number of people who experienced one or more fall‐related fractures Guide to the data. If 1000 people were followed over 1 year, the number of people who experienced one or more fall‐related fractures may be 29 (95% CI 16 to 49) compared with 64 in the group receiving usual care or attention control
	64 per 1000f	29 per 1000 (16 to 49)
Adverse events	See comment		Not estimable	4167 (15 RCTs)	⊕⊝⊝⊝h very low	Adverse events were reported on in 15 of the 48 trials with gait, balance, and functional (task) training as the primary intervention in exercise versus control analyses in trials. Adverse events were reported for both intervention and control groups (11 trials) or just the intervention group (4 trials). 200 adverse events were reported; most were non‐serious adverse events of a musculoskeletal nature; 173 were in a single study including 2 intervention groups. Other adverse events included shortness of breath in 4 participants; and 1 participant with palpitations. One study reported a pelvic stress fracture in an intervention group
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; RR: risk ratio
GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect

^aUsing Prevention of Falls Network Europe (ProFaNE) taxonomy, gait, balance, and functional [task] training is: gait training = specific correction of walking technique, and changes of pace, level and direction; balance training = transferring bodyweight from one part of the body to another or challenging specific aspects of the balance systems; functional training = functional activities, based on the concept of task specificity. Training is assessment‐based, tailored and progressed. Exercise programs included in this analysis contained a single primary exercise category (gait, balance, and functional [task] training); these exercise programs may also include secondary categories of exercise.
 ^bA control intervention is one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls.
 ^c The all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 59 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the 39 RCTs.
 ^dWe did not downgrade for risk of bias, as results were essentially unchanged with the removal of the trials with a high risk of bias in one or more items.
 ^eThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 63 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the 37 RCTs.

^fThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 10 all‐exercise types RCTs. Based on the number of events and the number of participants in the control group for this outcome over the seven RCTs, the assumed risk in the control group was 43 per 1000.
 ^gDowngraded by two levels due to risk of bias (removing studies with high risk of bias on one or more items had a marked impact on results), and imprecision (few events and wide CI due to small sample size).
 ^hDowngraded by three levels due to limitations in design of studies, suggesting a high likelihood of bias (no trials in this analysis measured the number of participants experiencing adverse events in both groups throughout the trial period).

Summary of findings 3. Summary of findings: resistance exercises versus control (e.g. usual activities).

Resistance exercises versus control (e.g. usual activities) for preventing falls in older people in the community
Patient or population: Older people living in the community (trials focusing on people recently discharged from hospital were not included) Settings: Community, either at home or in places of residence that, on the whole, do not provide residential health‐related care Intervention: Exercise, type = resistance traininga Comparison: Usual care (no change in usual activities) or a control (non‐active) interventionb
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Control	Exercise (resistance training)
Rate of falls (falls per person‐years) Follow‐up: range 4 to 12 months	All studies population		Rate ratio 1.14 (0.67 to 1.97)	327 (5 RCTs)	⊕⊝⊝⊝d very low	The evidence is of very low certainty, hence we are uncertain of the findings of an increase of 14% (95% CI 33% reduction to 97% increase) in the number of falls. Guide to the data based on the all‐studies estimate. If 1000 people were followed over 1 year, the number of falls would be 969 (95% CI 570 to 1675) compared with 850 in the group receiving usual care or attention control
	850 per 1000c	969 per 1000 (570 to 1675)
	Specific exercise population
	630 per 1000c	719 per 1000 (423 to 1242)
Number of people who experienced 1 or more falls Follow‐up: range 4 to 12 months	All studies population		RR 0.81 (0.57 to 1.15)	163 (2 RCTs)	⊕⊝⊝⊝f very low	The evidence is of very low certainty, hence we are uncertain of the findings of a decrease of 19% (95% CI 43% reduction to 15% increase) in the number of people who experienced one or more falls Guide to the data based on the all‐studies estimate. If 1000 people were followed over 1 year, the number of people who experienced one or more falls would be 389 (95% CI 274 to 552) compared with 480 in the group receiving usual care or attention control
	480 per 1000e	389 per 1000 (274 to 552)
	Specific exercise population
	864 per 1000e	700 per 1000 (493 to 994)
Number of people who experienced 1 or more fall‐related fractures	All studies population		RR 0.97 (0.14 to 6.49)	73 (1 RCT)	⊕⊝⊝⊝h very low	The evidence is of very low certainty, hence we are uncertain of the findings of a decrease of 3% (95% CI 86% reduction to 549% increase) The very small number of events (3 fractures in all) means that these data are not informative
	64 per 1000g	63 per 1000 (9 to 416)
Adverse events	See comment		Not estimable	64 (1 RCT)	⊕⊝⊝⊝i very low	Adverse events were reported on in one of the five trials with resistance training as the primary intervention in exercise versus control analyses. The study reported 10 musculoskeletal complaints in the intervention group and one musculoskeletal complaint in the control group.
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; RR: risk ratio
GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect

^aUsing Prevention of Falls Network Europe (ProFaNE) taxonomy, resistance training is any type of weight training (contraction of muscles against resistance to induce a training effect in the muscular system). Resistance is applied by body weight or external resistance. Training is assessment‐based, tailored and progressed. Exercise programmes included in this analysis had resistance training as the single primary exercise category; these exercise programmes may also include secondary categories of exercise.
 ^bA control intervention is one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls.
 ^cThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 59 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the 5 RCTs.
 ^dDowngraded by three levels due to risk of inconsistency (there was substantial heterogeneity (I² = 67%)), imprecision (wide CI due to small sample size), and risk of bias (removing studies with high risk of bias in one or more items had a marked impact on results).
 ^eThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 63 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the 2 RCTs.
 ^fDowngraded by one level due to risk of bias (removing studies with high risk of bias on one or more items had a marked impact on results), and downgraded by two levels due to imprecision (small number of trials and participants, wide CI).
 ^gThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 10 all‐exercise types RCTs. Based on the number of events and the number of participants in the control group for this outcome in the sole RCT, the assumed risk in the control group was 28 per 1000.
 ^hDowngraded by three levels for imprecision (wide CI, single study, very few events).
 ⁱDowngraded by three levels due to only one study reporting adverse events and limitations in design of studies, suggesting a high likelihood of bias (number of participants experiencing adverse events was not reported in the same manner in both groups throughout the trial period).

Summary of findings 4. Summary of findings: 3D (Tai Chi) exercise versus control (e.g. usual activities).

3D (Tai Chi) exercise versus control (e.g. usual activities) for preventing falls in older people in the community
Patient or population: Older people living in the community (trials focusing on people recently discharged from hospital were not included) Settings: Community, either at home or in places of residence that, on the whole, do not provide residential health‐related care Intervention: Exercise, type = 3D (Tai Chi) traininga Comparison: Usual care (no change in usual activities) or a control (non‐active) interventionb
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Control	Exercise (3D (Tai Chi))
Rate of falls (falls per person‐year) Follow‐up: range 6 to 17 months	All studies population		Rate ratio 0.81 (0.67 to 0.99)	2655 (7 RCTs)	⊕⊕⊝⊝d low	Overall, there may be a reduction of 19% (95% CI 1% to 33%) in the number of falls. Guide to the data based on the all‐studies estimate. If 1000 people were followed over 1 year, the number of falls may be 689 (95% CI 570 to 842) compared with 850 in the group receiving usual care or attention control
	850 per 1000c	689 per 1000 (570 to 842)
	Specific exercise population
	1020 per 1000c	827 per 1000 (684 to 1010)
Number of people who experienced one or more falls Follow‐up: range 5 to 17 months	All studies population		RR 0.80 (0.70 to 0.91)	2677 (8 RCTs)	⊕⊕⊕⊕f high	Overall, there is a reduction of 20% (95% CI 9% to 30%) in the number of people who experienced one or more falls Guide to the data based on the all‐studies estimate. If 1000 people were followed over 1 year, the number of people who experienced one or more falls would be 384 (95% CI 336 to 437) compared with 480 in the group receiving usual care or attention control
	480 per 1000e	384 per 1000 (336 to 437)
	Specific exercise population
	437 per 1000e	350 per 1000 (306 to 398)
Number of people who experienced one or more fall‐related fractures	See comment		Not estimable	See comment	‐	This outcomes was not reported
Adverse events	See comment		Not estimable	474 (2 RCTs)	⊕⊝⊝⊝g very low	Adverse events were reported in two of 10 trials (474 participants) with 3D (Tai Chi) as the primary intervention. There were no occurrences of adverse events
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; RR: risk ratio
GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect

^aUsing Prevention of Falls Network Europe (ProFaNE) taxonomy, 3D (Tai Chi) training uses upright posture, specific weight transferences and movements of the head and gaze, during constant movement in a fluid, repetitive, controlled manner through three spatial planes. Exercise programmes included in this analysis had 3D (Tai Chi) training as the single primary exercise category; these exercise programmes may also include secondary categories of exercise.
 ^bA control intervention is one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls.
 ^cThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 59 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the seven RCTs.
 ^dDowngraded by two levels due to inconsistency (there was substantial heterogeneity (I² = 74%)), and risk of bias (removing studies with high risk of bias in one or more items had a marked impact on results).
 ^eThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 63 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the eight RCTs.
 ^fWe did not downgrade for risk of bias, as results were essentially unchanged with removal of the trials with a high risk of bias in one or more items.
 ^gDowngraded by three levels due to only 30% of trials reporting adverse events to any degree, and limitations in the design of studies suggesting a high likelihood of bias (no trials in this analysis measured the number of participants experiencing adverse events in both groups throughout the trial period).

Summary of findings 5. Summary of findings: 3D (dance) exercise versus control (e.g. usual activities).

3D (dance) exercise versus control (e.g. usual activities) for preventing falls in older people in the community
Patient or population: Older people living in the community (trials focusing on people recently discharged from hospital were not included) Settings: Community, either at home or in places of residence that, on the whole, do not provide residential health‐related care Intervention: Exercise, type = 3D (dance) traininga Comparison: Usual care (no change in usual activities) or a control (non‐active) interventionb
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Control	Exercise (3D [dance])
Rate of falls (falls per person‐years) Follow‐up: 12 months	All studies population		Rate ratio 1.34 (0.98 to 1.83)	522 (1 RCT)	⊕⊝⊝⊝d very low	The evidence is of very low certainty, hence we are uncertain of the findings of an increase of 34% (95% CI 2% reduction to 83% increase) in the number of falls Guide to the data based on the all‐studies estimate If 1000 people were followed over 1 year, the number of falls may be 1139 (95% CI 833 to 1556) compared with 850 in the group receiving usual care or attention control
	850 per 1000c	1139 per 1000 (833 to 1556)
	Specific exercise population
	800 per 1000c	1072 per 1000 (784 to 1464)
Number of people who experienced one or more falls Follow‐up: 12 months	All studies population		RR 1.35 (0.83 to 2.20)	522 (1 RCT)	⊕⊝⊝⊝d very low	The evidence is of very low certainty, hence we are uncertain of the findings of an increase of 35% (95% CI 17% reduction to 120% increase) in the number of people who experienced one or more falls Guide to the data based on the all‐studies estimate If 1000 people were followed over 1 year, the number of people who experienced one or more falls may be 648 (95% CI 399 to 1056) compared with 480 in the group receiving usual care or attention control
	480 per 1000e	648 per 1000 (399 to 1056)
	Specific exercise population
	583 per 1000e	787 per 1000 (484 to 1283)
Number of people who experienced one or more fall‐related fractures	Not estimable		Not estimable	See comment	‐	This outcome was not reported
Adverse events	See comment		Not estimable	522 (1 RCT)	⊕⊝⊝⊝f very low	Adverse events were reported for the intervention group only (275 participants) in the one trial in this analysis. There were no occurrences of adverse events
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; RR: risk ratio
GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect

^aUsing Prevention of Falls Network Europe (ProFaNE) taxonomy, 3D (dance) training uses dynamic movement qualities, patterns and speeds whilst engaged in constant movement in a fluid, repetitive, controlled manner through three spatial planes. Exercise programmes included in this analysis had 3D (dance) training as the single primary exercise category; these exercise programmes may also include secondary categories of exercise.
 ^bA control intervention is one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls.
 ^cThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 59 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome in the sole RCT.
 ^dGraded very low due to serious imprecision (only one cluster‐RCT, with a wide CI due to small sample size).
 ^eThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 63 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome in the sole RCT.
 ^fDowngraded by three levels due to limitations in the design of studies, suggesting a high likelihood of bias (the trial measured the number of participants experiencing adverse events in the exercise group).

Summary of findings 6. Summary of findings: walking programme (general physical activity) versus control (e.g. usual activities).

General physical activity (including walking) training versus control (e.g. usual activities) for preventing falls in older people in the community
Patient or population: Older people living in the community (trials focusing on people recently discharged from hospital were not included) Settings: Community, either at home or in places of residence that, on the whole, do not provide residential health‐related care Intervention: Exercise, type = general physical activity (including walking) traininga Comparison: Usual care (no change in usual activities) or a control (non‐active) interventionb
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Control	Exercise (general physical activity [including walking])
Rate of falls (falls per person‐years) Follow‐up: range 12 to 24 months	All studies population		Rate ratio 1.14 (0.66 to 1.97)	441 (2 RCTs)	⊕⊝⊝⊝d very low	The evidence is of very low certainty, hence we are uncertain of the findings of an increase of 14% (95% CI 34% reduction to 97% increase) in the number of falls Guide to the data based on the all‐studies estimate If 1000 people were followed over 1 year, the number of falls may be 969 (95% CI 561 to 1675) compared with 850 in the group receiving usual care or attention control
	850 per 1000c	969 per 1000 (561 to 1675)
	Specific exercise population
	670 per 1000c	764 per 1000 (443 to 1320)
Number of people who experienced one or more falls Follow‐up: range 12 to 24 months	All studies population		RR 1.05 (0.71 to 1.54)	441 (2 RCTs)	⊕⊝⊝⊝f very low	The evidence is of very low certainty, hence we are uncertain of the findings of an increase of 5% (95% CI 29% reduction to 54% increase) in the number of people who experienced one or more falls Guide to the data based on the all‐studies estimate If 1000 people were followed over 1 year, the number of people who experienced one or more falls may be 504 (95% CI 341 to 740) compared with 480 in the group receiving usual care or attention control
	480 per 1000e	504 per 1000 (341 to 740)
	Specific exercise population
	374 per 1000e	393 per 1000 (266 to 576)
Number of people who experienced one or more fall‐related fractures	All studies population		RR 0.66 (0.11 to 3.76)	97 (1 RCT)	⊕⊝⊝⊝h very low	The evidence is of very low certainty, hence we are uncertain of the findings of a reduction of 34% (95% CI 89% reduction to 276% increase) in the number of people who experienced one or more fall‐related fractures Guide to the data If 1000 people were followed over 1 year, the number of people who experienced one or more fall‐related fractures may be 43 (95% CI 7 to 241) compared with 64 in the group receiving usual care or attention control
	64 per 1000g	43 per 1000 (7 to 241)
Adverse events	See comment		Not estimable	See comment	‐	This outcome was not reported
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; RR: risk ratio
GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect

^aUsing Prevention of Falls Network Europe (ProFaNE) taxonomy, physical activity is any movement of the body, produced by skeletal muscle, that causes energy expenditure to be substantially increased. Recommendations regarding intensity, frequency and duration are required in order to increase performance. Exercise programmes included in this analysis had general physical activity (including walking) training as the single primary exercise category; these exercise programmes may also include secondary categories of exercise.
 ^bA control intervention is one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls.
 ^cThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 59 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome in the two RCTs.
 ^dDowngraded by three levels due to inconsistency (there was substantial heterogeneity (I² = 67%)), imprecision (wide CI), and risk of bias (removing studies with high risk of bias on one or more items had a marked impact on results).
 ^eThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 63 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome in the two RCTs.
 ^fDowngraded by three levels due to inconsistency (there was moderate heterogeneity (I² = 50%), imprecision (wide CI), and risk of bias (removing studies with high risk of bias on one or more items had a marked impact on results).
 ^gThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 10 all‐exercise types RCTs. Based on the number of events and the number of participants in the control group for this outcome in the only RCT, the assumed risk in the control group was 84 per 1000.

^hDowngraded three levels due to risk of bias and imprecision (single study, wide CI).

Summary of findings 7. Summary of findings: multiple categories of exercise versus control (e.g. usual activities).

Multiple categories of exercise (often including, as primary interventions: gait, balance, and functional (task) training plus resistance training) versus control (e.g. usual activities) for preventing falls in older people in the community
Patient or population: Older people living in the community (trials focusing on people recently discharged from hospital were not included) Settings: Community, either at home or in places of residence that, on the whole, do not provide residential health‐related care Intervention: Exercise, type = Multiple types of exercisea Comparison: Usual care (no change in usual activities) or a control (non‐active) interventionb
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Control	Exercise (multiple types (including, as primary interventions: gait, balance, and functional (task) training, plus resistance training))
Rate of falls (falls per person‐years) Follow‐up: range 3 to 25 months	All studies population		Rate ratio 0.66 (0.50 to 0.88)d	1374 (11 RCTs)	⊕⊕⊕⊝e moderate	Overall, there is probably a reduction of 34% (95% CI 12% to 50%) in the number of falls Guide to the data based on the all‐studies estimate If 1000 people were followed over 1 year, the number of falls would probably be 561 (95% CI 425 to 748) compared with 850 in the group receiving usual care or attention control
	850 per 1000c	561 per 1000 (425 to 748)
	Specific exercise population
	1180 per 1000c	779 per 1000 (590 to 1039)
Number of people who experienced one or more falls Follow‐up: range 3 to 25 months	All studies population		RR 0.78 (0.64 to 0.96)	1623 (17 RCTs)	⊕⊕⊕⊝g moderate	Overall, there is probably a reduction of 22% (95% CI 4% to 36%) in the number of people who experienced one or more falls Guide to the data based on the all studies estimate. If 1000 people were followed over 1 year, the number of people who experienced one or more falls would probably be 375 (95% CI 308 to 461) compared with 480 in the group receiving usual care or attention control.
	480 per 1000f	375 per 1000 (308 to 461)
	Specific exercise population
	374 per 1000f	296 per 1000 (243 to 364)
Number of people who experienced one or more fall‐related fractures	64 per 1000h	55 per 1000 (40 to 75)	RR 0.85 (0.62 to 1.16)	1810 (3 RCTs)	⊕⊕⊝⊝i low	Overall, there may be a reduction of 15% (95% CI 38% reduction to 16% increase) in the number of people who experienced one or more fall‐related fractures Guide to the data If 1000 people were followed over 1 year, the number of people who experienced one or more fall‐related fractures would probably be 55 (95% CI 40 to 75) compared with 64 in the group receiving usual care or attention control
Adverse events	See comment		Not estimable	1177 (10 RCTs)	⊕⊝⊝⊝j very low	Adverse events were reported in 10 of the 21 trials with multiple primary intervention categories, in the exercise versus control analyses in these trials. Adverse events were reported for both intervention and control groups (5 trials), or the intervention group only (5 trials). There were a total of 43 adverse events reported. Most were non‐serious of a musculoskeletal nature. There was reported exacerbation of pre‐existing osteoarthritis conditions in one trial and inguinal hernia surgery was reported in one intervention arm of another trial
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; RR: risk ratio
GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect

^aExercise programmes included in this analysis had more than one primary exercise category. We categorised exercise based on the Prevention of Falls Network Europe (ProFaNE) taxonomy that classifies exercise type as: i) gait, balance, and functional (task) training; ii) strength/resistance (including power); iii) flexibility; iv) three‐dimensional (3D) exercise (e.g. Tai Chi, Qigong, dance); v) general physical activity; vi) endurance; and vii) other kind of exercises. The programmes often included, as the primary intervention, gait, balance, and functional (task) training plus resistance training. The exercise programmes may also include secondary categories of exercise.
 ^bA control intervention is one that is not thought to reduce falls, such as general health education, social visits, very gentle exercise, or 'sham' exercise not expected to impact on falls.
 ^cThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 59 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the 11 RCTs.
 ^dSensitivity analyses revealed little difference in the results when only trials that include the most common two components (balance and functional exercises plus resistance exercises) were pooled (RaR 0.69, 95% CI 0.48 to 0.97; 1084 participants; 8 studies; I² = 72%).
 ^eDowngraded by one level due to inconsistency (there was substantial heterogeneity (I² = 65%)). We did not downgrade for risk of bias, as results were essentially unchanged with removal of the trials at a high risk of bias in one or more items.
 ^fThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 63 all‐exercise types RCTs. The specific exercise population risk was based on the number of events and the number of participants in the control group for this outcome over the 17 RCTs.
 ^gDowngraded by one level due to risk of bias (removing studies with high risk of bias in one or more items had a marked impact on results).
 ^hThe all‐studies population risk was based on the number of events and the number of participants in the control group for this outcome over the 10 all‐exercise types RCTs. Based on the number of events and the number of participants in the control group for this outcome over three RCTs, the assumed risk in the control group was 87 per 1000.
 ⁱDowngraded by one level due to risk of bias and by one level due to imprecision.
 ^jDowngraded by three levels for limitations in the design of studies, suggesting a high likelihood of bias (no trials in this analysis measured the number of participants experiencing adverse events in both groups throughout the trial period).

Exercise (all types) versus control

Overview of results reporting format

For each outcome described below we report the overall pooled effects of all exercise interventions (including the subgroup analyses for age, baseline risk of falling, personnel, and group delivery, for the falls outcomes) then the effects in studies testing interventions within each exercise category of the ProFaNE taxonomy (Appendix 1; Appendix 5), as well as the results of studies of exercise interventions that included multiple categories. For analyses with more than 10 included comparisons (both rate of falls and number of people experiencing one or more falls comparisons for balance and functional exercises, and multiple categories of exercise) we also report the results of the three prespecified subgroup analyses (increased fall risk as a study entry criterion, exercise delivery by a health professional, group versus individual delivery).

The findings are summarised and the absolute impact of interventions illustrated in 'Summary of findings' tables for the overall 'exercise versus control' comparison and for separate primary exercise categories for which there are data. No trials compared primarily flexibility exercise, endurance exercise or other exercise type versus control.

The results for the four trials comparing exercise (all types) versus control in people who had been recently discharged from hospital are presented separately, after this main comparison.

Rate of falls (falls per person‐year)

Exercise (all types) reduces the rate of falls by 23% compared with control (rate ratio (RaR) 0.77, 95% confidence interval (CI) 0.71 to 0.83; 12,981 participants, 59 studies, I² = 55%; high‐certainty evidence; Analysis 1.1).

1.1. Analysis.

Comparison 1 Exercise versus control (rate of falls), Outcome 1 Rate of falls ‐ overall analysis.

Subgroup analysis by falls risk at baseline, found there was probably little or no difference in the effect of exercise (all types) on the rate of falls in trials where all participants were at an increased risk of falling (RaR 0.80, 95% CI 0.72 to 0.88; 6858 participants, 30 studies, I² = 56%) compared with trials that did not use increased risk of falling as an entry criterion (RaR 0.74, 95% CI 0.65 to 0.84; 6123 participants, 29 studies, I² = 53%); test for subgroup differences: Chi² = 0.90, df = 1, P = 0.34, I² = 0% (Analysis 1.2).

1.2. Analysis.

Comparison 1 Exercise versus control (rate of falls), Outcome 2 Rate of falls ‐ subgrouped by baseline falls risk.

Subgroup analysis by participant age found there was probably little or no difference in the effect of exercise (all types) on the rate of falls in trials where participants were aged 75 years or older (RaR 0.83, 95% CI 0.72 to 0.97; 3376 participants, 13 studies, I² = 54%) compared with trials where participants were aged less than 75 years (RaR 0.75, 95% CI 0.69 to 0.82; 9605 participants, 46 studies, I² = 55%); test for subgroup differences: Chi² = 1.36, df = 1, P = 0.24, I² = 27% (Analysis 1.3).

1.3. Analysis.

Comparison 1 Exercise versus control (rate of falls), Outcome 3 Rate of falls ‐ subgrouped by age (threshold 75 years).

Subgroup analyses found a larger effect of exercise (all types) in trials where interventions were delivered by a health professional (RaR 0.69, 95% CI 0.61 to 0.79; 4511 participants, 25 studies, I² = 47%) than in trials where the interventions were delivered by trained instructors who were not health professionals (RaR 0.82, 95% CI 0.75 to 0.90; 8470 participants, 34 studies, I² = 57%); test for subgroup differences: Chi² = 4.44, df = 1, P = 0.04, I² = 78% (Analysis 1.4). Notably, both approaches resulted in reductions in the rate of falls.

1.4. Analysis.

Comparison 1 Exercise versus control (rate of falls), Outcome 4 Rate of falls ‐ subgrouped by personnel.

Subgroup analyses found there may be no difference in the effect of exercise (all types) on the rate of falls where interventions were delivered in a group setting (RaR 0.76, 95% CI 0.69 to 0.85; 8163 participants, 40 studies, I² = 62%) compared with trials where interventions were delivered individually (RaR 0.79, 95% CI 0.71 to 0.88; 4818 participants, 21 studies, I² = 35%); test for subgroup differences: Chi² = 0.21, df = 1, P = 0.65, I² = 0% (Analysis 1.5). Two three‐group studies, appear in both subgroups (Iliffe 2015; Wolf 1996).

1.5. Analysis.

Comparison 1 Exercise versus control (rate of falls), Outcome 5 Rate of falls ‐ subgrouped by group or individual exercise.

Subgroup analysis by exercise type showed a variation in the effects of the different types of exercise on rate of falls, the visual impression being confirmed by the statistically significant test for subgroup differences: Chi² = 17.18, df = 5, P = 0.004, I² = 70.9% (Analysis 1.6).

1.6. Analysis.

Comparison 1 Exercise versus control (rate of falls), Outcome 6 Rate of falls ‐ subgrouped by exercise type.

Different categories of primary exercise versus control

Balance and functional exercises versus control

Exercise interventions that were classified as being primarily gait, balance, co‐ordination or functional task training using the ProFaNE taxonomy, reduce the rate of falls by 24% compared with control (RaR 0.76, 95% CI 0.70 to 0.81; 7920 participants, 39 studies, I² = 29%, high‐certainty evidence; Analysis 1.6).

Subgroup analyses found little or no difference in the effect of balance and functional exercises on the rate of falls in trials where all participants were at an increased risk of falling (RaR 0.72, 95% CI 0.65 to 0.80; 4602 participants, 21 studies, I² = 38%) compared with trials that did not use increased risk of falling as an entry criterion (RaR 0.80, 95% CI 0.72 to 0.90; 3355 participants, 18 studies, I² = 17%); test for subgroup differences: Chi² = 1.99, df = 1, P = 0.16, I² = 50% (Analysis 8.1).

8.1. Analysis.

Comparison 8 Balance and functional exercises versus control: subgroup analyses, Outcome 1 Rate of falls, subgrouped by baseline fall risk.

Subgroup analyses found a larger effect of balance and functional exercises in trials where interventions were delivered by a health professional (RaR 0.67, 95% CI 0.58 to 0.65; 2960 participants, 20 studies, I² = 37%) than in trials where the interventions were delivered by trained instructors who were not health professionals (RaR 0.82, 95% CI 0.76 to 0.88; 4997 participants, 19 studies, I² = 9%); test for subgroup differences: Chi² = 6.72, df = 1, P = 0.01, I² = 85% (Analysis 8.3). Notably, both approaches resulted in statistically significant reductions in the rate of falls.

8.3. Analysis.

Comparison 8 Balance and functional exercises versus control: subgroup analyses, Outcome 3 Rate of falls, subgrouped by personnel.

Subgroup analyses found little or no difference in the effect of balance and functional exercises on the rate of falls in trials where interventions were delivered in a group setting (RaR 0.73, 95% CI 0.65 to 0.82; 3620 participants, 20 studies, I² = 34%) compared with trials where interventions were delivered individually (RaR 0.77, 95% CI 0.70 to 0.85; 4589 participants, 20 studies, I² = 28%); test for subgroup differences: Chi² = 0.47, df = 1, P = 0.50, I² = 0% (Analysis 8.5).

8.5. Analysis.

Comparison 8 Balance and functional exercises versus control: subgroup analyses, Outcome 5 Rate of falls, subgrouped by group or individual exercise.

Resistance exercises versus control

We are uncertain whether exercises, classified as being primarily resistance or strength exercises using the ProFaNE taxonomy, reduce the rate of falls compared with control (RaR 1.14, 95% CI 0.67 to 1.97; 327 participants, 5 studies, I² = 67%; very low‐certainty evidence; Analysis 1.6).

3D (Tai Chi) exercise versus control

Exercise interventions that were classified as 3D (Tai Chi or similar) may reduce the rate of falls by 19% compared with control (RaR 0.81, 95% CI 0.67 to 0.99; 2655 participants, 7 studies, I² = 74%; low‐certainty evidence; Analysis 1.6).

3D (dance) exercise versus control

We are uncertain whether exercises, classified as being primarily 3D (dance) using the ProFaNE taxonomy, reduce the rate of falls compared with control (RaR 1.34, 95% CI 0.98 to 1.83; 522 participants, 1 study; very low‐certainty evidence; Analysis 1.6).

Walking programme versus control

We are uncertain whether exercises, classified as being primarily walking programmes using the ProFaNE taxonomy, reduce the rate of falls compared with control (RaR 1.14, 95% CI 0.66 to 1.97; 441 participants, 2 studies; I² = 67%; very low‐certainty evidence; Analysis 1.6).

Multiple categories of exercise versus control

Exercise interventions that include multiple categories of the ProFaNE taxonomy (most commonly balance and functional exercises plus resistance exercises) probably reduce the rate of falls by 34% compared with controls (RaR 0.66, 95% CI 0.50 to 0.88; 1374 participants, 11 studies; I² = 65%; moderate‐certainty evidence; Analysis 1.6).

Sensitivity analyses revealed little difference in the results when we pooled only trials that include the most common two components (balance and functional exercises plus resistance exercises) (RaR 0.69, 95% CI 0.48 to 0.97; 1084 participants, 8 studies; I² = 72%; Analysis 19.1).

19.1. Analysis.

Comparison 19 Sensitivity analysis 8: multiple categories of exercise versus control excluding trials that do not include balance and strength training, Outcome 1 Rate of falls.

Subgroup analyses found there is probably little or no difference in the effect of exercise interventions that included multiple categories on the rate of falls in trials where all participants were at an increased risk of falling (RaR 0.77, 95% CI 0.63 to 0.94; 618 participants, 5 studies, I² = 0%) compared with trials that did not use increased risk of falling as an entry criterion (RaR 0.54, 95% CI 0.29 to 0.99; 763 participants, 6 studies, I² = 79%); test for subgroup differences: Chi² = 1.19, df = 1, P = 0.27, I² = 16.2% (Analysis 9.1).

9.1. Analysis.

Comparison 9 Multiple categories of exercise versus control: subgroup analyses, Outcome 1 Rate of falls, subgrouped by baseline fall risk.

Subgroup analyses found there is probably little or no difference in the effect of exercise interventions that included multiple categories on rate of falls in trials where interventions were delivered by health professionals (RaR 0.65, 95% CI 0.43 to 0.99; 653 participants, 3 studies, I² = 72%) compared with trials where interventions were delivered by trained instructors who were not health professionals (RaR 0.66, 95% CI 0.44 to 0.99; 751 participants; 8 studies, I² = 67%); test for subgroup differences: Chi² = 0, df = 1, P = 0.96, I² = 0% (Analysis 9.3).

9.3. Analysis.

Comparison 9 Multiple categories of exercise versus control: subgroup analyses, Outcome 3 Rate of falls, subgrouped by personnel.

Subgroup analyses found there is probably little or no difference in the effect of exercise interventions that included multiple categories on the rate of falls in trials where interventions were delivered in a group setting (RaR 0.64, 95% CI 0.46 to 0.89; 1194 participants, 10 studies, I² = 67%) compared with trials where interventions were delivered individually (RaR 0.81, 95% CI 0.56 to 1.18; 210 participants, 1 study); test for subgroup differences: Chi² = 0.86, df = 1, P = 0.35, I² = 0% (Analysis 9.5).

9.5. Analysis.

Comparison 9 Multiple categories of exercise versus control: subgroup analyses, Outcome 5 Rate of falls, subgrouped by group or individual exercise.

Long‐term follow‐up rate of falls (secondary outcome)

Five studies reported the rate of falls at more than 18 months after randomisation. Data from four studies, pooled by exercise category, are presented in Analysis 1.7. Balance and functional exercises may reduce the rate of falls in the long term (RaR 0.82, 95% CI 0.66 to 1.01; 858 participants, 2 studies; I² = 41%; low‐certainty evidence). The long‐term effects of a walking programme tested in Ebrahim 1997 (97 participants) and a multiple exercise programme, including balance and strength training tested in Uusi‐Rasi 2015 (175 participants) are unclear (Analysis 1.7). Data from Iliffe 2015 were not included in Analysis 1.7 because the follow‐up period, which differed from the other four studies, was a one‐year period started six months after programme completion. There was no evidence of a difference in rate of falls for either exercise programme (FaME programme: RaR 0.94, 95% CI 0.62 to 1.41; 202 participants; Otago Exercise Program: RaR 1.04, 95% CI 0.69 to 1.55; 201 participants).

1.7. Analysis.

Comparison 1 Exercise versus control (rate of falls), Outcome 7 Rate of falls ‐ long‐term follow‐up by exercise type.

Number of people who experienced one or more falls (risk of falling)

Exercise (all types) reduces the number of people experiencing one or more falls by 15% compared with control (risk ratio (RR) 0.85, 95% CI 0.81 to 0.89; 13,518 participants, 63 studies, I² = 26%; high‐certainty evidence; Analysis 2.1).

2.1. Analysis.

Comparison 2 Exercise versus control (number of fallers), Outcome 1 Number of fallers ‐ overall analysis.

Subgroup analysis by falls risk at baseline found there was little or no difference in the effect of exercise (all types) on the number of people experiencing one or more falls in trials where all participants were at an increased risk of falling (RR 0.87, 95% CI 0.83 to 0.91; 7171 participants, 35 studies, I² = 1%) compared with trials that did not use increased risk of falling as an entry criterion (RR 0.82, 95% CI 0.73 to 0.92; 6347 participants, 28 studies, I² = 45%); test for subgroup differences: Chi² = 0.94, df = 1, P = 0.33, I² = 0% (Analysis 2.2).

2.2. Analysis.

Comparison 2 Exercise versus control (number of fallers), Outcome 2 Number of fallers ‐ subgrouped by baseline fall risk.

Subgroup analysis by participant age found there was little or no difference in the effect of exercise (all types) on the number of people experiencing one or more falls in trials where participants were aged 75 years or older (RR 0.86, 95% CI 0.80 to 0.92; 3172 participants, 13 studies, I² = 0%) compared with trials where participants were aged less than 75 years (RR 0.85, 95% CI 0.79 to 0.91; 10,346 participants, 50 studies, I² = 33%); test for subgroup differences: Chi² = 0.07, df = 1, P = 0.79, I² = 0% (Analysis 2.3).

2.3. Analysis.

Comparison 2 Exercise versus control (number of fallers), Outcome 3 Number of fallers ‐ subgrouped by age (threshold 75 years).

Subgroup analyses by personnel delivering exercise found there was little or no difference in the effect of exercise (all types) on the number of people experiencing one or more falls in trials where interventions were delivered by a health professional (RR 0.82, 95% CI 0.74 to 0.91; 3747 participants, 26 studies, I² = 25%) than in trials where the interventions were delivered by trained instructors who were not health professionals (RR 0.86, 95% CI 0.81 to 0.92; 9726 participants, 36 studies, I² = 29%); test for subgroup differences: Chi² = 0.63, df = 1 (P = 0.43), I² = 0% (Analysis 2.4). The personnel providing the exercise programme was not identified in Park 2008.

2.4. Analysis.

Comparison 2 Exercise versus control (number of fallers), Outcome 4 Number of fallers ‐ subgrouped by personnel.

Subgroup analyses found there may be no difference in the effect of exercise (all types) on the number of people experiencing one or more falls in trials where interventions were delivered in a group setting (RR 0.83, 95% CI 0.78 to 0.90; 9219 participants, 48 studies, I² = 33%) compared with trials where interventions were delivered individually (RR 0.88, 95% CI 0.83 to 0.93; 4299 participants, 16 studies; I² = 0%); test for subgroup differences: Chi² = 1.14, df = 1, P = 0.29, I² = 12% (Analysis 2.5). One three‐group study appears in both subgroups (Iliffe 2015).

2.5. Analysis.

Comparison 2 Exercise versus control (number of fallers), Outcome 5 Number of fallers ‐ subgrouped by group or individual exercise.

The subgroup analysis by exercise type provided a visual impression of potential subgroup differences of effect of different exercises on the numbers of fallers, but the test for subgroup differences did not show a statistically significant result: test for subgroup differences: Chi² = 6.45, df = 5, P = 0.26, I² = 22.5% (Analysis 2.6).

2.6. Analysis.

Comparison 2 Exercise versus control (number of fallers), Outcome 6 Number of fallers ‐ subgrouped by exercise type.

Different categories of primary exercise versus control

Balance and functional exercises versus control

Exercise interventions that were classified as being primarily gait, balance, co‐ordination or functional task training using the ProFaNE taxonomy, reduce the number of people experiencing one or more falls by 13% compared with control (RR 0.87, 95% CI 0.82 to 0.91; 8288 participants, 37 studies, I² = 9%; high‐certainty evidence; Analysis 2.6).

Subgroup analyses found little or no difference in the effect of balance and functional exercises on the number of people experiencing one or more falls in trials where all participants were at an increased risk of falling (RR 0.86, 95% CI 0.81 to 0.91; 4639 participants, 22 studies, I² = 6%) compared with trials that did not use increased risk of falling as an entry criterion (RR 0.88, 95% CI 0.80 to 97; 3649 participants, 15 studies, I² = 18%); test for subgroup differences: Chi² = 0.21, df = 1, P = 0.65, I² = 0% (Analysis 8.2).

8.2. Analysis.

Comparison 8 Balance and functional exercises versus control: subgroup analyses, Outcome 2 Number of fallers, subgrouped by baseline fall risk.

Subgroup analyses found little or no difference in the effect of balance and functional exercises on the number of people experiencing one or more falls in trials where interventions were delivered by health professionals (RR 0.82, 95% CI 0.75 to 0.90; 2894 participants, 19 studies, I² = 5%) compared with trials where interventions were delivered by trained instructors who were not health professionals (RR 0.89, 95% CI 0.84 to 0.94; 5394 participants, 18 studies, I² = 11%); test for subgroup differences: Chi² = 1.71, df = 1, P = 0.19, I² = 41% (Analysis 8.4).

8.4. Analysis.

Comparison 8 Balance and functional exercises versus control: subgroup analyses, Outcome 4 Number of fallers, subgrouped by personnel.

Subgroup analyses also found little or no difference in the effect of balance and functional exercises on the number of people experiencing one or more falls in trials where interventions were delivered in a group setting (RR 0.87, 95% CI 0.80 to 0.95; 4465 participants, 22 studies, I² = 19%) compared with trials where interventions were delivered individually (RR 0.87, 95% CI 0.82 to 0.92; 4075 participants, 16 studies, I² = 0%); test for subgroup differences: Chi² = 0.01, df = 1 (P = 0.92), I² = 0% (Analysis 8.6).

8.6. Analysis.

Comparison 8 Balance and functional exercises versus control: subgroup analyses, Outcome 6 Number of fallers, subgrouped by group or individual exercise.

Resistance exercises versus control

We are uncertain whether exercise, classified as being primarily resistance or strength exercises, reduces the number of people experiencing one or more falls compared with control (RR 0.81, 95% CI 0.57 to 1.15; 163 participants, 2 studies, I² = 0%; very low‐certainty evidence; Analysis 2.6).

3D (Tai Chi) exercise versus control

Exercise interventions that were classified as 3D (Tai Chi or similar) reduce the number of people experiencing one or more falls by 20% compared with control (RR 0.80, 95% CI 0.70 to 0.91; 2677 participants, 8 studies, I² = 42%; high‐certainty evidence; Analysis 2.6).

3D (dance) exercise versus control

We are uncertain whether exercise, classified as being primarily 3D (dance), reduces the number of people experiencing one or more falls compared with control (RR 1.35, 95% CI 0.83 to 2.20; 522 participants, 1 study; very low‐certainty evidence; Analysis 2.6). We assessed the certainty of the evidence as very low due to there being wide CIs in the single trial.

Walking programme versus control

We are uncertain whether exercise, classified as being primarily walking programmes, reduces the number of people experiencing one or more falls compared with control (RR 1.05, 95% CI 0.71 to 1.54; 441 participants, 2 studies, I² = 50%; Analysis 2.6), as we assessed the certainty of the evidence as very low.

Multiple categories of exercise versus control

Exercise interventions that included multiple categories of the ProFaNE taxonomy probably reduce the number of people experiencing one or more falls by 22% compared with control (RR 0.78, 95% CI 0.64 to 0.96; 1623 participants, 17 studies, I² = 48%; moderate‐certainty evidence; Analysis 2.6).

Sensitivity analyses revealed little difference in the results when we pooled only trials that included the two most common components (balance and functional exercises plus resistance exercises) (RR 0.76, 95% CI 0.61 to 0.95; 1375 participants, 13 studies; I² = 53%; Analysis 19.2).

19.2. Analysis.

Comparison 19 Sensitivity analysis 8: multiple categories of exercise versus control excluding trials that do not include balance and strength training, Outcome 2 Number of fallers.

Subgroup analyses found there may be little or no difference in the effect of exercise interventions that included multiple categories on the number of people experiencing one or more falls in trials where all participants were at an increased risk of falling (RR 0.84, 95% CI 0.71 to 1.00; 913 participants, 10 studies, I² = 19%) compared with trials that did not use increased risk of falling as an entry criterion (RR 0.70, 95% CI 0.41 to 1.19; 710 participants, 7 studies, I² = 67%); test for subgroup differences: Chi² = 0.42, df = 1, P = 0.52, I² = 0% (Analysis 9.2).

9.2. Analysis.

Comparison 9 Multiple categories of exercise versus control: subgroup analyses, Outcome 2 Number of fallers, subgrouped by baseline fall risk.

Subgroup analyses found there may be little or no difference in the effect of exercise interventions that included multiple categories on the number of people experiencing one or more falls in trials where interventions were delivered by health professionals (RR 0.81, 95% CI 0.65 to 1.02; 867 participants, 8 studies, I² = 50%) compared with trials where interventions were delivered by trained instructors who were not health professionals (RR 0.70, 95% CI 0.45 to 1.10; 711 participants, 8 studies, I² = 57%); test for subgroup differences: Chi² = 0.34, df = 1, P = 0.56, I² = 0% (Analysis 9.4).

9.4. Analysis.

Comparison 9 Multiple categories of exercise versus control: subgroup analyses, Outcome 4 Number of fallers, subgrouped by personnel.

Subgroup analyses found there may be little or no difference in the effect of exercise interventions that included multiple categories on the number of people experiencing one or more falls in trials where interventions were delivered in a group setting (RR 0.77, 95% CI 0.60 to 1.00; 1301 participants, 14 studies, I² = 57%) compared with trials where interventions were delivered individually (RR 0.86, 95% CI 0.72 to 1.03; 322 participants, 3 studies, I² = 0%); test for subgroup differences: Chi² = 0.45, df = 1 (P = 0.50), I² = 0% (Analysis 9.6).

9.6. Analysis.

Comparison 9 Multiple categories of exercise versus control: subgroup analyses, Outcome 6 Number of fallers, subgrouped by group or individual exercise.

Long‐term follow‐up

Data from the three studies reporting on the number of people experiencing one or more falls at more than 18 months after randomisation are shown in Analysis 2.7. Balance and functional exercises may reduce the number of fallers in the long term (RR 0.86, 95% CI 0.78 to 0.94; 1325 participants, 2 studies; I² = 0%; low‐certainty evidence) but there is no evidence of difference for a multiple exercise programme (including balance and strength training) tested in Uusi‐Rasi 2015 (RR 1.01, 95% CI 0.74 to 1.38; 175 participants; low‐certainty evidence).

2.7. Analysis.

Comparison 2 Exercise versus control (number of fallers), Outcome 7 Number of fallers ‐ long‐term follow‐up by exercise type.

Number of people who experienced one or more fall‐related fractures

Exercise (all types) may reduce the number of people experiencing one or more fall‐related fractures by 27% compared with control (RR 0.73, 95% CI 0.56 to 0.95; 4047 participants, 10 studies, I² = 0%; low‐certainty evidence; Analysis 3.1).

3.1. Analysis.

Comparison 3 Exercise versus control (number of people with fractures), Outcome 1 Number of people who experienced one or more fall‐related fractures‐ overall analysis.

Subgroup analysis by falls risk at baseline found there may be little or no difference in the effect of exercise (all types) on the number of people experiencing one or more fall‐related fractures in trials where all participants were at an increased risk of falling (RR 0.80, 95% CI 0.60 to 1.07; 2792 participants, 5 studies, I² = 0) compared with trials that did not use increased risk of falling as an entry criterion (RR 0.48, 95% CI 0.26 to 0.91; 1255 participants, 5 studies, I² = 0%); test for subgroup differences: Chi² = 2.05, df = 1, P = 0.15, I² = 50.6% (Analysis 3.2).

3.2. Analysis.

Comparison 3 Exercise versus control (number of people with fractures), Outcome 2 Number of people who experienced one or more fall‐related fractures ‐ subgrouped by baseline falls risk.

Subgroup analyses found there may be little or no difference in the effect of exercise (all types) on the number of people experiencing one or more fall‐related fractures in trials where participants were aged 75 years or older (RR 0.61, 95% CI 0.31 to 1.20; 2740 participants, 3 studies, I² = 42%) compared with trials where participants were aged less than 75 years (RR 0.53, 95% CI 0.29 to 0.96; 1308 participants, 7 studies, I² = 0%); test for subgroup differences: Chi² = 0.1, df = 1, P = 0.75, I² = 0% (Analysis 3.3).

3.3. Analysis.

Comparison 3 Exercise versus control (number of people with fractures), Outcome 3 Number of people who experienced one or more fall‐related fractures ‐ subgrouped by age (threshold 75 years).

The subgroup analysis by exercise type did not show subgroup differences on the effects on fall‐related fractures: test for subgroup differences: Chi² = 4.22, df = 3, P = 0.24, I² = 28.9% (Analysis 3.4).

3.4. Analysis.

Comparison 3 Exercise versus control (number of people with fractures), Outcome 4 Number of people who experienced one or more fall‐related fractures ‐ subgrouped by exercise type.

Different categories of primary exercise versus control

Balance and functional exercises versus control

Exercise interventions that were classified as being primarily gait, balance, co‐ordination or functional task training using the ProFaNE taxonomy, may reduce the number of people experiencing one or more fall‐related fractures by 56% compared with control (RR 0.44, 95% CI 0.25 to 0.76; 2139 participants, 7 studies, I² = 0%; low‐certainty evidence; Analysis 3.4).

Resistance exercises versus control

We are uncertain whether exercises, classified as being primarily resistance or strength exercises using the ProFaNE taxonomy, reduce the number of people experiencing one or more fall‐related fractures compared with control (RR 0.97, 95% CI 0.14 to 6.49; 73 participants; 1 study; very low‐certainty of evidence due to single study with very wide CI; Analysis 3.4).

3D exercise versus control

We did not find any studies that looked at the impact of 3D exercises (Tai Chi or dance) on the number of people experiencing one or more fall‐related fractures compared with control.

Walking programme versus control

We are uncertain whether exercises, classified as being primarily walking programmes using the ProFaNE taxonomy, reduce the number of people experiencing one or more fall‐related fractures compared with control (RR 0.66, 95% CI 0.11 to 3.76; 97 participants, 1 study; very low‐certainty evidence due to a single study with very wide CI; Analysis 3.4).

Multiple categories of exercise versus control

Exercise interventions that include multiple categories of the ProFaNE taxonomy, may slightly reduce the number of people experiencing one or more fall‐related fractures compared with control; however, the 95% CI includes the possibility of both reduced and increased numbers of people experiencing fall‐related fractures (RR 0.85, 95% CI 0.62 to 1.16; 1810 participants, 3 studies, I² = 0%; low‐certainty evidence; Analysis 3.4).

Long‐term follow‐up

Three studies, each testing a different exercise category, reported the number of people who experienced fractures more than 18 months after randomisation (Dangour 2011; Ebrahim 1997; Gill 2016). The effect of exercise on fractures at long‐term follow‐up is unclear (RR 0.93, 95% CI 0.69 to 1.25; 2351 participants, 3 studies; very low‐certainty; Analysis 3.5). Only the data (6 versus 4 fractures at 24 months compared with 2 versus 3 at 12 months) for Ebrahim 1997 differed from that presented in the main analysis (Analysis 3.1).

3.5. Analysis.

Comparison 3 Exercise versus control (number of people with fractures), Outcome 5 Number of people who experienced one or more fall‐related fractures ‐ long‐term follow‐up by exercise type.

Number of people who experienced one or more falls that resulted in hospital admission

Only two studies reported this outcome (Clegg 2014; Gill 2016). We are uncertain of the finding that exercise (all types) makes little or no difference to the number of people who experience one or more falls requiring hospital admission compared with control (RR 0.78, 95% CI 0.51 to 1.18; 1705 participants, 2 studies, I² = 0%; very low‐certainty evidence, downgraded three levels due to high risk of bias, imprecision (wide CI) and because a large number of studies included in the review do not contribute data to the outcome; Analysis 4.1).

4.1. Analysis.

Comparison 4 Exercise versus control (number of people with falls that resulted in hospital admission), Outcome 1 Number of people who experienced one or more falls that resulted in hospital admission ‐ overall analysis.

Number of people who experienced one or more falls that required medical attention

Exercise (all types) may reduce the number of people who experience one or more falls requiring medical attention by 39% compared with control (RR 0.61, 95% CI 0.47 to 0.79; 1019 participants, 5 studies (7 comparisons), I² = 3%; low‐certainty evidence downgraded due to imprecision and risk of publication bias; Analysis 5.1).

5.1. Analysis.

Comparison 5 Exercise versus control (number of people with falls that required medical attention), Outcome 1 Number of people who experienced one or more falls that required medical attention‐ overall analysis.

Different categories of primary exercise versus control

Balance and functional exercises versus control

Exercise interventions that were classified as being primarily gait, balance, co‐ordination or functional task training using the ProFaNE taxonomy, may make little or no difference to the number of people who experienced one or more falls requiring medical attention compared with control (RR 0.76, 95% CI 0.54 to 1.09; 583 participants, 3 studies, I² = 0%; low‐certainty evidence; Analysis 5.2).

5.2. Analysis.

Comparison 5 Exercise versus control (number of people with falls that required medical attention), Outcome 2 Number of people who experienced one or more falls that required medical attention ‐ subgrouped by exercise type.

Resistance exercises versus control

Exercises classified as being primarily resistance or strength exercises using the ProFaNE taxonomy, may make little or no difference to the number of falls requiring medical attention compared with control (RR 0.92, 95% CI 0.47 to 1.80; 73 participants, 1 study; very low‐certainty evidence; Analysis 5.2).

3D (Tai Chi) exercise versus control

Exercise interventions that were classified as 3D (Tai Chi or similar) may reduce the number of falls requiring medical attention by 65% compared with control (RR 0.35, 95% CI 0.13 to 0.93; 188 participants, 1 study; low‐certainty evidence; Analysis 5.2).

Walking programme versus control

This outcome was not reported.

Multiple categories of exercise versus control

Exercise interventions that include multiple categories of the ProFaNE taxonomy, may reduce the rate of falls requiring medical attention (RR 0.44, 95% CI 0.29 to 0.66; 247 participants, 2 studies, I² = 0%; low‐certainty evidence; Analysis 5.2).

Long‐term follow‐up

Two studies reported on this outcome at more than 18 months after randomisation (Karinkanta 2007; Uusi‐Rasi 2015). Pooled data from these two studies showed exercise (all types) may reduce the number of people who experience one or more falls requiring medical attention in the long term (RR 0.54, 95% CI 0.37 to 0.78; 319 participants, 2 studies; low‐certainty evidence; Analysis 5.3). The same data from both studies were used in Analysis 5.1 and Analysis 5.3.

5.3. Analysis.

Comparison 5 Exercise versus control (number of people with falls that required medical attention), Outcome 3 Number of people who experienced one or more falls that required medical attention ‐ long‐term follow‐up pooled.

Health‐related quality of life

We were able to pool data from 15 of the 23 trials that assessed health‐related quality of life in people not recently discharged from hospital. Based on pooled standardised mean difference (SMD) results from the 15 trials (17 comparisons) that reported final scores, exercise interventions may make little or no difference to people's reported health‐related quality of life compared with those who received usual care or an attention control; however, the 95% CI includes the possibility of both increased and reduced quality of life (SMD ‐0.03, 95% CI ‐0.10 to 0.04; 3172 participants, 15 studies; I² = 76%; low‐quality evidence downgraded two levels due to inconsistency (there was considerable heterogeneity, 76%), and risk of bias (removing studies with high risk of bias on two or more items had a marked impact on results; Analysis 6.1).

6.1. Analysis.

Comparison 6 Exercise versus control (health‐related quality of life), Outcome 1 Health‐related quality of life‐ overall analysis.

Four trials (6 comparisons) reported end point scores using the EQ‐5D; the SMD converted back to mean difference (MD) ‐0.0026 points (95% CI ‐0.0086 to 0.0034) on the 0 to 1 EQ‐5D scale, which is less than the minimally important difference of 0.074 (Walters 2005). For the five trials that measured health‐related quality of life using SF‐36, converting these data to the SF‐36 scale (0 worst to 100 best) indicates that the estimated MD of 0.36 (95% CI ‐1.20 to 0.47) is not clinically important, as the minimally important difference is usually 3 to 5 (Walters 2003).

Appendix 16 provides summary information for all 23 trials including three post‐hospital studies and those which we could not include in the meta‐analysis (e.g. because they used unique outcome measures or reported median, IQR or P value), the results of which are similar to the above.

Number of people who experienced one or more adverse events

Twenty‐seven trials reported on adverse event to some degree (Appendix 13). Fourteen of the trials reporting on adverse events stated there were no adverse events.

Iliffe 2015 measured the number of people experiencing adverse events in both groups throughout the trial period and reported 59 events classified as 'adverse reactions' or 'possible adverse reactions' in the group receiving FaME intervention, 60 in the OEP group and 45 in the control group; the majority were reports of musculoskeletal pain and none were serious. No other studies reported adverse events that were monitored closely in all groups over the entire study period. A serious adverse effect was a pelvic stress fracture reported in Clemson 2012. The remaining trials reported non‐serious adverse events of a musculoskeletal nature, with a median of three events (range 1 to 26) in the intervention group. The majority of reported adverse events were of a musculoskeletal nature and not serious. Of the studies that reported adverse events, a greater proportion of the strength‐only exercises were associated with adverse events than in the gait, balance and functional training or multiple exercise categories.

Different categories of primary exercise

Balance and functional exercises versus control

Adverse events were reported in 15 of the 48 trials, including exercise interventions that were classified as being primarily gait, balance, co‐ordination or functional task training using the ProFaNE taxonomy. Two hundred adverse events were reported; most were non‐serious adverse events of a musculoskeletal nature, one trial (two intervention arms) reported 128 of these adverse events (Iliffe 2015), one intervention arm reported shortness of breath in four participants (Liu‐Ambrose 2004), another trial reported palpitations in a participant (Sakamoto 2013), and one trial reported a pelvic stress fracture (Clemson 2012). See Appendix 13.

Resistance exercises versus control

Adverse events were reported in one trial, including exercises classified as being primarily resistance or strength exercises using the ProFaNE taxonomy (Liu‐Ambrose 2004). The study reported 10 musculoskeletal complaints in the intervention group and one musculoskeletal complaint in the control group.

3D (Tai Chi) exercise versus control

Adverse events were reported in two of 10 trials with 3D (Tai Chi) as the primary intervention. There were zero occurrences of adverse events.

3D (dance) exercise versus control

Adverse events were reported in the one trial in this analysis, in the intervention group only. There were zero occurrences of adverse events.

Walking programme versus control

This outcome was not reported.

Multiple categories of exercise versus control

Adverse events were reported in 10 of the 21 trials of exercise interventions that include multiple categories of the ProFaNE taxonomy. Adverse events were reported for both intervention and control groups (5 trials), or the intervention group only (5 trials). There was a total of 43 adverse events reported. The majority were non‐serious and of a musculoskeletal nature. There was reported exacerbation of pre‐existing osteoarthritis conditions (Uusi‐Rasi 2015), and inguinal hernia surgery was reported in one intervention arm (Clemson 2012).

Number of people who died

Death was primarily reported as a reason for loss to follow‐up in all 30 trials with separate group data. Exercise (all types) may reduce the number of people who died compared with control; however, the 95% CI includes the possibility of both reduced death and increased death with exercise (RR 0.86, 95% CI 0.66 to 1.12; 10,037 participants, 30 studies, I² = 0%; low‐certainty evidence (downgraded one level due to risk of bias, as results changed, becoming statistically significant, with removal of the 14 trials with a high risk of bias on one or more items; and one level for indirectness, as the outcome was assessed indirectly as a reason for loss to follow‐up; Analysis 7.1). The risk of death did not differ between the trials including people selected or not‐selected for risk of falling: test for subgroup differences: Chi² = 0.19, df = 1, P = 0.67, I² = 0% (Analysis 7.2). None of the deaths were explicitly linked to trial participation.

7.1. Analysis.

Comparison 7 Exercise versus control (number of people who died), Outcome 1 Number of people who died‐ overall analysis.

7.2. Analysis.

Comparison 7 Exercise versus control (number of people who died), Outcome 2 Number of people who died ‐ subgrouped by baseline fall risk.

Exercise (all types) versus control tested in people who had recently been discharged from hospital

Four studies investigated outcomes in people who had recently been discharged from hospital (Haines 2009; Latham 2003; Sherrington 2014; Vogler 2009). Results of individual studies for rate of falls (3 trials) are shown in Analysis 10.1; number of falls (4 trials) in Analysis 10.2; health‐related quality of life (3 trials) in Analysis 10.3; and mortality (4 trials) in Analysis 10.4. Given the diversity of interventions, we did not pool data. It is noted that overall, the effects of exercise on falls appear smaller (or in the opposite direction in the case of Sherrington 2014) in these studies compared with studies in the general older population (very low‐certainty evidence).

10.1. Analysis.

Comparison 10 Exercise versus control (by exercise type, in people after hospital stays), Outcome 1 Rate of falls.

10.2. Analysis.

Comparison 10 Exercise versus control (by exercise type, in people after hospital stays), Outcome 2 Number of fallers.

10.3. Analysis.

Comparison 10 Exercise versus control (by exercise type, in people after hospital stays), Outcome 3 Health‐related quality of life.

10.4. Analysis.

Comparison 10 Exercise versus control (by exercise type, in people after hospital stays), Outcome 4 Number of people who died.

All four studies reported on adverse events to some degree (Appendix 13). Latham 2003 measured the number of people experiencing adverse events in both groups throughout the trial period and reported that 18 participants had back and knee pain directly attributable to the exercise programme; there were no details of the five participants with adverse events in the control group. The remaining trials reported non‐serious adverse events of a musculoskeletal nature.

Exercise versus exercise

Comparisons of different types of exercise

The results of individual trials directly comparing different types of exercise are shown for rate of falls in Analysis 11.1, with long‐term rate of falls data in Analysis 11.2; number of fallers in Analysis 11.3; number with fall‐related fractures in Analysis 11.4; number requiring medical attention in Analysis 11.5; quality of life in Analysis 11.6; and mortality in Analysis 11.7. Given the variability between programmes, we did not undertake any meta‐analyses for these comparisons for any of the outcomes. Overall there is very low‐certainty evidence for each comparison.

11.1. Analysis.

Comparison 11 Exercise versus exercise, Outcome 1 Rate of falls, different types of exercise compared.

11.2. Analysis.

Comparison 11 Exercise versus exercise, Outcome 2 Rate of falls >18 months, different types of exercise compared.

11.3. Analysis.

Comparison 11 Exercise versus exercise, Outcome 3 Number of fallers, different types of exercise compared.

11.4. Analysis.

Comparison 11 Exercise versus exercise, Outcome 4 Number of people who experienced one or more fall‐related fractures, different types of exercise compared.

11.5. Analysis.

Comparison 11 Exercise versus exercise, Outcome 5 Number of people who experienced one or more falls that required medical attention, different types of exercise compared.

11.6. Analysis.

Comparison 11 Exercise versus exercise, Outcome 6 Quality of life, different types of exercise compared.

11.7. Analysis.

Comparison 11 Exercise versus exercise, Outcome 7 Number of people who died, different types of exercise compared.

Most of the trials in these analyses did not find significant differences in the fall prevention effects of different programmes, but most were not likely to be adequately powered to detect differences between different exercise programmes.

A few studies did find greater effects of particular programmes. For example, Kemmler 2010 found greater effects on the rate of falls of a more intensive programme delivered twice a week compared with a low intensity programme delivered once a week. Studies by Yamada et al found greater fall prevention effects of complex obstacle negotiation training compared with simple training (Yamada 2012), and greater effects of multidimensional stepping compared with walking (Yamada 2013). Both these interventions were delivered in addition to group exercise primarily targeting balance. Hwang 2016 found greater effects of Tai Chi than supervised balance and strength training on the rate of falls and the number of people falling. All these findings require confirmation in different and larger studies.

Different modes of delivery (e.g. group versus individual) of the same type of exercise

The results of individual trials that provided direct comparisons between the same programmes being delivered in group‐based settings and individually are shown for rate of falls in Analysis 11.8; number of fallers in Analysis 11.9; number requiring hospital admission in Analysis 11.10; quality of life in Analysis 11.11; and mortality in Analysis 11.12. All results were inconclusive; the five trials were too small to draw conclusions (Barker 2016; Helbostad 2004; Iliffe 2015; Kyrdalen 2014; Wu 2010).

11.8. Analysis.

Comparison 11 Exercise versus exercise, Outcome 8 Rate of falls, group vs individual exercise delivery within the same type of exercise.

11.9. Analysis.

Comparison 11 Exercise versus exercise, Outcome 9 Number of fallers, group vs individual exercise delivery within the same type of exercise.

11.10. Analysis.

Comparison 11 Exercise versus exercise, Outcome 10 Number of people who experienced one or more falls requiring hospital admission, group vs individual exercise delivery within the same type of exercise.

11.11. Analysis.

Comparison 11 Exercise versus exercise, Outcome 11 Health‐related quality of life, group vs individual exercise delivery within the same type of exercise.

11.12. Analysis.

Comparison 11 Exercise versus exercise, Outcome 12 Number of people who died, group vs individual exercise delivery within the same type of exercise.

Different doses (e.g. higher intensity versus lower intensity) of the same type of exercise

The results of the individual trials that directly compared higher with lower doses of the same type of exercise are shown for rate of falls in Analysis 11.13, number of fallers in Analysis 11.14, and mortality in Analysis 11.15. Taylor 2012 found a greater impact on the rate of falls when Tai Chi classes were delivered twice rather than once per week. The other two trials were too small to draw conclusions (Ballard 2004; Davis 2011).

11.13. Analysis.

Comparison 11 Exercise versus exercise, Outcome 13 Rate of falls, higher vs lower dose within the same type of exercise.

11.14. Analysis.

Comparison 11 Exercise versus exercise, Outcome 14 Number of fallers, higher vs lower dose within the same type of exercise.

11.15. Analysis.

Comparison 11 Exercise versus exercise, Outcome 15 Number of people who died, higher vs lower dose within the same type of exercise.

Number of people who experienced one or more adverse events

No studies reported adverse events that were monitored closely in all groups over the entire study period. Adverse events reported to any degree are described in Appendix 13. Three of the 10 trials reporting on adverse events stated there were no adverse events. The remaining trials reported non‐serious adverse events of a musculoskeletal nature.

Economic data

We identified 12 out of the 108 studies that reported economic data. These included reports of costs of intervention or health service use and/or the results of trial‐based cost‐effectiveness or cost‐utility analyses (Appendix 17).

As in Gillespie 2012, the perspectives taken, the cost items measured and valued, and the type of healthcare resources included in the calculation of incremental cost‐effectiveness ratios (ICERs) all varied, so that comparison of ICERs for the interventions remains difficult even for evaluations carried out within similar health systems.
 
 Nonetheless, the results from several studies demonstrate the potential cost‐effectiveness of fall prevention interventions. One trial of the Otago Exercise Program showed cost savings in those aged 80 years and over resulting from fewer hospital admissions (Robertson 2001a). Davis 2011 reported that both once and twice weekly resistance training dominated control (balance and tone) classes in terms of both falls and quality‐adjusted life years (i.e. were less costly and more effective).

Other studies provide information on the cost per fall prevented from the delivery of exercise interventions. For example, Voukelatos 2007 reported AUD 1683 per fall prevented from group‐based Tai Chi and Davis 2009 reports a cost of CAD 247 per fall prevented from a group‐based exercise programme compared with guideline‐based care.

Sensitivity analyses

For each of these, the impact on the pooled exercise versus control fall rate outcome is summarised in Appendix 18. The results of the sensitivity analyses can be seen in Analyses 12 to 20.

1. Sensitivity analysis 1, removing trials that included participants aged < 65 years: Analysis 12.1 (rate of falls: pooled data); Analysis 12.2 (rate of falls: grouped by exercise); Analysis 12.3 (number of fallers: pooled data); Analysis 12.4 (number of fallers: grouped by exercise); Analysis 12.5 (fracture: pooled data); Analysis 12.6 (fracture: grouped by exercise type); Analysis 12.7 (medical attention: pooled data); Analysis 12.8 (medical attention: subgrouped by exercise).

2. Sensitivity analysis 2, removing trials with high risk of bias on any item: Analysis 13.1 (rate of falls: pooled data); Analysis 13.2 (rate of falls: subgrouped by exercise); Analysis 13.3 (number of fallers: pooled data); Analysis 13.4 (number of fallers: subgrouped by exercise type); Analysis 13.5 (fracture: pooled data).

3. Sensitivity analysis 3, removing trials with unclear or high risk of bias on allocation concealment: Analysis 14.1 (rate of falls: pooled data).

4. Sensitivity analysis 4, removing trials with unclear or high risk of bias on assessor blinding: Analysis 15.1 (rate of falls: pooled data).

5. Sensitivity analysis 5, removing trials with unclear or high risk of bias on incomplete outcome data: Analysis 16.1 (rate of falls: pooled data).

6. Sensitivity analysis 6, removing cluster‐randomised trials: Analysis 17.1 (rate of falls: pooled data).

7. Sensitivity analysis 7, all trials, fixed‐effect meta‐analysis: Analysis 18.1 (rate of falls: pooled data).

8. Sensitivity analysis 8, multiple categories of exercise versus control, removing trials that do not include balance and strength training: Analysis 19.1 (rate of falls: pooled data); Analysis 19.2 (number of fallers: pooled data).

9. Sensitivity analysis 9a, classification of interventions based on the Otago Exercise Program as multiple categories of exercise: Analysis 20.1 (rate of falls: pooled data); Analysis 20.2 (number of fallers: pooled data).

10. Sensitivity analysis 9b, classification of interventions that included balance and functional exercises plus strength exercises as multiple categories of exercise: Analysis 20.3 (rate of falls: pooled data); Analysis 20.4 (number of fallers: pooled data).

12.1. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 1 Rate of falls: pooled data.

12.2. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 2 Rate of falls: grouped by exercise type.

12.3. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 3 Number of fallers: pooled data.

12.4. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 4 Number of fallers: grouped by exercise type.

12.5. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 5 Number of people who experienced one or more fall‐related fractures: pooled data.

12.6. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 6 Number of people who experienced one or more fall‐related fractures: by exercise type.

12.7. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 7 Number of people who experienced one or more falls requiring medical attention: pooled data.

12.8. Analysis.

Comparison 12 Sensitivity analysis 1: exercise versus control excluding studies that included people < 65 years, Outcome 8 Number of people who experienced one or more falls requiring medical attention ‐ subgrouped by exercise type.

13.1. Analysis.

Comparison 13 Sensitivity analysis 2: exercise versus control excluding studies at a high risk of bias, Outcome 1 Rate of falls ‐ overall analysis.

13.2. Analysis.

Comparison 13 Sensitivity analysis 2: exercise versus control excluding studies at a high risk of bias, Outcome 2 Rate of falls ‐ subgrouped by exercise type.

13.3. Analysis.

Comparison 13 Sensitivity analysis 2: exercise versus control excluding studies at a high risk of bias, Outcome 3 Number of fallers ‐ overall analysis.

13.4. Analysis.

Comparison 13 Sensitivity analysis 2: exercise versus control excluding studies at a high risk of bias, Outcome 4 Number of fallers ‐ subgrouped by exercise type.

13.5. Analysis.

Comparison 13 Sensitivity analysis 2: exercise versus control excluding studies at a high risk of bias, Outcome 5 Number of people who experienced one or more fall‐related fractures ‐ overall analysis.

14.1. Analysis.

Comparison 14 Sensitivity analysis 3: exercise versus control excluding studies with unclear or high risk of bias due to allocation concealment (rate of falls), Outcome 1 Rate of falls ‐ overall analysis.

15.1. Analysis.

Comparison 15 Sensitivity analysis 4: exercise versus control excluding studies with unclear or high risk of bias due to assessor blinding (rate of falls), Outcome 1 Rate of falls ‐ overall analysis.

16.1. Analysis.

Comparison 16 Sensitivity analysis 5: exercise versus control excluding studies with unclear or high risk of bias due to incomplete outcome data (rate of falls), Outcome 1 Rate of falls ‐ overall analysis.

17.1. Analysis.

Comparison 17 Sensitivity analysis 6: exercise versus control excluding cluster trials (rate of falls), Outcome 1 Rate of falls ‐ overall analysis.

18.1. Analysis.

Comparison 18 Sensitivity analysis 7: exercise versus control with fixed‐effect meta‐analysis (rate of falls), Outcome 1 Rate of falls ‐ overall analysis.

20.1. Analysis.

Comparison 20 Sensitivity analysis 9: different exercise type coding, Outcome 1 Rate of falls ‐ subgrouped by exercise type (OEP as multiple intervention).

20.2. Analysis.

Comparison 20 Sensitivity analysis 9: different exercise type coding, Outcome 2 Number of fallers ‐ subgrouped by exercise type (OEP as multiple intervention).

20.3. Analysis.

Comparison 20 Sensitivity analysis 9: different exercise type coding, Outcome 3 Rate of falls ‐ subgrouped by exercise type (any balance+strength as multiple intervention).

20.4. Analysis.

Comparison 20 Sensitivity analysis 9: different exercise type coding, Outcome 4 Number of fallers ‐ subgrouped by exercise type (any balance+strength as multiple intervention).

As shown in Appendix 18; the nine sensitivity analyses (based on age of included participants, risk of bias, cluster trials, fixed‐effect analyses, and categorisation of interventions) made little difference to the results of the primary pooled analysis. This indicates the robustness of the review's primary findings and methods.

In undertaking the GRADE assessment we downgraded the certainty of evidence based on sensitivity analysis (removal of trials with one or more items at high risk of bias) for the following comparisons.

1. Fall outcome: resistance exercises versus control, Tai Chi versus control, walking programme versus control.

2. Faller outcome: resistance exercises versus control, walking programme versus control, multiple categories of exercise versus control.

3. Fracture outcome: exercise (all types) versus control, balance and functional exercises versus control, multiple versus control.

4. Health‐related quality of life outcome: exercise (all types) versus control.

Heterogeneity

This review's primary analyses display minimal to substantial heterogeneity with P < 0.05 for the Chi² test and I² values up to 74%. This variability was not explained by our subgroup analyses. We consider this likely to represent between‐study differences in the exact nature of programmes (e.g. dose, intensity, adherence) and target populations, which requires ongoing investigation. Given the overall positive impact of the programmes and the stability of results, we do not consider this to preclude the meta‐analyses we have undertaken.

Funnel plots

The funnel plots in Figure 4; Figure 5; Figure 6; Figure 7; Figure 8 and Figure 9 do show some asymmetry, particularly for the fracture outcomes. We used this information in the GRADE assessment to downgrade the strength of the evidence for the fracture outcomes but did not consider the asymmetry sufficient to downgrade the level of evidence for the other outcomes.

4.

Funnel plot of comparison: 1 Exercise versus control (rate of falls), outcome: 1.1 Rate of falls ‐ overall analysis.

5.

Funnel plot of comparison: 2 Exercise versus control (number of fallers), outcome: 2.1 Number of fallers ‐ overall analysis.

6.

Funnel plot of comparison: 3 Exercise versus control (number of people with fractures), outcome: 3.1 Number of people who experienced one or more fall‐related fractures‐ overall analysis.

7.

Funnel plot of comparison: 6 Exercise versus control (health‐related quality of life), outcome: 6.1 Health‐related quality of life‐ overall analysis.

8.

Funnel plot of comparison: 1 Exercise versus control (rate of falls), outcome: 1.6 Rate of falls ‐ subgrouped by exercise type.

9.

Funnel plot of comparison: 2 Exercise versus control (number of fallers), outcome: 2.6 Number of fallers ‐ subgrouped by exercise type.

Discussion

Summary of main results

This review includes 108 trials with 23,407 participants, who were older people living in the community. Of these, 81 trials (19,684 participants) contributed the evidence for the main 'exercise versus control' intervention (one that is not thought to reduce falls) comparison; these did not include the four trials that included only people who had been recently discharged from hospital. After summarising the results for this comparison, we summarised the evidence for the primary exercise categories versus control comparisons, where data were available. Our illustrative risks for dichotomous outcomes presented in Table 1, are based on counts (number of events divided by the number of participants) for those trials included in the analysis for that outcome. In Table 1, we also based our illustrative risks for falls outcomes on the median values obtained from the subgroups of trials for which: a) an increased risk of falls was not an inclusion criterion (not selected population); or b) increased risk of falls was an inclusion criterion. In the other 'Summary of findings' tables, we used the 'all‐exercise versus control' studies risks to illustrate the absolute risks for falls and fracture outcomes; we supplemented the falls outcomes by illustrative risks based on count data for the specific exercise category summarised.

Exercise (all types) versus control

There is high‐certainty evidence that falls can be prevented by exercise programmes, as summarised in Table 1. Exercise reduces both the rate of falls (reported in 59 randomised controlled trials (RCTs)) and the number of people experiencing falls (reported in 63 RCTs). Subgroup analyses did not reveal differences in effect on both falls outcomes according to whether trials were selected for high risk of falling or not. Hence, the overall rate of falls and number of fallers results were applied when estimating absolute risks in the following lower and higher risk of falls categories. As shown below, the absolute numbers of falls or numbers of fallers prevented are greater in the higher risk populations.

1. For the overall risk category, based on an illustrative risk of 850 falls per 1000 person‐years in the control group, there were 195 (23%) fewer falls per 1000 person‐years in the exercise group (95% confidence interval (CI) 144 (17%) to 246 (29%) fewer). Based on an illustrative risk of 480 fallers per 1000 older people in the control group, there were 72 (15%) fewer fallers per 1000 older people in the exercise group (95% CI 52 (11%) to 91 (19%) fewer).

2. For the non‐selected lower risk category, based on an illustrative risk of 605 falls per 1000 person‐years in the control group, there were 139 (23%) fewer falls per 1000 person‐years in the exercise group (95% CI 102 (17%) to 175 (29%) fewer). Based on an illustrative risk of 380 fallers per 1000 older people in the control group, there were 57 (15%) fewer fallers per 1000 older people in the exercise group (95% CI 41 (11%) to 72 (19%) fewer).

3. For the selected higher risk category, based on an illustrative risk of 1200 falls per 1000 person‐years in the control group, there were 276 (23%) fewer falls per 1000 person‐years in the exercise group (95% CI 204 (17%) to 348 (29%) fewer). Based on an illustrative risk of 500 fallers per 1000 older people in the control group, there were 75 (15%) fewer fallers per 1000 older people in the exercise group (95% CI 55 (11%) to 95 (19%) fewer).

Subgroup analyses did not reveal differences in effect on both falls outcomes according to whether trials included younger and older populations based on a 75 year cut‐off. There was, however, a greater reduction on the rate of falls from exercises (all types) in trials where interventions were delivered by a health professional than in trials where trained instructors who were not health professionals delivered the interventions; however, both approaches reduced the rate of falls. This finding did not apply to the subgroup analysis for number of fallers. Subgroup analyses did not reveal differences in effect on both falls outcomes according to whether interventions were delivered in a group setting or delivered individually.

The test for subgroup differences for when subgrouped by exercise type revealed significant subgroup differences for rate of falls, a finding that endorsed our prespecified intention to report separate analyses by primary exercise type (see below).

Far fewer studies reported on number of people who experienced fall‐related fractures (10 RCTs), fall‐related hospital admission (2 RCTs) and medical attention (5 RCTs). Exercise may reduce the number of people with fall‐related fractures: 27% reduction, 95% CI 5% to 44% reduction. Based on an illustrative risk, derived from the study data, of 64 people with fall‐related fractures per 1000 older people in the control group, there were 17 fewer people with fall‐related fractures per 1000 older people in the exercise group (95% CI 3 to 28 fewer). Exercise may make little or no difference to the number of people who experience one or more falls requiring hospital admission; reduction 22%, 95% CI 49% reduction to 18% increase. Based on an illustrative risk of 57 people with fall‐related hospital admission per 1000 older people in the control group, there were 12 fewer people with fall‐related hospital admissions per 1000 older people in the exercise group (95% CI 28 fewer to 11 more). Exercise may reduce the number of people who experience one or more falls requiring medical attention: 39% reduction, 95% CI 21% to 53% reduction. Based on an illustrative risk of 211 people with falls that required medical attention per 1000 older people in the control group, there were 82 fewer people with fall‐related medical attention per 1000 older people in the exercise group (95% CI 44 to 111 fewer).

Exercise may make little important difference to people‐reported health‐related quality of life compared with control: conversion of the pooled result (standardised mean difference (SMD) ‐0.03, 95% CI ‐0.10 to 0.04; 15 RCTs) to the EQ‐5D and SF‐36 scores showed the respective 95% CIs were much smaller than minimally important differences for both scales.

We are uncertain of the evidence for adverse events, which were incompletely reported and mainly for the exercise groups only in 27 RCTs (6019 participants). Fourteen trials reported no adverse events. Aside from two serious adverse events (1 pelvic stress fracture and 1 inguinal hernia surgery) reported in one trial, the remainder were non‐serious adverse events, primarily of a musculoskeletal nature.

Different exercise types versus control

'Summary of findings' tables, summarising the evidence for the rate of falls, risk of falling, fall‐related fractures and adverse events, are presented for the primary exercise categories for which data are available. There are no data available for flexibility exercise or endurance exercise versus control. The following should be viewed in terms of the data available for each exercise type. The few direct comparisons of different exercise types were clinically heterogeneous and we did not undertake any meta‐analyses for these comparisons for any of the outcomes.

Balance and functional exercises

This was compared with control in 48 trials. As summarised in Table 2, there is high‐certainty evidence that balance and functional exercises reduce the rate of falls and the number of people who experience falls. There is low‐certainty evidence that this type of exercise programme may help prevent fall‐related fractures. Adverse events, which were incompletely reported, were mainly non‐serious adverse events of a musculoskeletal nature.

Resistance (strength) exercises

This was compared with control in seven trials. As summarised in Table 3, we are uncertain of the effects of resistance training on the rate of falls and number of fallers. We are uncertain of the effects on fall‐related fractures; only three participants had fractures in the single trial reporting this outcome. Adverse events, which were incompletely reported, were non‐serious adverse events of a musculoskeletal nature.

3D exercise: Tai Chi

This was compared with control in 10 trials. As summarised in Table 4, there is low‐certainty evidence that Tai Chi may reduce the rate of falls and high‐certainty evidence that Tai Chi reduces the number of people who experience falls. Fall‐related fractures were not reported. The two trials reporting on adverse events, reported none.

3D exercise: dance

This was compared with control in one trial. As summarised in Table 5, we uncertain of findings of little effect of dance training on rate of falls or numbers of fallers. Fall‐related fractures were not reported. The trials reported there had been no adverse events in the dance group.

General physical activity: walking programme

This was compared with control in three trials. As summarised in Table 6, we are uncertain of the effects of walking programmes on rate of falls and number of people who experience falls. We are uncertain of the effects on fall‐related fractures; only 10 participants had fractures in the single trial reporting this outcome. All three trials reported there had been no adverse events.

Multiple categories of exercise

Multiple categories of exercise (most commonly balance and functional exercises plus resistance exercises) were compared with control in 21 trials. As summarised in Table 7, there is moderate‐certainty evidence that these interventions probably reduce rate of falls and number of fallers. Sensitivity analyses revealed little difference in the results when only the trials that included the most commonly two components (balance and functional exercises plus resistance exercises) as primary outcomes were pooled. Sensitivity analyses also revealed little difference in the results when any intervention that included balance and functional exercises plus strength exercises, as primary or secondary interventions, was classified as multiple types of exercise (Appendix 18). There is low‐certainty evidence that these interventions may have little effect on fall‐related fractures. Adverse events, which were incompletely reported, were mainly non‐serious adverse events of a musculoskeletal nature.

Subgroup analyses

Our prespecified subgroup analyses were performed on falls outcomes for balance and functional exercises and multiple categories of exercise. As for the overall exercise versus control comparison, subgroup analysis did not suggest a difference in effects on falls outcomes between trials that used increased risk of falls as an inclusion criterion to those in trials that did not. Also consistent with the overall exercise versus control comparison, there was greater reduction on the rate of falls from balance and functional exercises in trials where interventions were delivered by a health professional than in trials where the interventions were delivered by trained instructors who were not health professionals; although both approaches resulted in reductions in the rate of falls. There was no difference in the reduction on rate of falls from multiple primary types of exercise in trials where interventions were delivered by a health professional than in trials where the intervention was not delivered by a health professional. Other subgroup analyses did not detect differences in effects of exercises in trials where interventions were delivered in a group setting compared with trials where interventions were delivered individually. We did not explore the interaction between subgroups. For example, higher risk people are likely to require health professional input for safe exercise prescription.

Adverse events

Forty‐one of the 108 included trials reported on adverse events to some degree (31 exercise versus control trials, of which four trials included people recently discharged from hospital, and 10 exercise versus exercise trials). Seventeen trials reported an absence of adverse events, one trial reported a pelvic fracture and an inguinal hernia surgery (Clemson 2012), and the remaining trials primarily reported non‐serious musculoskeletal events. Only two trials, one of which included post‐discharge from hospital participants, reported adverse events in both exercise and control groups over the whole trial period, perhaps reflecting the cost and complexity of such monitoring.

Exercise (all types) versus control in people who had recently been discharged from hospital

Four heterogeneous studies investigated outcomes in people who had recently been discharged from hospital. We did not pool the data available for rate of falls, number of fallers and health‐related quality of life given the small numbers of trials and diversity of the interventions. Overall, the very low‐certainty evidence, downgraded for risk of bias, inconsistency and imprecision evidence is insufficient to draw any conclusions.

Comparisons of different types, modes of delivery and doses of exercise

Given the variability between programmes, we did not undertake any meta‐analyses of comparisons between different types of exercise. Most of the trials in these analyses did not find significant differences in the fall prevention effects of different programmes, but most were not likely to be adequately powered to detect differences between different exercise programmes. When comparing different exercise types delivered within the same studies we found some indication that higher doses of exercise were associated with a greater impact on the rate of falls and the number of people falling.

Economic data

Of the 12 studies included in this review that reported economic evaluation, some give an indication of value for money for the interventions tested. Variations in the methods used, however, made comparisons across studies difficult. There was some, although limited, evidence that fall prevention strategies can be cost‐saving during the trial period, and may also be cost‐effective over the participants’ remaining lifetime; however, it should be noted that these analyses usually fail to include the cost of identifying the target population, which can be substantial and can impact on cost‐effectiveness measures (Eldridge 2005). Additional studies have modelled the impact and cost‐effectiveness of a public health falls prevention programme in Australia (Farag 2015), undertaken secondary analyses to estimate cost‐effectiveness of implementing the Otago Exercise Program in Norway (Hektoen 2009), performed cost–benefit analysis of fall prevention interventions (Campbell 1999; Carande‐Kulis 2015; Clemson 2004a; Li 2005), and undertaken a literature review and developed a tool to estimate the cost‐effectiveness of fall prevention interventions in the community (Public Health England 2018).

Overall completeness and applicability of evidence

Trial design and participants

The 108 trials included in this review included 23,407 community‐dwelling older people, who were predominantly women (77%). A wide range of ages were included as few trials set upper age limits. Participant characteristics varied greatly due to the recruitment methods used, and the inclusion and exclusion criteria applied. Participants in most trials were healthy volunteers; however, some trials recruited people who were attending outpatient clinics. Sixty trials (56%) recruited participants with a history of falls or one or more risk factors for falling.

We excluded trials that tested exercise interventions for preventing falls in people affected by particular conditions, such as stroke, Parkinson’s disease, multiple sclerosis, hip fracture and dementia from this review as we considered that the results of interventions for these conditions were not necessarily applicable to older people as a whole. Fall prevention trials in these populations also often include a wider age range which would result in some being excluded from this review; Cochrane Reviews for each of these specific groups (including all age groups) would be preferable for summarising the evidence. The majority of trials (67%) excluded older people who were cognitively impaired, therefore the results of this review may not be applicable to this high risk group.

Most trials were relatively small (median = 134 participants), with a mean age of 76 (ranging from a mean age of 65 to a maximum mean age of 88 years). Thirty‐seven trials reported 12‐month follow‐up, with 49 reporting less than 12 months and 22 reporting more than 12 months follow‐up. Trials were undertaken over 25 years from 1992 to 2017.

Setting

Exercise‐based fall prevention interventions tested in a further 58 RCTs were included in this review compared with Gillespie 2012. The included trials were conducted in 25 countries using a variety of healthcare models. These different healthcare systems and structures may have impacted upon the effectiveness of some interventions. There remains a paucity of studies undertaken in low‐income economies.

Interventions

We classified the exercise interventions using the ProFaNE guidelines. This classification system is clearly described(Lamb 2011; Appendix 1); however, we acknowledge there is a degree of subjectivity in the classification of exercise interventions based on brief descriptions in trial reports. We conducted post‐hoc sensitivity analyses to explore the effects of recategorising trials with a secondary component of strength training as having multiple primarily exercise categories and found this made little difference to the results (Appendix 18). The duration of exercise intervention in the 81 exercise versus control trials ranged from 5 to 130 weeks; it being one year or more in 30% of these.

Outcomes

We sought data for rate of falls, number of people falling, number of people sustaining a fall‐related fracture, number of people who experienced falls leading to medical attention, number of people who had a fall‐related hospital admission, health‐related quality of life and number of people who experienced adverse events. However, few studies provided fracture, medical attention, hospital admission, health‐related quality of life and full adverse events data. As the analyses and Appendix 10 demonstrate, some studies provided data for both falls and fallers, as recommended in Lamb 2005, and others provided data for one or other falls outcomes.

The outcome of interest, falling, was not always clearly defined, which is a source of concern. Comparability of future research findings would be enhanced by the adoption of the consensus definition of a fall developed for trials in community‐dwelling populations by the Prevention of Falls Network Europe (ProFaNE) (Lamb 2005). The included studies also varied in the methods used for falls ascertainment, recording, analysing and reporting. Studies should use accepted protocols for recording of falls data, including daily recording of falls with monthly or more frequent follow‐up by the researchers who are blind to group allocation (Lamb 2005). At least 26% of included trials did not do this despite evidence of a 25% underreporting of falls when data were collected retrospectively by telephone at the end of a three‐month period, compared with data collected daily and returned monthly over the same period (Hannan 2010). There are difficulties in using fall diaries over long time periods however, with trial dropouts due to over‐burden of paperwork reported by Iliffe 2015.

The lack of consistent measurement of adverse events in trials requires attention by trialists. We found just two studies that measured adverse events in both groups throughout the trial period. Although it is worth noting that the burden on trial resources and participants of full documentation of adverse events is probably a key reason this has not been done to date. Trials of exercise interventions do not tend to be as well‐resourced as trials of pharmacological interventions in which adverse event monitoring is routine.

This review only included data for the risk of fractures and injurious falls, rather than for the rates of fractures and injurious falls; however, it is important to note that several trials have identified an impact of exercise on rates of fall‐related fracture (Karinkanta 2007; Korpelainen 2006; Kemmler 2010), as well as rates of injurious falls (Uusi‐Rasi 2015). There is also evidence of an impact of exercise on the rate of falls requiring medical care, over and above the impact from other types of interventions (Fitzharris 2010).

Other considerations relating to applicability

We decided not to pool studies undertaken in people who had recently been discharged from hospital with studies undertaken among general older populations. It is well documented that people who have recently been discharged form hospital are at a particularly high risk of falls (Mahoney 1994), and as such may require different intervention approaches. There is increasing awareness that many older people deteriorate physically during a hospital admission (Oliver 2017). We note that a number of recent studies of interventions have been undertaken in this population and among emergency department attendees (Harper 2017; Matchar 2017; Oliver 2017); however, there is still uncertainly of the best treatment for this population and a separate review may be needed in future.

For the control groups of the trials that did not have increased risk of falls as an inclusion criterion, the median rate of falls (if 1000 people were followed over 1 year, there would be 605 falls) and the median proportion of fallers (if 1000 people were followed over 1 year, 380 would experience one or more falls) are similar to estimates of fall risk and rate in the general community derived from large population studies (AIHW 2018; Lord 2011; NICE 2018). This indicates that participants in trials that do not recruit based on fall risk, are representative of the general community, rather than being at low risk of falls.

Subgroup analyses comparing the effects on falls outcomes in trials with predominantly older populations and those with predominantly younger populations should be interpreted with some caution, as implementation of one of the categorisation criteria (mean age minus 1 SD > 75) may result in some younger people in the older group and vice versa.

Ongoing studies

The 16 identified ongoing studies may contribute to research priorities. Six ongoing studies, two of which have a larger sample size (exceeding 400 participants), will evaluate the relative impact of different exercise programmes (NCT02126488; NCT03211429; NCT03404830; NCT03455179; n > 400 (NCT02287740; NCT02926105). Two studies will investigate individual versus group delivery of the LiFE programme (NCT03462654), and Otago Exercise Program (NCT03320668). Also, one large trial awaiting classification studied the difference between three types of exercise, including flexibility exercise (Li 2018b). Fall‐related fractures are listed as outcomes in only two trials (ISRCTN71002650; NCT02617303). Two trials, in India (CTRI/2018/01/011214), and Columbia (NCT03211429), will contribute to the understanding of the effect of exercise on falls in emerging economies. In addition, research is underway to investigate strategies for optimal translation of effective exercise programmes from the research setting to clinical and community settings (Carpenter 2018; Hawley‐Hague 2017).

Certainty of the evidence

This review, containing 108 trials (23,407 participants) provides moderate‐ to high‐certainty evidence of the effectiveness of exercise‐based interventions for preventing falls among community‐dwelling people aged 60 years and over.

We have summarised the GRADE certainty of evidence in seven 'Summary of findings' tables: Table 1 (Exercise (all types) versus control); Table 2 (Balance and functional exercises versus control); Table 3 (Resistance exercises versus control); Table 4 (3D (Tai Chi) exercise versus control)); Table 5 (3D (dance) exercise versus control)); Table 6 (Walking programme versus control); Table 7 (Multiple categories of exercise versus control).

The certainty of the evidence ranged from high to very low. We downgraded outcomes by one level for risk of bias if the results changed with removal of the trials with a high risk of bias on one or more items. We downgraded one level for inconsistency where heterogeneity was greater than 60%. In addition, we downgraded the level of evidence for imprecision by one or two levels due to the wide confidence intervals, often reflecting the small number of participants and trials. We downgraded where the risk of small sample bias was evident on funnel plot and downgraded one level for fall‐related hospital admission and fall‐related medical attention because a large number of studies included in the review do not contribute to the outcome.

Sensitivity analyses revealed the results for the falls outcomes to be stable (see Appendix 18) suggesting that the results are robust to key risks of bias and essentially unchanged by methodological choices in the conduct of the review. In undertaking the GRADE assessment we downgraded the certainty of evidence based on sensitivity analysis (removal of trials with one or more items at high risk of bias) for one or both falls outcomes for several types of exercise (resistance, Tai Chi, walking, multiple) and for the overall fracture and quality of life outcomes. It is noteworthy that many of the sensitivity analyses undertaken regarding risk of bias revealed a stability of the results of this review.

Rates of fractures and injurious falls were not prespecified outcomes in this review. More trials reported the outcome in this way than anticipated. We would be in favour of reporting these outcomes in future versions of this review.

Potential biases in the review process

We conducted a comprehensive search of the published literature using multiple databases and also searched clinical trial registries for completed trials for which full reports had not been identified. Two review authors who were blinded to each other's results performed screening and data extraction in duplicate to minimise bias. Despite this thorough search strategy, we acknowledge the possibility that some relevant trials may have been missed, especially if they were published in languages other than English.

Two review authors independently classified the exercise interventions using the ProFaNE guidelines (Lamb 2011), including assigning intervention categories to primary or secondary status. We recognise there is some subjectivity in this classification system, particularly for those interventions containing more than one category of exercise. Sensitivity analyses that tested the effects of recategorising primary balance and functional exercise trials with a secondary component of strength training indicated that this did not importantly affect the results.

We recorded and reported data on fracture, hospitalisation, medical attention and health‐related quality of life only where it was reported by intervention group. To check whether this could be a source of potential bias, we conducted an audit of fracture reporting in the 48 trials with balance, function and gait exercise interventions. Of the 10 trials reporting fracture outcomes, we included seven reporting fracture outcomes by intervention group in the analysis. We did not include the three other studies in the analysis because they either did not report fractures by group (Skelton 2005), they reported fractures during the intervention period but not during follow‐up (Iliffe 2014), or they just reported a fracture (1 pelvic stress fracture) as an adverse event (Clemson 2012). This provided some reassurance that our approach for these secondary and generally under‐reported outcomes did not have an important impact on the results.

Agreements and disagreements with other studies or reviews

Our review adds to the existing body of evidence and supports the findings of Gillespie 2012, whereby multiple component group‐based exercise was found to reduce the rate of falls (rate ratio (RaR) 0.71, 95% confidence interval (CI) 0.63 to 0.82; 16 trials, 3622 participants) and the risk of falling (risk ratio (RR) 0.85, 95% CI 0.76 to 0.96; 22 trials, 5333 participants). Similar results were found for individually‐delivered multiple component exercise that reduced the rate of falls (RaR 0.68, 95% CI 0.58 to 0.80; 951 participants, 7 trials) and the number of people falling (RR 0.78, 95% CI 0.64 to 0.94; 714 participants, 6 trials). The review by Gillespie 2012, also found that Tai Chi reduced the rate of falls (RaR 0.72, 95% CI 0.52 to 1.00; 1563 participants, 5 trials) and the number of people falling (RR 0.71, 95% CI 0.57 to 0.87; 1625 participants, 6 trials). Group‐based balance or functional exercises also demonstrated a statistically significant reduction in the rate of falls (RaR 0.72, 95% CI 0.55 to 0.94; 519 participants, 4 trials) but not in the number of people falling (RR 0.81, 0.62 to 1.07; 453 participants, 3 trials). This influential review has informed, and been the basis of, many guidelines and policy documents internationally.

We extended the findings of Gillespie 2012 by recoding intervention programmes (Appendix 1), in an attempt to identify a primary exercise component for each included study and reserving the 'multiple component' category for trials in which the intervention programme had an equal focus on each of the multiple components. As a result, more studies in our review are classified as balance and functional exercises and fewer as multiple component programmes. We hope that this change will be of assistance to those seeking to design exercise intervention programmes.

The present review also adds to our previous non‐Cochrane review (Sherrington 2017), that used different methodology (multivariable metaregression) yet reached similar conclusions about the importance of the inclusion of exercises that safely challenge balance in fall prevention exercise programmes. Other recent analyses have reached similar findings, including a large network meta‐analysis (Tricco 2017).

The importance of exercise in fall prevention suggests that greater attention be given to the widespread implementation of a life course approach to healthy ageing, i.e. lifelong exercise to maximise physical functioning in older age, as suggested by the World Health Organization (WHO 2015).

Authors' conclusions

Implications for practice.

Well‐designed exercise programmes reduce the rate of falls and the number of people experiencing falls amongst older people living in the community (high‐certainty evidence).

The effects of exercise programmes are uncertain for other non‐falls outcomes, mainly reflecting the considerable under‐reporting of these outcomes in the included trials. Exercise may reduce the number of people experiencing one or more fall‐related fractures and the number of people experiencing one or more falls requiring medical attention (low‐certainty evidence). We are uncertain about the effect of exercise programmes on the number of people who experience one or more falls requiring hospital admission. Exercise may make little important difference to health‐related quality of life (low‐certainty evidence). The reporting of adverse events was poor; where reported these were usually non‐serious and predominantly musculoskeletal.

Effective exercise programmes that reduce both falls outcomes primarily involve balance and functional exercises (high‐certainty evidence) or include multiple exercise categories, most commonly balance and functional exercises plus resistance exercises (moderate‐certainty evidence). Tai Chi reduces the number of people experiencing falls (high‐certainty evidence) and may reduce the rate of falls (low‐certainty evidence). We are uncertain about the effect of programmes involving primarily resistance exercises, dance or walking, as there is insufficient evidence on these. There are no data available for flexibility exercise or endurance exercise versus control.

Exercise programmes were effective regardless of whether they were delivered individually or in groups, by health professionals or trained non‐health professionals, to younger or older populations (based on a 75 year age threshold) or to those identified at a higher risk of falls or not selected for risk of falls. There is likely to be a greater absolute impact in people identified at increased risk of falling, but there is benefit also for those who are at more general risk in the community. Although trial follow‐up ranged from 3 to 18 months in the main comparison, there may also be longer‐term benefits of introducing fall prevention exercise habits in people in the general community. Notably too, the duration of most of the exercise programmes was 12 weeks or over and nearly one‐third lasted a year or more. These findings highlight the importance of primary prevention.

There is currently insufficient evidence to determine the effects of exercise programmes for people recently discharged from hospital. There is also insufficient information from direct comparisons to determine whether there are differences in the effectiveness of different types, modes of delivery and doses of exercise.

Implications for research.

Further work is needed to understand the relative impact of different exercise programmes. Such studies will need to be very large to be adequately powered to detect effects between interventions.

Large studies are also needed to establish the impact of fall prevention interventions on fall‐related fractures and falls requiring medical attention, as such falls are particularly costly to health systems and impactful for individuals.

During the development of priority topics for future research, the current evidence base should be considered in conjunction with the areas studied in the ongoing trials.

Individual participant data meta‐analysis could contribute further to the investigation of differential effects of exercise in people of different ages and baseline fall risks, as these are individual‐level rather than trial‐level characteristics. We recommend researchers follow the Prevention of Falls Network Europe (ProFaNE) guidelines for the conduct of falls trials (Lamb 2005).

Further research is required to establish the effectiveness of fall prevention programmes in emerging economies, where the burden of falls is increasing more rapidly than in high‐income countries due to rapidly ageing populations (WHO 2015).

There is an urgent need to investigate strategies to enhance implementation of effective exercise‐based fall prevention interventions into routine care of older people by healthcare professionals and community organisations.

As it is possible that interventions designed to increase physical activity could increase falls due to increased exposure to risk, we suggest that those undertaking trials of physical activity interventions in older people consider monitoring falls.

Future studies should use the consensus definition of a fall developed for trials in community‐dwelling populations by ProFaNE (Lamb 2005), consistent methods of falls ascertainment, and consistent measurement of adverse events in both groups throughout the trial period. Future research should use the ProFaNE descriptors to categorise interventions (Lamb 2011), but should be clear how this was operationalised. Appendix 1 outlines how this guide was operationalised in the present review and may provide a useful resource.

Notes

This review provides updated evidence for one of the intervention categories (exercise) covered in the Cochrane Review 'Interventions for preventing falls in older people living in the community' (Gillespie 2012). Some of the wording in several sections of the protocol, such as Background/Description of the condition, was taken from Gillespie 2012. This reflects shared authorship of the two publications, but also attempts to maintain a continuity with the Gillespie 2012 review, as well as links between our review and other proposed reviews that will cover other intervention categories, such as multifactorial and multiple component interventions (Hopewell 2018).

Editorial management and appraisal for this review were conducted by the Cochrane Fast‐Track Service (Managing Editor: Helen Wakeford; Associate Editor: Liz Bickerdike; Information Specialist Advisor: Ruth Foxlee) with additional oversight and appraisal by the Cochrane Bone, Joint and Muscle Trauma Group (Managing Editor: Joanne Elliott; Co‐ordinating Editor: Helen Handoll). Approval for publication given by Helen Handoll. This review was copy‐edited by Kate Cahill and Clare Dooley.

Support to the authors for implementing the requirement by NICE for additional analyses to inform the update of their guideline on Falls in older people was provided by Helen Handoll and Liz Bickerdike, with facilitation by Joanne Elliott and Helen Wakeford. This aspect was under the aegis of Michael Brown, Senior Editor of the Cochrane Acute and Emergency Care Network.

Appendices

Appendix 1. Categories of exercise (ProFaNE): definitions and application

Exercise category	ProFaNE description	How the category criteria were applied in this reviewa
Gait, balance, and functional training	Gait training involves specific correction of walking technique (e.g. posture, stride length and cadence) and changes of pace, level and direction. Balance training involves the efficient transfer of bodyweight from one part of the body to another or challenges specific aspects of the balance systems (e.g. vestibular systems). Balance retraining activities range from the re‐education of basic functional movement patterns to a wide variety of dynamic activities that target more sophisticated aspects of balance. Functional training uses functional activities as the training stimulus, and is based on the theoretical concept of task specificity. All gait, balance and functional training should be based on an assessment of the participant’s abilities prior to starting the programme; tailoring of the intervention to the individuals abilities; and progression of the exercise programme as ability improves	Selected as exercise category if the intervention met the baseline assessment, tailoring and progression criteria. Selected as primary category for interventions where most exercises were conducted standing and where the intervention focus and most time spent was on exercise in this category
Strength/resistance (including power)	The term 'resistance training' covers all types of weight training i.e. contracting the muscles against a resistance to ‘overload’ and bring about a training effect in the muscular system. The resistance is an external force, which can be one’s own body placed in an unusual relationship to gravity (e.g. prone back extension) or an external resistance (e.g. free weight). All strength/resistance training should be based on an assessment of the participant’s abilities prior to starting the programme; tailoring the intervention to the individual's abilities; and progression of the exercise programme as ability improves	Selected as exercise category if the intervention met the baseline assessment, tailoring and progression criteria. Selected as primary category for interventions where additional resistance was used or where it was clear that overload was sufficient without external resistance and where the intervention focus and most time spent was on exercise in this category
Flexibility	Flexibility training is the planned process by which stretching exercises are practised and progressed to restore or maintain the optimal range of movement (ROM) available to a joint or joints. The ranges of motion used by flexibility programmes may vary from restoration/maintenance of the entire physiological range of motion, or alternatively, maintenance of range that is essential to mobility or other functions	Selected as exercise category if the intervention met the progression of stretching criterion. Selected as primary category for interventions where flexibility training was a stated aim of the intervention and where the intervention focus and most time spent was on exercise in this category
3D	3D training involves constant movement in a controlled, fluid, repetitive way through all three spatial planes or dimensions (forward and back, side to side, and up and down). Tai Chi and Qi Gong incorporate specific weight transferences and require upright posture and subtle changes of head position and gaze direction. Dance involves a wide range of dynamic movement qualities, speeds and patterns	Selected as exercise category if the intervention involved Tai Chi or dance. Selected as primary category for interventions where the intervention focus and most time spent was on exercise in this category
General physical activity	Physical activity is any bodily movement produced by skeletal muscle contraction resulting in a substantial increase in energy expenditure. Physical activity has both occupational, transportation and recreational components and includes pursuits like golf, tennis, and swimming. It also includes other active pastimes like gardening, cutting wood, and carpentry. Physical activity can provide progressive health benefits and is a catalyst for improving health attitudes, health habits, and lifestyle. Increasing habitual physical activity should be with specific recommendations as to duration, frequency and intensity if a physical or mental health improvement is indicated	Selected as exercise category if the intervention included unstructured physical activity. We classed programmes that included unstructured walking as this category. Selected as primary category for interventions where the intervention focus and most time spent was on exercise in this category
Endurance	Endurance training is aimed at cardiovascular conditioning and is aerobic in nature and simultaneously increases the heart rate and the return of blood to the heart	Selected as exercise category if the intervention focused on structured aerobic training. We classed programmes that included treadmill walking as this category. Selected as primary category for interventions where the intervention focus and most time spent was on exercise in this category
Other	Other kinds of exercises not described	Selected as exercise category if the intervention did not meet the other categories listed and where the intervention focus and most time spent was on exercise in this category
aInterventions were allocated a secondary category if some but not all criteria were met by the intervention or where the category was not the primary focus of the intervention, or both

Appendix 2. Search strategies (February 2012 to 2 May 2018)

CENTRAL (CRS Online)

#1 MESH DESCRIPTOR Accidental Falls EXPLODE ALL TREES
 #2 (falls or faller*):TI,AB,KY
 #3 #1 or #2
 #4 MESH DESCRIPTOR Aged EXPLODE ALL TREES
 #5 (senior* or elder* or old* or aged or ag?ing or postmenopausal or community dwelling):TI,AB,KY
 #6 #4 or #5
 #7 #3 and #6

MEDLINE (Ovid Interface)

1 Accidental Falls/
 2 (falls or faller*1).tw.
 3 or/1‐2
 4 exp Aged/
 5 (senior*1 or elder* or old* or aged or ag?ing or postmenopausal or community dwelling).tw.
 6 or/4‐5
 7 3 and 6
 8 Randomized controlled trial.pt.
 9 Controlled clinical trial.pt.
 10 randomized.ab.
 11 placebo.ab.
 12 Clinical trials as topic/
 13 randomly.ab.
 14 trial.ti.
 15 8 or 9 or 10 or 11 or 12 or 13 or 14
 16 exp Animals/ not Humans/
 17 15 not 16
 18 7 and 17

Embase (Ovid Interface)

1 Falling/
 2 (falls or fallers).tw.
 3 or/1‐2
 4 exp Aged/
 5 (senior*1 or elder* or old* or aged or ag?ing or postmenopausal or community dwelling).tw.
 6 or/4‐5
 7 3 and 6
 8 exp Randomized Controlled Trial/ or exp Single Blind Procedure/ or exp Double Blind Procedure/ or Crossover Procedure/
 9 (random* or RCT or placebo or allocat* or crossover* or 'cross over' or trial or (doubl* adj1 blind*) or (singl* adj1 blind*)).ti,ab.
 10 8 or 9
 11 (exp Animal/ or animal.hw. or Nonhuman/) not (exp Human/ or Human cell/ or (human or humans).ti.)
 12 10 not 11
 13 7 and 12

CINAHL (Ebsco)

S1 (MH "Accidental Falls")
 S2 TI ( falls or faller* ) OR AB ( falls or faller* )
 S3 S1 OR S2
 S4 (MH "Aged+")
 S5 TI ( senior* or elder* or old* or aged or ag?ing or postmenopausal or community dwelling ) OR AB ( senior* or elder* or old* or aged or ag?ing or postmenopausal or community dwelling )
 S6 S4 OR S5
 S7 S3 AND S6
 S8 PT Clinical Trial
 S9 (MH "Clinical Trials+")
 S10 TI clinical trial* OR AB clinical trial*
 S11 TI ( (single blind* or double blind*) ) OR AB ( (single blind* or double blind*) )
 S12 TI random* OR AB random*
 S13 S8 OR S9 OR S10 OR S11 OR S12
 S14 S7 AND S13

PEDro

Advanced search option selected

Abstract and Title: fall*
 Method: clinical trial
 Sub discipline: gerontology

New record added since: (date of last review entered here)

ClinicalTrials.gov

(prevent OR reduce OR reduction OR risk) AND (fall OR fallers) AND (exercise OR training)

WHO ICTRP

prevent* AND fall* AND exercise* OR reduc* AND fall* AND exercise* OR risk* AND fall* AND exercise* OR prevent* AND fall* AND train* OR reduc* AND fall* AND train* OR risk* AND fall* AND exercise*

Appendix 3. 'Risk of bias' assessment tool

Domain	Criteria for judging risk of bias
Random sequence generation relating to selection bias (biased allocation to interventions) due to inadequate generation of a randomised sequence	Judgement of 'low risk' if the trial authors described a random component in the sequence generation, e.g. referring to a random‐number table; using a computer random‐number generator; coin‐tossing; shuffling cards or envelopes; throwing dice; drawing of lots; minimisation Judgement of 'high risk' if the trial used a systematic non‐random method, e.g. date of admission; odd or even date of birth; case record number; clinician judgement; participant preference; patient risk factor score or test results; availability of intervention Judgement of 'unclear risk' if there is insufficient information about the sequence generation process to permit judgement of 'low risk' or 'high risk'.
Allocation concealment relating to selection bias (biased allocation to interventions) due to inadequate concealment of allocations prior to assignment	Judgement of 'low risk' in studies using: individual randomisation if the trial described allocation concealment as by central allocation (telephone, internet‐based or pharmacy‐controlled randomisation); sequentially‐numbered identical drug containers; sequentially‐numbered, opaque, sealed envelopes cluster‐randomisation if allocation of all cluster units performed at the start of the study and individual participant recruitment was completed prior to assignment of the cluster, and the same participants were followed up over time or individual participants were recruited after cluster assignment, but recruitment carried out by a person unaware of group allocation and participant characteristics (e.g. fall history) or individual participants in intervention and control arms were invited by mail questionnaire with identical information Judgement of 'high risk' in studies using: individual randomisation if investigators enrolling participants could possibly foresee assignments and thus introduce selection bias, e.g. using an open random allocation schedule (e.g. a list of random numbers); assignment envelopes unsealed, non‐opaque, or not sequentially‐numbered; alternation or rotation; date of birth; case record number; or any other explicitly unconcealed procedure cluster‐randomisation if individual participant recruitment was undertaken after group allocation by a person who was unblinded and may have had knowledge of participant characteristics Judgement of 'unclear risk' if insufficient information to permit judgement of 'low risk' or 'high risk'. This is usually the case if the method of concealment is not described or not described in sufficient detail to allow a definite judgement, e.g. if the use of assignment envelopes is described, but it remains unclear whether envelopes were sequentially numbered, opaque and sealed
Blinding of participants and personnel relating to performance bias due to knowledge of the allocated interventions by participants and personnel carrying out the interventions	Judgement of 'low risk' if blinding of participants and personnel implementing the interventions was ensured, and unlikely that the blinding could have been broken but the review authors judge that the outcomes (falls and fractures) are unlikely to be influenced by lack of blinding Judgement of 'high risk' if participants or intervention delivery personnel, or both, were not blinded to group allocation (e.g. exercise intervention), and the outcomes (falls and fractures) are likely to be influenced by lack of blinding Judgement of 'unclear risk' if there is insufficient information to make a judgement of 'low risk' or 'high risk'
Blinding of outcome assessment relating to detection bias due to knowledge of the allocated interventions by outcome assessors	Falls judgement of 'low risk' if outcomes were recorded/confirmed in all allocated groups using the same method and the personnel recording/confirming outcomes were blind to group allocation judgement of 'high risk' if outcomes were not recorded/confirmed in all allocated groups using the same method or the personnel recording/confirming outcomes were NOT blind to group allocation judgement of 'unclear risk' if there is insufficient information to make a judgement of 'low risk' or 'high risk' Fractures: judgement of 'low risk' if fractures were recorded/confirmed in all allocated groups using the same method and fractures were confirmed by the results of radiological examination or from primary‐care case records and the personnel recording/confirming fractures were blind to group allocation judgement of 'high risk' if fractures were not recorded/confirmed in all allocated groups using the same method or the only evidence for fractures was from self‐reports from participants or carers judgement of 'unclear risk' if there is insufficient information to make a judgement of 'low risk' or 'high risk' Hospital admission, medical attention and adverse events: judgement of 'low risk' if requiring hospital admission/medical attention as a result of a fall was recorded/confirmed in all allocated groups using the same method (e.g. from primary‐care records) judgement of 'high risk' if requiring hospital admission/medical attention as a result of a fall was not recorded/confirmed in all allocated groups using the same method (e.g. from primary‐care records) or the only evidence for requiring medical attention was from self‐reports from participants or carers judgement of 'unclear risk' if there is insufficient information to make a judgement of 'low risk' or 'high risk' Health‐related quality of life (self‐reported outcome): judgement of 'low risk' if trial participants were blind to group allocation judgement of 'high risk' if trial participants were not blind to group allocation judgement of 'unclear risk' if blinding was reported and thus trial participants may have been unaware of group
Incomplete outcome data relating to attrition bias due to amount, nature or handling of incomplete outcome data	Judgement of 'low risk' if there are no missing outcome data, or less than 20% of outcome data are missing and losses are balanced in numbers across intervention groups with similar reasons for missing data across groups or missing data have been imputed using appropriate methods Judgement of 'high risk' if greater than 20% of outcome data missing, or reason for missing outcome data likely to be related to true outcome, with either imbalance in numbers or reasons for missing data across intervention groups, or ‘as‐treated’ analysis done with substantial departure of the intervention received from that assigned at randomisation or potentially inappropriate application of simple imputation Judgement of 'unclear risk' if there is insufficient information to make a judgement of 'low risk' or 'high risk'
Selective outcome reporting relating to bias due to the selective reporting or non‐reporting of findings	Judgement of 'low risk' if the trial reports rate of falls, risk of falls and adverse events (minimum set of expected outcomes) and the prospective trial registration or the study protocol are available and prespecify the same fall outcomes as those in the trial report Judgement of ‘high risk’ if there is evidence of selective outcome reporting, with important disparity between prespecified falls outcomes if the prospective trial registration or the study protocol are available; or the lack of appropriate data for both falls outcomes Judgement of 'unclear risk' if the trial does not report on one or more of the minimum set of expected outcomes or if there is insufficient information to make a judgement of ‘low risk’ or ‘high risk’
Method of ascertaining falls relating to bias in the recall of falls due to unreliable methods of ascertainment	Judgement of 'low risk' if the study used some form of concurrent collection of data about falling, e.g. participants given postcards to fill in daily and mail back monthly, calendar to mark monthly, or more frequent follow‐up by the researchers Judgement of 'high risk' if ascertainment relied on participant recall at longer intervals than one month during the study or at its conclusion Judgement of 'unclear risk' if there was retrospective recall over a short period only, or if the trial authors did not describe details of ascertainment, i.e. insufficient information was provided to allow a judgement of 'low risk' or 'high risk'
Cluster‐randomised trials relating to bias due to factors particular to cluster‐randomised trials	Judgement of ’low risk’ if the study predominantly had the following characteristics: i) individuals were recruited to the trial prior to randomisation of the clusters; ii) baseline comparability of clusters was reported or there was statistical adjustment for baseline characteristics; iii) no loss of clusters or missing outcomes for individuals within specific clusters; iv) clustering is accounted for in the analyses; v) results are comparable with individually‐randomised trials Judgement of ’high risk’ if the study predominantly had the following characteristics: i) individuals were recruited to the trial after the randomisation of the clusters; ii) baseline comparability of clusters was not reported and there was no statistical adjustment for baseline characteristics; iii) loss of entire clusters or missing outcomes for individuals within clusters; iv) no account for clustering in analyses; v) results not comparable with individually‐randomised trials Judgement of ’unclear risk’ if there is insufficient information to make a judgement of ’low risk’ or ’high risk’

We adapted this from Table 8.5.a 'The Cochrane Collaboration's tool for assessing risk of bias' and Table 8.5.d 'Criteria for judging risk of bias in the 'Risk of bias' assessment tool' (Higgins 2011).

Appendix 4. Description of included studies: reference links

Study description	Links to references
Setting (country)	Australia: Barker 2016; Barnett 2003; Brown 2002; Clemson 2010; Clemson 2012; Day 2002; Day 2015; Duque 2013; Haines 2009; Lord 1995; Lord 2003; Merom 2016; Nitz 2004; Sales 2017; Sherrington 2014; Vogler 2009; Voukelatos 2015; Voukelatos 2007; Yang 2012 Australia, New Zeland:Latham 2003 Australia, Spain and Germany:Gschwind 2015 Belgium, Israel, Italy, Netherlands, and United Kingdom:Mirelman 2016 Brazil: Almeida 2013; Ansai 2015; Arantes 2015; Madureira 2007 Canada: Carter 2002; Davis 2011; Liu‐Ambrose 2004; Liu‐Ambrose 2008 Chile: Bunout 2005; Dangour 2011 China: Woo 2007 Denmark: Beyer 2007   Finland: Karinkanta 2007; Korpelainen 2006; Lehtola 2000; Luukinen 2007; Uusi‐Rasi 2015 France: Cornillon 2002; El‐Khoury 2015 Germany: Freiberger 2007; Hauer 2001; Kemmler 2010; Siegrist 2016 Hungary: Kovacs 2013; Miko 2017 Iran: Dadgari 2016 Italy: Morone 2016; Verrusio 2017 Japan:Hirase 2015; Iwamoto 2009; Kamide 2009; Kim 2014; Okubo 2016; Sakamoto 2013; Shigematsu 2008; Suzuki 2004; Yamada 2010; Yamada 2012; Yamada 2013 Korea:Park 2008 Netherlands: Logghe 2009; Smulders 2010; Weerdesteyn 2006 New Zealand: Campbell 1997; Kerse 2010; Robertson 2001a; Taylor 2012 Norway: Helbostad 2004; Kyrdalen 2014 Singapore: Kwok 2016; Ng 2015 Sweden: Arkkukangas 2015; Grahn Kronhed 2009; Halvarsson 2013; Halvarsson 2016 Switzerland: Trombetti 2011 Taiwan: Huang 2010; Hwang 2016; Lin 2007 Thailand: Boongrid 2017 Turkey: Irez 2011 United Kingdom: Clegg 2014; Ebrahim 1997; Iliffe 2015; Liston 2014; McMurdo 1997; Skelton 2005; Steadman 2003 USA: Ballard 2004; Buchner 1997; Cerny 1998; Fiatarone 1997; Gill 2016; Hamrick 2017; LaStayo 2017; Li 2005; Lurie 2013; Means 2005; Morgan 2004; Morrison 2018; Reinsch 1992; Resnick 2002; Rubenstein 2000; Wolf 1996; Wolf 2003; Wu 2010
Participants
Trials in which all participants were women	Arantes 2015; Ballard 2004; Beyer 2007; Campbell 1997; Carter 2002; Davis 2011; Ebrahim 1997; El‐Khoury 2015; Grahn Kronhed 2009; Hauer 2001; Irez 2011; Kamide 2009; Karinkanta 2007; Kemmler 2010; Kim 2014; Korpelainen 2006; Kovacs 2013; Liu‐Ambrose 2004; Lord 1995; Madureira 2007; McMurdo 1997; Miko 2017; Morone 2016; Park 2008; Resnick 2002; Skelton 2005; Suzuki 2004; Uusi‐Rasi 2015
Trials that would have been excluded if the review inclusion criteria had been set at 65+ years of age	Barker 2016; Dadgari 2016; Hamrick 2017; Hwang 2016; Iwamoto 2009; Kovacs 2013; Kwok 2016; Mirelman 2016; Morgan 2004; Nitz 2004; Reinsch 1992; Sales 2017; Sherrington 2014; Steadman 2003; Verrusio 2017; Voukelatos 2007
Trials recruiting on the basis of identified falls history or one or more risk factors	Almeida 2013; Ansai 2015; Arantes 2015; Barker 2016; Ballard 2004; Barnett 2003; Beyer 2007; Boongrid 2017; Buchner 1997; Campbell 1997; Clegg 2014; Clemson 2010; Clemson 2012; Dadgari 2016; Day 2015; Duque 2013; El‐Khoury 2015; Fiatarone 1997; Freiberger 2007; Gill 2016; Haines 2009; Halvarsson 2013; Halvarsson 2016; Hauer 2001; Helbostad 2004; Hirase 2015; Hwang 2016; Kim 2014; Kyrdalen 2014; Kwok 2016; LaStayo 2017; Latham 2003; Lin 2007; Liston 2014; Liu‐Ambrose 2008; Logghe 2009; Lurie 2013; Luukinen 2007; Mirelman 2016; Morgan 2004; Morone 2016; Ng 2015; Nitz 2004; Rubenstein 2000; Sakamoto 2013; Sales 2017; Sherrington 2014; Siegrist 2016; Skelton 2005; Smulders 2010; Steadman 2003; Taylor 2012; Trombetti 2011; Uusi‐Rasi 2015; Verrusio 2017; Vogler 2009; Weerdesteyn 2006; Wolf 2003; Wu 2010; Yang 2012
Trials excluding participants with cognitive impairment (either defined as an exclusion criterion or implied by the stated requirement to be able to give informed consent and/or to follow instructions)	Almeida 2013; Ansai 2015; Arantes 2015; Arkkukangas 2015; Barker 2016; Barnett 2003; Beyer 2007; Boongrid 2017; Brown 2002; Bunout 2005; Campbell 1997; Clegg 2014; Clemson 2010; Clemson 2012; Cornillon 2002; Dangour 2011; Davis 2011; Day 2002; Day 2015; Duque 2013; Ebrahim 1997; Freiberger 2007; Gill 2016; Grahn Kronhed 2009; Gschwind 2015; Haines 2009; Halvarsson 2013; Halvarsson 2016; Hamrick 2017; Hauer 2001; Helbostad 2004; Hirase 2015; Hwang 2016; Iliffe 2015; Irez 2011; Kamide 2009; Karinkanta 2007; Kerse 2010; Kim 2014; Korpelainen 2006; Kwok 2016; Kyrdalen 2014; Latham 2003; LaStayo 2017; Li 2005; Liu‐Ambrose 2004; Liu‐Ambrose 2008; Lord 1995; Lord 2003; McMurdo 1997; Means 2005; Merom 2016; Mirelman 2016; Morgan 2004; Ng 2015; Park 2008; Resnick 2002; Robertson 2001a; Rubenstein 2000; Sakamoto 2013; Sherrington 2014; Skelton 2005; Steadman 2003; Taylor 2012; Vogler 2009; Voukelatos 2007; Voukelatos 2015; Wolf 1996; Wolf 2003; Yamada 2010; Yamada 2012; Yamada 2013
Interventions
Comparisons	Exercise versus control (not recently discharged): Almeida 2013*; Ansai 2015*; Arantes 2015; Arkkukangas 2015; Barnett 2003; Beyer 2007; Boongrid 2017; Brown 2002; Buchner 1997; Bunout 2005; Campbell 1997; Carter 2002; Cerny 1998; Clegg 2014; Clemson 2010; Clemson 2012*; Cornillon 2002; Dadgari 2016; Dangour 2011; Day 2002; Duque 2013; Ebrahim 1997; El‐Khoury 2015; Fiatarone 1997; Gill 2016; Grahn Kronhed 2009; Gschwind 2015; Halvarsson 2013; Halvarsson 2016*; Hamrick 2017; Hauer 2001; Hirase 2015; Huang 2010; Iliffe 2015*; Irez 2011; Iwamoto 2009; Kamide 2009; Karinkanta 2007; Kerse 2010; Kim 2014; Korpelainen 2006; Kovacs 2013; Day 2015; Lehtola 2000; Li 2005; Lin 2007; Liu‐Ambrose 2004*; Liu‐Ambrose 2008; Logghe 2009; Lord 1995; Lord 2003; Luukinen 2007; Madureira 2007; McMurdo 1997; Means 2005; Merom 2016; Miko 2017; Morgan 2004; Ng 2015; Nitz 2004; Park 2008; Reinsch 1992; Resnick 2002; Robertson 2001a; Rubenstein 2000; Sakamoto 2013; Sales 2017; Siegrist 2016; Skelton 2005; Smulders 2010; Suzuki 2004; Taylor 2012; Trombetti 2011; Uusi‐Rasi 2015; Voukelatos 2007; Voukelatos 2015; Weerdesteyn 2006; Wolf 1996*; Wolf 2003; Woo 2007*; Yamada 2012 Exercise versus control (recently discharged): Haines 2009; Latham 2003; Sherrington 2014 Different types of exercise (not recently discharged): Ballard 2004; Barker 2016; Davis 2011*; Freiberger 2007; Helbostad 2004; Hwang 2016; Kemmler 2010; Kwok 2016; Kyrdalen 2014; LaStayo 2017; Liston 2014; Lurie 2013; Mirelman 2016; Morone 2016; Morrison 2018; Okubo 2016; Shigematsu 2008; Steadman 2003; Taylor 2012; Verrusio 2017; Wu 2010*; Yamada 2010; Yamada 2012; Yamada 2013 Different types of exercise (recently discharged): Vogler 2009* Group versus individual exercise: Barker 2016; Helbostad 2004; Iliffe 2015*; Kyrdalen 2014; Wu 2010* High‐ versus low‐dose exercise: Ballard 2004; Davis 2011*; Taylor 2012
Exercises	Predominantly group‐based: Almeida 2013 (supervised group); Ansai 2015; Arantes 2015; Ballard 2004; Beyer 2007; Brown 2002; Buchner 1997; Bunout 2005; Carter 2002; Cerny 1998; Dangour 2011; Davis 2011; Day 2002; Day 2015; Freiberger 2007; Grahn Kronhed 2009; Halvarsson 2013; Halvarsson 2016; Hauer 2001; Huang 2010; Irez 2011; Karinkanta 2007; Kemmler 2010 (low intensity group); Kim 2014; Kovacs 2013; Kwok 2016; LaStayo 2017; Li 2005; Liu‐Ambrose 2004; Lord 1995; Lord 2003; Madureira 2007; McMurdo 1997; Means 2005; Merom 2016; Miko 2017; Morgan 2004; Morone 2016; Morrison 2018 (balance group); Nitz 2004; Okubo 2016; Park 2008; Reinsch 1992; Rubenstein 2000; Sales 2017; Shigematsu 2008; Siegrist 2016; Smulders 2010; Steadman 2003; Suzuki 2004; Taylor 2012; Trombetti 2011; Voukelatos 2007; Weerdesteyn 2006; Wolf 2003; Woo 2007; Wu 2010 (community group); Yamada 2010; Yamada 2012 Combination of group‐based and individual exercise: Almeida 2013 (home based group); Barker 2016 (pilates group); Barnett 2003; El‐Khoury 2015; Gill 2016; Hamrick 2017; Helbostad 2004; Iliffe 2015 (FaME group); Kemmler 2010 (high intensity group); Kyrdalen 2014 (OEP group); Lehtola 2000; Liston 2014; Logghe 2009; Resnick 2002; Skelton 2005 (FaME group); Uusi‐Rasi 2015; Wolf 1996 (Tai Chi group) Predominantly individual exercise: Arkkukangas 2015; Barker 2016 (home program group); Boongrid 2017; Campbell 1997; Clegg 2014; Clemson 2010; Clemson 2012; Cornillon 2002; Dadgari 2016; Duque 2013; Ebrahim 1997; Fiatarone 1997; Gschwind 2015; Haines 2009; Hirase 2015; Hwang 2016; Iliffe 2015 (OEP group); Iwamoto 2009; Kamide 2009; Kerse 2010; Korpelainen 2006; Kyrdalen 2014 (home based group); Latham 2003; Lin 2007; Liu‐Ambrose 2008; Lurie 2013; Luukinen 2007; Mirelman 2016; Morrison 2018 (Wii group); Ng 2015; Robertson 2001a; Sakamoto 2013; Sherrington 2014; Verrusio 2017; Vogler 2009; Voukelatos 2015; Wolf 1996 (balance group); Wu 2010 (telephone and home exercise groups); Yamada 2013; Yang 2012
Personnel delivering intervention	Health professional delivering intervention: Almeida 2013; Arantes 2015; Arkkukangas 2015; Barker 2016; Beyer 2007; Boongrid 2017; Brown 2002; Campbell 1997; Clegg 2014; Clemson 2010; Clemson 2012; Duque 2013; Ebrahim 1997; Grahn Kronhed 2009; Haines 2009; Halvarsson 2013; Halvarsson 2016; Hauer 2001; Helbostad 2004; Hirase 2015; Hwang 2016 (lower limb group); Kamide 2009; Korpelainen 2006; Kovacs 2013; Kwok 2016; Kyrdalen 2014; LaStayo 2017; Latham 2003; Lin 2007; Liston 2014; Liu‐Ambrose 2008; Lurie 2013; Luukinen 2007; Madureira 2007; Means 2005; Miko 2017; Morgan 2004; Morone 2016; Nitz 2004; Robertson 2001a; Sales 2017; Sherrington 2014; Siegrist 2016; Smulders 2010; Steadman 2003; Uusi‐Rasi 2015; Verrusio 2017; Vogler 2009; Yamada 2010; Yamada 2012; Yang 2012 No health professional delivering intervention: Ansai 2015; Ballard 2004; Barnett 2003; Buchner 1997; Bunout 2005; Carter 2002; Cerny 1998; Cornillon 2002; Dadgari 2016; Dangour 2011; Davis 2011; Day 2002; Day 2015; El‐Khoury 2015; Fiatarone 1997; Freiberger 2007; Gill 2016; Gschwind 2015; Hamrick 2017; Huang 2010; Hwang 2016 (Tai Chi group); Iliffe 2015; Irez 2011; Iwamoto 2009; Karinkanta 2007; Kemmler 2010; Kerse 2010; Kim 2014; Lehtola 2000; Li 2005; Liu‐Ambrose 2004; Logghe 2009; Lord 1995; Lord 2003; McMurdo 1997; Merom 2016; Mirelman 2016; Morrison 2018; Ng 2015; Okubo 2016; Reinsch 1992; Resnick 2002; Rubenstein 2000; Sakamoto 2013; Shigematsu 2008; Skelton 2005; Suzuki 2004; Taylor 2012; Trombetti 2011; Voukelatos 2007; Voukelatos 2015; Weerdesteyn 2006; Wolf 1996; Wolf 2003; Woo 2007; Wu 2010; Yamada 2013 Not reported: Park 2008
* = multigroup trial appearing in more than one category

Appendix 5. Categories of exercise (ProFaNE) in interventions in the included trials

Study ID	Gait/balance/functional training	Strength/resistance training	Flexibility	3D (Tai Chi, dance etc)	General physical activity	Endurance	Other
Almeida 2013 Fully supervised group‐based balance and strength training	Primary	Secondary	Secondary	‐	‐	‐	‐
Almeida 2013 Minimally supervised group‐based balance and strength training	Primary	Secondary	Secondary	‐	‐	‐	‐
Ansai 2015 Group‐based balance, strength and aerobic training	Primary	Primary	‐	‐	‐	Primary	‐
Ansai 2015 Group‐based progressive strength training		Primary	‐	‐	‐	‐	‐
Arantes 2015 Group‐based balance training	Primary	‐	‐	‐	‐	‐	‐
Arkkukangas 2015 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Ballard 2004 Group‐based balance, strength and aerobic training for 15 weeks	Primary	Secondary	‐	‐	‐	Secondary	‐
Ballard 2004 Group‐based balance, strength and aerobic training for 2 weeks	Primary	Secondary	‐	‐	‐	Secondary	‐
Barker 2016 Group‐based Pilates focused on balance and strength plus home practice	Primary	Secondary	‐	‐	‐	‐	‐
Barker 2016 Individual strength and balance	Primary	Secondary	‐	‐	‐	‐	‐
Barnett 2003 Group‐based balance, strength and aerobic training	Primary	Secondary	‐	‐	‐	Secondary	‐
Beyer 2007 Group‐based balance, strength and flexibility training	Primary	Primary	Primary	‐	‐	‐	‐
Boongrid 2017 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐‐
Brown 2002 Group‐based balance, strength and aerobic training	Primary	Primary	‐	‐	‐	Secondary	Secondary ‐ co‐ordination activities
Buchner 1997 Group‐based strength training (combined with endurance and combined groups in analysis)*		Primary
Buchner 1997 Group‐based stationary cycling (combined with resistance and combined groups in analysis)*	‐	‐	‐	‐	‐	Primary	‐
Buchner 1997 Group‐based stationary cycling + strength training (combined with endurance and resistance groups in analysis)*	‐	Primary	‐	‐	‐	Primary	‐
Bunout 2005 Group‐based balance, strength and walking	Primary	Primary	‐	‐	‐	Primary	‐
Campbell 1997 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Carter 2002 Group‐based Osteofit strength training	Secondary	Primary	‐	‐	‐	‐	‐
Cerny 1998 Group‐based balance, strength, flexibility, aerobic training and brisk walking	Primary	Primary	Primary	‐	‐	Primary	‐
Clegg 2014 Individual balance and strength training	Primary	Secondary
Clemson 2010 LiFE (Lifestyle approach to reducing Falls through Exercise) programme‐ balance and strength training embedded in daily life activities	Primary	Secondary	‐	‐	‐	‐	‐
Clemson 2012 LiFE (Lifestyle approach to reducing Falls through Exercise) programme‐ balance and strength training embedded in daily life activities	Primary	Secondary	‐	‐	‐	‐	‐
Clemson 2012 Individual balance and strength training	Primary	Primary	‐	‐	‐	‐	‐
Cornillon 2002 Group‐based balance and gait training	Primary	‐	‐	‐	‐	‐	‐
Dadgari 2016 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Dangour 2011 Group‐based balance and strength	Primary	Secondary	‐	‐	Secondary	‐	‐
Davis 2011 Group‐based progressive high intensity resistance training once weekly	‐	Primary	‐	‐	‐	‐	‐
Davis 2011 Group‐based progressive high intensity resistance training twice weekly	‐	Primary	‐	‐	‐	‐	‐
Davis 2011 Group‐based balance and tone	Primary		Secondary
Day 2002 Group‐based balance and strength	Primary	Secondary	Secondary	‐	‐	‐	‐
Day 2015 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Duque 2013 Virtual reality balance training	Primary	‐	‐	‐	‐	‐	Secondary‐ visual‐vestibular rehabilitation
Ebrahim 1997 Brisk walking	‐	‐	‐	‐	Primary	‐	‐
El‐Khoury 2015 Group‐based balance and strength plus home practice	Primary	Secondary	Secondary	‐	‐	‐	‐
Fiatarone 1997 Individual high‐intensity progressive resistance training	‐	Primary	‐	‐	‐	‐	‐
Freiberger 2007 Group‐based psychomotor programme	Primary	Primary	‐	‐	‐	‐	Primary‐ perceptual training
Freiberger 2007 Group‐based balance, strength, flexibility, endurance	Primary	Primary	Primary			Primary
Gill 2016 Group and home‐based balance, strength, flexibility and walking training	Primary	Primary	Secondary		Primary	‐	‐
Grahn Kronhed 2009 Group‐based strength and balance training	Secondary	Primary	Secondary	‐	‐	Secondary	‐
Gschwind 2015 Individual balance and strength training using exergames	Primary	Secondary	‐	‐	‐	‐	‐
Haines 2009 Home strength and balance program with DVD/workbook	Primary	Primary	‐	Primary	‐	‐	‐
Halvarsson 2013 Group‐based progressive balance training	Primary	‐	‐	‐	‐	‐	‐
Halvarsson 2016 Group‐based progressive balance training	Primary	‐	‐	‐	‐	‐	‐
Halvarsson 2016 Group‐based progressive balance training plus walking	Primary	‐	‐	‐	Primary	‐	‐
Hamrick 2017 Home exercise group	Primary	‐	Secondary	‐	‐	‐	‐
Hauer 2001 Group‐based progressive strength and balance training	Primary	Primary	‐	‐	Primary	‐	‐
Helbostad 2004 Combined group and home‐based balance and strength training	Primary	Secondary	‐	‐	‐	‐	‐
Helbostad 2004 Individual home balance and strength training	Primary	Secondary	‐	‐	‐	‐	‐
Hirase 2015 Group‐based balance training on foam rubber	Primary	‐	‐	‐	‐	‐	‐
Hirase 2015 Group‐based balance training on stable surface	Primary	‐	‐	‐	‐	‐	‐
Huang 2010 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Hwang 2016 Individually supervised balance and strength training	Primary	Secondary	Secondary	‐	‐	‐	‐
Hwang 2016 Individually supervised Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Iliffe 2015 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Iliffe 2015 Group‐based FaME plus home training based on Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Irez 2011 Group‐based pilates	Primary	Primary	‐	‐	‐	‐	‐
Iwamoto 2009 Group‐based balance and gait training	Primary	‐	‐	‐	‐	‐	‐
Kamide 2009 Individual balance and strength training	Primary	Primary	‐	‐	‐	‐	‐
Karinkanta 2007 Group‐based balance and agility training	Primary	‐	‐	‐	‐	‐	‐
Karinkanta 2007 Group‐based resistance training	‐	Primary	‐	‐	‐	‐	‐
Karinkanta 2007 Combined group‐based balance, agility and resistance training	Primary	Primary	‐	‐	‐	‐	‐
Kemmler 2010 Group‐based balance, gait flexibility and strength training plus home practice	Primary	Primary	Primary	‐	‐	Secondary	‐
Kemmler 2010 Group‐based low intensity, low frequency balance and endurance training	Primary	‐	Primary	‐	‐	Secondary	‐
Kerse 2010 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Kim 2014 Group‐based balance and strength training	Primary	Primary	‐	‐	‐	‐	‐
Korpelainen 2006 Group‐based balance and strength training plus home practice	Primary	Secondary	‐	‐	‐	‐	‐
Kovacs 2013 Group‐based balance and strength based on Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Kwok 2016 Group‐based balance, strength and aerobic training plus home practice	Primary	Primary	‐	‐	‐	Primary	‐
Kwok 2016 Balance, strength and aerobic training using the Nintendo WiiActive	Primary	Primary	‐	‐	‐	Primary	‐
Kyrdalen 2014 Group‐based Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Kyrdalen 2014 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
LaStayo 2017 Resisted lower limb exercise in standing and leg press	Primary	Primary	Secondary	‐	Secondary	‐	‐
LaStayo 2017 Resisted lower limb exercise using recumbent stepper‐ergometer	Primary	Primary	Secondary	‐	Secondary	‐	‐
Latham 2003 Resistance exercise	‐	Primary	‐	‐	‐	‐	‐
Lehtola 2000 Group‐based balance and flexibility training plus walking and home practice	Primary	‐	Primary	‐	Primary	‐	‐
Li 2005 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Lin 2007 Individual balance, strength and flexibility training	Primary	Secondary	Secondary	‐	‐	‐	‐
Liston 2014 Group‐based modified Otago Exercise Program plus individual, partially supervised multisensory balance training	Primary	Secondary	‐ ‐	‐	Secondary	‐	‐
Liston 2014 Group‐based modified Otago Exercise Program plus individual, partially supervised flexibility training	Primary	Secondary	Secondary	‐	Secondary	‐	‐
Liu‐Ambrose 2004 Supervised, high‐intensity resistance training	‐	Primary	‐	‐	‐	‐	‐
Liu‐Ambrose 2004 Supervised agility training	Primary	‐	‐	‐	‐	‐	‐
Liu‐Ambrose 2008 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Logghe 2009 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Lord 1995 Group‐based balance, strength, gait training	Primary	Secondary	Secondary	‐	‐	‐	‐
Lord 2003 Group‐based balance, strength, gait training	Primary	Secondary	Secondary	‐	‐	‐	‐
Lurie 2013 Standard Physical Therapy programme + surface perturbation treadmill training	Primary	Secondary	Secondary	‐	‐	‐	Secondary‐ slip and trip training
Lurie 2013 Standard Physical Therapy programme	Primary	Secondary	‐	‐	‐	‐	‐
Luukinen 2007 Individual balance and gait training	Primary	‐	‐	‐	Secondary	‐	‐
Madureira 2007 Group‐based balance training and walking plus home practice	Primary	‐	‐	‐	Secondary	‐	‐
McMurdo 1997 Group‐based balance training	Primary	‐	‐	‐	‐	‐	‐
Means 2005 Group‐based balance, strength, flexibility, gait training and walking	Primary	Primary	Primary	‐	Secondary	‐	‐
Merom 2016 Group‐based social dancing	‐	‐	‐	Primary	‐	Secondary	‐
Miko 2017 Individual, partially supervised balance training	Primary	‐	‐	‐	‐	‐	‐
Mirelman 2016 Individual, supervised treadmill training	Primary	‐	‐	‐	‐	Primary	‐
Mirelman 2016 Individual, supervised treadmill training plus virtual reality	Primary	‐	‐	‐	‐	Secondary	‐
Morgan 2004 Group‐based strength, balance and gait training	Primary	Secondary	Secondary	‐	‐	‐	‐
Morone 2016 Group‐based balance training using Wii‐Fit	Primary	‐	‐	‐	Secondary	‐	‐
Morone 2016 Group‐based balance training	Secondary	‐	Primary	‐	‐	‐	‐
Morrison 2018 Group‐based balance training	Primary	‐	‐	‐	‐	‐	‐
Morrison 2018 Home‐based strength, balance and aerobic Wii Fit programme	Primary	‐	‐	‐	‐	Secondary	‐
Ng 2015 Group‐based strength and balance training plus home practice	Primary	Primary	‐	‐	‐	‐	‐
Nitz 2004 Group‐based balance	Primary	‐	‐	‐	‐	‐	‐
Okubo 2016 Group‐based Tai Chi and Otago Exercise Program plus home practice	Secondary	Secondary	‐	Primary	Secondary	‐	‐
Okubo 2016 Group‐based brisk walking	‐	‐	‐	‐	Primary	‐	‐
Park 2008 Strength and balance and endurance training	Primary	Secondary	Secondary	‐	‐	Primary	‐
Reinsch 1992 Group‐based balance and strength training	Primary	Secondary	‐	‐	‐	‐	‐
Resnick 2002 Individual or group‐based walking	‐	‐	‐	‐	Primary	‐	‐
Robertson 2001a Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐
Rubenstein 2000 Group‐based balance, strength and endurance	Secondary	Primary	‐	‐	‐	Primary	‐
Sakamoto 2013 One leg stand balance training	Primary	‐	‐	‐	‐	‐
Sales 2017 Group‐based strength, balance, co‐ordination, mobility and flexibility	Primary	Secondary	‐	‐	‐	‐	‐
Sherrington 2014 home‐based strength and balance programme	Primary	Primary	‐	‐		‐
Shigematsu 2008 Group‐based stepping training on felt mat	Primary	‐	‐	‐	‐	‐	‐
Shigematsu 2008 Group‐based walking	Primary	‐	‐	‐	Primary	‐
Siegrist 2016 Group‐based balance, strength, power and gait training plus home practice	Primary	Secondary	‐	‐	‐	‐	‐
Skelton 2005 Group‐based FaME balance and strength training plus home practice	Primary	Secondary	‐	‐	Secondary	‐	‐
Smulders 2010 Group‐based balance and gait training using an obstacle avoidance course	Primary	‐	‐	‐	Secondary	‐	Secondary‐ training in fall techniques, lifting techniques
Steadman 2003 Standard, individualised physiotherapy focused on functional training plus balance training	Primary	‐	‐	‐	‐	‐	‐
Steadman 2003 Standard, individualised physiotherapy focused on functional training	Primary	‐	‐	‐	‐	‐	‐
Suzuki 2004 Group‐based strength, balance and gait training plus home practice	Primary	Primary	Primary	Primary	‐	‐	‐
Taylor 2012 Group‐based Tai Chi, 2x/ week	‐	‐	‐	Primary	‐	‐	‐
Taylor 2012 Group‐based Tai Chi, 1x/ week	‐	‐	‐	Primary	‐	‐	‐
Trombetti 2011 Group‐based balance and gait training	Primary	‐	‐	‐	‐	‐	‐
Uusi‐Rasi 2015 Group‐based balance and strength training plus home practice	Primary	Primary	‐	‐	‐	‐	‐
Verrusio 2017 Individual, supervised balance and gait training using exoskeleton human body posturizer	Primary	‐	‐	‐	‐	‐	‐
Verrusio 2017 Individual, supervised balance and gait training	Primary	‐	‐	‐	‐	‐	‐
Vogler 2009 home‐based seated lower limb strength exercises	‐	Primary	‐	‐	‐	‐	‐
Vogler 2009 home‐based strength training with weightbearing, functional tasks	Primary	‐	‐	‐	‐	‐	‐
Voukelatos 2007 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Voukelatos 2015 Individual walking programme	‐	‐	‐	‐	Primary	‐	‐
Weerdesteyn 2006 Group‐based balance and gait training using an obstacle avoidance course	Primary	‐	‐	‐	Secondary	‐	‐
Wolf 1996 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Wolf 1996 Individual, computerised balance training on force platform.	Primary	‐	‐	‐	‐	‐	‐
Wolf 2003 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Woo 2007 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Woo 2007 Group‐based resistance training	Secondary	Primary	‐	‐	‐	‐	‐
Wu 2010 Individual, supervised Tai Chi delivered via videoconferencing	‐	‐	‐	Primary	‐	‐	‐
Wu 2010 Group‐based Tai Chi	‐	‐	‐	Primary	‐	‐	‐
Wu 2010 Individual Tai Chi with DVD instruction	‐	‐	‐	Primary	‐	‐	‐
Yamada 2010 Group‐based trail walking	Primary	Secondary	Secondary	‐	Secondary	‐	‐
Yamada 2010 Group‐based indoor walking	Secondary	Secondary	Secondary	‐	Primary	‐	‐
Yamada 2012 Group‐based balance, strength, flexibility and gait training involving complex obstacle course	Primary	Secondary	Secondary	‐	Secondary	‐	‐
Yamada 2012 Group‐based balance, strength, flexibility and gait training involving simple obstacle course	Primary	Secondary	Secondary	‐	Secondary	‐	‐
Yamada 2013 Group‐based balance, strength, flexibility and gait training including stepping mat	Primary	Secondary	Secondary	‐	Secondary	‐	‐
Yamada 2013 Group‐based balance, strength, flexibility and gait training plus indoor walking	Primary	Secondary	Secondary	‐	Secondary	‐	‐
Yang 2012 Individual Otago Exercise Program	Primary	Secondary	‐	‐	Secondary	‐	‐

* Intervention groups combined due to fall data not being available for individual intervention groups.

Appendix 6. Study IDs for the 81 studies included in the exercise (all types) versus control comparison

Comparison	Study IDs
Exercise (all types) versus control	Almeida 2013*; Ansai 2015*; Arantes 2015; Arkkukangas 2015; Barnett 2003; Beyer 2007; Boongrid 2017; Brown 2002; Buchner 1997; Bunout 2005; Campbell 1997; Carter 2002; Cerny 1998; Clegg 2014; Clemson 2010; Clemson 2012*; Cornillon 2002; Dadgari 2016; Dangour 2011; Day 2002; Day 2015; Duque 2013; Ebrahim 1997; El‐Khoury 2015; Fiatarone 1997; Gill 2016; Grahn Kronhed 2009; Gschwind 2015; Halvarsson 2013; Halvarsson 2016*; Hamrick 2017; Hauer 2001; Hirase 2015*; Huang 2010; Iliffe 2015*; Irez 2011; Iwamoto 2009; Kamide 2009; Karinkanta 2007*; Kerse 2010; Kim 2014; Korpelainen 2006; Kovacs 2013; Lehtola 2000; Li 2005; Lin 2007; Liu‐Ambrose 2004*; Liu‐Ambrose 2008; Logghe 2009; Lord 1995; Lord 2003; Luukinen 2007; Madureira 2007; McMurdo 1997; Means 2005; Merom 2016; Miko 2017; Morgan 2004; Ng 2015; Nitz 2004; Park 2008; Reinsch 1992; Resnick 2002; Robertson 2001a; Rubenstein 2000; Sakamoto 2013; Sales 2017; Siegrist 2016; Skelton 2005; Smulders 2010; Suzuki 2004; Taylor 2012; Trombetti 2011; Uusi‐Rasi 2015; Voukelatos 2007; Voukelatos 2015; Weerdesteyn 2006; Wolf 1996*; Wolf 2003; Woo 2007*; Yang 2012
Balance and functional exercises versus control	Almeida 2013*; Arantes 2015; Arkkukangas 2015; Barnett 2003; Boongrid 2017; Campbell 1997; Clegg 2014; Clemson 2010; Clemson 2012*; Cornillon 2002; Dadgari 2016; Dangour 2011; Day 2002; Duque 2013; El‐Khoury 2015; Gschwind 2015; Halvarsson 2013; Halvarsson 2016*; Hamrick 2017; Hirase 2015*; Iliffe 2015*; Iwamoto 2009; Karinkanta 2007*; Kerse 2010; Korpelainen 2006; Kovacs 2013; Lin 2007; Liu‐Ambrose 2004*; Liu‐Ambrose 2008; Lord 1995; Lord 2003; Luukinen 2007; Madureira 2007; McMurdo 1997; Miko 2017; Morgan 2004; Nitz 2004; Reinsch 1992; Robertson 2001a; Sakamoto 2013; Sales 2017; Siegrist 2016; Skelton 2005; Smulders 2010; Trombetti 2011; Weerdesteyn 2006; Wolf 1996*; Yang 2012
Resistance exercises versus control	Ansai 2015*; Carter 2002; Fiatarone 1997; Grahn Kronhed 2009; Karinkanta 2007*; Liu‐Ambrose 2004*; Woo 2007*
Flexibility versus control	‐
3D exercise (Tai Chi) versus control	Day 2015; Huang 2010; Li 2005; Logghe 2009; Taylor 2012; Voukelatos 2007; Wolf 1996*; Wolf 2003; Woo 2007*
3D exercise (dance) versus control	Merom 2016
General physical activity (walking programme) versus control	Ebrahim 1997; Resnick 2002; Voukelatos 2015
Endurance training versus control	‐
Other kinds of exercise versus control	‐
Multiple categories of exercise versus control	Ansai 2015*; Beyer 2007; Brown 2002; Buchner 1997; Bunout 2005; Cerny 1998; Clemson 2012*; Gill 2016; Halvarsson 2016*; Hauer 2001; Irez 2011; Kamide 2009; Karinkanta 2007*; Kim 2014; Lehtola 2000; Means 2005; Ng 2015; Park 2008; Rubenstein 2000; Suzuki 2004; Uusi‐Rasi 2015

* = multigroup trial appearing in more than one category.

Appendix 7. Study IDs for the 59 exercise versus control studies included in rate of falls analysis

Comparison	Study IDs
Exercise (all types) versus control	Ansai 2015*; Arkkukangas 2015; Barnett 2003; Boongrid 2017; Buchner 1997; Bunout 2005; Campbell 1997; Carter 2002; Clegg 2014; Clemson 2010; Clemson 2012*; Cornillon 2002; Dadgari 2016; Day 2002; Day 2015; Duque 2013; Ebrahim 1997; El‐Khoury 2015; Grahn Kronhed 2009; Gschwind 2015; Hamrick 2017; Hirase 2015*; Iliffe 2015*; Irez 2011; Karinkanta 2007*; Kerse 2010; Korpelainen 2006; Kovacs 2013; Lehtola 2000; Li 2005; Lin 2007; Liu‐Ambrose 2004*; Liu‐Ambrose 2008; Logghe 2009; Lord 1995; Lord 2003; Luukinen 2007; Madureira 2007; McMurdo 1997; Means 2005; Merom 2016; Miko 2017; Nitz 2004; Robertson 2001a; Rubenstein 2000; Sakamoto 2013; Sales 2017; Siegrist 2016; Skelton 2005; Smulders 2010; Suzuki 2004; Taylor 2012; Trombetti 2011; Uusi‐Rasi 2015; Voukelatos 2007; Voukelatos 2015; Weerdesteyn 2006; Wolf 1996*
Balance and functional exercises versus control	Arkkukangas 2015; Barnett 2003; Boongrid 2017; Campbell 1997; Clegg 2014; Clemson 2010; Clemson 2012*; Cornillon 2002; Dadgari 2016; Day 2002; Duque 2013; El‐Khoury 2015; Gschwind 2015; Hamrick 2017; Hirase 2015*; Iliffe 2015*; Karinkanta 2007*; Kerse 2010; Korpelainen 2006; Kovacs 2013; Lin 2007; Liu‐Ambrose 2004*; Liu‐Ambrose 2008; Lord 1995; Lord 2003; Luukinen 2007; Madureira 2007; McMurdo 1997; Miko 2017; Nitz 2004; Robertson 2001a; Sakamoto 2013; Sales 2017; Siegrist 2016; Skelton 2005; Smulders 2010; Trombetti 2011; Weerdesteyn 2006; Wolf 1996*
Resistance exercises versus control	Ansai 2015*; Carter 2002; Grahn Kronhed 2009; Karinkanta 2007*; Liu‐Ambrose 2004*
Flexibility versus control	‐
3D exercise (Tai Chi) versus control	Day 2015; Li 2005; Logghe 2009; Taylor 2012; Voukelatos 2007; Wolf 1996*; Wolf 2003
3D exercise (dance) versus control	Merom 2016
General physical activity (walking program) versus control	Ebrahim 1997; Voukelatos 2015
Endurance training versus control	‐
Other kinds of exercise versus control	‐
Multiple categories of exercise versus control	Ansai 2015*; Buchner 1997; Bunout 2005; Clemson 2012*; Irez 2011; Karinkanta 2007*; Lehtola 2000; Means 2005; Rubenstein 2000; Suzuki 2004; Uusi‐Rasi 2015

* = multigroup trial appearing in more than one category

Appendix 8. Number of studies and participants in primary analysis (exercise versus control on rate of falls), by primary category of exercise

Comparisona	Number of trials (cluster)b	Number of trials with no secondary exercise categoriesc	Number of participants randomised	Number of participants analysedd
Exercise (all types) versus control	59 (6)	15	16363	12981
Balance and functional exercises versus control	39 (4)	7	9815	7920
Resistance exercises versus control	5	3	480	327
Flexibility versus control	0	0	0	0
3D exercise (Tai Chi) versus control	7 (1)	6	2794	2655
3D exercise (dance) versus control	1 (1)	0	530	522
General physical activity (walking program) versus control	2	2	551	441
Endurance training versus control	0	0	0	0
Other kinds of exercise versus control	0	0	0	0
Multiple categories of exercise versus control	11	N/A	1783	1374

^aExercise (all types) combines all categories of exercise. Multiple categories of exercise include studies containing two or more primary categories of exercise, as categorised using the ProFaNE taxonomy. The remaining analyses include only one primary category of exercise, as categorised using the ProFaNE taxonomy.
 ^bStudy IDs are shown in Appendix 7.
 ^cThe number of trials where the intervention programme did not include a secondary exercise from another exercise category using the ProFaNE taxonomy.
 ^dThese data apply to the follow‐up (at the time point included in main analysis) for the primary outcome (rate of falls) for the individual trials

Appendix 9. Source of data for generic inverse variance analysis (see footnotes for explanation of codes)

Study ID	Source for rate ratio: falls	Source for risk ratio: fallers	Source for risk ratio: number with fractures	Source for risk ratio: number with falls requiring medical attention	Source for risk ratio: number with adverse events	Source for risk ratio: hospitalisation	Source for risk ratio: death
Almeida 2013	ND	ND	NA	NA	ND	NA	NA
Ansai 2015	NA	7	NA	NA	NA	NA	NA
Arantes 2015	NA	7	NA	NA	NA	NA	NA
Arkkukangas 2015	3	7	NA	NA	ND	NA	NA
Ballard 2004	3	NA	NA	NA	ND	NA	NA
Barker 2016	1	7	NA	ND	ND	NA	NA
Barnett 2003	1	5	NA	NA	NA	NA	7
Beyer 2007	NA	7	NA	NA	ND	NA	NA
Boongrid 2017	1	4	NA	NA	ND	NA	7
Brown 2002	NA	7	NA	NA	NA	NA	7
Buchner 1997	1	4	NA	NA	NA	NA	NA
Bunout 2005	3	7	NA	NA	NA	NA	7
Campbell 1997	2	4	NA	NA	NA	NA	NA
Carter 2002	3	NA	NA	NA	ND	NA	NA
Cerny 1998	NA	7	NA	NA	NA	NA	NA
Clegg 2014	3	5	NA	NA	NA	7	7
Clemson 2010	1	7	NA	NA	NA	NA	NA
Clemson 2012	1 (ex v control), 3 (ex v ex)	7	NA	NA	ND	NA	7
Cornillon 2002	3	7	NA	NA	NA	NA	7
Dadgari 2016	3c	7c	NA	NA	NA	NA	7
Dangour 2011	NA	7c	7	NA	NA	NA	NA
Davis 2011	1, 3	NA	NA	NA	ND	NA	NA
Day 2002	1, 3	4	NA	7	NA	NA	7
Day 2015	3	7	NA	ND	NA	NA	7
Duque 2013	3	NA	NA	NA	NA	NA	NA
Ebrahim 1997	3	7	7	NA	NA	NA	NA
El‐Khoury 2015	2b	7	ND	ND	ND	NA	7
Fiatarone 1997	NA	ND	NA	NA	NA	NA	NA
Freiberger 2007	1	7	NA	NA	NA	NA	NA
Gill 2016	NA	NA	7	NA	NA	7	7
Grahn Kronhed 2009	3	NA	NA	NA	NA	NA	NA
Gschwind 2015	3	NA	NA	NA	ND	NA	NA
Haines 2009	1	7	7	ND	ND	NA	NA
Halvarsson 2013	NA	7	NA	NA	NA	NA	NA
Halvarsson 2016	NA	7	ND	NA	NA	NA	NA
Hamrick 2017	3	7	NA	NA	NA	NA	NA
Hauer 2001	NA	5	NA	NA	ND	NA	NA
Helbostad 2004	3	7	NA	NA	NA	NA	NA
Hirase 2015	3	NA	NA	NA	NA	NA	NA
Huang 2010	NA	7c	NA	NA	NA	NA	NA
Hwang 2016	1	7	NA	NA	NA	NA	7
Iliffe 2015	3	7	NA	NA	NA	NA	7
Irez 2011	3	NA	NA	NA	NA	NA	NA
Iwamoto 2009	ND	7	ND	ND	ND	NA	NA
Kamide 2009	ND	7	NA	NA	NA	NA	NA
Karinkanta 2007	3	NA	7	7	NA	NA	7
Kemmler 2010	3	5	ND	NA	ND	NA	7
Kerse 2010	3	7	NA	NA	NA	NA	7
Kim 2014	NA	7	7	NA	NA	NA	NA
Korpelainen 2006	3	NA	7	NA	ND	NA	NA
Kovacs 2013	3	7	ND	NA	NA	NA	NA
Kwok 2016	1a	7	NA	NA	ND	NA	NA
Kyrdalen 2014	NA	7	NA	NA	NA	ND	7
LaStayo 2017	3	7	NA	NA	NA	NA	NA
Latham 2003	3	4	NA	NA	7	NA	7
Lehtola 2000	1	NA	NA	NA	NA	NA	NA
Li 2005	2a	4	NA	7	ND	NA	NA
Lin 2007	3	NA	NA	NA	NA	NA	7
Liston 2014	3	ND	NA	NA	NA	NA	NA
Liu‐Ambrose 2004	3	ND	NA	NA	ND	NA	NA
Liu‐Ambrose 2008	1	7	NA	NA	NA	NA	7
Logghe 2009	2	7	NA	NA	NA	NA	7
Lord 1995	3	5	NA	NA	NA	NA	7
Lord 2003	1a	7	NA	NA	NA	NA	7
Lurie 2013	NA	7	NA	NA	NA	NA	NA
Luukinen 2007	2	7	NA	NA	NA	NA	NA
Madureira 2007	3	NA	NA	NA	NA	NA	NA
McMurdo 1997	3	7	7	NA	NA	NA	NA
Means 2005	3	7	NA	NA	ND	NA	7
Merom 2016	1b	5b	NA	NA	ND	NA	7
Miko 2017	3	7	NA	NA	NA	NA	NA
Mirelman 2016	ND	NA	NA	NA	ND	NA	NA
Morgan 2004	NA	7	NA	NA	NA	NA	NA
Morone 2016	ND	ND	NA	NA	NA	NA	NA
Morrison 2018	NA	ND	NA	NA	NA	NA	NA
Ng 2015	NA	7	NA	NA	ND	NA	7
Nitz 2004	3	NA	NA	NA	ND	NA	NA
Okubo 2016	NA	NA	NA	NA	NA	NA	NA
Reinsch 1992	NA	7c	NA	ND	ND	NA	NA
Resnick 2002	ND	NA	NA	NA	NA	NA	NA
Robertson 2001a	1	7	7	7	NA	NA	7
Rubenstein 2000	3	7	NA	NA	ND	NA	NA
Sakamoto 2013	3	7	7	NA	ND	NA	NA
Sherrington 2014	1	5	ND	ND	ND	NA	7
Sales 2017	3	7	NA	NA	ND	NA	7
Shigematsu 2008	3	7	NA	NA	ND	NA	NA
Siegrist 2016	1b	7b	ND	NA	ND	NA	7
Skelton 2005	1	7	ND	NA	ND	NA	7
Smulders 2010	1	7	7	NA	NA	NA	NA
Steadman 2003	3	NA	NA	NA	NA	NA	NA
Suzuki 2004	3	7	NA	NA	NA	NA	NA
Taylor 2012	3	7	NA	NA	NA	NA	7
Trombetti 2011	1	4	NA	NA	ND	NA	7
Uusi‐Rasi 2015	1	4	NA	4	ND	NA	7
Verrusio 2017	NA	7	ND	NA	NA	NA	NA
Vogler 2009	NA	7	NA	NA	ND	NA	7
Voukelatos 2007	1	4	NA	NA	NA	NA	NA
Voukelatos 2015	1	5	NA	NA	NA	NA	7
Weerdesteyn 2006	3	7	NA	NA	NA	NA	NA
Wolf 1996	3	NA	NA	NA	NA	NA	NA
Wolf 2003	2b	7c	NA	NA	ND	NA	7
Woo 2007	NA	7	NA	NA	NA	NA	NA
Wu 2010	3	NA	NA	NA	NA	NA	NA
Yamada 2010	1	7	NA	NA	ND	NA	NA
Yamada 2012	1	7	ND	NA	ND	NA	NA
Yamada 2013	1	7	ND	NA	ND	NA	NA
Yang 2012	NA	7	NA	NA	NA	NA	7

Abbreviations:

Codes for source of rate ratio
 1: incidence rate ratio reported by trial authors
 2: hazard ratio/relative hazard (multiple events) reported by trial authors
 3: incidence rate ratio calculated by review authors
 a: adjusted for confounders by trial authors
 b: adjusted for clustering by trial authors
 c: adjusted for clustering by review authors
 Codes for source of risk ratio:
 4: hazard ratio/relative hazard (first fall only) reported by trial authors
 5: relative risk reported by trial authors
 6: odds ratio reported by trial authors
 7: relative risk calculated by review authors
 a: adjusted for confounders by trial authors
 b: adjusted for clustering by trial authors
 c: adjusted for clustering by review authors
 NA: not applicable. Falls (for rate ratio) or fallers (for risk ratio) or number of people sustaining a fracture (for risk ratio) or number with falls requiring medical attention (for risk ratio) or number with adverse events (for risk ratio) or number of people with falls requiring hospital admission (for risk ratio) or death (for risk ratio) not reported as an outcome in the trial
 ND: outcomes relating to falls or fallers or fractures or falls requiring medical attention or adverse events or hospital admission or death were reported, but there were no useable data

Appendix 10. Raw data for rate of falls and number of fallers when available

Study ID	Intervention group: falls per person‐year	Control group: falls per person‐year	Intervention group: number of fallers	Intervention group: number in analysis	Control group: number of fallers	Control group: number in analysis	Follow‐up
Almeida 2013	0	0	0	28	0	26	4 mo
Ansai 2015 multiple/resistance vs control	4.06/10.14	4.88	4/8	22/23	8	22	4 mo
Arantes 2015	‐	‐	2	15	5	13	3 mo
Arkkukangas 2015	0.89	1.23	5	27	3	13	3 mo
Ballard 2004	0.16	0.41	‐	20	‐	19	16 mo
Barker 2016	1.17	1.16	6	20	9	24	6 mo
Barnett 2003	0.61	0.95	27	76	37	74	12 mo
Beyer 2007	‐	‐	12	24	14	29	12 mo
Boongrid 2017	0.30	0.40	51	218	61	219	12 mo
Brown 2002	‐	‐	20	39	21	32	14 mo
Buchner 1997	0.49	0.81	29	70	18	30	25 mo
Bunout 2005	0.23	0.18	23	111	16	130	12 mo
Campbell 1997 12 mo/24 mo vs control	0.87/0.83	1.34/1.19	53	116/71	62	117/81	24 mo
Carter 2002	0.46	0.52	‐	40	‐	40	5 mo
Cerny 1998	‐	‐	3	15	3	13	6 mo
Clegg 2014	0.70	0.93	7	40	8	30	3 mo
Clemson 2010	‐	‐	8	18	9	16	6 mo
Clemson 2012 LiFE/ structured vs control	1.66/1.90	2.28	60/65	105/107	71	106	12 mo
Cornillon 2002	0.39	0.47	39	150	48	153	12 mo
Dadgari 2016	2.52	3.28	138	160	154	157	6 mo
Dangour 2011	‐	‐	189	325	198	294	24 mo
Davis 2011 x1/x2 wkly resistance vs balance/tone	0.74/0.82	1.06	‐	52/54	‐	49	9 mo
Day 2002	1.05	1.20	76	135	87	137	18 mo
Day 2015	0.62	0.58	65	204	64	205	12 mo
Duque 2013	2.2	4	‐	30	—	40	9 mo
Ebrahim 1997 12 mo/24 mo vs control	0.80/0.70	0.52/0.55	25/‐	52/49	18/‐	50/48	24 mo
El‐Khoury 2015	0.79	0.92	189	352	222	354	24 mo
Freiberger 2007 12 mo/24 mo Fitness vs strength & balance	0.90/—	1.22/‐	19/‐	65/48	29/—	62/49	24 mo
Gill 2016	‐	‐	‐	818	‐	817	42 mo
Grahn Kronhed 2009	0.6	0.8	‐	31	‐	34	12 mo
Gschwind 2015	0.25	0.50	‐	71	‐	65	6 mo
Haines 2009	‐	‐	11	19	20	34	6 mo
Halvarsson 2013	‐	‐	18	30	2	18	15 mo
Halvarsson 2016 balance/ balance+walking vs control	‐	‐	4/5	25/18	4	26	3 mo
Hamrick 2017	0.63	0.84	4	19	7	19	6 mo
Hauer 2001	‐	‐	14	31	15	25	6 mo
Helbostad 2004	1.45	1.33	20	34	18	34	12 mo
Hirase 2015 foam/ stable surface vs control	0.72/1.77	2.7	‐	29/29	‐	28 (14 in analysis)	4 mo
Huang 2010	‐	‐	0	31	2	47	5 mo
Hwang 2016 Tai Chi vs lower extremity	0.08	0.16	72	167	99	167	18 mo
Iliffe 2015 FAME/ OEP vs control, (18 mo/30mo)	0.64/0.66	0.76	‐	230/227	‐	252	30 mo
Irez 2011	1.48	5.2	‐	30	‐	30	3 mo
Iwamoto 2009	0.00	0.29	0	34	4	33	5 mo
Kamide 2009	‐	‐	0	20	1	23	6 mo
Karinkanta 2007 balance/resistance/bal+resistance vs control	0.51/0.21/0.53	0.36	‐	35/37/36	‐	36	12 mo
Kemmler 2010	0.17	0.28	‐	112	—	115	18 mo
Kerse 2010	0.48	0.41	47	98	39	95	12 mo
Kim 2014	‐	‐	10	51	21	52	12 mo
Korpelainen 2006	0.42	0.53	‐	84	‐	76	30 mo
Kovacs 2013	0.42	0.17	6	36	15	36	12 mo
Kwok 2016	‐	‐	8	40	11	40	12 mo
Kyrdalen 2014	‐	‐	19	47	17	47	3 mo
Latham 2003	1.02	1.07	60	112	64	110	6 mo
LaStayo 2017 stepper‐ergometer resistance vs traditional resistance	2.78	1.40	36	54	32	58	12 mo
Lehtola 2000	0.15	0.24	‐	92	‐	39	10 mo
Li 2005	0.80	1.57	27	95	43	93	6 mo
Lin 2007	0.58	0.88	‐	50	‐	50	6 mo
Liston 2014	2.29	2.25	‐	7	‐	8	6 mo
Liu‐Ambrose 2004 resistance/agility vs stretching	1.13/0.65	0.63	‐	32/34	‐	32	6 mo
Liu‐Ambrose 2008	‐	‐	12	31	16	28	12 mo
Logghe 2009	‐	‐	58	138	59	131	12 mo
Lord 1995	0.53	0.63	26	75	33	94	12 mo
Lord 2003	0.67	0.85	109	259	117	249	12 mo
Lurie 2013	‐	‐	5	26	11	33	3 mo
Luukinen 2007	1.23	1.15	126	217	136	220	16 mo
Madureira 2007	0.96	0.40	‐	30	‐	30	12 mo
McMurdo 1997	0.17	0.32	13	44	21	48	24 mo
Means 2005	0.48	1.18	22	144	36	94	6 mo
Merom 2016	1.03	0.80	133	275	144	247	12 mo
Miko 2017	0.14	0.33	6	49	11	48	12 mo
Morgan 2004	‐	‐	34	119	34	110	12 mo
Morrison 2018 Wii vs balance	0	0	0	14	0	32	3 mo
Ng 2015	‐	‐	3	46	5	46	12 mo
Nitz 2004	1.00	1.24	‐	24	‐	21	6 mo
Okubo 2016 walking vs balance	‐	‐	‐	50	‐	40	16 mo
Park 2008	‐	‐	4	22	4	23	11 mo
Reinsch 1992	‐	‐	55	129	34	101	12 mo
Resnick 2002	‐	‐	‐	10	‐	7	6 mo
Robertson 2001a	0.69	1.01	38	121	51	119	12 mo
Rubenstein 2000	1.68	2.00	12	28	9	31	3 mo
Sales 2017	0.89	0.76	11	27	10	21	12 mo
Sakamoto 2013	0.28	0.33	‐	410	‐	455	6 mo
Sherrington 2014	‐	‐	11	169	15	171	4 mo
Shigematsu 2008	0.23	0.33	4	32	7	36	8 mo
Siegrist 2016	1.3	2.4	93	222	77	156	12 mo
Skelton 2005	‐	‐	35	50	23	31	9 mo
Smulders 2010	0.72	1.18	21	47	23	45	12 mo
Steadman 2003	7.13	7.13	‐	69	‐	64	1 mo
Suzuki 2004	0.16	0.46	3	22	12	22	20 mo
Taylor 2012 Tai Chi x1 week/ x2 week v low level ex.	1.55/1.16	1.38	132/111	233/220	140 (70 for analysis)	231 (115 for analysis)	17 mo
Trombetti 2011	0.7	1.6	19	66	32	68	6 mo
Uusi‐Rasi 2015	1.21	1.18	‐	91	‐	95	24 mo
Verrusio 2017	—	—	6	73	19	74	12 mo
Vogler 2009
Voukelatos 2007	0.50	0.75	71	347	81	337	6 mo
Voukelatos 2015	‐	‐	54	159	68	180	12 mo
Weerdesteyn 2006	0.89	1.68	10	30	9	28	7 mo
Wolf 1996 Tai Chi/ balance training vs education	0.86/1.53	1.29	‐	72/64	‐	64	8 mo
Wolf 2003	‐	‐	69	145	85	141	11 mo
Woo 2007 Tai Chi/ resistance vs control	‐	‐	15/24	58/59	31/31	59	12 mo
Wu 2010 Telecommunication‐based Tai Chi/ home‐based Tai Chi vs group Tai Chi	0.47/0.94	0.35	‐	22/22	‐	20	4 mo
Yamada 2010	‐	‐	5	29	11	29	12 mo
Yamada 2012	‐	‐	19	73	2	72	12 mo
Yamada 2013	‐	‐	13	112	39	118	12 mo
Yang 2012	‐	‐	12	59	18	62	6 mo

mo: months

Appendix 11. Raw data for fall‐related fracture, falls requiring medical attention, falls requiring hospital admission and death, when available

Study ID	Intervention group: fall‐ related fracture	Control group: fall‐related fracture	Intervention group: falls requiring medical attention	Control group: falls requiring medical attention	Intervention group: falls requiring hospital admission	Control group: falls requiring hospital admission	Intervention group: number in analysis	Control group: number in analysis	Follow‐up
Almeida 2013	‐	‐	‐	‐	‐	‐	‐	‐	‐
Ansai 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Arantes 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Arkkukangas 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Ballard 2004	‐	‐	‐	‐	‐	‐	‐	‐	‐
Barker 2016	‐	‐	3	8	‐	‐	20	24	6 mo
Barnett 2003	‐	‐	28	38	‐	‐	76	74	12 mo
Beyer 2007	‐	‐	‐	‐	‐	‐	‐	‐	‐
Boongrid 2017	‐	‐	‐	‐	‐	‐	‐	‐	‐
Brown 2002	‐	‐	‐	‐	‐	‐	‐	‐	‐
Buchner 1997	‐	‐	‐	‐	‐	‐	‐	‐	‐
Bunout 2005	‐	‐	‐	‐	‐	‐	‐	‐	‐
Campbell 1997 12 mo/24 mo	‐	‐	27/103	43/110	‐	‐	116/71	117/81	24 mo
Carter 2002	‐	‐	‐	‐	‐	‐	‐	‐	‐
Cerny 1998	0	0	‐	‐	‐	‐	15	13	6 mo
Clegg 2014	‐	‐	‐	‐	2	4	41	33	3 mo
Clemson 2010	‐	‐	‐	‐	‐	‐	‐	‐	‐
Clemson 2012 LiFE/ structured v control	‐	‐	‐	‐	‐	‐	‐	‐	‐
Cornillon 2002	‐	‐	‐	‐	‐	‐	‐	‐	‐
Dadgari 2016	‐	‐	‐	‐	‐	‐	‐	‐	‐
Dangour 2011	10	5	‐	‐	‐	‐	412	406	24 mo
Davis 2011 x1/x2 wkly resistance v balance/tone	‐	‐	0/0	0	‐	‐	52/54	49	9 mo
Day 2002	‐	‐	16	18	‐	‐	135	137	18 mo
Day 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Duque 2013	‐	‐	‐	‐	‐	‐	‐	‐	‐
Ebrahim 1997	6	4	‐	‐	‐	‐	49	48	24 mo
El‐Khoury 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Freiberger 2007	‐	‐	‐	‐	‐	‐	‐	‐	‐
Gill 2016	66	76	‐	‐	46	44	818	817	42 mo
Grahn Kronhed 2009	0	0	‐	‐	‐	‐	31	34	12 mo
Gschwind 2015	0	0	0	0	‐	‐	71	65	6 mo
Haines 2009	1	2	5	26	‐	‐	19	34	6 mo
Halvarsson 2013	‐	‐	‐	‐	‐	‐	‐	‐	‐
Halvarsson 2016 balance/ balance + walking v control	0	0	0	0	‐	‐	25/18	26	3 mo
Hamrick 2017	0	0	0	0	‐	‐	19	19	6 mo
Hauer 2001	0	0	0	0	‐	‐	31	25	6 mo
Helbostad 2004	‐	‐	‐	‐	‐	‐	‐	‐	‐
Hirase 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Huang 2010	‐	‐	‐	‐	‐	‐	‐	‐	‐
Hwang 2016 Tai Chi v lower extremity	‐	‐	‐	‐	‐	‐	‐	‐	‐
Iliffe 2015 FAME/ OEP v control	‐	‐	‐	‐	‐	‐	‐	‐	‐
Irez 2011	‐	‐	‐	‐	‐	‐	‐	‐	‐
Iwamoto 2009	‐	‐	‐	‐	‐	‐	‐	‐	‐
Kamide 2009	‐	‐	‐	‐	‐	‐	‐	‐	‐
Karinkanta 2007 balance/resistance/bal+resistance	0/2/1	1	17/16/11	17	‐	‐	36/37/36	36	12 mo
Kemmler 2010 high intensity / low intensity	‐	‐	0	0	‐	‐	115	113	18 mo
Kerse 2010	‐	‐	‐	‐	‐	‐	‐	‐	‐
Kim 2014	1	2	‐	‐	‐	‐	51	52	12 mo
Korpelainen 2006	6	16	‐	‐	‐	‐	84	76	30 mo
Kovacs 2013	‐	‐	‐	‐	‐	‐	‐	‐	‐
Kwok 2016	‐	‐	‐	‐	‐	‐	‐	‐	‐
Kyrdalen 2014 OEP group / OEP home	‐	‐	‐	‐	3	4	62	63	3 mo
Latham 2003	‐	‐	‐	‐	‐	‐	‐	‐	‐
LaStayo 2017	‐	‐	‐	‐	‐	‐	‐	‐	‐
Lehtola 2000	‐	‐	‐	‐	‐	‐	‐	‐	‐
Li 2005	‐	‐	5	14	‐	‐	95	93	6 mo
Lin 2007	‐	‐	‐	‐	‐	‐	‐	‐	‐
Liston 2014	‐	‐	‐	‐	‐	‐	‐	‐	‐
Liu‐Ambrose 2004	‐	‐	‐	‐	‐	‐	‐	‐	‐
Liu‐Ambrose 2008	‐	‐	‐	‐	‐	‐	‐	‐	‐
Logghe 2009	‐	‐	‐	‐	‐	‐	‐	‐	‐
Lord 1995	‐	‐	‐	‐	‐	‐	‐	‐	‐
Lord 2003	‐	‐	‐	‐	‐	‐	‐	‐	‐
Lurie 2013	‐	‐	‐	‐	‐	‐	‐	‐	‐
Luukinen 2007	‐	‐	‐	‐	‐	‐	‐	‐	‐
Madureira 2007	‐	‐	‐	‐	‐	‐	‐	‐	‐
McMurdo 1997	0	2	‐	‐	‐	‐	44	48	24 mo
Means 2005	‐	‐	‐	‐	‐	‐	‐	‐	‐
Merom 2016	‐	‐	‐	‐	‐	‐	‐	‐	‐
Miko 2017	‐	‐	‐	‐	‐	‐	‐	‐	‐
Morgan 2004	‐	‐	‐	‐	‐	‐	‐	‐	‐
Morrison 2018	‐	‐	‐	‐	‐	‐	‐	‐	‐
Ng 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Nitz 2004	‐	‐	‐	‐	‐	‐	‐	‐	‐
Okubo 2016	‐	‐	‐	‐	‐	‐	‐	‐
Park 2008	‐	‐	‐	‐	‐	‐	‐	‐	‐
Reinsch 1992	‐	‐	4	1	‐	‐	129	101	12 mo
Resnick 2002	‐	‐	‐	‐	‐	‐	‐	‐	‐
Robertson 2001a	2	7	18	26	‐	‐	114	104	12 mo
Rubenstein 2000	0	0	0	0	‐	‐	28	31	3 mo
Sales 2017	‐	‐	‐	‐	‐	‐	‐	‐	‐
Sakamoto 2013	4	11	‐	‐	‐	‐	410	455	6 mo
Sherrington 2014	14	15	61	53	‐	‐	171	169	4 mo
Shigematsu 2008	‐	‐	‐	‐	‐	‐	‐	‐	‐
Siegrist 2016	‐	‐	‐	‐	‐	‐	‐	‐	‐
Skelton 2005	NDa	NDa	‐	‐	‐	‐	‐	‐	9 mo
Smulders 2010	1	3	0	2	‐	‐	47	45	12 mo
Steadman 2003	‐	‐	‐	‐	‐	‐	‐	‐	‐
Suzuki 2004	0	0	‐	‐	‐	‐	22	22	20 mo
Taylor 2012 Tai Chi x1 week/ x2 week v low level ex.	‐	‐	‐	‐	‐	‐	‐	‐	‐
Trombetti 2011	‐	‐	‐	‐	‐	‐	‐	‐	‐
Uusi‐Rasi 2015	‐	‐	HR	HR	‐	‐	91	97	24 mo
Verrusio 2017	‐	‐	‐	‐	‐	‐	‐	‐	‐
Vogler 2009 seated v weightbearing training	‐	‐	‐	‐	‐	‐	‐	‐	‐
Voukelatos 2007	‐	‐	‐	‐	‐	‐	‐	‐	‐
Voukelatos 2015	‐	‐	‐	‐	‐	‐	‐	‐	‐
Weerdesteyn 2006	‐	‐	‐	‐	‐	‐	‐	‐	‐
Wolf 1996	‐	‐	‐	‐	‐	‐	‐	‐	‐
Wolf 2003	‐	‐	‐	‐	‐	‐	‐	‐	‐
Woo 2007	‐	‐	‐	‐	‐	‐	‐	‐	‐
Wu 2010	‐	‐	‐	‐	‐	‐	‐	‐	‐
Yamada 2010	‐	‐	‐	‐	‐	‐	‐	‐	‐
Yamada 2012	1	8	‐	‐	‐	‐	73	72	12 mo
Yamada 2013	3	13	‐	‐	‐	‐	112	118	12 mo
Yang 2012	‐	‐	‐	‐	‐	‐	‐	‐	‐

mo: months; HR: hazard ratio data only; NDa: no data presented by group

Appendix 12. Raw data for death, when available

Study ID	Intervention group: death	Control group: death	Intervention group: number in analysis	Control group: number in analysis	Follow‐up
Barnett 2003	0	3	76	74	12 mo
Boongrid 2017	0	1	219	220	12 mo
Brown 2002	0	3	46	47	14 mo
Bunout 2005	3	3	111	133	12 mo
Clegg 2014	1	3	41	33	3 mo
Clemson 2012 LiFE/structured vs control	1/3	3	100/99	94	12 mo
Cornillon 2002	1	0	150	153	12 mo
Dangour 2011	9	6	412	406	24 mo
Day 2002	NRa	NRa	135	137	18 mo
Day 2015	1	4	204	205	12 mo
El‐Khoury 2015	5	6	352	354	24 mo
Gill 2016	42	37	818	817	42 mo
Haines 2009c	0	3	19	34	6 mo
Hwang 2016 Tai Chi vs lower extremity	2	3	169	170	18 mo
Iliffe 2015 FAME/OEP vs control	3/3	4	243/256	274	18 mo
Karinkanta 2007 balance/resistance/bal+resistance	1/0/0	1	36/37/36	36	12 mo
Kemmler 2010 high intensity/low intensity	0	1	115	113	18 mo
Kerse 2010	1	4	92	95	12 mo
Kyrdalen 2014 OEP group/OEP home	6	3	62	63	3 mo
Latham 2003c	6	8	118	118	6 mo
Lin 2007	2	0	45	45	6 mo
Liu‐Ambrose 2008	1	2	31	27	12 mo
Logghe 2009	1	0	127	117	12 mo
Lord 1995	NRb	NRb	75	94	12 mo
Lord 2003	5	1	264	250	12 mo
Means 2005	4	4	148	98	6 mo
Merom 2016	3	2	278	249	12 mo
Ng 2015	0	1	46	47	12 mo
Robertson 2001a	1	6	114	104	12 mo
Sales 2017	0	1	31	22	12 mo
Sherrington 2014c	10	9	171	169	4 mo
Siegrist 2016	8	10	222	156	12 mo
Skelton 2005	1	1	44	28	9 mo
Taylor 2012 Tai Chi x 1 week/ x 2 week vs low‐level exercise	2/0	7	182/174	181	17 mo
Trombetti 2011	2	2	57	52	6 mo
Uusi‐Rasi 2015	0	2	91	97	24 mo
Vogler 2009c seated vs weight‐bearing training	1	1	58	58	3 mo
Voukelatos 2015	4	0	180	189	12 mo
Wolf 2003	2	4	147	141	11 mo
Yang 2012	0	1	59	62	6 mo

mo: months

NR: not reported.
 ^adata not presented by group; total deaths = 15.
 ^bdata not presented by group; total deaths = 3.
 ^cPost‐hospital discharge trials.

Appendix 13. Adverse events. Studies that reported on adverse events

Study IDa	Group in which adverse events were reported	Adverse events reported in intervention group(s)b		Adverse events reported in control groupb
		Adverse events reported	Number in intervention group(s)	Adverse events reported	Number in control group
Gait, balance, and functional training
Almeida 2013	Two intervention groups and control	0, 0	28	0	26
Boongrid 2017	Intervention and control	Knee pain (n = 2)	218	Knee pain (n = 2)	219
Clemson 2012 LiFEc	Intervention only	Pelvic stress fracture (n = 1)	105	‐	106
El‐Khoury 2015	Intervention only	Painful wrist (n = 1), twisted ankle (n = 1), bruises (n = 5), lumbago (n = 1)	352	‐	354
Gschwind 2015	Intervention and control	0	71	0	65
Iliffe 2015 FaME/OEP groupsd	Two intervention groups and control	FaME: 59 (including 'pulled muscles', exacerbation of back/knee pain, muscle/joint soreness) OEP: 69 (including 'pulled muscles', venous problems, exacerbation of back/knee/hip pain and sciatica)	230/227	45 (including exacerbation of back pain)	252
Iwamoto 2009	Intervention and control	0	34	0	33
Liu‐Ambrose 2004 agility groupc	Two intervention groups and control	Agility intervention group: Musculoskeletal complaints (n = 3), shortness of breath (n = 4)	34	Musculoskeletal complaint (n = 1)	32
Nitz 2004	Intervention and control	0	24	0	21
Reinsch 1992	Intervention and control	Pain, bruise, minor injury	129	Pain, bruise, minor injury	101
Sakamoto 2013	Intervention only	Knee pain (n = 2), lower limb pain (n = 1), palpitations (n = 1)	410	‐	455
Sales 2017	Intervention only	Falls during exercise session, no injury (n = 2)	27	‐	21
Siegrist 2016	Intervention and control	0	222	0	156
Skelton 2005	Intervention and control	0	50	0	31
Trombetti 2011	Intervention and control	0	66	0	68
Strength/resistance (including power)
Latham 2003f	Intervention and control	Back and knee pain directly attributable to the exercise programme (n = 18)	112	n = 5 (no further details)	110
Liu‐Ambrose 2004 Resistance groupc	Two intervention groups and control	Resistance intervention group: Musculoskeletal complaints (n = 10)	32	Musculoskeltal complaint (n = 1)4	32
Vogler 2009 Seated groupf	Two intervention groups and control	Musculoskeletal symptoms in all groups: lower back, hip, knee pain			All groups n = 171
3D (Tai Chi)
Li 2005	Intervention and control	0	95	0	93
Wolf 2003	Intervention and control	0	145	0	141
3D (Dance)
Merom 2016	Intervention only	0	275	‐	247
Multiple primary exercise categories
Arkkukangas 2015	Intervention only	0	27	‐	13
Beyer 2007e	Intervention only	Mild and transient pain symptoms: knee (n = 3), hip (n = 1), thigh/gluteal/groin/hamstrings (n = 3), back (n = 2), ankle (n = 1)	24	‐	29
Carter 2002	Intervention and control	0	40	Grade 1 quadriceps strain (n = 1)	40
Clemson 2012 structuredc,e	Intervention only	Groin strain and inguinal hernia surgery (n = 1)	107	‐	106
Haines 2009e,f	Intervention only	Muscle soreness (n = 1)	19	‐	34
Hauer 2001e	Intervention and control	0	31	0	25
Korpelainen 2006	Intervention only	Musculoskeletal problems (n = 3)	84	‐	76
Means 2005e	Intervention only	0	144	‐	94
Ng 2015e	Intervention and control	Joint pain, hip and knee (n = 2); relieved after adjusting training regimen	46	0	46
Rubenstein 2000	Intervention and control	0	28	0	31
Sherrington 2014e,f	Intervention only	Finger pain following grip strength assessment; thigh pain after assessment, low back pain, calf pain, knee pain, exacerbation of hernia symptoms, pre‐existing conditions (mainly musculoskeletal) limited progression of exercises (n = 12)	169	‐	171
Uusi‐Rasi 2015	Intervention and control	Mild musculoskeletal overuse symptoms, pre‐existing osteoarthritic symptoms (n = 25)	91	Mild musculoskeletal overuse symptoms (n = 1)	95
Exercise versus exercise only	Group in which adverse events were reported	Adverse events reported in intervention group		Adverse events reported in intervention group
		Adverse events reported	Number in intervention group	Adverse events reported	Number in intervention group
Ballard 2004	One intervention group	15 weeks ex group: Hip pain (n = 1)	20		19
Barker 2016	One intervention group	Pilates group: Hip pain (n = 1)	20		24
Davis 2011	Two intervention groups and control	1x/week group: Musculoskeletal complaints (n = 14)	52;	2 a week group: Musculoskeletal complaints (n = 5) Balance and tone group: Musculoskeletal complaints (n = 4)	2 a week group = 54; Balance and tone group = 49
Kemmler 2010	Two intervention groups	0	112	0	115
Kwok 2016	Two intervention groups	0	40	0	40
Mirelman 2016	Two intervention groups	Treadmill group: Death from natural causes (n = 1), myocardial infarctions (n = 2), exacerbated orthopaedic‐related pain or arthritis (n = 3), rhabdomyolysis (n = 4). (All deemed not to be caused by the interventions)	136	Virtual reality group: Stroke (n = 1), exacerbated orthopaedic‐related pain or arthritis (n = 4), herpes zoster (n = 1) (All deemed not to be caused by the interventions)	146
Shigematsu 2008	Two intervention groups	0	32	0	36
Yamada 2010	Two intervention groups	Muscle ache and fatigue	29	Muscle ache and fatigue	29
Yamada 2012	Two intervention groups	Muscle ache and fatigue	73	Muscle ache and fatigue	72
Yamada 2013	Two intervention groups	Muscle ache and fatigue	112	Muscle ache and fatigue	118

^aCategorised by primary exercise category.
 ^bAt time point used in falls analysis (if available).
 ^cStudy with two intervention groups plus a control group; intervention groups reported across multiple rows.
 ^dIncluded events classified as adverse reactions and possible adverse reactions.
 ^eIndicates the primary interventions include gait, balance, and functional training plus strength/resistance training.
 ^fParticipants recently discharged from hospital.

Appendix 14. Adherence

Study IDa	Adherence was measured	Adherence data were reported	Measurement of adherence	Reported adherence resultsb
Gait, balance, and functional training
Almeida 2013	No	No	‐	‐
Arantes 2015	Yes	Yes	Adherence to exercise programme	Mean (range) number of sessions attended: exercise group = 22.1 (20 to 24), control group = 10.8 (10 to 12)
Barnett 2003	Yes	Yes	Attendance	33.7% of participants attended > 30/37 classes
Boongrid 2017	Yes	Yes	Repetitions, sets, duration	56.8% exercised ≥ 120 minutes a week at 12 months
Campbell 1997	Yes	No	‐	‐
Clegg 2014	Yes	Yes	Attendance	Mean attendance = 46%
Clemson 2010	No	No	‐	‐
Clemson 2012 LiFE	Yes	Yes	Adherence to exercise programme	76% still exercising at 6 months
Cornillon 2002	No	No	‐	‐
Dadgari 2016	No	No	‐	‐
Dangour 2011	Yes	Yes	Attendance	Adherence: 38%
Day 2002	Yes	Yes	Attendance	Mean (SD) number of sessions attended = 10 (3.8) of 15 sessions
Duque 2013	Yes	Yes	Adherence to exercise programme	Adherence: 97%
El‐Khoury 2015	Yes	Yes	Started exercise programme	Started the programme = 84%. Attended > 1 month = 73%
Gschwind 2015	Yes	Yes	Adherence to exercise programme	Median (IQR): number of times iStoppFalls system used = 42 (57); duration = 11.7 hours (22); number of times balance exergames were performed = 24 (30); duration = 4.0 hours (6.9); number of strength exercises performed = 20 (31); duration = 7.9 hours (13.4)
Halvarsson 2013	Yes	Yes	Attendance	Mean (range) adherence to the training sessions, intervention group: 87% (71% to 100%)
Halvarsson 2016 balance	Yes	Yes	Attendance	Mean (range) attendance, intervention group: 89% (66% to 100%)
Hamrick 2017	Yes	Yes	Attendance	Mean attendance at yoga classes: 92%
Hirase 2015	Yes	Yes	Attendance	Proportion of classes attended, foam rubber: 95.5%; stable surface: 93.3%; control: 91.2%
Iliffe 2015	Yes	Yes	Attendance	Proportion attended ≥ 75% classes, group ex + OEP group: 40%. Attained ≥ 75% home exercise prescription of 90 minutes/week, OEP: 37%
Iwamoto 2009	Yes	Yes	Attendance	Attendance at 3‐week programme: 100%
Karinkanta 2007 balance	Yes	Yes	Attendance	Mean attendance: 59%
Kerse 2010	Yes	Yes	Adherence to exercise programme	Intervention group: exercised ≥ 2 a week = 55% of participants; walked ≥ 2 a week = 59%; exercised ≥ 3 a week = 25%; walked ≥ 3 a week = 37%; programme almost daily = 20%. Control group: completed all visits = 86% of participants
Kovacs 2013	Yes	Yes	Adherence to exercise programme	Mean (range) attendance (/50 sessions): 80.6% (56% to 100%)
Lin 2007	No	No	‐	‐
Liu‐Ambrose 2008	Yes	Yes	Adherence to exercise programme	Intervention group. Completed programme ≥ 1 a week = 68% of participants; ≥ 2 a week = 57% of participants; ≥ 3 a week = 25% of participants
Liu‐Ambrose 2004 agility group	Yes	Yes	Attendance	Attendance, agility training group: 87.3%; stretching group: 78.8%
Lord 1995	Yes	Yes	Attendance	Attendance at ≥ 60% classes: 75%. For these attendees, mean (range) number of classes attended: 60 (26 ‐ 82)
Lord 2003	Yes	Yes	Attendance	Mean proportion of sessions attended: 42.3%
Luukinen 2007	No	No	‐	‐
Madureira 2007	Yes	Yes	Adherence to exercise programme	Proportion of participants who attended 100% of sessions: 60%. Proportion of participants who did home exercise ≥ 1 a week: 76.7%; ≥ 1 to 4 a week: 36.7%; every day: 40%
McMurdo 1997	Yes	Yes	Attendance	Mean (range) proportion of sessions attended: 76% (46% to 100%)
Miko 2017	No	No	‐	‐
Morgan 2004	Yes	Yes	Attendance	Mean proportion of the 24 scheduled sessions attended: 70%. Participants who dropped out of the study completed an average of 31.7% of the exercise sessions compared with 82.9% completed session by those who did not drop out
Nitz 2004	No	No	‐	‐
Reinsch 1992	No	No	‐	‐
Robertson 2001a	Yes	Yes	Adherence to exercise programme	Performed exercises ≥ 2x/week = 72% of participants; ≥ 3x/week = 43% of participants. Walked ≥ 2x/week = 71% of participants
Sakamoto 2013	Yes	No	Adherence to exercise programme	No data
Sales 2017	Yes	Yes	Attendance	Mean adherence: 79.6%
Siegrist 2016	Yes	Yes	Attendance	Proportion of participants who attended > 10 sessions: 82%. Proportion of participants who performed home exercise programme ≥ 10x: 46%
Smulders 2010	Yes	Yes	Attendance	Proportion of sessions completed: 92.8%. Proportion of participants who completed 100% of sessions: 53.2%
Trombetti 2011	Yes	Yes	Attendance	Mean attendance at exercise programme: 78%
Weerdesteyn 2006	Yes	Yes	Attendance	Mean attendance at exercise sessions: 87%
Wolf 1996 balance	No	No	‐	‐
Yang 2012	Yes	Yes	Adherence to exercise programme	Proportion of intervention participants with full adherence: 44.1%; exercised < 2x/week: 13.6%
Strength/resistance (including power)
Ansai 2015 resistance	Yes	Yes	Adherence to exercise programme	56.5% performed ≥ 24 sessions for 16 weeks (50% intervention)
Fiatarone 1997	No	No	‐	‐
Grahn Kronhed 2009	Yes	Yes	Attendance	Mean attendance at scheduled sessions, exercise group: 24/30 sessions (median (range) 25 (13 ‐ 30)
Karinkanta 2007 resistance	Yes	Yes	Attendance	Mean attendance: 74%
Latham 2003	Yes	Yes	Attendance, exercise intensity	Mean adherence: 82% of prescribed sessions. Mean (SD) exercise intensity at the end of training: 51% (13%) of 1 RM; 25% of participants reached the high intensity desired by the intervention
Liu‐Ambrose 2004 resistance	Yes	Yes	Attendance	Attendance, agility training group: 87.3%; stretching group: 78.8%
Vogler 2009 seated group	Yes	Yes	Attendance	Proportion of sessions completed: 70%
Woo 2007 resistance	No	No	‐	‐
3D
Day 2015	Yes	Yes	Attendance, hours	Mean (SD) class attendance (/96 classes offered): intervention 34.4 (SD 26.9); control 41.3 (26.1). Mean intervention dose = 25.8 hours
Huang 2010	No	No	‐	‐
Li 2005	Yes	Yes	Attendance	‐
Logghe 2009	Yes	Yes	Attendance	Attendance at ≥ 80% of lessons: 47%
Merom 2016	Yes	Yes	Attendance	Median (IQR) attendance to sessions was 56% (IQR 26% to 77%) (approximately 45 sessions)
Voukelatos 2007	No	No	‐	‐
Wolf 2003	Yes	Yes	Attendance	Mean (SD) attendance in the Tai Chi group: 76% (19); control group 81% (17)
Wolf 1996Tai Chi	No	No	‐	‐
Woo 2007 Tai Chi	No	No	‐	‐
Wu 2010 Com‐ex	No	No	‐	‐
Wu 2010 Home‐ex	No	No	‐	‐
Wu 2010 Tel‐ex	Yes	Yes	Attendance	Mean (SD) attendance in Tel‐ex group: 69% (27); Comm‐ex: 71% (27); Home‐ex: 38% (46). Mean (SD) total exercise time (hours): Tel‐ex: 30 (12); Comm‐ex: 31 (12); Home‐ex 17 (21)
General physical activity
Ebrahim 1997	No	No	‐	‐
Resnick 2002	Yes	Yes	Adherence to exercise programme	7/10 intervention participants adhered to the recommended walking programme (20 minutes 3 a week). 2/10 engaged in a regular walking programme but did not meet the recommended dose. 1 did not engage in any exercise. None of the 7 control group participants started an exercise programme during the course of the study
Voukelatos 2015	No	No	‐	‐
Exercise versus exercise
Ballard 2004	No	No	‐	‐
Barker 2016	Yes	Yes	Adherence to exercise programme	Proportion attended over 75% of the classes: 95%
Davis 2011	No	No	‐	‐
Helbostad 2004	No	No	‐	‐
Hwang 2016	Yes	Yes	Attendance	Proportion attended >20 sessions: Tai Chi group 78%; lower limb group 72%
Kemmler 2010	No	No	‐	‐
Kwok 2016	No	‐	‐	‐
Kyrdalen 2014	Yes	Yes	Attendance	Mean(SD) attendance, OEP group: 21.9 (SD 2.7) out of the possible 24 exercise sessions; OEP home: 32.8 (2.8) out of the recommended 36 exercise sessions
LaStayo 2017	Yes	Yes	Attendance	In both groups, all participants completed ≥ requisite minimum 18/36 exercise classes and > 90% of participants who > 28/36 exercise classes
Liston 2014	No	No	‐	‐
Lurie 2013	No	No	‐	‐
Mirelman 2016	No	Yes	‐	‐
Morone 2016	No	No	‐	‐
Morrison 2018	Yes	Yes	Adherence to exercise programme	Proportion who completed the training or all sessions in Wii group: < 50%
Okubo 2016	Yes	Yes	Repetitions, sets, duration	Mean (SD) exercise, balance group: 1.4 (0.5) sets/day, for 4.6 (2.0) days/week; walking group: 45.2 (24.5) min/day of walking for 4.3 (1.7) days week
Shigematsu 2008	No	No	‐	‐
Skelton 2005	Yes	Yes	Started exercise programme	Proportion of intervention participants who completed > 1 intervention session: 73%
Steadman 2003	No	No	‐	‐
Taylor 2012	Yes	Yes	Attendance	Median (IQR) attendance at exercise programme: 79% (49% to 90%)
Verrusio 2017	No	No	‐	‐
Yamada 2010	Yes	Yes	Attendance	Median (IQR) adherence: 100% (74% to 100%) for each group
Yamada 2012	Yes	Yes	Attendance	Median (IQR) adherence in complex course group: 96% (88% to 100%); simple course group: 96% (88% to 100%)
Yamada 2013	Yes	Yes	Attendance	Median (IQR) adherence, multitarget stepping programme: 93% (83% to 96%); walking programme: 92% (83% to 96%)
Multiple primary exercise categories
Ansai 2015 multicomponent*	Yes	Yes	Adherence to exercise programme	34.7% performed ≥ 24 sessions for 16 weeks (50% intervention)
Arkkukangas 2015	Yes	Yes	Adherence to exercise programme	Adherent = 73, not adherent = 27. Definition of adherence unclear
Beyer 2007c	Yes	Yes	Attendance	Training compliance was on average 79% (42 ‐ 100%)
Brown 2002c	Yes	Yes	Attendance	Mean attendance 84.6% (22/26 sessions), range 62% to 100%
Buchner 1997	No	No	‐	‐
Bunout 2005c	Yes	Yes	Attendance	58% attended > 50% of sessions
Carter 2002	Yes	Yes	Attendance	Attendance: 89%
Cerny 1998c	No	No	‐	‐
Clemson 2012 structuredc	Yes	Yes	Adherence to exercise programme	71% still exercising at 6 months
Freiberger 2007c	Yes	Yes	Attendance	Proportion of intervention participants participating in > 75% of sessions: 77%
Gill 2016c	Yes	Yes	Attendance	Mean attendance at scheduled sessions, physical activity group: 68%; median (IQR) 71% (50% to 83%)
Haines 2009c	Yes	Yes	Adherence to exercise programme	Number of intervention participants who adhered to exercise in week 8: ≥ 1 a week = 8/19; ≥ 2 a week = 4/19
Hauer 2001c	Yes	Yes	Adherence to exercise programme	Mean adherence, intervention group: 85.4%; control group: 84.2%
Irez 2011c	Yes	Yes	Attendance	Proportion of sessions completed: 92%
Kamide 2009*	Yes	Yes	Adherence to exercise programme	Intervention participants. Completed intervention > 3 a week, 19/23 (82.6%) participants; completed intervention > 2 a wk, 21/23 (91.3%) participants
Karinkanta 2007 resistance and balance groupsc	Yes	Yes	Attendance	Mean attendance: 67%
Kim 2014c	Yes	Yes	Attendance; exercise sessions at home	Intervention group. Mean (range) attendance at sessions: 75.3% (64% ‐ 86%); mean frequency of home exercise programme: 3.4 a week; mean exercise time: 24.9 minutes
Korpelainen 2006	Yes	Yes	Attendance	Intervention group. Mean attendance at sessions: 75%; mean frequency of home exercise programme: 3 a week
Lehtola 2000	Yes	Yes	Adherence to exercise programme	"Active participants": 52 participants; "Passive participants": 20
Means 2005c	No	No	‐	‐
Ng 2015c	Yes	Yes	Attendance	Mean attendance: physical training 85%, control 94%
Park 2008	No	No	‐	‐
Rubenstein 2000	Yes	Yes	Attendance	Exercise participants attended 84% of the exercise sessions
Sherrington 2014c	Yes	Yes	Reps, sets, duration	Proportion of prescribed repetitions completed in 12th month: 47%
Suzuki 2004c	Yes	Yes	Attendance	Mean attendance at exercise classes: 75.3%
Uusi‐Rasi 2015c	Yes	Yes	Attendance	Mean (range) attendance at group training: 72.8% (0% to 97.4%); home training sessions: 66.1% (0% to 100%)
Vogler 2009 Weight‐bearing group	Yes	Yes	Attendance	Proportion of sessions completed: 62%

^aCategorised by primary exercise category.
 ^bAt time point used in falls analysis (if available).
 ^cIndicates the primary interventions include gait, balance, plus functional training and strength/resistance training.

Appendix 15. Description of excluded studies: reference links

Reason for exclusion	Links to references
Types of participants
Not meeting age criteria	N = 1: Pereira 1998
In a single diagnostic group with increased risk of falls	N = 1: Hsu 2017
Not predominantly community‐dwelling	N = 1: DeSure 2013
Types of intervention
Not exercise as a single intervention	N = 15: Alkan 2011; Beling 2009; Clemson 2004b; Fahlström 2017; Gianoudis 2014; Iwamoto 2012; Lee 2013; Leung 2014; Li 2018a; Olsen 2014; Pai 2014; Rossi‐Izquierdo 2017; Steinberg 2000; Swanenburg 2007; Ueda 2017
Type of control
Control did not meet inclusion criteria	N = 1: Ohtake 2013
Type of outcome
Falls not measured	N = 1: Hinrichs 2016
Participants with injurious falls excluded	N = 1: Morris 2008

Appendix 16. Raw data for quality of life outcome where available

Study ID	Outcome measure	Outcome format	Intervention group quality of life	Intervention group number in analysis	Control group quality of life	Control group number in analysis	Data included in analysis
Boongrid 2017	Thai EQ‐5D	Mean (SD) Baseline 6 month	7.37 (?) 7.7 (?)	219	7.35 (?) 7.4 (?)	220	None
Carter 2002	Osteoporosis‐specific health‐related quality of life	Mean (95% CI) change 5 month‐ baseline (adjusted)	‐0.31 (−2.98 to 2.37)	40	−0.48 (−3.00 to 2.37)	40	None
Clegg 2014	EQ‐5D	Mean (SD) Baseline 3 month	0.53 (0.30) 0.51 (0.34)	40	0.52 (0.25) 0.46 (0.26)	30	EQ‐5D
Clemson 2010	SF‐36 ‐ physical SF‐36 ‐ mental	Median (IQR) change 0 to 6 months	0.6 (−5.0 to 10.1) −1.1 (−8.4 to 0)	17	2.3 (−5/3 to 6.3) −2.9 (−10.9 to 5.7)	14	None
Clemson 2012 (LIFE)	EQ‐5D	Mean (SD) Baseline 6 month 12 month	7.1 (1.4) 6.6 (1.3) 6.7 (1.5)	99	7.0 (1.4) 7.2 (1.6) 6.7 (1.3)	91	EQ‐5D 12 months
Clemson 2012 (Structured)	EQ‐5D	Mean (SD) Baseline 6 month 12 month	6.9 (1.5) 6.9 (1.5) 6.7 (1.6)	96	7.0 (1.4) 7.2 (1.6) 6.7 (1.3)	91	EQ‐5D 12 months
Dangour 2011	SF‐36 ‐ physical SF‐36 ‐ mental	Mean (SD) Baseline 24 month Baseline 24 month	51.2 (6.7) 51.1 (14.3) 49.3 (9.1) 49.2 (6.3)	325	49.8 (6.3) 50.6 (8.9) 49.4 (7.9) 48.3 (6.3)	294	SF‐36 physical 24 months
Grahn Kronhed 2009	SF‐36 ‐ physical SF‐36 ‐ mental QUALEFFO‐41	Mean (SD) Baseline 12 month Baseline 12 month mean (SD) change	44.8 (9.3) 46.9 (8.8) 49.2 (9.7) 53.0 (8.0) ‐0.7 (5.0)	31	36.7 (10.8) 35.7 (9.4) 48.9 (10.3) 47.6 (11.0) −0.2 (5.5)	34	SF‐36 physical 12 months
Gschwind 2015	EQ‐ 5D utility score EQ‐5D VAS	Mean (SD) Baseline 6 month Baseline 6 month	0.86 (0.11) 0.86 (0.15) 79.2 (14.7) 80.9 (13.7)	71	0.86 (0.13) 0.87 (0.13) 81.7 (12.7) 79.9 (14.6)	65	EQ‐ 5D utility score 6 months
Haines 2009	EQ‐5D utility score EQ‐5D VAS	Mean (SD) Baseline 6 month Baseline 6 month	0.58 (0.32) 0.48 (0.35) 66.7 (14.3) 58.9 (21.4)	19	0.65 (0.25) 0.52 (0.36) 67.5 (18.9) 58.1 (25.0)	31	EQ‐5D utility 6 months
Iliffe 2015 FAME	EQ‐5D SF‐12 physical SF‐12 mental OPQOL	Mean (SD) Baseline 12 month Baseline 12 month Baseline 12 month Baseline 12 month	0.67 (0.09) 0.67 (0.07) 38.7 (5.6) 38.9 (4.9) 49.6 (6.0) 48.7 (5.8) 129.4 (13.5) 132.3 (16.0)	179	0.68 (0.08) 0.68 (0.07) 38.7 (5.5) 39.1 (5.0) 49.9 (6.1) 49.2 (5.6) 130.8 (13.5) 134.8 (14.8)	212	EQ‐5D 12 months
Iliffe 2015 OEP	EQ‐5D SF‐12 physical SF‐12 mental OPQOL	Mean (SD) Baseline 12 month Baseline 12 month Baseline 12 month Baseline 12 month	0.68 (0.09) 0.68 (0.07) 38.8 (5.6) 39.3(4.7) 50.2 (5.9) 49.05 (5.1) 129.4 (12.7) 133.7 (15.0)	176	0.68 (0.08) 0.68 (0.07) 38.7 (5.5) 39.1 (5.0) 49.9 (6.1) 49.2 (5.6) 130.8 (13.5) 134.8 (14.8)	212	EQ‐5D 12 months
Kerse 2010	SF‐36 physical SF‐36 mental	Mean (SD) Baseline 6 month 12 month Baseline 6 month 12 month	39.0 (1.2) 39.5 (1.2) 38.3 (1.2) 51.2 (0.9) 54.7 (0.7) 55.4 (0.7)	94	39.3 (1.1) 37.9 (1.3) 39.4 (1.2) 48.7 (1.0) 53.7 (0.9) 52.7 (0.0)	87	SF‐36 physical 12 months
Kyrdalen 2014 (group versus home OEP)	SF‐36 physical SF‐36 mental	Mean (95%CI) Baseline 3 month 6 month Baseline 3 month 6 month	178.2 (158.6 to 197.7) 232.9 (211.0 to 254.8) 218.0 (194.5 to 241.1) 237.3 (217.2 to 257.3) 286.4 (263.6 to 309.2) 269.1 (244.4 to 293.9)	47	192.3 (172.4 to 212.2) 202.1 (179.6 to 224.6) 212.2 (188.4 to 234.1) 254.3 (233.9 to 274.7) 276.0 (252.4 to 299.5) 289.2 (265.2 to 313.2)	47	SF‐36 physical 6 months
Latham 2003	SF‐36 physical	Mean (95%CI) 3 month 6 month	34 (32 to 36) 35 (33 to 37)	112	35 (33 to 37) 37 (35 to 39)	110	SF‐36 physical 6 months
Lin 2007	WHOQOL‐BREF Physical Psychological Social Environmental	Mean (SD) Baseline 6 month 8 month Baseline 6 month 8 month Baseline 6 month 8 month Baseline 6 month 8 month	51.0 (17.9) 59.0 (12.5) 62.8 (9.9) 55.2 (13.6) 62.9 (13.2) 64.4 (12.6) 69.9 (11.4) 71.9 (10.0) 75.4 (9.4) 64.1 (12.5) 70.2 (9.4) 74.9 (6.8)	39	48.9 (17.3) 52.6 (13.8) 55.5 (15.3) 55.7 (16.0) 53.8 (17.0) 56.3 (17.6) 68.8 (10.6) 63.8 (14.8) 66.3 (13.3) 62.5 (9.8) 62.1 (14.4) 65.1 (14.3)	40	WHOQOL‐BREF physical 8 months
Merom 2016	SF‐12 Physical Mental	Mean (SD) Baseline 12 month Baseline 12 month	43.2 (8.6) 41.8 (10.3) 53.0 (8.1) 52.7 (8.7)	274	44.6 (8.7) 42.6 (9.9) 51.9 (7.4) 51.8 (8.2)	247	SF‐12 Physical 12 months
Resnick 2002	SF‐12 Physical Mental	Mean (SD) Baseline 2 month 6 month Baseline 2 month 6 month	31.1 (5.8) 33.7 (4.7) 33.4 (4.8) 48.3 (3.0) 48.4 (2.6) 47.0 (5.2)	10	32.7 (6.7) 32.2 (7.3) 31.2 (4.9) 46.9 (3.0) 47.1 (3.4) 46.8 (3.2)	7	SF‐12 Physical 6 months
Rubenstein 2000	SF‐36 Physical functioning Physcial role limits Health perceptions Health question	Mean (SD) Baseline 3 month Baseline 3 month Baseline 3 month Baseline 3 month	59.6 (24.8) 65.0 (17.4) 66.9 (36.7) 75.0 (34.0) 60.0 (19.1) 64.3 (18.2) 51.8 (26.3) 67.9 (21.4)	28	62.2 (21.0) 60.6 (20.3) 53.7 (38.4) 57.4 (35.2) 58.9 (19.5) 61.1 (19.9) 50.9 (20.2) 46.3 (22.7)	27	SF‐36 Physical functioning 3 months
Sales 2017	SF‐12 Physical Mental	Mean (SD) Baseline 12 month Baseline 12 month	46.9 (7.6) 49.6 (8.3) 53.1 (9.8) 54.5 (7.0)	27	49.1 (7.9) 48.9 (7.6) 51.4 (6.1) 51.6 (7.9)	21	SF‐12 Physical, 12 months
Sherrington 2014	EQ‐5D utility SF‐12 Physical Mental	Mean (SD) Baseline 12 month Baseline 12 month Baseline 12 month	0.63 (0.23) 0.66 (0.27) 37.44 (8.9) 40.37 (8.29) 54.71 (6.5) 55.87 (5.02)	157	0.62 (0.23) 0.60 (0.33) 38.17 (8.36) 39.27 (9.26) 54.70 (6.79) 55.19 (7.09)	155	EQ‐5D utility 12 months
Smulders 2010	QUALEFFO‐41	Mean (SD) Baseline 6 weeks 12 month	25.2 (10.0) 25.4 (10.9) 26.2 (10.6)	47	28.7 (10.9) 26.3 (10.6) 27.3 (11.0)	45	QUALEFFO‐41 12 months
Steadman 2003 (balance vs physio)	Euroqol VAS	Mean (SD) Baseline 6 weeks 3 month 6 month	57.8 (19.7) 65.1 (19.6) 65.1 (17.7) 64.4 (19.9)	69	59.4 (17.2) 64.9 (17.3) 65.7 (16.9) 64.5 (17.4)	64	Euroqol VAS 6 months
Verrusio 2017 (HBP v physio)	SF‐36 Physical Mental	Mean (SD) Baseline 3 month Baseline 3 month	52.1 (6.0) 52.2 (5.4)	73	52.7 (7.1) 53.1 (5.3)	74	None (too hard to read follow‐up data from figure)
Voukelatos 2015	Australian QoVL	Mean (95% CI) Baseline 12 month	0.81 (0.79 to 0.83) 0.84 (0.82 to 0.86)	144	0.81 (0.79 to 0.83) 0.83 (0.81 to 0.85)	169	AQoL 12 months
Wu 2010 Telecommunication‐based Tai Chi vs group Tai Chi	SF‐36 Physical Mental	Mean change (SD)	7.3 (16,3) 2.9 (18.1)	22	9.0 (15.8) 6.2 (11.9)	20	None
Wu 2010 home‐based Tai Chi vs group Tai Chi	SF‐36 Physical Mental	Mean change (SD)	6.7 (14.7) ‐0.2 (8.0)	22	9.0 (15.8) 6.2 (11.9)	20	None
Yang 2012	Assessment of quality of life	Mean (SD) Baseline 6 months	24.8 (4.8) 23.4 (4.1)	59	25.0 (4.5) 24.6 (5.2)	62	QoL, 6 months

Appendix 17. Studies reporting cost‐effectiveness, cost‐utility, or costs (intervention and/or healthcare resource use) related to fall outcomes

Study ID (source if not primary reference), sample, efficacy analyses, type of evaluation	Intervention(s) and comparator (N in analysis)	Perspective(s), type of currency, price year, time horizon	Cost items measured	Mean (SD) intervention cost per person	Healthcare service costs	Incremental cost per fall prevented/per QALY gained
•Buchner 1997 •Patients from a HMO, mild deficits in strength and balance, mean age 75 years •Analysis •Cost analysis	•Centre‐based endurance training or strength training, or both, supervised for 24 to 26 weeks then self‐supervised (N = 75) vs no intervention (N = 30)	•HMO •US dollar •Not specified (presumed 1992) •Period 7 to 18 months after randomisation	•Hospital costs, ancillary outpatient costs (from HMO computerised records)	‐	•Hospitalised control participants more likely to have hospital costs > USD 5000 (P < 0.05) •Ancillary outpatient costs 7 ‐ 18 months after randomisation: Exercise: USD 270 Control: USD 285 (no significant difference)	‐
•Campbell 1997 and Campbell 1999 (Robertson 2001b) •Women aged ≥ 80 years from 17 general practices, mean age (SD) 84.1 (3.3) years •Analysis •Cost‐effectiveness analysis	•Specific set of muscle strengthening and balance retraining exercises individually prescribed at home (OEP) by physiotherapist, 4 home visits and monthly phone calls in year 1, phone contact only in year 2 (N = 116) vs social visits and usual care (N = 117)	•Societal •New Zealand dollar •1995 •During participation in trial (up to 2 years)	•Intervention costs (recruitment, programme delivery, overheads) •Healthcare costs resulting from falls (actual costs of hospital admissions and outpatient services, estimates of GP visits and other costs) •Total healthcare resource use (actual costs of hospital admissions and outpatient services)	In research setting: •NZD 173 (0) in year 1 •NZD 22 (0) in year 2	•No difference between the 2 groups for healthcare costs resulting from falls or for total healthcare costs •27% of hospital admission costs during trial resulted from falls	At 1 year: •NZD 314 per fall prevented (programme implementation costs only) At 2 years: •NZD 265 per fall prevented (programme implementation costs only)
•Dangour 2011 (Walker 2009) •People aged 65 to 67.9 years living in low‐middle socioeconomic status municipalities in Santiago, Chile •Analysis •Cost analysis	•Multicomponent exercise classes, 2 x 1‐hour supervised classes a week for 24 months (10 health centres, N = 854) vs remainder (10 health centres, N = 811)	•Societal and health system •Chilean peso converted to US dollar •2007 •During 2‐year trial	From 93 exit interviews: •Physical activity intervention	•USD 164 for physical activity intervention	‐	•Not calculated (neither intervention reduced risk of falling; cost‐effectiveness of physical activity intervention reported as USD 4.84 per extra metre walked)
•Davis 2011 (Liu‐Ambrose 2010) •Community‐living women aged 65 to 75 years •Analysis •Cost‐effectiveness analysis, cost‐utility analysis	•Once weekly resistance training (N = 54) vs twice‐weekly balance and tone classes (N = 49) •Twice‐weekly resistance training (N = 51) vs twice‐weekly balance and tone classes (N = 49)	•Health service •Canadian dollar •2008 •9 months	•Costs of delivering the interventions (staff time, room use, equipment, building overhead costs); visits to health professionals; all visits, admissions, and procedures in hospital; laboratory and diagnostic tests	•CAD 353 once‐weekly resistance training •CAD 706 twice‐weekly resistance training •CAD 706 twice‐weekly balance and tone classes	•Mean healthcare costs resulting from falls, mean total healthcare costs respectively: CAD 547, CAD 1379 once‐weekly resistance training •CAD 184, CAD 1684 twice‐weekly resistance training •CAD 162, CAD 1772 twice‐weekly balance and tone classes	•Both once‐ and twice‐weekly resistance training dominated balance and tone classes in terms of both falls and QALYs (i.e. less costly, more effective)
•Day 2002 (McLean 2015) •Community‐dwelling people identified from the electoral roll, mean age 76.1 years •Analysis •Cost‐effectiveness analysis Cost‐utility analysis	Exercise group, 1‐hour class a week, 15 weeks, plus daily home exercises designed by physiotherapist (N = 135) vs no intervention (N = 137)	•Healthcare •Australian dollar (costs converted from Australian Ddllar to GBP using 2010 purchasing‐power parity) •2010 •18 months	•Intervention cost (labour, equipment, venue hire, music and consumables) •Healthcare costs: (General Practitioner, ambulance services, emergency department visits, hospital admissions)	•AUD 52	•AUD 33. for exercise group; AUD 39. for control group	ICER per: •Fall prevented 652 •Injurious fall prevented 1176 •Fracture prevented 26,236 •QALY 51,483
•Iliffe 2014 and Iliffe 2015 •Community‐dwelling people with mean age 73 years •Analysis •Cost‐effectiveness analysis Cost‐utility analysis	1. home‐based Otago exercise programme (OEP) (N = 410) 30 minutes, 3 a week, 24 weeks vs Control group: no intervention (N = 457) 2. Community centre‐based Falls Management Exercise (FaME) group (N = 387) 1 hour, weekly + home exercises based on OEP 30 minutes, 2 a week for 24 weeks vs Control group: no intervention (N = 457) 3. OEP vs FaME	•Healthcare •GBP •2011 •52 weeks	•Cost of delivering the intervention (venue hire, procurement of exercise equipment, instructors, training and reimbursement of instructors and mentors). •Cost of primary care service use (GP, practice nurse, out‐of‐hours, other).	OEP London = GBP 88, Nottingham = GBP 117 FaME: London = GBP 269; Nottingham = GBP 218	OEP GBP 404; FaME GBP 412.; usual care GBP 367	Cost‐effectiveness analysis not conducted due to no between‐group difference in QALY
•Kemmler 2010 •Women aged ≥ 65 living independently •Analysis 4.1, 4.2 •Cost analysis	•Multicomponent exercise, 2 60‐minute classes and 2 20‐minute home training sessions weekly for 18 months (N = 115) vs control (low‐intensity exercise classes 60 minutes once‐weekly for 10 weeks followed by 10 weeks of rest) (N = 112) •All participants received calcium (1500 m/d) and cholecalciferol (500 IU/d) supplements	•Health system •Euro (Germany) •Not specified •During participation in 18‐month trial	•Total healthcare costs (no details provided)	‐	•EUR 2255 (2596) exercise group and EUR 2780 (3318) control group for mean total healthcare costs (P = 0.20)	‐
• Liu‐Ambrose 2008 (Davis 2009) •Women and men aged ≥ 70 years recruited from 2 referral‐based falls clinics •Analysis •Cost‐effectiveness analysis	•Specific set of muscle strengthening and balance retraining exercises individually prescribed at home (OEP) by trained physiotherapist for 1 year (N = 36) vs guideline care (N = 38) •All participants received falls risk assessment, comprehensive geriatric assessment and treatment	•Health system •Canadian dollar •Not specified •12 months	•Cost of delivering the intervention •Cost of the falls clinic	•CAD 14,285	‐	•CAD 247 per fall prevented (comparable to incremental cost‐effectiveness ratios in New Zealand studies of the Otago Exercise Program)
•Robertson 2001a •Men and women aged ≥ 75 years from 17 general practices, mean (SD) age 80.9 (4.2) years •Analysis •Cost‐effectiveness analysis	•Specific set of muscle‐strengthening and balance‐retraining exercises individually prescribed at home (OEP) by trained district nurse, supervised by physiotherapist, 5 home visits and monthly phone calls for 1 year (N = 121) vs usual care (N = 119)	•Health system •New Zealand dollar •1998 •During participation in 1‐year trial	•Intervention costs (training, recruitment, programme delivery, supervision of exercise instructor, overheads) •Hospital admission costs resulting from fall injuries during trial (actual costs of hospital admissions)	In community health service setting: •NZD 432 (0) for 1 year	•5 hospital admissions due to fall injuries in control group, none in exercise group (cost savings of NZD 47,818)	•NZD 1803 per fall prevented (programme implementation costs only) ‐ NZD 7471 per injurious fall prevented (programme implementation costs only) •NZD 155 per fall prevented (programme implementation costs and hospital admission cost savings) ‐ NZD 640 per injurious fall prevented (programme implementation costs and hospital admission cost savings)
•Sherrington 2014 (Farag 2015a) •Community‐dwelling people aged 60 years and over, discharged from hospital •Analysis •Cost‐effectiveness analysis Cost‐utility analysis	•Weight‐bearing Exercise for Better Balance (WEBB) programme, 15 – 20 minutes up to 6 times weekly for 12 months (N = 171) vs usual care (N = 169)	•Health and community care funder perspective (Australia) •Australian Dollar •2012 •1 year	•Costs of delivering the interventions (travel, staff, equipment, phone calls) •Cost of health service use (respite care, residential aged care, hospital admission, emergency department presentation, general practitioner, specialist and nursing services, allied health, social support services)	AUD 751 for WEBB AUD 0 for usual care	AUD 12,029 for WEBB AUD 10,327 for usual care	AUD 77,403 per QALY gained
•Uusi‐Rasi 2015 (Patil 2016) •Community‐dwelling women with mean age •74 years • Analysis •Cost‐effectiveness analysis	•No exercise + placebo •No exercise + vitamin D 800 IU/day •Exercise + placebo: supervised group training classes 2 a week for first year, and 1 a week for second year (N = 91) vs No exercise + placebo (control) (N = 95) •Exercise + vitamin D 800 IU/day	•Societal •Euros (Finland) •2011 •2 years	•Intervention costs (salaries, administration costs) •Healthcare costs (fall‐related health care costs for all injurious falls reported during the intervention period)	Total costs (intervention and healthcare): EUR 30.9 (95) for no exercise + placebo; EUR 206.9 (786) for no exercise + vitamin D 800IU/day; EUR 73.4 (104) for exercise + placebo; EUR 188.0 (454) for exercise + vitamin D 800IU/day	‐	ICER all intervention (excluding outliers): EUR 220.7 (220.7) for no exercise + placebo EUR 17,600 (exc) for no exercise + vitamin D 800 IU/day EUR 2670 (708.3) for exercise + placebo EUR 3820 (3820) for exercise + vitamin D 800IU/day
•Voukelatos 2007 (Haas 2006) •Healthy community‐living people aged ≥ 60 years, mean (SD) age 69 (6.5) years •Analysis •Cost‐effectiveness analysis	•Tai Chi classes 1 hour weekly for 16 weeks (N = 347) vs no intervention (N = 337)	•Public health system (NSW Health) •Australian dollar •Not specified (presumed 2001) •During 24‐week trial period	•Intervention costs (cost of venues, advertising, instructors) •Health service use related to falls from healthcare use diary and hospital records, valued at standard costs (GP, specialist, tests, hospitalisations, medications)	•AUD 245 (0) intervention group plus charge AUD 44 per participant	•Mean total healthcare costs higher for Tai Chi group (AUD 55) than control group (AUD 17) (P < 0.001)	•AUD 1683 per fall prevented (includes cost offset by charging AUD 44 per instruction course)

See also Davis 2010
 GP: general practitioner; HMO: health maintenance organisation; OEP: Otago Exercise Program; QALY: quality‐adjusted life‐year

Appendix 18. Sensitivity analyses: exploring impact on results (rate of falls outcome)

Sensitivity analysis	Pooled impact of exercise on fall rate, Rate ratio, 95% CI
Primary analysis, all trials, random‐effects meta‐analysis	0.77, 0.71 to 0.83; participants = 12,981; studies = 59
Sensitivity analysis 1, removing trials that included participants aged < 65 years	0.77, 0.71 to 0.84; participants = 11,807; studies = 53
Sensitivity analysis 2, removing trials with high risk of bias on any itema	0.78, 0.71 to 0.87; participants = 6757; studies = 25
Sensitivity analysis 3, removing trials with unclear or high risk of bias on allocation concealment	0.85, 0.77 to 0.95; participants = 6092; studies = 22
Sensitivity analysis 4, removing trials with unclear or high risk of bias on assessor blinding (falls outcome)	0.76, 0.69 to 0.85; participants = 6996; studies = 27
Sensitivity analysis 5, removing trials with unclear or high risk of bias on incomplete outcome data	0.77, 0.69 to 0.85; participants = 7646; studies = 36
Sensitivity analysis 6, removing cluster‐randomised trials	0.76, CI 0.70 to 0.83; participants = 10,261; studies = 53
Sensitivity analysis 7, all trials, fixed‐effect meta‐analysis	0.82, 0.79 to 0.86; participants = 12,981; studies = 59
Sensitivity analysis 8, multiple categories of exercise versus control, excluding trials that do not include balance and strength training	0.69, 0.48 to 0.97; participants = 1084; studies = 8
Primary analysis, subgrouped by exercise type Balance and functional exercises versus control Multiple categories of exercise versus control	0.76, CI 0.70 to 0.81; participants = 7920; studies = 39 0.66, CI 0.50 to 0.88; participants = 1374; studies = 11
Sensitivity analysis 9a, classification of interventions based on the Otago Exercise Program as multiple categories of exercise Balance and functional exercises versus control Multiple categories of exercise versus control	0.75, 0.68 to 0.82; participants = 5556; studies = 30 0.72, 0.62 to 0.83; participants = 3738; studies = 20
Sensitivity analysis 9b, classification of interventions that included balance and functional exercises plus strength exercises as multiple categories of exercise Balance and functional exercises versus control Multiple categories of exercise versus control	0.72, 0.62 to 0.84; participants = 2718; studies = 16 0.74, 0.67 to 0.81; participants = 6721; studies = 35

^aAfter removing trials assessed as high risk of bias in one or more key domains: random sequence generation (selection bias), allocation concealment (selection bias), blinding of outcome assessors (detection bias), and incomplete outcome data (attrition bias).

Data and analyses

Comparison 1. Exercise versus control (rate of falls).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Rate of falls ‐ overall analysis	59	12981	Rate Ratio (Random, 95% CI)	0.77 [0.71, 0.83]
2 Rate of falls ‐ subgrouped by baseline falls risk	59		Rate Ratio (Random, 95% CI)	Subtotals only
2.1 Not selected for high risk of falling	29	6123	Rate Ratio (Random, 95% CI)	0.74 [0.65, 0.84]
2.2 Selected for high risk of falling	30	6858	Rate Ratio (Random, 95% CI)	0.80 [0.72, 0.88]
3 Rate of falls ‐ subgrouped by age (threshold 75 years)	59		Rate Ratio (Random, 95% CI)	Subtotals only
3.1 Age < 75	46	9605	Rate Ratio (Random, 95% CI)	0.75 [0.69, 0.82]
3.2 Age 75+	13	3376	Rate Ratio (Random, 95% CI)	0.83 [0.72, 0.97]
4 Rate of falls ‐ subgrouped by personnel	59	12981	Rate Ratio (Random, 95% CI)	0.77 [0.71, 0.83]
4.1 Health professional delivering intervention	25	4511	Rate Ratio (Random, 95% CI)	0.69 [0.61, 0.79]
4.2 No health professional delivering intervention	34	8470	Rate Ratio (Random, 95% CI)	0.82 [0.75, 0.90]
5 Rate of falls ‐ subgrouped by group or individual exercise	59	12981	Rate Ratio (Random, 95% CI)	0.77 [0.71, 0.83]
5.1 Group exercise	40	8163	Rate Ratio (Random, 95% CI)	0.76 [0.69, 0.85]
5.2 Not group exercise	21	4818	Rate Ratio (Random, 95% CI)	0.79 [0.71, 0.88]
6 Rate of falls ‐ subgrouped by exercise type	59		Rate Ratio (Random, 95% CI)	Subtotals only
6.1 Balance and functional exercises vs control	39	7920	Rate Ratio (Random, 95% CI)	0.76 [0.70, 0.81]
6.2 Resistance exercise vs control	5	327	Rate Ratio (Random, 95% CI)	1.14 [0.67, 1.97]
6.3 3D exercise (Tai Chi) vs control	7	2655	Rate Ratio (Random, 95% CI)	0.81 [0.67, 0.99]
6.4 3D exercise (dance) vs control	1	522	Rate Ratio (Random, 95% CI)	1.34 [0.98, 1.83]
6.5 Walking programme vs control	2	441	Rate Ratio (Random, 95% CI)	1.14 [0.66, 1.97]
6.6 Multiple categories of exercise vs control	11	1374	Rate Ratio (Random, 95% CI)	0.66 [0.50, 0.88]
7 Rate of falls ‐ long‐term follow‐up by exercise type	4		Rate Ratio (Random, 95% CI)	Subtotals only
7.1 Balance and functional exercises vs control	2	858	Rate Ratio (Random, 95% CI)	0.82 [0.66, 1.01]
7.2 Walking programme vs control	1	97	Rate Ratio (Random, 95% CI)	1.27 [0.89, 1.81]
7.3 Multiple categories of exercise vs control	1	175	Rate Ratio (Random, 95% CI)	0.80 [0.55, 1.16]

Comparison 2. Exercise versus control (number of fallers).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Number of fallers ‐ overall analysis	63	13518	Risk Ratio (Random, 95% CI)	0.85 [0.81, 0.89]
2 Number of fallers ‐ subgrouped by baseline fall risk	63		Risk Ratio (Random, 95% CI)	Subtotals only
2.1 Not selected for high risk of falling	28	6347	Risk Ratio (Random, 95% CI)	0.82 [0.73, 0.92]
2.2 Selected for high risk of falling	35	7171	Risk Ratio (Random, 95% CI)	0.87 [0.83, 0.91]
3 Number of fallers ‐ subgrouped by age (threshold 75 years)	63		Risk Ratio (Random, 95% CI)	Subtotals only
3.1 Age < 75	50	10346	Risk Ratio (Random, 95% CI)	0.85 [0.79, 0.91]
3.2 Age 75+	13	3172	Risk Ratio (Random, 95% CI)	0.86 [0.80, 0.92]
4 Number of fallers ‐ subgrouped by personnel	62	13473	Risk Ratio (Random, 95% CI)	0.85 [0.81, 0.89]
4.1 Health professional delivering intervention	26	3747	Risk Ratio (Random, 95% CI)	0.82 [0.74, 0.91]
4.2 No health professional delivering intervention	36	9726	Risk Ratio (Random, 95% CI)	0.86 [0.81, 0.92]
5 Number of fallers ‐ subgrouped by group or individual exercise	63	13518	Risk Ratio (Random, 95% CI)	0.85 [0.81, 0.89]
5.1 Group exercise	48	9219	Risk Ratio (Random, 95% CI)	0.83 [0.78, 0.90]
5.2 Not group exercise	16	4299	Risk Ratio (Random, 95% CI)	0.88 [0.83, 0.93]
6 Number of fallers ‐ subgrouped by exercise type	63		Risk Ratio (Random, 95% CI)	Subtotals only
6.1 Balance and functional exercises vs control	37	8288	Risk Ratio (Random, 95% CI)	0.87 [0.82, 0.91]
6.2 Resistance exercise vs control	2	163	Risk Ratio (Random, 95% CI)	0.81 [0.57, 1.15]
6.3 3D exercise (Tai Chi) vs control	8	2677	Risk Ratio (Random, 95% CI)	0.80 [0.70, 0.91]
6.4 3D exercise (dance) vs control	1	522	Risk Ratio (Random, 95% CI)	1.35 [0.83, 2.20]
6.5 Multiple categories of exercise vs control	17	1623	Risk Ratio (Random, 95% CI)	0.78 [0.64, 0.96]
6.6 Walking programme vs control	2	441	Risk Ratio (Random, 95% CI)	1.05 [0.71, 1.54]
7 Number of fallers ‐ long‐term follow‐up by exercise type	3		Risk Ratio (Fixed, 95% CI)	Subtotals only
7.1 Balance and functional exercises vs control	2	1325	Risk Ratio (Fixed, 95% CI)	0.86 [0.78, 0.94]
7.2 Multiple categories of exercise vs control	1	175	Risk Ratio (Fixed, 95% CI)	1.01 [0.74, 1.38]

Comparison 3. Exercise versus control (number of people with fractures).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Number of people who experienced one or more fall‐related fractures‐ overall analysis	10	4047	Risk Ratio (Random, 95% CI)	0.73 [0.56, 0.95]
2 Number of people who experienced one or more fall‐related fractures ‐ subgrouped by baseline falls risk	10		Risk Ratio (Random, 95% CI)	Subtotals only
2.1 Not selected for high risk of falling	5	1255	Risk Ratio (Random, 95% CI)	0.48 [0.26, 0.91]
2.2 Selected for high risk of falling	5	2792	Risk Ratio (Random, 95% CI)	0.80 [0.60, 1.07]
3 Number of people who experienced one or more fall‐related fractures ‐ subgrouped by age (threshold 75 years)	10		Risk Ratio (Random, 95% CI)	Subtotals only
3.1 Age < 75	7	1307	Risk Ratio (Random, 95% CI)	0.53 [0.29, 0.96]
3.2 Age 75+	3	2740	Risk Ratio (Random, 95% CI)	0.61 [0.31, 1.20]
4 Number of people who experienced one or more fall‐related fractures ‐ subgrouped by exercise type	10		Risk Ratio (Random, 95% CI)	Subtotals only
4.1 Balance and functional exercises vs control	7	2139	Risk Ratio (Random, 95% CI)	0.44 [0.25, 0.76]
4.2 Resistance exercise vs control	1	73	Risk Ratio (Random, 95% CI)	0.97 [0.14, 6.49]
4.3 Walking programme vs control	1	97	Risk Ratio (Random, 95% CI)	0.66 [0.11, 3.76]
4.4 Multiple categories of exercise vs control	3	1810	Risk Ratio (Random, 95% CI)	0.85 [0.62, 1.16]
5 Number of people who experienced one or more fall‐related fractures ‐ long‐term follow‐up by exercise type	3	2351	Risk Ratio (Fixed, 95% CI)	0.93 [0.69, 1.25]
5.1 Balance and functional exercises vs control	1	619	Risk Ratio (Fixed, 95% CI)	1.80 [0.46, 7.11]
5.2 Walking programme vs control	1	97	Risk Ratio (Fixed, 95% CI)	1.46 [0.44, 4.83]
5.3 Multiple categories of exercise vs control	1	1635	Risk Ratio (Fixed, 95% CI)	0.87 [0.64, 1.19]

Comparison 4. Exercise versus control (number of people with falls that resulted in hospital admission).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Number of people who experienced one or more falls that resulted in hospital admission ‐ overall analysis	2	1705	Risk Ratio (IV, Random, 95% CI)	0.78 [0.51, 1.18]

Comparison 5. Exercise versus control (number of people with falls that required medical attention).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Number of people who experienced one or more falls that required medical attention‐ overall analysis	5	1019	Risk Ratio (Random, 95% CI)	0.61 [0.47, 0.79]
2 Number of people who experienced one or more falls that required medical attention ‐ subgrouped by exercise type	5		Risk Ratio (Random, 95% CI)	Subtotals only
2.1 Balance and functional exercises vs control	3	583	Risk Ratio (Random, 95% CI)	0.76 [0.54, 1.09]
2.2 Resistance exercises vs control	1	73	Risk Ratio (Random, 95% CI)	0.92 [0.47, 1.80]
2.3 3D exercise (Tai Chi) vs control	1	188	Risk Ratio (Random, 95% CI)	0.35 [0.13, 0.93]
2.4 Multiple categories of exercise vs control	2	247	Risk Ratio (Random, 95% CI)	0.44 [0.29, 0.66]
3 Number of people who experienced one or more falls that required medical attention ‐ long‐term follow‐up pooled	2	319	Risk Ratio (Random, 95% CI)	0.54 [0.37, 0.78]

Comparison 6. Exercise versus control (health‐related quality of life).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Health‐related quality of life‐ overall analysis	15	3172	Std. Mean Difference (IV, Fixed, 95% CI)	‐0.03 [‐0.10, 0.04]
2 Health‐related quality of life ‐ subgrouped by baseline fall risk	15		Std. Mean Difference (IV, Random, 95% CI)	Subtotals only
2.1 Not selected for high risk of falling	8	2420	Std. Mean Difference (IV, Random, 95% CI)	‐0.01 [‐0.24, 0.23]
2.2 Selected for high risk of falling	7	752	Std. Mean Difference (IV, Random, 95% CI)	0.05 [‐0.12, 0.22]

6.2. Analysis.

Comparison 6 Exercise versus control (health‐related quality of life), Outcome 2 Health‐related quality of life ‐ subgrouped by baseline fall risk.

Comparison 7. Exercise versus control (number of people who died).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Number of people who died‐ overall analysis	30	10037	Risk Ratio (IV, Random, 95% CI)	0.86 [0.66, 1.12]
2 Number of people who died ‐ subgrouped by baseline fall risk	30		Risk Ratio (IV, Random, 95% CI)	Subtotals only
2.1 Not selected for high risk of falling	12	4606	Risk Ratio (IV, Random, 95% CI)	0.94 [0.54, 1.67]
2.2 Selected for high risk of falling	18	5421	Risk Ratio (IV, Random, 95% CI)	0.82 [0.60, 1.12]

Comparison 8. Balance and functional exercises versus control: subgroup analyses.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Rate of falls, subgrouped by baseline fall risk	39		Rate Ratio (Random, 95% CI)	Subtotals only
1.1 Not selected for higher risk of falling	18	3355	Rate Ratio (Random, 95% CI)	0.80 [0.72, 0.90]
1.2 Selected for higher risk of falling	21	4602	Rate Ratio (Random, 95% CI)	0.72 [0.65, 0.80]
2 Number of fallers, subgrouped by baseline fall risk	37		Risk Ratio (Random, 95% CI)	Subtotals only
2.1 Not selected for higher risk of falling	15	3649	Risk Ratio (Random, 95% CI)	0.88 [0.80, 0.97]
2.2 Selected for higher risk of falling	22	4639	Risk Ratio (Random, 95% CI)	0.86 [0.81, 0.91]
3 Rate of falls, subgrouped by personnel	39		Rate Ratio (Random, 95% CI)	Subtotals only
3.1 Health professional delivering intervention	20	2960	Rate Ratio (Random, 95% CI)	0.67 [0.58, 0.76]
3.2 No health professional delivering intervention	19	4997	Rate Ratio (Random, 95% CI)	0.82 [0.76, 0.88]
4 Number of fallers, subgrouped by personnel	37		Risk Ratio (Random, 95% CI)	Subtotals only
4.1 Health professional delivering intervention	19	2894	Risk Ratio (Random, 95% CI)	0.82 [0.75, 0.90]
4.2 No health professional delivering intervention	18	5394	Risk Ratio (Random, 95% CI)	0.89 [0.84, 0.94]
5 Rate of falls, subgrouped by group or individual exercise	39		Rate Ratio (Random, 95% CI)	Subtotals only
5.1 Group exercise	20	3620	Rate Ratio (Random, 95% CI)	0.73 [0.65, 0.82]
5.2 Not group exercise	20	4589	Rate Ratio (Random, 95% CI)	0.77 [0.70, 0.85]
6 Number of fallers, subgrouped by group or individual exercise	37		Risk Ratio (Random, 95% CI)	Subtotals only
6.1 Group exercise	22	4465	Risk Ratio (Random, 95% CI)	0.87 [0.80, 0.95]
6.2 Not group exercise	16	4075	Risk Ratio (Random, 95% CI)	0.87 [0.82, 0.92]

Comparison 9. Multiple categories of exercise versus control: subgroup analyses.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Rate of falls, subgrouped by baseline fall risk	11		Rate Ratio (Random, 95% CI)	Subtotals only
1.1 Not selected for higher risk of falling	6	786	Rate Ratio (Random, 95% CI)	0.54 [0.29, 0.99]
1.2 Selected for higher risk of falling	5	618	Rate Ratio (Random, 95% CI)	0.77 [0.63, 0.94]
2 Number of fallers, subgrouped by baseline fall risk	17	1623	Risk Ratio (Random, 95% CI)	0.78 [0.64, 0.96]
2.1 Not selected for higher risk of falling	7	710	Risk Ratio (Random, 95% CI)	0.70 [0.41, 1.19]
2.2 Selected for higher risk of falling	10	913	Risk Ratio (Random, 95% CI)	0.84 [0.71, 1.00]
3 Rate of falls, subgrouped by personnel