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).