Dual combination therapy versus long‐acting bronchodilators alone for chronic obstructive pulmonary disease (COPD): a systematic review and network meta‐analysis

Abstract

Background

Long‐acting bronchodilators such as long‐acting β‐agonist (LABA), long‐acting muscarinic antagonist (LAMA), and LABA/inhaled corticosteroid (ICS) combinations have been used in people with moderate to severe chronic obstructive pulmonary disease (COPD) to control symptoms such as dyspnoea and cough, and prevent exacerbations. A number of LABA/LAMA combinations are now available for clinical use in COPD. However, it is not clear which group of above mentioned inhalers is most effective or if any specific formulation works better than the others within the same group or class.

Objectives

To compare the efficacy and safety of available formulations from four different groups of inhalers (i.e. LABA/LAMA combination, LABA/ICS combination, LAMA and LABA) in people with moderate to severe COPD. The review will update previous systematic reviews on dual combination inhalers and long‐acting bronchodilators to answer the questions described above using the strength of a network meta‐analysis (NMA).

Search methods

We identified studies from the Cochrane Airways Specialised Register, which contains several databases. We also conducted a search of ClinicalTrials.gov and manufacturers’ websites. The most recent searches were conducted on 6 April 2018.

Selection criteria

We included randomised controlled trials (RCTs) that recruited people aged 35 years or older with a diagnosis of COPD and a baseline forced expiratory volume in one second (FEV1) of less than 80% of predicted. We included studies of at least 12 weeks' duration including at least two active comparators from one of the four inhaler groups.

Data collection and analysis

We conducted NMAs using a Bayesian Markov chain Monte Carlo method. We considered a study as high risk if recruited participants had at least one COPD exacerbation within the 12 months before study entry and as low risk otherwise. Primary outcomes were COPD exacerbations (moderate to severe and severe), and secondary outcomes included symptom and quality‐of‐life scores, safety outcomes, and lung function. We collected data only for active comparators and did not consider placebo was not considered. We assumed a class/group effect when a fixed‐class model fitted well. Otherwise we used a random‐class model to assess intraclass/group differences. We supplemented the NMAs with pairwise meta‐analyses.

Main results

We included a total of 101,311 participants from 99 studies (26 studies with 32,265 participants in the high‐risk population and 73 studies with 69,046 participants in the low‐risk population) in our systematic review. The median duration of studies was 52 weeks in the high‐risk population and 26 weeks in the low‐risk population (range 12 to 156 for both populations). We considered the quality of included studies generally to be good.

The NMAs suggested that the LABA/LAMA combination was the highest ranked treatment group to reduce COPD exacerbations followed by LAMA in the both populations.

There is evidence that the LABA/LAMA combination decreases moderate to severe exacerbations compared to LABA/ICS combination, LAMA, and LABA in the high‐risk population (network hazard ratios (HRs) 0.86 (95% credible interval (CrI) 0.76 to 0.99), 0.87 (95% CrI 0.78 to 0.99), and 0.70 (95% CrI 0.61 to 0.8) respectively), and that LAMA decreases moderate to severe exacerbations compared to LABA in the high‐ and low‐risk populations (network HR 0.80 (95% CrI 0.71 to 0.88) and 0.87 (95% CrI 0.78 to 0.97), respectively). There is evidence that the LABA/LAMA combination reduces severe exacerbations compared to LABA/ICS combination and LABA in the high‐risk population (network HR 0.78 (95% CrI 0.64 to 0.93) and 0.64 (95% CrI 0.51 to 0.81), respectively).

There was a general trend towards a greater improvement in symptom and quality‐of‐life scores with the combination therapies compared to monotherapies, and the combination therapies were generally ranked higher than monotherapies.

The LABA/ICS combination was the lowest ranked in pneumonia serious adverse events (SAEs) in both populations. There is evidence that the LABA/ICS combination increases the odds of pneumonia compared to LAMA/LABA combination, LAMA and LABA (network ORs: 1.69 (95% CrI 1.20 to 2.44), 1.78 (95% CrI 1.33 to 2.39), and 1.50 (95% CrI 1.17 to 1.92) in the high‐risk population and network or pairwise OR: 2.33 (95% CI 1.03 to 5.26), 2.02 (95% CrI 1.16 to 3.72), and 1.93 (95% CrI 1.29 to 3.22) in the low‐risk population respectively). There were significant overlaps in the rank statistics in the other safety outcomes including mortality, total, COPD, and cardiac SAEs, and dropouts due to adverse events.

None of the differences in lung function met a minimal clinically important difference criterion except for LABA/LAMA combination versus LABA in the high‐risk population (network mean difference 0.13 L (95% CrI 0.10 to 0.15). The results of pairwise meta‐analyses generally agreed with those of the NMAs. There is no evidence to suggest intraclass/group differences except for lung function at 12 months in the high‐risk population.

Authors' conclusions

The LABA/LAMA combination was the highest ranked treatment group to reduce COPD exacerbations although there was some uncertainty in the results. LAMA containing inhalers may have an advantage over those without a LAMA for preventing COPD exacerbations based on the rank statistics. Combination therapies appear more effective than monotherapies for improving symptom and quality‐of‐life scores. ICS‐containing inhalers are associated with an increased risk of pneumonia.

Our most comprehensive review including intraclass/group comparisons, free combination therapies, 99 studies, and 20 outcomes for each high‐ and low‐risk population summarises the current literature and could help with updating existing COPD guidelines.

Plain language summary

Which long‐acting inhalers are the most effective and safest for people with advanced chronic obstructive pulmonary disease (COPD)?

What is COPD and why does a doctor prescribe an inhaler?

Chronic obstructive lung disease (COPD) is usually caused by smoking or other airway irritants. COPD damages the lungs and causes airways to narrow which makes it difficult to breathe.

There are two types of inhalers for COPD: rescue and maintenance. A rescue inhaler is short‐ and fast‐acting, and used as needed for quick relief of symptoms, whereas a maintenance inhaler is long‐acting and used on a daily basis to relieve daily symptoms and reduce flare‐ups. The long‐acting inhalers are usually reserved for more advanced COPD.

Does it matter which long‐acting inhaler is used in people with advanced COPD?

Commonly used maintenance inhalers are grouped into four different groups: long‐acting beta2‐agonists (LABAs); long‐acting muscarinic antagonists (LAMAs); LABA/inhaled corticosteroid (ICS) combinations; and LABA/LAMA combinations. Combination inhalers are usually reserved for individuals whose single‐maintenance inhaler, such as LAMA or LABA fails. There are not many head‐to‐head comparisons to determine which treatment group or individual inhaler is better compared to the others. Preventing severe flare‐ups and hospital admissions is especially important to people with COPD, healthcare providers, policy makers and society.

How did we answer the question?

We collected and analysed data from 99 studies, including a total of 101,311 participants with advanced COPD, using a special method called network meta‐analysis, which enabled us to simultaneously compare the four inhaler groups and 28 individual inhalers (4 LABAs, 5 LAMAs, 9 LABA/ICS combinations, and 10 LABA/LAMA combinations).

What did we find?

The LABA/LAMA combination was the best treatment, followed by LAMA, in preventing flare‐ups although there was some uncertainty in the results. Combination inhalers (LABA/LAMA and LABA/ICS), are more effective for controlling symptoms than single‐agent therapies (LAMA and LABA), in general. The LABA/LAMA combination was better than LABA/ICS combination, especially in people with a prior episode of flare‐ups. The LABA/ICS combination had a higher incidence of severe pneumonia compared to the others. We did not find a difference in benefits and harms, including side effects, among individual inhalers within the same treatment groups.

Conclusion

The LABA/LAMA combination is likely the best treatment in preventing COPD flare‐ups. LAMA‐containing inhalers appear to have an advantage over those without LAMA for preventing flare‐ups. Combination inhalers (LABA/LAMA and LABA/ICS), appear more effective for controlling symptoms than single‐agent therapies (LAMA and LABA). Inhaled steroids carry an increased risk of pneumonia.

Background

Description of the condition

Chronic obstructive pulmonary disease (COPD) is a globally prevalent illness, characterised by chronic airway inflammation leading to slow progression of airflow limitation (GOLD 2018). The inflammatory nature of the disease leads to variable degrees of small airway obstruction and destruction of lung parenchyma. COPD accounts for more than three million deaths annually and is the third leading cause of death worldwide. This disease is due primarily to tobacco smoke in high‐income countries; tobacco smoking is also the primary cause of COPD in low‐income countries, but air pollution and indoor biomass fuel consumption are more frequent causes compared to high‐income countries. The disease affects men and women equally (WHO 2016). Despite the worldwide prevalence of the disease, it remains largely under‐recognised and underdiagnosed. COPD is a costly disease, with an estimated annual cost of USD 49.9 billion, including an indirect cost estimated at approximately 41% of the total cost in the USA and a total cost of EUR 38.7 billion in Europe (Patel 2014; WHO 2016). Clinically, the disease is characterised by chronic dyspnoea, productive cough and exposure to a risk factor such as smoking. The post‐bronchodilator forced expiratory volume in one second (FEV1)/forced vital capacity (FVC) is required to be less than 0.7 for this diagnosis (GOLD 2018). The disease course is usually interrupted by episodes of acute exacerbation, the frequency of which contributes to overall morbidity and mortality (Suissa 2012).

Description of the intervention

Management of stable COPD

Once COPD has been diagnosed, the main goals of therapy include alleviation of symptoms and prevention of disease progression and acute exacerbations. Smoking cessation is one of the most important non‐pharmacological interventions. Annual influenza vaccination is recommended for everyone with COPD. In observational studies, influenza vaccination was associated with fewer outpatient visits, hospitalisations and deaths (Trucchi 2015). Pulmonary rehabilitation has been proven to improve exercise tolerance while reducing symptoms and exacerbations (McCarthy 2015; Rochester 2015). Inhaled medications, the mainstay of pharmacological therapies, are used to improve lung function, symptoms and quality of life, as well as to reduce acute exacerbations. Short‐acting bronchodilators are given on an as‐needed basis to provide immediate relief, and long‐acting bronchodilators are used as maintenance therapy in people with moderate to very severe disease (Decramer 2012). The Global Initiative for Chronic Obstructive Lung Disease (GOLD), recommends long‐acting bronchodilators as maintenance therapy in people experiencing long‐term respiratory symptoms or exacerbations.(GOLD 2018).

How the intervention might work

Combination bronchodilators

Dual combination inhalers include long‐acting beta‐adrenoceptor agonist/inhaled corticosteroid (LABA/ICS) and LABA/long‐acting muscarinic antagonist (LAMA) combinations. An ICS has anti‐inflammatory effects and may reduce airway inflammation as well as systemic inflammation, as evidenced by a reduction in C‐reactive protein (Heidari 2012). ICSs and LABAs have synergistic effects when used in combination. Corticosteroids upregulate beta₂‐receptors and beta₂‐agnoists and facilitate translocation of steroid receptors from the cytoplasm to the nucleus (Falk 2008). In vitro synergistic effects mentioned above may translate into clinical benefit. Clinical studies have suggested that a LABA/ICS combination significantly improved lung function, health status and rate of exacerbation compared with placebo, LABA alone or ICS alone (Nannini 2012).

Preclinical studies have suggested drug synergy between a beta₂‐adrenoreceptor agonist and a muscarinic agonist. A possible mechanism for this synergism is that a muscarinic agonist causes less suppression of potassium channel opening, leading to relaxation of the airway smooth muscle, which further promotes beta₂‐mediated smooth muscle relaxation by activating ion channels and other intracellular signalling pathways (Kume 2014). Clinical studies have demonstrated that LABA/LAMA combinations were superior to monotherapies with regard to lung function improvement and in a recent network meta‐analysis (NMA), were associated with improved quality of life and symptom scores, and reduced COPD exacerbations as compared with LABA or LAMA alone (Oba 2016a).

Guidelines recommend a LABA/LAMA combination for people whose symptoms are not well controlled with a single long‐acting bronchodilator, and a LABA/LAMA or LABA/ICS combination for those with frequent exacerbations (i.e. two or more exacerbations per year or one hospitalisation per year for an exacerbation). A LABA/LAMA combination may be preferred to a LABA/ICS combination, as ICSs are associated with increased risk of pneumonia (GOLD 2018; Oba 2016b; Wedzicha 2016).

Why it is important to do this review

Data on the efficacy and safety of LABA/LAMA combinations are accumulating (Huisman 2015; Oba 2016a; Schlueter 2016). However, an important clinical question is how do the efficacy and safety of LABA/LAMA combinations compare with those of LABA/ICS combinations for people with uncontrolled symptoms or frequent exacerbations, or both. Additional clinical studies, including several head‐to‐head studies comparing LABA/LAMA and LABA/ICS combinations (Donohue 2015; Singh 2015d; Vogelmeier 2013a; Vogelmeier 2015; Wedzicha 2016; Zhong 2015), have been published since an NMA comparing combination inhalers focused on studies up to December 2013 (Tricco 2015). Our review updates previous systematic reviews on dual combination inhalers and long‐acting bronchodilators using the strength of an NMA.

Objectives

To compare the efficacy and safety of available formulations from four different groups of inhalers (i.e. LABA/LAMA combination, LABA/ICS combination, LAMA and LABA) in people with moderate to severe COPD. The review will update previous systematic reviews on dual combination inhalers and long‐acting bronchodilators to answer the questions described above using the strength of a network meta‐analysis (NMA).

Methods

Criteria for considering studies for this review

Types of studies

We included parallel, randomised controlled trials (RCTs), of at least 12 weeks' duration, published or unpublished. We did not consider cross‐over studies.

Types of participants

We included studies that recruited people aged 35 years or older with a diagnosis of COPD, in accordance with American Thoracic Society‐European Respiratory Society (ATS/ERS 2004), GOLD report (GOLD 2018), or equivalent criteria. Obstructive ventilatory defect should be at least moderate, with a baseline FEV1 less than 80% of predicted. We excluded studies that enrolled participants with a history of asthma or other respiratory disease.

Types of interventions

We included studies comparing at least two of the following therapies. We limited treatment arms to drug formulations and doses that were licensed in the USA or EU countries, or both, for clinical use. We did not consider triple combination therapy (i.e. LABA/LAMA/ICS) because it was out of scope for this review.

1. LAMA monotherapy

2. LABA monotherapy

3. Fixed‐dose or free combination of LABA/ICS

4. Fixed‐dose or free combination of LABA/LAMA

We allowed the use of a short‐acting bronchodilator, such as salbutamol( also known as albuterol), and ipratropium as rescue treatment.

Types of outcome measures

Primary outcomes

1. COPD exacerbations (moderate to severe and severe)

Secondary outcomes

1. Change from baseline in St George's Respiratory Questionnaire (SGRQ) score and decrease in SGRQ score by 4 units or more (SGRQ responder)

2. Transition Dyspnea Index (TDI)

3. Mortality

4. Total serious adverse events (SAEs)

5. Cardiac and COPD SAEs

6. Dropouts due to adverse events

7. Change from baseline in trough FEV1

8. Pneumonia reported as SAE

We used an end‐point score for dichotomous outcomes. For continuous outcomes, we used a change score reported at 3, 6, 12 months and the end of the study, when available. We defined 'moderate exacerbation' as worsening of respiratory status that requires treatment with systemic corticosteroids or antibiotics, or both; we defined 'severe exacerbation' as rapid deterioration that requires hospitalisation. The above‐mentioned outcomes and their definitions are well established and widely used across the medical literature.

Search methods for identification of studies

Electronic searches

We identified studies from the Cochrane Airways Trials Register, which is maintained by the Information Specialist for the Group. The Register contains trial reports identified through systematic searches of the following bibliographic databases:

1. monthly searches of the Cochrane Central Register of Controlled Trials (CENTRAL), through the Cochrane Register of Studies (CRS);

2. weekly searches of MEDLINE Ovid SP 1946 to date;

3. weekly searches of Embase Ovid SP 1974 to date;

4. Monthly searches of PsycINFO Ovid SP 1967 to date;

5. Monthly searches of CINAHL EBSCO (Cumulative Index to Nursing and Allied Health Literature) 1937 to date;

6. Monthly searches of AMED EBSCO (Allied and Complementary Medicine) all years to date;

7. handsearches of the proceedings of major respiratory conferences.

Studies contained in the Trials Register are identified through search strategies based on the scope of Cochrane Airways. Details of these strategies, as well as a list of handsearched conference proceedings are in Appendix 1. See Appendix 2 for search terms used to identify studies for this review.

We also conducted a search of ClinicalTrials.gov (www.ClinicalTrials.gov) and manufacturers’ websites. We searched all sources from their inception to 6 April 2018, and we imposed no restriction on language of publication.

Searching other resources

We checked the reference lists of all primary studies and review articles for additional references. We searched relevant manufacturers' websites for study information. We searched for errata or retractions from included studies published in full text on PubMed (www.ncbi.nlm.nih.gov/pubmed) and reported within the review the date this was done.

Data collection and analysis

Selection of studies

Two review authors (YO, NG) independently screened studies by title and abstract to evaluate whether a study met the inclusion and exclusion criteria. We selected studies that evaluated the clinical efficacy and safety of any of the following therapies in people with COPD: LABA/LAMA, LABA/ICS, LABA and LAMA. We resolved disagreements by involving a third contributor Joe V Devasahayam (JVD). We recorded the selection process in sufficient detail to complete a PRISMA flow diagram and a 'Characteristics of excluded studies' table (Moher 2009).

Data extraction and management

Two review authors (YO, NG), independently extracted information on study design, study size, population, interventions (drug, dose, inhaler type, allowed co‐medications), severity of illness and end points of interest. We gathered information on whether a participant had been unsuccessfully treated with a long‐acting bronchodilator before entry into clinical studies. We extracted and verified data from each of the existing reviews, which were cross‐checked and verified by at least two review authors. We resolved disagreements regarding values, inconsistencies and uncertainties by involving a third contributor. Two review authors (YO, NG) independently extracted outcome data from the included studies. We noted in the 'Characteristics of included studies' table if outcome data were not reported in a useable way. We resolved disagreements by reaching consensus or by involving a third contributor (JVD). One review author (YO) transferred data into the Review Manager 5 file (Review Manager 2014). We double‐checked that data had been entered correctly by comparing data presented in the systematic review versus study reports. A second review author (NG) spot‐checked study characteristics for accuracy against the study report.

Assessment of risk of bias in included studies

Two review authors (YO, NG) independently assessed risk of bias for each study using the criteria outlined in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2017). We resolved disagreements by discussion or by consultation with another contributor (JVD). We assessed risk of bias according to the following domains.

1. Random sequence generation

2. Allocation concealment

3. Blinding of participants and personnel

4. Blinding of outcome assessment

5. Incomplete outcome data

6. Selective outcome reporting

7. Other bias

We graded each potential source of bias as high, low or unclear and provided a quote from the study report together with a justification for our judgement in the 'Risk of bias' table. We summarised 'Risk of bias' judgements across different studies for each of the domains listed. We considered blinding separately for different key outcomes when necessary (e.g. for unblinded outcome assessment, risk of bias for all‐cause mortality may have been very different than for a patient‐reported dyspnoea scale). When information on risk of bias related to unpublished data, we noted this in the 'Risk of bias' table. When considering treatment effects, we took into account the risk of bias for studies that contributes to that outcome.

Assessment of bias in conducting the systematic review

We conducted the review according to this published protocol and reported deviations from it in the 'Differences between protocol and review' section of the systematic review.

Measures of treatment effect

Network meta‐analysis

We conducted NMAs using a Bayesian Markov chain Monte Carlo method and fitted in WinBUGS (version 1.4.3.), using code adapted from Dias 2018, which correctly accounts for correlations in studies with more than two arms and allows the specific data structures being considered. We compared each pair of treatments by estimating an odds ratio (OR) or hazard ratio (HR) for dichotomous outcomes, and a difference in mean or median for continuous outcomes, along with their 95% credible intervals (CrIs). We used a normal likelihood with an identity link for continuous outcomes (FEV1, TDI and SGRQ) and a binomial likelihood with a logit link for mortality, SAEs (total, cardiac and COPD), dropouts due to adverse events, SGRQ responders and pneumonia. We used a shared parameter model for exacerbation outcomes, whereby data on the log hazard ratio (lnHR and standard error) were modelled with the assumption that continuous treatment differences (lnHR) had a normal likelihood. When lnHR data were not available, or when appropriate covariance matrices could not be extracted or calculated for studies with more than two arms, we modelled data on the number of participants with at least one exacerbation out of the total number of participants at a given time as lnHR by using a binomial likelihood with Cloglog link. We used lnHR data in preference to dichotomous data when available and considered only the HR for the first event. We assessed model fit by comparing residual deviance to the number of data points, and by assessing the size of the between‐study standard deviation (SD).

Direct pairwise meta‐analysis

We conducted pairwise meta‐analyses (MAs) considering only direct evidence. We analysed dichotomous data as ORs and continuous data as mean differences (MDs) along with their 95% confidence intervals (CIs). We undertook MAs only when this was meaningful (i.e. if treatments, participants and the underlying clinical question were similar enough for pooling to make sense). When a single study reported multiple study arms, we included only the relevant arms.

Unit of analysis issues

We analysed dichotomous data by using number of participants (rather than events), as the unit of analysis to avoid multiple counting of data from the same participant.

Dealing with missing data

We requested additional data from the responsible author of the included studies to verify key study characteristics and to obtain missing numerical outcome data when possible (e.g. when a study was identified as an abstract only). When this was not possible, and when the missing data were thought to introduce serious bias, we explored the impact of including such studies in the overall assessment of results by performing a sensitivity analysis.

Assessment of heterogeneity

Assessment of similarity of participants, interventions and study methods

We assessed similarity of participants, interventions, potential effect modifiers and study methods in all studies and across pairwise comparisons to examine heterogeneity and inconsistency in the NMAs. The initial editorial review for study protocol had questioned the similarity of patient populations across clinical studies owing to the presence of potential effect modifiers. After a preliminary search of clinical studies and a review of inclusion/exclusion criteria, participant characteristics and study methods, we decided to divide the study populations into those with and without a history of COPD exacerbations within 12 months before study entry, which we viewed as a potential effect modifier (Table 8). This is consistent with the GOLD 2018 update, which recommends treatment options based on an exacerbation history.

1. Study characteristics of included trials.

High‐risk group
Study	Number of participants	Study duration (weeks)	Arms included (drug, dose in μg, dosing frequency)	Mean age (years)	Male (%)	Current smoker (%)	Prebronchodilator FEV1 (L)	Bronchial reversibility (%)
Aaron 2007	304	52	Tio 18 once daily + SAL 50 twice daily Tio 18 once daily	68	56	26	1.01	NR
Agusti 2014	528	12	FP/SAL 500/50 twice daily FF/VI 100/25 once daily	63	82	NR	1.29	11.8
Anzueto 2009	797	52	FP/SAL 250/50 twice daily SAL 50 twice daily	65	54	43	0.98	21
Calverley 2003	509	52	BUD/FM 320/9 twice daily FM 9 twice daily	63	76	35	0.99	NR
Calverley 2003 TRISTAN	730	52	FP/SAL 500/50 twice daily SAL 50 twice daily	63	75	51	1.28	7.8
Calverley 2010	703	48	BDP/FM 200/12 twice daily BUD/FM 400/12 twice daily FM 12 twice daily	64	81	37	1.15	NR
COMBINE 2017	222	24	FP 250 twice daily + SAL 50 twice daily BUD 400 twice daily + IND 150 once daily	67	57	NR	NR	NR
Decramer 2013	3439	52	IND 150 once daily Tio 18 once daily	64	77	34	NR	NR
Ferguson 2008	776	52	FP/SAL 250/50 twice daily SAL 50 twice daily	65	55	39	0.94	24.2
Ferguson 2017	1219	26	BUD/FM 320/9 twice daily FM 9 twice daily	64	57	NR	NR	NR
Fukuchi 2013	1293	12	BUD/FM 320/9 twice daily FM 9 twice daily	65	89	34	0.96	13.6
Hagedorn 2013	213	52	FP/SAL 500/50 twice daily FP 500 + SAL 50 twice daily	65	71	29	1.05	NR
Kardos 2007	994	44	FP/SAL 500/50 twice daily SAL 50 twice daily	64	76	42	1.13	7
Ohar 2014	639	26	FP/SAL 250/50 twice daily SAL 50 twice daily	63	91	NR	1.11	13.6
Pepin 2014	257	12	FF/VI 100/25 once daily Tio 18 once daily	67	86	46	1.27	8.5
Rennard 2009	1483	52	BUD/FM 320/9 twice daily BUD/FM 160/9 twice daily FM 9 twice daily	63	64	42	1.00	NR
Sarac 2016	44	52	FP/SAL 500/50 twice daily Tio 18 once daily	67	95	NR	NR	NR
SCO40041 2008	186	156	FP/SAL 250/50 twice daily SAL 50 twice daily	66	61	42	1.14	15.2
Sharafkhaneh 2012	1218	52	BUD/FM 320/9 twice daily BUD/FM 160/9 twice daily FM 9 twice daily	63	62	36	1.00	NR
Szafranski 2003	409	52	BUD/FM 320/9 twice daily FM 9 twice daily	64	76	34	0.98	NR
Tashkin 2008	842	24	BUD/FM 320/9 twice daily BUD/FM 160/9 twice daily FM 9 twice daily	63	66	45	1.04	NR
Vogelmeier 2011	7376	52	SAL 50 twice daily Tio 18 once daily	63	75	48	NR	NR
Wedzicha 2008	1323	104	FP/SAL 250/50 twice daily Tio 18 once daily	65	83	38	1.05	6.7
Wedzicha 2013	2206	64	IND/Glyco 110/50 once daily Glyco 50 once daily Tio 18 once daily	63	75	38	0.90	18.3
Wedzicha 2014	1197	48	BDP/FM 200/12 twice daily FM 12 twice daily	64	69	40	1.05	10.8
Wedzicha 2016	3358	52	IND/Glyco 110/50 once daily FP/SAL 500/50 twice daily	65	76	40	1.00	22.4
Low‐risk group
Study	Number of participants	Study duration (weeks)	Arms included (drug, dose in μg, dosing frequency)	Mean age (years)	Male (%)	Current smoker (%)	Prebronchodilator FEV1 (L)	Bronchial reversibility (%)
Asai 2013	158	52	IND/Glyco 110/50 once daily Tio 18 once daily	69	96	NR	NR	NR
BI 205.137 2001	385	12	SAL 50 twice daily Tio 18 once daily	NR	NR	NR	NR	NR
Bateman 2013	1903	26	IND/Glyco 110/50 once daily Glyco 50 once daily Tio 18 once daily IND 150 once daily	64	75	40	1.30	20.4
Bogdan 2011	405	12	FM 4.5 twice daily FM 9 twice daily	67	87	NR	1.30	10.6
Briggs 2005	653	12	SAL 50 twice daily Tio 18 once daily	64	67	36	1.05	NR
Brusasco 2003	807	24	SAL 50 twice daily Tio 18 once daily	64	76	NR	1.09	NR
Buhl 2011	1598	12	IND 150 once daily Tio 18 once daily	64	69	45	1.33	13.9
Buhl 2015a&b	3100	52	Tio/Olo 5/5 once daily Tio 5 once daily Olo 5 once daily	64	73	37	1.20	14.2
Buhl 2015c	934	26	IND/Glyco 110/50 once daily Tio 18 once daily + FM 12 twice daily	63	66	49	1.33	19.4
Calverley 2007	3054	156	FP/SAL 500/50 twice daily SAL 50 twice daily	65	75	43	1.11	10.2
Cazzola 2007	52	12	FP/SAL 500/50 twice daily Tio 18 once daily	65	90	38	NR	12.3
Chapman 2014	657	12	Glyco 50 once daily Tio 18 once daily	64	74	45	NR	NR
COSMOS‐J 2016	262	24	FP/SAL 250/50 twice daily Tio 18 once daily	68	95	40	NR	NR
Covelli 2016	623	12	FF/VI 100/25 once daily TIO 18 once daily	63	65	52	1.35	13
D'Urzo 2014	994	24	ACL/FM 400/12 twice daily ACL 400 twice daily FM 12 twice daily	64	52	51	1.35	17.4
D'Urzo 2017	568	52	ACL/FM 400/12 twice daily ACL 400 twice daily FM 12 twice daily	63	50	56	1.34	18.3
Dahl 2010	871	52	IND 300 once daily FM 12 twice daily	64	80	NR	1.29	10
Decramer 2014a	420	24	UMEC/VI 62.5/25 once daily Tio 18 once daily	63	69	47	1.31	11.6
Decramer 2014b	432	24	UMEC/VI 62.5/25 once daily Tio 18 once daily	65	68	45	1.17	15.2
Donohue 2010	1247	26	IND150 once daily IND 300 once daily Tio 18 once daily	64	63	NR	1.50	15.5
Donohue 2013	831	24	UMEC/VI 62.5/25 once daily UMEC 62.5 once daily	63	71	50	1.23	13.9
Donohue 2015a	706	12	UMEC/VI 62.5/25 once daily FP/SAL 250/50 twice daily	63	70	43	1.32	11.3
Donohue 2015b	697	12	UMEC/VI 62.5/25 once daily FP/SAL 250/50 twice daily	64	76	52	1.34	13.3
Donohue 2016a	590	56	ACL/FM 400/12 twice daily FM 12 twice daily	64	55	46	1.31	NR
Dransfield 2014	1858	12	FP/SAL 250/50 twice daily FF/VI 100/25 once daily	61	69	55	1.34	12
Feldman 2016	1017	12	UMEC 62.5 once daily Tio 18 once daily	64	72	51	1.36	12.1
Ferguson 2016	410	52	IND/Glyco 27.5/15.6 twice daily IND 75 once daily	63	68	51	1.25	22.4
GLOW4 2012	163	52	Glyco 50 once daily Tio 18 once daily	69	98	NR	NR	NR
Hanania 2003	355	24	FP/SAL 250/50 twice daily SAL 50 twice daily	64	60	47	1.21	20.7
Hoshino 2013	45	16	FP/SAL 250/50 twice daily Tio 18 once daily SAL 50 twice daily	71	87	NR	1.35	NR
Hoshino 2014	54	16	TIO 18 once daily + IND 150 once daily IND 150 once daily Tio 18 once daily	71	93	NR	1.53	NR
Hoshino 2015	43	16	TIO 18 once daily + IND 150 once daily FP/SAL 250/50 twice daily	71	84	NR	1.37	NR
Kalberg 2016	961	12	UMEC/VI 62.5/25 once daily Tio 18 once daily + IND 150 once daily	64	73	43	1.23	12.3
Kerwin 2012a	792	52	Glyco 50 once daily Tio 18 once daily	64	64	45	1.30	16.3
Kerwin 2017	494	12	UMEC/VI 62.5/25 once daily Tio 18 once daily	64	66	50	1.65	7.9
Koch 2014	919	48	Olo 5 once daily FM 12 twice daily	64	80	34	1.26	12.3
Kornmann 2011	667	26	IND 150 once daily SAL 50 twice daily	63	74	46	1.35	11.5
Koser 2010	247	12	FP/SAL 250/50 twice daily FP/SAL 230/42 twice daily	63	53	62	1.27	12.7
Mahler 2002	325	24	FP/SAL 500/50 twice daily SAL 50 twice daily	63	63	46	1.25	20.9
Mahler 2012a	1131	12	Tio 18 once daily + IND 150 once daily Tio 18 once daily	64	69	38	1.15	16.9
Mahler 2012b	1142	12	Tio 18 once daily + IND 150 once daily Tio 18 once daily	63	66	40	1.14	16.4
Mahler 2015a; Mahler 2015b	1530	12	IND/Glyco 27.5/15.6 twice daily Glyco 15.6 twice daily	64	64	52	1.27	22.8
Mahler 2016	507	52	IND 75 once daily Glyco 15.6 twice daily	63	57	55	1.25	21.2
Maleki‐Yazdi 2014	905	24	UMEC/VI 62.5/25 once daily Tio 18 once daily	62	68	57	1.26	13.4
Martinez 2017a	1880	24	Glyco/FM 18/9.6 twice daily Glyco 18 twice daily Tio 18 once daily FM 9.6 twice daily	63	56	54	1.25	19.8
Martinez 2017b	1387	24	Glyco/FM 18/9.6 twice daily Glyco 18 twice daily FM 9.6 twice daily	63	55	54	NR	19.2
NCT00876694 2011	186	52	IND 300 once daily SAL 50 twice daily	69	95	NR	NR	NR
NCT01536262 2014	82	52	Tio/Olo 5/5 once daily Olo 5 once daily	70	96	NR	NR	NR
Perng 2009	67	12	FP/SAL 500/50 twice daily Tio 18 once daily	73	94	61	1.21	NR
Hanania 2017	3267	52	Glyco/FM 18/9.6 twice daily Glyco 18 twice daily Tio 18 once daily FM 9.6 twice daily	63	56	54	NR	19.6
RADIATE 2016	812	52	IND/Glyco 110/50 once daily Tio 18 once daily	64	72	NR	NR	NR
Rheault 2016	1034	12	UMEC 62.5 once daily Glyco 50 once daily	64	69	48	1.34	13.2
Rossi 2014	581	26	FP/SAL 500/50 twice daily IND 150 once daily	66	69	36	1.54	9.7
SCO100470 2006	1050	24	FP/SAL 500/50 twice daily SAL 50 twice daily	64	78	43	1.67	NR
SCO40034 2005	125	12	FP/SAL 500/50 twice daily Tio 18 once daily	65	74	NR	1.37	NR
Singh 2014	1154	24	ACL/FM 400/12 twice daily ACL 400 twice daily FM 12 twice daily	63	67	47	1.41	NR
Singh 2015a	406	12	Tio/Olo 5/5 once daily Tio 5 once daily	65	59	52	1.31	14.5
Singh 2015b	405	12	Tio/Olo 5/5 once daily Tio 5 once daily	65	65	45	1.38	14.5
Singh 2015c	716	12	UMEC/VI 62.5/25 once daily FP/SAL 250/50 twice daily	62	72	59	1.44	10.8
Tashkin 2009	255	12	Tio 18 once daily + FM 12 twice daily Tio 18 once daily	64	66	47	NR	NR
Tashkin 2012a&b	1340	26‐52	MF/FM 400/10 twice daily MF/FM 200/10 twice daily FM 10 twice daily	60	75	49	1.21	8.9
To 2012	230	12	IND 150 once daily IND 300 once daily	67	97	34	1.24	15
Troosters 2016	152	12	Tio/Olo 5/5 once daily Tio 5 once daily	65	68	NR	NR	NR
Vincken 2014	447	12	IND/Glyco 110/50 once daily IND 150 once daily	64	81	42	1.46	19.5
Vogelmeier 2008	638	24	Tio 18 once daily + FM 10 twice daily Tio 18 once daily FM 10 twice daily	63	78	NR	1.50	10.8
Vogelmeier 2013a	522	26	IND/Glyco 110/50 once daily FP/SAL 500/50 twice daily	63	71	48	1.45	20.4
Vogelmeier 2016	933	24	ACL/FM 400/12 twice daily FP/SAL 500/50 twice daily	63	65	NR	1.38	11.8
Vogelmeier 2017	1080	12	IND/Glyco 110/50 once daily ICS/LABA free or fixed	65	64	49	NR	NR
Wise 2013	11392	120	Tio 5 once daily Tio 18 once daily	65	72	38	NR	NR
Yao 2014	375	26	IND 150 once daily IND 300 once daily	66	95	22	1.13	14.7
Zhong 2015	741	26	IND/Glyco 110/50 once daily FP/SAL 500/50 twice daily	65	91	26	1.08	24.1
ZuWallack 2014a&b	2267	12	Tio 18 once daily + Olo 5 once daily Tio 18 once daily	64	52	49	1.25	16

ACL: aclidinium; BDP: beclomethasone; BUD: budesonide; FEV1: forced expiratory volume in 1 second; FF: fluticasone furoate; FM: formoterol; Glyco: glycopyrrolate; FP: fluticasone propionate; IND: indacaterol; MF: mometasone furoate; NR: not reported; Olo: olodaterol; SAL: salmeterol; Tio: tiotropium; UMEC: umeclidinium; VI: vilanterol

We assessed if there was any difference in effect modifiers across the group pairwise comparisons especially when there was a discrepancy between the NMA and pairwise MA results and interpreted the results accordingly.

Assessment of heterogeneity and statistical consistency

We assessed heterogeneity by comparing the between‐study SD to the size of relative treatment effects, on the log‐scale for OR and HR. We assessed consistency by comparing the model fit and between‐study heterogeneity from the NMA models versus those from an unrelated mean‐effects (inconsistency) model (Dias 2013a; Dias 2013b). We used this to determine the presence and area of inconsistency. We also qualitatively compared the results from direct pairwise MA versus NMA estimates to check for broad agreement. If we identified substantial inconsistency, we explored factors, including participant and design characteristics that may have contributed to inconsistency (Table 9; Table 10; Table 11; Table 12; Table 13). For the pairwise MA, we tested heterogeneity among studies with I² statistics greater than 30%, indicating substantial heterogeneity (Higgins 2003). We used optimal information size calculations as an objective measure of imprecision for grading evidence, with an α of 0.05 and a β of 0.80 (Guyatt 2011a). We addressed heterogeneity in the pairwise MAs according to the GRADE criteria (Guyatt 2011b).

2. Study characteristics of treatment group pair‐wise comparisons and clinical homogeneity assessment in moderate to severe exacerbations in the high‐risk population.

Comparison	Comparisons	Number of participants	Mean age (years)	Male (%)	Current smoker (%)	Baseline FEV1 (L) prebronchodilator	Baseline FEV1 (L) postbronchodilator	Bronchial reversibility %
LABA/LAMA vs LABA/ICS	1	3372	65	76	40	NA	1.2	NA
LABA/LAMA vs LAMA	1	2206	63	75	38	0.9	1.04	18.3
LABA/LAMA vs LABA	0	0	NA	NA	NA	NA	NA	NA
LABA/ICS vs LAMA	2	1580	65	83	39	1.09	1.16	7
LABA/ICS vs LABA	10	9049	64	69	40	1.05	1.19	13.6
LAMA vs LABA	2	10,815	63	76	44	NA	1.32	NA

FEV1: forced expiratory volume in 1 second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; NA: not applicable

3. Study characteristics of treatment group pair‐wise comparisons and clinical homogeneity assessment in moderate to severe exacerbations in the low‐risk population.

Comparison	Comparisons	Number of participants	Mean age (years)	Male %	Current smoker %	Baseline FEV1 (L) prebronchodilator	Bronchial reversibility (%)
LABA/LAMA vs LABA/ICS	6	4315	63	74	45	1.33	14.9
LABA/LAMA vs LAMA	8	5192	63	71	47	1.32	14.7
LABA/LAMA vs LABA	5	2488	64	68	44	1.36	17.5
LABA/ICS vs LAMA	1	623	63	65	52	1.35	13
LABA/ICS vs LABA	6	6689	64	74	44	1.27	11.1
LAMA vs LABA	5	4567	64	71	39	1.3	17.1

FEV1: forced expiratory volume in 1 second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

4. Study characteristics of treatment group pair‐wise comparisons and clinical homogeneity assessment in severe exacerbations in the high‐risk population.

Comparison	Comparisons	Number of participants	Mean age (years)	Male (%)	Current smoker (%)	Baseline FEV1 (L) postbronchodilator	Bronchial reversibility (%)
LABA/LAMA vs LABA/ICS	1	3354	65	76	40	1	22.4
LABA/LAMA vs LAMA	1	304	68	56	26	1.01	NA
LABA/LAMA vs LABA	0	0	NA	NA	NA	NA	NA
LABA/ICS vs LAMA	2	1580	65	83	39	1.09	7
LABA/ICS vs LABA	5	4216	64	74	41	1.04	15.9
LAMA vs LABA	1	7376	63	76	48	NA	NA

FEV1: forced expiratory volume in 1 second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; NA: not applicable

5. Study characteristics of treatment group pair‐wise comparisons and clinical homogeneity assessment in severe exacerbations in the low‐risk population.

Comparison	Comparisons	Number of participants	Mean age (years)	Male (%)	Current smoker (%)	Baseline FEV1 (L) % prebronchodilator	Bronchial reversibility (%)	Baseline FEV1 (L) postbronchodilator
LABA/LAMA vs LABA/ICS	6	2860	63	74	45	1.33	14.9	1.5
LABA/LAMA vs LAMA	7	4973	63	72	41	1.33	15.1	1.49
LABA/LAMA vs LABA	6	2898	64	67	45	1.35	18.3	1.55
LABA/ICS vs LAMA	1	623	63	65	52	1.35	13	1.48
LABA/ICS vs LABA	6	6482	64	74	44	1.27	11.1	1.32
LAMA vs LABA	4	3320	64	74	39	1.23	18.2	1.54

FEV1: forced expiratory volume in 1 second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

6. Study characteristics of treatment group pair‐wise comparisons and clinical homogeneity assessment in pneumonia in the low‐risk population.

Comparison	Comparisons	Number of participants	Mean age (years)	Male (%)	Current smoker (%)	Baseline FEV1 (L) prebronchodilator	Bronchial reversibility %
LABA/LAMA vs LABA/ICS	7	5395	64	72	46	1.33	14.9
LABA/LAMA vs LAMA	21	19,043	64	68	47	1.27	16.7
LABA/LAMA vs LABA	11	8556	64	65	43	1.30	15.8
LABA/ICS vs LAMA	4	2465	65	80	43	1.16	8.7
LABA/ICS vs LABA	16	15,992	64	72	41	1.14	11
LAMA vs LABA	12	22,351	63	70	43	1.34	16.8

FEV1: forced expiratory volume in 1 second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

Assessment of reporting biases

We tried to minimise reporting biases from unpublished studies or selective outcome reporting by using a broad search strategy and by checking references of included studies and relevant systematic reviews. For each outcome, we reported the number of studies contributing data to the NMAs. For the pairwise MA, we assessed small study and publication bias through visual inspection of a funnel plot and performance of the Egger test (Egger 1997), if more than 10 studies were being pooled. We assumed the presence of small study bias when the number of participants was fewer than 50 per study, 1000 per pooled analysis or 100 per arm, when no more than 10 studies could be pooled (Dechartres 2013; Nüesch 2010). We assumed a selective reporting bias if a clinical study was not registered (Mathieu 2009).

Data synthesis

We based model comparison on deviance information criterion (DIC) (Spiegelhalter 2002). Differences of three points or more were considered meaningful. If models differed by less than three points, we selected the simplest model. We also calculated the posterior mean of the residual deviance to assess model fit. We considered this adequate when the posterior mean of the residual deviance approximated the number of unconstrained data points (Dias 2013c).

We chose a model and considered it as the primary analysis for NMAs using the following strategy:

1. Start with fixed‐class models (random‐ and fixed‐treatment‐effects). If both fit well, choose model with lowest DIC (if difference less than 3 choose fixed‐effect model) and stop.

2. If the fixed‐treatment‐effect, fixed‐class model does not fit well, try the fixed‐treatment‐effect, random‐class model – assess fit and choose the model with the lowest DIC.

3. If neither fixed‐ nor random‐treatment‐effect models with fixed‐class fit well, try also random‐treatment‐effects with random‐class.

4. Choose a final model based on DIC, but interpret with caution if model fit is poor.

We estimated the probability that each treatment group ranked at one of the four possible positions in the class model NMAs with rank 1 meaning that group is best for that outcome.

GRADE and 'Summary of findings' table

We used GRADE to assess the quality of evidence as it related to studies that contributed data to the pairwise MAs. We created a 'Summary of findings' table including the primary outcomes and pneumonia. We used the five GRADE considerations (study limitations, consistency of effect, imprecision, indirectness and publication bias), to assess the certainty of a body of evidence as it related to studies that contributed data to pairwise MAs for prespecified outcomes. We used methods and recommendations described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2017), and used GRADEpro GDT 2015 software. We justified all decisions to downgrade or upgrade the certainty of evidence by using footnotes, and we made comments to aid the reader's understanding of the review when necessary.

Subgroup analysis and investigation of heterogeneity

We combined the high‐ and low‐risk populations (presence or absence of a history of COPD exacerbation within the previous year), and performed subgroup analyses investigating if there was a substantial difference between them. We analysed studies of different duration separately (3, 6, and 12 months), for symptom and quality‐of‐life scores and change from baseline in FEV1 to minimise intransitivity because a previous study (Oba 2016a), suggested different durations could influence treatment effects on these outcomes. We used a formal test for subgroup interactions provided in Review Manager 2014.

Sensitivity analysis

We used a model not used in the primary analysis (fixed‐effect or random‐effects), as a sensitivity analysis for both NMAs and pairwise MAs.

Results

Description of studies

The study and patient characteristics including study duration, treatment arms, and baseline pulmonary function are presented in Table 8 and details of each study are shown in Characteristics of included studies.

Results of the search

We identified 870 plus 166 records (original and updated search respectively), from the Cochrane Airways Specialised Register (CAGR) of studies, and 28 references through other sources, such as manufactures' websites. We searched all records in the CAGR using the search strategy in Appendix 2 in March 2017 and again on 6 April 2018 for the updated search. We excluded 119 studies on abstract review. We reviewed the remaining 156 studies for further details and excluded an additional 57 studies for various reasons as shown in Figure 1.

Included studies

We included 26 studies with 32,265 participants in the high‐risk group (one or more exacerbations in the previous 12 months), and 73 studies with 69,046 participants in the low‐risk group, totaling 99 studies with a total of 101,311 randomised participants. The numbers of included studies varied with each outcome due to data availability and are summarised in Figure 1. Four in the low‐risk group (Hoshino 2013; Hoshino 2014; Hoshino 2015; Perng 2009), and one in the high‐risk group (Sarac 2016), were single‐centre studies and the rest were multicenter studies. They were all industry‐funded studies except for Aaron 2007, Cazzola 2007, Hoshino 2013, Hoshino 2014, Hoshino 2015, Perng 2009, and Sarac 2016.

1.

Study flow diagram
 AEs: adverse events; CAGR: Cochrane Airways Group Specialised Register; CFB: change from baseline; H: high−risk group; L: low−risk group; NA: not applicable; NMA: network meta−analysis; SAE: serious adverse event; SGRQ: St George's Respiratory Questionnaire; TDI: Transition Dyspnea Index

Table 9, Table 10, Table 11, Table 12, and Table 13 show comparisons of study characteristics among pairwise MAs in the relevant outcomes. The median duration of study was 52 (range 12 to 156) and 24 (range 12 to 156) weeks in the high‐ and low‐risk groups respectively.

Table 14; and Table 15 present the distribution of treatment arms across all 99 included studies, categorised by the four treatment groups. Vilanterol is available only as a component of combination inhalers for clinical use (i.e. it is not available as a single inhaler), therefore we did not include vilanterol as a node in the review. Indacaterol 27.5 μg and 600 μg twice daily, indacaterol/glycopyrronium 27.5 μg/25 μg twice daily, umeclidinium/vilanterol 125 μg/25 μg once daily, tiotropium/olodaterol 2.5 μg/5 μg once daily, and aclidinium/formoterol 400 μg/6 μg twice daily were also excluded from the analysis because they were not approved or available for clinical use at the time of data extraction. The network of treatments for each outcome is displayed in a corresponding figure. The treatments formed a closed network, which was amenable to a NMA except for SGRQ responders at 3 and 6 months, and TDI at 3, 6, and 12 months in the high‐risk population, and SGRQ responders at 12 months in the low‐risk population. When fixed‐ or random‐class models were considered, all networks were connected and could be analysed.

7. Distribution of studies by individual treatment node in the high‐risk population.

Class	Treatment node (drug, dose μg, dosing frequency)	Studies
LABA	Salmeterol 50 twice daily	Anzueto 2009; Calverley 2003 TRISTAN; Ferguson 2008; Kardos 2007; Ohar 2014; SCO40041 2008; Vogelmeier 2011
	Formoterol 9‐12 twice daily	Calverley 2003; Calverley 2010; Ferguson 2017; Fukuchi 2013; Rennard 2009; Sharafkhaneh 2012; Szafranski 2003; Tashkin 2008; Wedzicha 2014
	Indacaterol 150 once daily	Bateman 2013; Decramer 2013
LAMA	Tiotripium 18 once daily	Aaron 2007; Asai 2013; Covelli 2016; Decramer 2013; Pepin 2014; Sarac 2016; Vogelmeier 2011; Wedzicha 2008; Wedzicha 2013
	Glycopyrrolate 50 once daily	Bateman 2013; Wedzicha 2013
LABA/ICS	Salmetrol/fluticasone 50/250 twice daily	Anzueto 2009; Ferguson 2008; Ohar 2014; SCO40041 2008; Wedzicha 2008
	Salmetrol/fluticasone 50/500 twice daily	Agusti 2014; Calverley 2003; Hagedorn 2013; Kardos 2007; Sarac 2016; Wedzicha 2016
	Formoterol/budesonide 9/160 twice daily	Rennard 2009; Sharafkhaneh 2012; Tashkin 2008
	Formoterol/budesonide 9/320 twice daily	Calverley 2003; Ferguson 2017; Fukuchi 2013; Rennard 2009; Sharafkhaneh 2012; Szafranski 2003; Tashkin 2008
	Formoterol/budesonide 12/400 twice daily DPI	Calverley 2010
	Formoterol/beclomethasone 12/200 twice daily	Calverley 2010; Wedzicha 2014
	Salmeterol 50 twice daily + fluticasone 250 twice dailya	COMBINE 2017
	Salmeterol 50 twice daily + fluticasone 500 twice dailya	Hagedorn 2013
	Vilanterol/fluticasone 25/100 once daily	Agusti 2014; Covelli 2016; Pepin 2014;
	Indacaterol 150 once daily + budesonide 400 twice dailya	COMBINE 2017
LABA/LAMA	Indacaterol/glycopyrrolate 27.5/15.6 twice daily	Ferguson 2016
	Indacaterol/glycopyrrolate 110/50 once daily	Asai 2013; Bateman 2013; Wedzicha 2013; Wedzicha 2016
	Salmeterol 50 twice daily + tiotropium 18 once dailya	Aaron 2007

^aFree combination

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

8. Distribution of studies by individual treatment node in the low‐risk population.

Class	Treatment node (drug, dose μg, dosing frequency)	Studies
LABA	Salmeterol 50 twice daily	BI 205.137 2001; Briggs 2005; Brusasco 2003; Calverley 2007; Hanania 2003; Hoshino 2013; Jones 2011; Kornmann 2011; Mahler 2002; NCT00876694 2011; SCO100470 2006
	Formoterol 4.5 twice daily	Bogdan 2011
	Formoterol 9‐12 twice daily	Bogdan 2011; Calverley 2010; Dahl 2010; Donohue 2016a; D'Urzo 2014; D'Urzo 2017; Hanania 2017; Jones 2011; Koch 2014; Martinez 2017a; Martinez 2017b; Singh 2014; Tashkin 2012a&b; Vogelmeier 2008
	Indacaterol 75 once daily	Ferguson 2016; Mahler 2016
	Indacaterol 150 once daily	Buhl 2011; Donohue 2010; Hoshino 2014; Jones 2011; Kornmann 2011; Rossi 2014; To 2012; Yao 2014; Vincken 2014
	Indacaterol 300 once daily	Dahl 2010; Donohue 2010; Jones 2011; NCT00876694 2011; To 2012; Yao 2014
	Olodaterol 5 once daily	Buhl 2015a&b; NCT01536262 2014; Koch 2014
LAMA	Tiotripium 18 once daily	BI 205.137 2001; Briggs 2005; Brusasco 2003; Buhl 2011; Cazzola 2007; Chapman 2014; COSMOS‐J 2016; Covelli 2016; Decramer 2014a; Decramer 2014b; Donohue 2010; Fang 2008; Feldman 2016; GLOW4 2012; Hanania 2017; Hoshino 2013; Hoshino 2014; Kerwin 2012a; Kerwin 2017; Mahler 2012a; Mahler 2012b; Maleki‐Yazdi 2014; Martinez 2017a; Perng 2009; RADIATE 2016; SCO40034 2005; Tashkin 2009; Vogelmeier 2008; Wise 2013; ZuWallack 2014a&b
	Tiotripium 5 once daily	Buhl 2015a; Buhl 2015b; Singh 2015a&b; Troosters 2016; Wise 2013
	Aclidinium 400 twice daily	D'Urzo 2014; D'Urzo 2017; Singh 2014
	Umeclidinium 62.5 once daily	Donohue 2013; Feldman 2016; Rheault 2016
	Glycopyrrolate 15.6 twice daily	Hanania 2017; Mahler 2015a; Mahler 2015b; Mahler 2016; Martinez 2017a; Martinez 2017b
	Glycopyrrolate 50 once daily	Chapman 2014; GLOW4 2012; Kerwin 2012a; Rheault 2016
LABA/ICS	Salmetrol/fluticasone 50/250 twice daily	COSMOS‐J 2016; Donohue 2015a; Donohue 2015b; Dransfield 2014; Fang 2008; Hanania 2003; Hoshino 2013 ; Hoshino 2015; Koser 2010; Singh 2015d
	Salmetrol/fluticasone 50/500 twice daily	Calverley 2007; Cazzola 2007; Mahler 2002; Perng 2009; Rossi 2014; SCO100470 2006; SCO40034 2005; Vogelmeier 2013a; Vogelmeier 2016; Zhong 2015
	Salmetrol/fluticasone 42/230 (HFA) twice daily	Koser 2010
	Formoterol/budesonide 9/320 twice daily	Calverley 2010
	Formoterol/mometasone 200/10 twice daily	Tashkin 2012a&b
	Formoterol/mometasone 400/10 twice daily	Tashkin 2012a&b
	Vilanterol/fluticasone 25/100 once daily	Covelli 2016; Dransfield 2014
LABA/LAMA	Vilaterol/umeclidinium 25/62.5 once daily	Decramer 2014a; Decramer 2014b; Donohue 2013; Donohue 2015a; Donohue 2015b; Kalberg 2016; Kerwin 2017; Maleki‐Yazdi 2014; Singh 2015d
	Formoterol/glycopyrrolate 9.6/18 twice daily	Hanania 2017; Martinez 2017a; Martinez 2017b
	Indacaterol/glycopyrrolate 27.5/15.6 twice daily	Ferguson 2016; Mahler 2015a; Mahler 2015b
	Indacaterol/glycopyrrolate 110/50 once daily	Buhl 2015c; RADIATE 2016; Vogelmeier 2013a; Vogelmeier 2017; Zhong 2015
	Olodaterol/tiotropium 5/5 once daily	Buhl 2015a&b; NCT01536262 2014; Singh 2015a&b; Troosters 2016
	Formterol/aclidinium 12/400 twice daily	Donohue 2016a; D'Urzo 2014; D'Urzo 2017; Singh 2014; Vogelmeier 2016
	Indacaterol 150 once daily + tiotropium 18 once dailya	Hoshino 2014; Hoshino 2015; Kalberg 2016; Mahler 2012a; Mahler 2012b
	Formoterol 10‐12 twice daily + tiotropium 18 once dailya	Buhl 2015c; Tashkin 2009; Vogelmeier 2008
	Olodaterol 5 once daily + tiotropium 18 once dailya	ZuWallack 2014a&b
	Indacaterol 110 once daily + glycopyrrolate 50 once dailya	Vincken 2014

^aFree combination

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

Participants

The mean age, proportion of male participants and current smokers, and pre‐bronchodilator baseline FEV1, were 64.5 years (SD 1.5), 72.5% (SD 11.7), 39.0% (SD 6.0), and 1.06 L (SD 0.11), in the high‐risk group and 64.6 years (SD 2.4), 72.5% (SD 12.3), 46.0% (SD 8.1), and 1.31 L (SD 0.13), in the low‐risk group. The median bronchial reversibility at the baseline was 13.6% (range 7.0 to 22.4), and 14.2% (range 7.9 to 24.1), in the high‐ and low‐risk groups respectively.

Excluded studies

We excluded 57 studies after full‐text review and we recorded them in Characteristics of excluded studies, with reasons for exclusion. We excluded 27 studies because, after we had excluded an unapproved or unavailable dosage, there were no valid comparisons. Two studies became available after data extraction (Calverley 2018; Papi 2017), and we did not included them in the analysis. We would have excluded Calverley 2018 anyway because they included participants with coexisting reactive airway disease.

Risk of bias in included studies

We have presented 'Risk of bias' judgements for individual studies in the Characteristics of included studies and a summary overview of the findings in Figure 2. Generally, we deemed the risk of bias in the included studies to be moderate to low. There were no studies that we should clearly have excluded from the analysis because of differences in baseline characteristics or poor quality.

2.

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

Allocation

All studies were randomised trials and most of them were industry funded. We confirmed a random allocation sequence using a validated computerised system in 60 out of 92 industry‐funded studies, and assumed an industry‐standard method for the rest and considered them to be at low risk for random sequence generation and allocation concealment (concealment assumed by automatisation). We could not confirm a random allocation sequence in four out of seven non‐industry studies (Hoshino 2013; Hoshino 2014: Hoshino 2015: Sarac 2016), and we considered them to be at unclear risk.

Blinding

The following studies were open‐label or partially blinded, with tiotropium being administered open‐label, and considered to be at a high risk of bias: Asai 2013, Bateman 2013, COMBINE 2017, Donohue 2010, Hagedorn 2013, Hanania 2017, Hoshino 2013, Hoshino 2014, Hoshino 2015, Kerwin 2012a, Martinez 2017a, NCT00876694 2011, Perng 2009, Sarac 2016, Vogelmeier 2008,Vogelmeier 2017, Wedzicha 2013. They consisted of 15.4% and 17.8% of studies in the high‐ and low‐risk populations. The rest of the studies were double‐blinded (82.8%), and rated as having low risk of bias (blinding of participants, personnel and outcome assessors).

Incomplete outcome data

We rated 18 studies (18.1%), at high risk due to high attrition or unbalanced dropouts. We gave an unclear rating to four studies (4.0%), because of high but balanced attrition (Calverley 2003 TRISTAN), imbalanced but relatively low attrition (Ferguson 2017; Hanania 2017), and a small sample size with unknown attrition (Sarac 2016). We tested whether the above studies compromised the validity of the results by excluding them one by one or all together in each outcome. The results are described in 'Summary of findings' tables in the selected outcomes.

Selective reporting

We were able to locate a study protocol, and most studies reported confirmed expected outcomes in publications. We could not locate a preregistered protocol for five studies (Briggs 2005; Cazzola 2007: Hoshino 2013: Perng 2009: Sarac 2016), and rated them as unclear risk of bias. Two studies reported outcomes of interest but in an insufficient form to be incorporated into a meta‐analysis and we rated them as having high risk of bias (Hoshino 2015; Vogelmeier 2008).

Other potential sources of bias

The vast majority of the included studies were designed, sponsored and conducted by pharmaceutical companies. Industry sponsorship bias cannot be excluded.

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. LABA/LAMA compared to LABA/ICS for chronic obstructive pulmonary disease.

LABA/LAMA compared to LABA/ICS for chronic obstructive pulmonary disease
Patient or population: chronic obstructive pulmonary disease with predicted FEV1 of less than 80% Setting: outpatient Intervention: LABA/LAMA Comparison: LABA/ICS
Outcomes	Anticipated absolute effects* (95% CI)		Relative effect (95% CI)	Number of participants (studies)	Certainty of the evidence (GRADE)
	Risk with LABA/ICS	Risk difference with LABA/LAMA
Moderate to severe exacerbations: high‐risk population	443 per 1000	34 fewer per 1000 (66 fewer to 0 fewer)	OR 0.87 (0.76 to 1.00)	3372 (1 RCT)	⊕⊕⊕⊝ Moderate1,2
Moderate to severe exacerbations: low‐risk population	89 per 1000	11 fewer per 1000 (29 fewer to 11 more)	OR 0.86 (0.65 to 1.14)	4315 (6 RCTs)	⊕⊕⊕⊝ Moderate1,3
Severe exacerbations: high‐risk population	172 per 1000	17 fewer per 1000 (39 fewer to 8 more)	OR 0.88 (0.74 to 1.06)	3354 (1 RCT)	⊕⊕⊕⊝ Moderate1,3
Severe exacerbations: low‐risk population	17 per 1000	6 fewer per 1000 (12 fewer to 10 more)	OR 0.66 (0.27 to 1.63)	2860 (4 RCTs)	⊕⊕⊕⊝ Moderate1,3
Pneumonia: high‐risk population	32 per 1000	12 fewer per 1000 (19 fewer to 1 fewer)	OR 0.62 (0.40 to 0.96)	3358 (1 RCT)	⊕⊕⊕⊝ Moderate1
Pneumonia: low‐risk population	8 per 1000	4 fewer per 1000 (6 fewer to 0 fewer)	OR 0.43 (0.19 to 0.97)	5395 (7 RCTs)	⊕⊕⊕⊝ Moderate1
*The risk in the intervention group (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; FEV1: forced expiratory volume‐one second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; OR: odds ratio; RCT: randomised controlled trial
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.

¹Optimal information size was not met.
 ²95% CI contains the line of no difference.
 ³ We could not exclude the possibility of a clinically important difference due to a wide 95% CI.

Summary of findings 2. LABA/LAMA compared to LAMA for chronic obstructive pulmonary disease.

LABA/LAMA compared to LAMA for chronic obstructive pulmonary disease
Patient or population: chronic obstructive pulmonary disease with predicted FEV1 of less than 80% Setting: outpatient Intervention: LABA/LAMA Comparison: LAMA
Outcomes	Anticipated absolute effects* (95% CI)		Relative effect (95% CI)	Number of participants (studies)	Certainty of the evidence (GRADE)
	Risk with LAMA	Risk difference with LABA/LAMA
Moderate to severe exacerbations: high‐risk population	561 per 1000	14 more per 1000 (29 fewer to 58 more)	OR 1.06 (0.89 to 1.27)	2206 (1 RCT)	⊕⊕⊕⊝ Moderate1,2,3
Moderate to severe exacerbations: low‐risk population	108 per 1000	7 fewer per 1000 (34 fewer to 28 more)	OR 0.93 (0.66 to 1.30)	5192 (8 RCTs)	⊕⊕⊝⊝ Low2,3,4,5
Severe exacerbations: high‐risk population	397 per 1000	72 fewer per 1000 (169 fewer to 36 more)	OR 0.73 (0.45 to 1.16)	304 (1 RCT)	⊕⊕⊕⊝ Moderate2,3
Severe exacerbations: low‐risk population	17 per 1000	0 fewer per 1000 (7 fewer to 12 more)	OR 0.99 (0.57 to 1.72)	4937 (7 RCTs)	⊕⊕⊕⊝ Moderate2,3,4
Pneumonia: high‐risk population	30 per 1000	1 fewer per 1000 (12 fewer to 17 more)	OR 0.98 (0.59 to 1.61)	2510 (2 RCTs)	⊕⊕⊕⊝ Moderate2,3,4
Pneumonia: low‐risk population	6 per 1000	1 more per 1000 (1 fewer to 4 more)	OR 1.23 (0.84 to 1.81)	18,538 (22 RCTs)	⊕⊕⊕⊝ Moderate3,4,6
*The risk in the intervention group (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; FEV1: forced expiratory volume‐one second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; OR: odds ratio; RCT: randomised controlled trial
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.

¹Results were unchanged when open tiotropium arm was excluded.
 ²Optimal information size was not met.
 ³ We could not exclude the possibility of a clinically important difference due to a wide 95% CI.
 ⁴Results were unchanged when studies with open tiotropium arm were excluded one by one or all together.
 ⁵Moderate heterogeneity (I² = 30% to 60%).
 ⁶Results were unchanged when studies with uneven and/or high dropouts were excluded one by one or all together.

Summary of findings 3. LABA/LAMA compared to LABA for chronic obstructive pulmonary disease.

LABA/LAMA compared to LABA for chronic obstructive pulmonary disease
Patient or population: chronic obstructive pulmonary disease with predicted FEV1 of less than 80% Setting: outpatient Intervention: LABA/LAMA Comparison: LABA
Outcomes	Anticipated absolute effects* (95% CI)		Relative effect (95% CI)	Number of participants (studies)	Certainty of the evidence (GRADE)
	Risk with LABA	Risk difference with LABA/LAMA
Moderate to severe exacerbations: high‐risk population	‐	‐	‐	0 (0 RCTs)	‐
Moderate to severe exacerbations: low‐risk population	166 per 1000	33 fewer per 1000 (56 fewer to 4 fewer)	OR 0.77 (0.62 to 0.97)	2488 (5 RCTs)	⊕⊕⊕⊝ Moderate1
Severe exacerbations: high‐risk population	‐	‐	‐	0 (0 RCTs)	‐
Severe exacerbations: low‐risk population	59 per 1000	12 fewer per 1000 (25 fewer to 7 more)	OR 0.78 (0.55 to 1.12)	2898 (6 RCTs)	⊕⊕⊕⊝ Moderate1,2
Pneumonia: high‐risk population	‐	‐	‐	0 (0 RCTs)	‐
Pneumonia: low‐risk population	7 per 1000	4 more per 1000 (0 fewer to 10 more)	OR 1.54 (0.95 to 2.49)	8252 (10 RCTs)	⊕⊕⊕⊝ Moderate2
*The risk in the intervention group (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; FEV1: forced expiratory volume‐one second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; OR: odds ratio; RCT: randomised controlled trial
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.

¹Optimal information size was not met.
 ²A clinically important difference cannot be excluded due to a wide 95% CI.

Summary of findings 4. LABA/ICS compared to LAMA for chronic obstructive pulmonary disease.

LABA/ICS compared to LAMA for chronic obstructive pulmonary disease (COPD)
Patient or population: chronic obstructive pulmonary disease with predicted FEV1 of less than 80% Setting: outpatient Intervention: LABA/ICS Comparison: LAMA
Outcomes	Anticipated absolute effects* (95% CI)		Relative effect (95% CI)	Number of participants (studies)	Certainty of the evidence (GRADE)
	Risk with LAMA	Risk difference with LABA/ICS
Moderate to severe exacerbations: high‐risk population	504 per 1000	28 more per 1000 (26 fewer to 81 more)	OR 1.12 (0.90 to 1.39)	1580 (2 RCTs)	⊕⊕⊕⊝ Moderate1,2
Moderate to severe exacerbations: low‐risk population	35 per 1000	13 fewer per 1000 (26 fewer to 22 more)	OR 0.63 (0.24 to 1.66)	623 (1 RCT)	⊕⊕⊝⊝ Low1,3
Severe exacerbations: high‐risk population	112 per 1000	27 more per 1000 (5 fewer to 67 more)	OR 1.28 (0.95 to 1.73)	1580 (2 RCTs)	⊕⊕⊕⊝ Moderate1,2
Severe exacerbations: low‐risk population	3 per 1000	6 more per 1000 (2 fewer to 83 more)	OR 3.05 (0.32 to 29.47)	623 (1 RCT)	⊕⊕⊝⊝ Low1,2
Pneumonia: high‐risk population	28 per 1000	21 more per 1000 (2 more to 52 more)	OR 1.80 (1.06 to 3.06)	1580 (2 RCTs)	⊕⊕⊕⊝ Moderate1
Pneumonia: low‐risk population	0 per 1000	0 fewer per 1000 (0 fewer to 0 fewer)	OR 5.82 (0.70 to 48.80)	885 (2 RCTs)	⊕⊕⊝⊝ Low1,2,3
*The risk in the intervention group (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; FEV1: forced expiratory volume‐one second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; OR: odds ratio; RCT: randomised controlled trial
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.

¹Optimal information size was not met.
 ² We could not exclude the possibility of a clinically important difference due to a wide 95% CI.
 ³Significant small study effects are possible due to small sample sizes in the included studies.

Summary of findings 5. LABA/ICS compared to LABA for chronic obstructive pulmonary disease.

LABA/ICS compared to LABA for chronic obstructive pulmonary disease (COPD): a network meta‐analysis
Patient or population: chronic obstructive pulmonary disease with predicted FEV1 of less than 80% Setting: outpatient Intervention: LABA/ICS Comparison: LABA
Outcomes	Anticipated absolute effects* (95% CI)		Relative effect (95% CI)	Number of participants (studies)	Certainty of the evidence (GRADE)
	Risk with LABA	Risk difference with LABA/ICS
Moderate to severe exacerbations: high‐risk population	430 per 1000	51 fewer per 1000 (69 fewer to 28 fewer)	OR 0.81 (0.75 to 0.89)	9041 (10 RCTs)	⊕⊕⊕⊕ High1
Moderate to severe exacerbations: low‐risk population	454 per 1000	46 fewer per 1000 (86 fewer to 5 fewer)	OR 0.83 (0.70 to 0.98)	6689 (6 RCTs)	⊕⊕⊕⊝ Moderate2
Severe exacerbations: high‐risk population	94 per 1000	8 fewer per 1000 (23 fewer to 11 more)	OR 0.91 (0.74 to 1.13)	4216 (5 RCTs)	⊕⊕⊕⊝ Moderate1,3,4
Severe exacerbations: low‐risk population	130 per 1000	7 more per 1000 (11 fewer to 26 more)	OR 1.06 (0.90 to 1.24)	6482 (6 RCTs)	⊕⊕⊕⊕ High
Pneumonia: high‐risk population	14 per 1000	6 more per 1000 (0 fewer to 15 more)	OR 1.46 (1.03 to 2.08)	12586 (14 RCTs)	⊕⊕⊕⊝ Moderate5
Pneumonia: low‐risk population	29 per 1000	18 more per 1000 (7 more to 31 more)	OR 1.64 (1.25 to 2.14)	6705 (6 RCTs)	⊕⊕⊕⊕ High
*The risk in the intervention group (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; FEV1: forced expiratory volume‐one second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; OR: odds ratio; RCT: randomised controlled trial
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.

¹Results were unchanged when we excluded studies with uneven dropouts, one by one or all together.
 ²Moderate heterogeneity (I² = 30% to 60%).
 ³Optimal information size not met.
 ⁴ We could not exclude the possibility of a clinically important difference due to a wide 95% CI.
 ⁵Several studies had a high dropout rate and 95% CI crossed/uncrossed the line of no difference when we excluded a study with a high dropout rate.

Summary of findings 6. LAMA compared to LABA for chronic obstructive pulmonary disease.

LAMA compared to LABA for chronic obstructive pulmonary disease
Patient or population: chronic obstructive pulmonary disease with predicted FEV1 of less than 80% Setting: outpatient Intervention: LAMA Comparison: LABA
Outcomes	Anticipated absolute effects* (95% CI)		Relative effect (95% CI)	Number of participants (studies)	Certainty of the evidence (GRADE)
	Risk with LABA	Risk difference with LAMA
Moderate to severe exacerbations: high‐risk population	385 per 1000	40 fewer per 1000 (63 fewer to 20 fewer)	OR 0.84 (0.76 to 0.92)	7376 (1 RCT)	⊕⊕⊕⊕ High
Moderate to severe exacerbations: low‐risk population	198 per 1000	13 fewer per 1000 (35 fewer to 11 more)	OR 0.92 (0.79 to 1.07)	4567 (5 RCTs)	⊕⊕⊕⊝ Moderate1,2
Severe exacerbations: high‐risk population	151 per 1000	16 fewer per 1000 (29 fewer to 1 more)	OR 0.88 (0.78 to 1.01)	7376 (1 RCT)	⊕⊕⊕⊝ Moderate2
Severe exacerbations: low‐risk population	30 per 1000	10 fewer per 1000 (19 fewer to 4 more)	OR 0.64 (0.36 to 1.13)	3320 (4 RCTs)	⊕⊕⊝⊝ Low2,3,4
Pneumonia: high‐risk population	17 per 1000	3 fewer per 1000 (7 fewer to 2 more)	OR 0.83 (0.61 to 1.13)	10,815 (2 RCTs)	⊕⊕⊕⊝ Moderate4
Pneumonia: low‐risk population	7 per 1000	0 fewer per 1000 (3 fewer to 5 more)	OR 1.01 (0.61 to 1.69)	11,338 (10 RCTs)	⊕⊕⊕⊝ Moderate4
*The risk in the intervention group (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; FEV1: forced expiratory volume‐one second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; OR: odds ratio; RCT: randomised controlled trial
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.

¹Results were unchanged when we excluded studies with open‐label tiotropium arm, one by one or all together.
 ²Optimal information size was not met.
 ³95% CI no longer contained the line of no difference when we excluded a study with open‐label tiotropium arm.
 ⁴A clinically important difference cannot be excluded due to a wide 95% CI.

Summary of findings 7. Summary of findings for network meta‐analyses.

Patient or population: chronic obstructive pulmonary disease with predicted FEV1 of less than 80%. Settings: outpatient
Outcomes	Anticipated absolute effects* (95% CrI)		Relative effect (95% CrI)	No of participants (studies)
	Risk with LABA	Risk difference with LABA/LAMA
Moderate to severe exacerbations: high‐risk population	427 per 1000	106 fewer per 1000 (139 fewer to 68 fewer)	HR 0.70 (0.61 to 0.80)	11,113 (21 RCTs)
Moderate to severe exacerbations: low‐risk population	250 per 1000	52 fewer per 1000 (76 fewer to 25 more)	HR 0.78 (0.67 to 0.90)	14,450 (28 RCTs)
Severe exacerbations: high‐risk population	142 per 1000	48 fewer per 1000 (66 fewer to 26 fewer)	HR 0.64 (0.51 to 0.81)	9,045 (13 RCTs)
Severe exacerbations: low‐risk population	92 per 1000	24 fewer per 1000 (44 fewer to 2 more)	HR 0.72 (0.48 to 1.02)	11,127 (31 RCTs)
	Risk with LABA	Risk difference with LABA/ICS	Relative effect (95% CrI)	No of participants (studies)
Moderate to severe exacerbations: high‐risk population	427 per 1000	66 fewer per 1000 (87 fewer to 46 fewer)	HR 0.80 (0.75 to 0.86)	18,561 (21 RCTs)
Moderate to severe exacerbations: low‐risk population	250 per 1000	24 fewer per 1000 (37 fewer to 10 fewer)	HR 0.89 (0.84 to 0.96)	16,437 (28 RCTs)
Severe exacerbations: high‐risk population	142 per 1000	23 fewer per 1000 (39 fewer to 4 fewer)	HR 0.83 (0.71 to 0.97)	12,447 (13 RCTs)
Severe exacerbations: low‐risk population	92 per 1000	2 more per 1000 (10 fewer to 15 more)	HR 1.01 (0.72 to 1.28)	12,265 (31 RCTs)
	Risk with LABA	Risk difference with LAMA	Relative effect (95% CrI)	No of participants (studies)
Moderate to severe exacerbations: high‐risk population	427 per 1000	69 fewer per 1000 (99 fewer to 40 fewer)	HR 0.80 (0.71 to 0.88)	16,655 (21 RCTs)
Moderate to severe exacerbations: low‐risk population	250 per 1000	27 fewer per 1000 (48 fewer to 5 fewer)	HR 0.87 (0.78 to 0.97)	14,209 (28 RCTs)
Severe exacerbations: high‐risk population	142 per 1000	37 fewer per 1000 (49 fewer to 24 fewer)	HR 0.72 (0.63 to 0.82)	15,205 (13 RCTs)
Severe exacerbations: low‐risk population	92 per 1000	15 fewer per 1000 (29 fewer to 2 more)	HR HR 0.80 (0.56 to 1.05)	22,819 (31 RCTs)
	Risk with LABA/ICS	Risk difference with LABA/LAMA	Relative effect (95% CrI)	No of participants (studies)
Pneumonia: high‐risk population	24 per 1000	10 fewer per 1000 (14 fewer to 4 fewer)	OR 1.69 (1.2 to 2.44)	13,546 (24 RCTs)
Pneumonia: low‐risk population	24 per 1000	8 fewer per 1000 (13 fewer to 0 fewer)	OR 1.64 (0.99 to 2.94)	27,043 (61 RCTs)
	Risk with LABA/ICS	Risk difference with LAMA	Relative effect (95% CrI)	No of participants (studies)
Pneumonia: high‐risk population	24 per 1000	10 fewer per 1000 (14 fewer to 6 fewer)	OR 1.78 (1.33 to 2.39)	18,844 (24 RCTs)
Pneumonia: low‐risk population	24 per 1000	11 fewer per 1000 (16 fewer to 4 fewer)	OR 2.02 (1.16 to 3.72)	39,236 (31 RCTs)
	Risk with LABA/ICS	Risk difference with LABA	Relative effect (95% CrI)	No of participants (studies)
Pneumonia: high‐risk population	24 per 1000	8 fewer per 1000 (11 fewer to 3 fewer)	OR 1.50 (1.17 to 1.92)	21,404 (24 RCTs)
Pneumonia: low‐risk population	24 per 1000	11 fewer per 1000 (14 fewer to 7 fewer)	OR 1.93 (1.29 to 3.22)	20,158 (61 RCTs)
*The risk in the intervention group (and its 95% credible interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CrI). CrI: credible interval; FEV1: forced expiratory volume‐one second; HR: hazard ratio; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; OR: odds ratio; RCT: randomised controlled trial

1. Results: high‐risk population

1.1 Outcome: exacerbations

1.1.1 Outcome: moderate to severe exacerbations

We included 21 studies of 14 interventions and four treatment groups for this outcome (Appendix 3; Figure 3).

3.

Moderate to severe exacerbations in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

bid: twice daily; BDP: beclomethasone; BUD: budesonide; FF: fluticasone furoate; FM: formoterol; FP: fluticasone propionate; Glyco: glycopyrronium; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; qd: once daily; SAL: salmeterol; Tio: tiotropium; VI: vilanterol

1.1.1.1 Model selection and inconsistency checking

We chose a random‐treatment‐effects model with fixed‐class effects, assuming consistency (Appendix 4).

1.1.1.2 NMA results

The NMA included a total of 25,771 participants (LABA: 10,279, LAMA: 6376, LABA/ICS: 8282, LABA/LAMA: 834). The median duration of follow‐up was 52 weeks (range 12 to 156 weeks). Figure 3 and Table 16 show the HR for moderate to severe exacerbations for each group compared to every other. The NMA suggested that LABA/LAMA combination was the highest ranked treatment group to reduce moderate to severe exacerbations (95% CrI 1st to 2nd), followed by LAMA (95% CrI 2nd to 3rd), (Appendix 5; Table 17). HRs against LABA/ICS, LAMA, and LABA were 0.86 (95% CrI 0.76 to 0.99), 0.87 (95% CrI 0.78 to 0.99) and 0.70 (95% CrI 0.61 to 0.80), respectively (Appendix 6). LABA is the worst ranked treatment group for this outcome (95% CrI 4th to 4th), and all groups of interventions decrease the rate of moderate to severe exacerbations compared to LABA. HRs for other treatment groups versus LABA were 0.70 (95% CrI 0.61 to 0.80), 0.80 (95% CrI 0.75 to 0.86) and 0.80 (95% CrI 0.71 to 0.88) for LABA/LAMA, LABA/ICS, and LAMA respectively (Appendix 6; Table 7).

9. Relative effects: moderate to severe exacerbations in the high‐risk population.

Treatment comparison	Hazard ratios: random‐effects
	Median	95% CrI
LABA/LAMA v LABA/ICS	0.86	0.76 to 0.99
LABA/LAMA v LAMA	0.87	0.78 to 0.99
LABA/LAMA v LABA	0.70	0.61 to 0.80
LABA/ICS v LAMA	1.01	0.91 to 1.13
LABA/ICS v LABA	0.80	0.75 to 0.86
LAMA v LABA	0.80	0.71 to 0.88

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

10. Mean and median ranks: moderate to severe exacerbations in the high‐risk population.

Treatment group	Rank (from random‐effects model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 2
LAMA	2.4	2	2 to 3
LABA/ICS	2.6	3	2 to 3
LABA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.1.1.3 Clinical homogeneity assessment

Table 9 shows the clinical homogeneity assessment (or transitivity), across the available comparisons. Bronchial reversibility ranged from 7.0% to 18.3%. The mean bronchial reversibility for LABA/ICS versus LAMA comparison was 7%, which could have underestimated the effects of LABA/ICS. The NMA results should be interpreted with caution because of the difference in bronchial reversibility across the pairwise comparisons.

1.1.1.4 Pairwise meta‐analyses

There was no direct comparison for LABA/LAMA versus LABA. The results from pairwise MAs were consistent with the NMAs except for LABA/LAMA versus LABA/ICS or LAMA, in which the 95% CI contained the line of no difference (OR 0.87, 95% CI 0.76 to 1.00, and OR 1.06, 95% CI 0.89 to 1.27), unlike the NMAs (HR 0.86, 95% CrI 0.76 to 0.99, and HR 0.87, 95% CrI 0.78 to 0.99; Appendix 6). The certainty of evidence was moderate for LABA/LAMA versus LABA/ICS or LAMA due to a suboptimal sample size, which could explain the discrepancy between the NMAs and pairwise MAs. Otherwise, it was moderate for LABA/ICS versus LAMA and high for LABA/ICS versus LABA and LAMA versus LABA (see 'Summary of findings' tables). There was no difference between random and fixed analyses.

1.1.2 Outcomes: severe exacerbations

We included 13 studies of nine interventions and four treatment groups for this outcome (Appendix 3; Figure 4 a and b).

4.

Severe exacerbations in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.1.2.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐effects model for comparison (Appendix 4).

1.1.2.2 NMA results

This NMA included a total of 21,733 participants (LABA: 7482, LAMA: 7723, LABA/ICS: 4965, LABA/LAMA: 1563). The median duration of follow‐up was 52 weeks (range 12 to 104 weeks). Figure 4 and Table 18 show the HR for severe exacerbations for each treatment group compared to every other. The NMA suggested that LABA/LAMA combination was the highest ranked treatment group to reduce severe exacerbations (95% CrI 1st to 2nd), followed by LAMA (95% CrI 1st to 3rd; Appendix 5; Table 19). HRs against LABA/ICS, LAMA, and LABA were 0.78 (95% CrI 0.64 to 0.93), 0.89 (95% CrI 0.71 to 1.11), and 0.64 (95% CrI 0.51to 0.81), respectively. Results using the fixed‐ or random‐treatment‐effects assumption are very similar. There is evidence that all treatment groups decrease the rate of severe exacerbations compared to LABA (HRs against LABA: 0.64 (95% CrI 0.51 to 0.81), 0.83 (95% CrI 0.71 to 0.97), and 0.72 (95% CrI 0.63 to 0.82), for LABA/LAMA, LABA/ICS and LAMA respectively), and that LABA/LAMA decreases the rate of severe exacerbations compared to LABA/ICS (HR 0.78, 95% CrI 0.64 to 0.93; Appendix 6; Table 7).

11. Relative effects: severe exacerbations in the high‐risk population.

Treatment comparison	Hazard ratios: fixed‐effect		Hazard ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	0.78	0.64 to 0.93	0.78	0.62 to 0.98
LABA/LAMA v LAMA	0.89	0.71 to 1.11	0.91	0.73 to 1.13
LABA/LAMA v LABA	0.64	0.51 to 0.81	0.65	0.50 to 0.84
LABA/ICS v LAMA	1.15	0.97 to 1.36	1.16	0.94 to 1.41
LABA/ICS v LABA	0.83	0.71 to 0.97	0.83	0.69 to 1.00
LAMA v LABA	0.72	0.63 to 0.82	0.72	0.60 to 0.86

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

12. Mean and median ranks: severe exacerbations in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.2	1	1 to 2
LAMA	1.9	2	1 to 3
LABA/ICS	3.0	3	2 to 3
LABA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.1.2.3 Clinical homogeneity assessment

Table 11 shows the clinical homogeneity assessment across the available comparisons. Bronchial reversibility ranged from 7.0% to 22.4% and was not available in three comparisons, which could have introduced a bias favouring an ICS‐containing inhaler in a population with a significant bronchodilator response. The NMA results should be interpreted with caution because of the difference in and lack of data on bronchial reversibility.

1.1.2.4 Pairwise meta‐analyses

Contrary to the NMAs, the pairwise MAs showed no evidence that any treatment group was better than the others. There was no direct comparison for LABA/LAMA versus LABA (Appendix 6). The certainty of evidence was moderate for all comparisons due to a suboptimal information size, which could explain the discrepancy between the NMAs and pairwise MAs (See 'Summary of findings' tables). There was no difference between random and fixed analyses.

1.1.3 Rank probabilities for exacerbations

Figure 5 plots the ranks of each treatment group for severe exacerbations and moderate to severe exacerbations. The vertical axis shows the probability of being ranked best, second best, third best, or worst treatment group for each of the treatment groups. LABA/LAMA has a high probability of being the best intervention for both severe and moderate to severe exacerbations in the high‐risk population, with a probability of nearly 100% of being the best treatment group to reduce moderate to severe exacerbations. LABA has a very high probability of being the worst treatment group for reducing both severe and moderate to severe exacerbations.

5.

Plot of rank probabilities for each treatment group 
 Severe exacerbations (solid line), and moderate to severe exacerbations (dashed line), in the high‐risk population

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.2 Outcome: St George's Respiratory Questionnaire (SGRQ) responders

1.2.1 Outcome: SGRQ responders at three and six months

There were insufficient data to perform a NMA for SGRQ responders at three and six months. The results were based on one study for the following comparisons: LABA/LAMA versus LAMA at six months; LABA/ICS versus LAMA at three and six months; and LAMA versus LABA at three and six months. There is no evidence to suggest any treatment group is associated with a higher proportion of SGRQ responders compared to the others except for LABA/LAMA versus LAMA at six months, in which LABA/LAMA had a significantly greater proportion of SGRQ responders compared to LAMA (OR 1.30, 95% CI 1.08 to 1.56; Appendix 6). The certainty of evidence was low to moderate.

1.2.2 Outcome: SGRQ responders at 12 months

Seven studies of 10 interventions and four treatment groups were available for this outcome (Appendix 3; Figure 6 a and b). Note that interventions formoterol 12 μg twice daily, formoterol/budesonide 400μg/12 μg twice daily, and formoterol/beclomethasone 200 μg/12 μg twice daily are disconnected from the main treatment network (Figure 6a), but we included them in a class/group model.

6.

St George's Respiratory Questionnaire responders at 12 months in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.2.2.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.2.2.2 NMA results

The NMA included a total of 11,089 participants (LABA: 2313, LAMA: 3078, LABA/ICS: 3496, LABA/LAMA: 2202). Figure 6d and Table 20 show the ORs of SGRQ responders at 12 months for each treatment group compared to every other. There is evidence to suggest that LABA/ICS increases the odds of response at 12 months compared to LABA (OR 1.17, 95% CrI 1.02 to 1.34), and that LABA/LAMA increases the odds of response compared to all other treatment groups (OR 1.21, 95% CrI 1.07 to 1.36; OR 1.36, 95% CrI 1.18 to 1.58, and OR 1.41, 95% CrI 1.20 to 1.66, against LABA/ICS, LAMA and LABA respectively), using the fixed‐treatment‐effect model. Results are more uncertain when random‐treatment effects are assumed. Table 21 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st).

13. Relative effects: St. George's Respiratory Questionnaire responders at 12 months in the high‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	1.21	1.07 to 1.36	1.19	0.83 to 1.71
LABA/LAMA v LAMA	1.36	1.18 to 1.58	1.34	0.93 to 1.88
LABA/LAMA v LABA	1.41	1.20 to 1.66	1.38	0.89 to 2.04
LABA/ICS v LAMA	1.13	0.98 to 1.30	1.12	0.81 to 1.54
LABA/ICS v LABA	1.17	1.02 to 1.34	1.15	0.87 to 1.49
LAMA v LABA	1.03	0.91 to 1.18	1.03	0.72 to 1.44

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

14. Mean and median ranks: St. George's Respiratory Questionnaire responders at 12 months in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LABA/ICS	2.1	2	2 to 3
LAMA	3.3	3	2 to 4
LABA	3.7	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.2.2.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the fixed‐effect NMA except for LABA/ICS versus LABA, in which LABA/ICS significantly increased the odds of SGRQ response compared to LABA with the fixed‐effect model (OR 1.22, 95% CI 1.03 to 1.46), but not with the random‐effects model (OR 1.15, 95% CI 0.78 to 1.72). There was no direct comparison for LABA/LAMA versus LABA. The certainty of evidence was high for LABA/LAMA versus LABA/ICS, moderate for LABA/ICS versus LAMA or LABA and LAMA versus LABA, and low for LABA/LAMA versus LAMA. There was no difference between random and fixed analyses except for LABA/ICS versus LABA, in which the difference was significant with the fixed model but not with the random model (Appendix 6).

1.3 Change from baseline in SGRQ score

1.3.1 Outcome: change from baseline in SGRQ score at three months

We included nine studies of 12 interventions and four treatment groups for this outcome (Appendix 3; Figure 7 a and b). Note that interventions salmeterol 50 μg twice daily, formoterol 9 μg twice daily, salmeterol 50 μg twice daily + fluticasone 250 μg twice daily, salmeterol/fluticasone 50 μg/250 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, and formoterol/budesonide 9 μg/320 μg twice daily are disconnected from the main treatment network (Figure 7a), but we included them in a class/group model.

7.

Change from baseline in St George's Respiratory Questionnaire score at 3 months in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 0 favour the first named treatment group.

BUD: budesonide; FM: formoterol; FP: fluticasone propionate; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.3.1.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.3.1.2 NMA results

The NMA included a total of 11,263 participants (LABA: 2764, LAMA: 2992, LABA/ICS: 3220, LABA/LAMA: 2287). Figure 7d and Table 22 show the mean difference in change from baseline in SGRQ score at three months for each treatment group compared to every other. There is evidence to suggest that both LABA/LAMA and LABA/ICS improve SGRQ score at three months compared to LABA (MD −3.21, 95% CrI −4.52 to −1.92; MD −1.82, 95% CrI −2.86 to −0.78), and LAMA monotherapies (MD −3.31, 95% CrI −4.67to −1.97; MD −1.92, 95% CrI −3.11 to −0.74) and that LABA/LAMA improves the score compared to LABA/ICS, when the fixed‐treatment‐effect model is used (MD −1.39, 95% CrI −2.37 to −0.42). The 95% CI exceeding minimal clinically important difference (MCID) of 4 suggests a possibility of clinically significant improvement favouring LABA/LAMA over LAMA and LABA. Results are more uncertain when considering the random‐treatment‐effects model although there is evidence that LABA/LAMA improves the score compare to LABA and LAMA monotherapies. Table 23 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st).

15. Relative effects: change from baseline in St. George's Respiratory Questionnaire score at 3 months in the high‐risk population.

Treatment comparison	Mean differences ‐ fixed effects		Mean differences ‐ random effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	‐1.39	(‐2.37, ‐0.42)	‐1.47	(‐3.74, 0.45)
LABA/LAMA v LAMA	‐3.31	(‐4.67, ‐1.97)	‐3.32	(‐5.52, ‐1.12)
LABA/LAMA v LABA	‐3.21	(‐4.52, ‐1.92)	‐3.21	(‐5.63, ‐0.81)
LABA/ICS v LAMA	‐1.92	(‐3.11, ‐0.74)	‐1.83	(‐3.76, 0.35)
LABA/ICS v LABA	‐1.82	(‐2.86, ‐0.78)	‐1.73	(‐3.25, 0.05)
LAMA v LABA	0.1	(‐0.76, 0.96)	0.1	(‐1.86, 2.09)

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

16. Mean and median ranks: change from baseline in St. George's Respiratory Questionnaire score at 3 months in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LABA/ICS	2.0	2	2 to 2
LABA	3.4	3	3 to 4
LAMA	3.6	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.1.3 Pairwise meta‐analyses

There was no direct comparison for LABA/LAMA versus LABA. Otherwise, the results from pairwise MAs were consistent with the NMAs, except for LABA/ICS versus LAMA, in which the 95% CI crossed the line of no difference with the pairwise MA (MD −1.06, 95% CI −4.39 to 2.27) and the random‐effects NMA (MD −1.83, 95% CrI −3.76 to 0.35)) but not with the fixed‐effect NMA (MD −1.92, 95% CrI −3.11 to −0.74; Appendix 6 and Table 22). The certainty of evidence for LAMA/ICS versus LAMA was low, as in the NMAs. A clinically important improvement cannot be excluded with LABA/LAMA compared to LAMA (MD −3.68, 95% CI −5.84 to −1.52), as well as with LABA/ICS compared to LAMA (MD −1.06, 95% CI −4.39 to 2.27), because the 95% CIs crossed the line of MCID of 4. Otherwise, there is no evidence of a clinically significant difference in treatment effects between treatment groups. The certainty of evidence was high for LABA/LAMA versus LABA/ICS and LAMA versus LABA, moderate for LABA/LAMA versus LAMA, and low for LABA/ICS versus LABA. There was no difference between random and fixed analyses.

1.3.2 Outcome: change from baseline in SGRQ score at six months

We included 10 studies of 12 interventions and four treatment groups for this outcome (Appendix 3, Figure 8 a and b). Note that interventions formoterol 9 μg twice daily, salmeterol 50 μg twice daily + fluticasone 250 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily and formoterol/budesonide 9 μg/320 μg twice daily are disconnected from the main treatment network (Figure 8a), but we included them in a class/group model.

8.

Change from baseline in St George's Respiratory Questionnaire score at 6 months in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 0 favour the first named treatment group.

BUD: budesonide; FM: formoterol; FP: fluticasone propionate; Glyco: glycopyrronium; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.3.2.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Table 24).

17. Relative effects: change from baseline in St. George's Respiratory Questionnaire score at 6 months in the high‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	−1.27	−2.26 to −0.29	−1.29	−3.03 to 0.46
LABA/LAMA v LAMA	−2.48	−3.72 to −1.24	−2.6	−4.52 to −0.75
LABA/LAMA v LABA	−2.88	−4.03 to −1.73	−2.9	−4.79 to −0.93
LABA/ICS v LAMA	−1.21	−2.16 to −0.25	−1.31	−2.90 to 0.17
LABA/ICS v LABA	−1.60	−2.27 to −0.93	−1.61	−2.61 to −0.54
LAMA v LABA	−0.39	−1.27 to 0.47	−0.3	−1.74 to 1.34

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.2.2 NMA results

The NMA included a total of 12,967 participants (LABA: 3091, LAMA: 3273, LABA/ICS: 4317, LABA/LAMA: 2286). Figure 8d and Table 24 show the mean difference in change from baseline in SGRQ score at six months for each treatment group compared to every other. There is evidence to suggest that both LABA/LAMA and LABA/ICS improve SGRQ score at six months compared to LABA (MD −2.88, 95% CrI −4.03 to −1.73; MD −1.60, 95% CrI −2.27 to −0.93), and LAMA monotherapies (MD −2.48, 95% CrI −3.72 to −1.24), and that LABA/LAMA improves the score compared to LABA/ICS (MD −1.27, 95% CrI −2.26 to −0.29), using a fixed‐treatment‐effect model. The 95% CI exceeding MCID of 4 suggests a possibility of clinically significant improvement favouring LABA/LAMA over LABA. Results are more uncertain when considering the random‐treatment‐effects model although there is evidence that LABA/ICS and LABA/LAMA improve the score compare to LABA. Table 25 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st).

18. Mean and median ranks: change from baseline in St. George's Respiratory Questionnaire score at 6 months in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LABA/ICS	2.0	2	2 to 2
LAMA	3.2	3	3 to 4
LABA	3.8	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.2.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the fixed‐treatment‐effect NMA. There was no direct comparison for LABA/LAMA versus LABA. A clinically important improvement could not be excluded with LABA/LAMA compared to LAMA because the 95% CIs crossed the line of MCID of 4 (MD −2.79, 95% CI −5.02 to −0.56). Otherwise, there is no evidence of a clinically significant difference in treatment effects between treatment groups although no clear difference was seen in the all comparisons except for LAMA versus LABA (MD −0.70, 95% CI −1.74 to 0.34; Appendix 6). The certainty of evidence was high for LABA/LAMA versus LABA/ICS and LAMA versus LABA, moderate for LABA/LAMA versus LAMA, low for LABA/ICS versus LAMA, and very low for LABA/ICS versus LABA. There was no difference between random and fixed analyses.

1.3.3 Outcome: change from baseline in SGRQ score at 12 months

We included 14 studies of 15 interventions and four treatment groups for this outcome (Appendix 3; Figure 9 a and b). Note that interventions formoterol 9 to 12 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, formoterol/budesonide 12 μg/400 μg twice daily, formoterol/beclomethasone 12 μg/200 μg twice daily, and formoterol/budesonide 9 μg/320 μg twice daily are disconnected from the main treatment network (Figure 9a) but we included them in a class/group model.

9.

Change from baseline in St George's Respiratory Questionnaire score at 12 months in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 0 favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.3.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐effects‐model for comparison (Appendix 4).

1.3.3.2 NMA results

The NMA included a total of 15,459 participants (LABA: 4021, LAMA: 3216, LABA/ICS: 5891, LABA/LAMA: 2331). Figure 9d and Table 26 show the mean difference in change from baseline in SGRQ score at 12 months for each treatment group compared to every other. There is evidence to suggest that all treatment groups improve SGRQ score at 12 months compared to LABA (MD −2.10, 95% CrI −3.08 to −1.13; MD −1.57, 95% CrI −2.23 to −0.92; MD −0.98, 95% CrI −1.86 to −0.08 for LABA/LAMA, LABA/ICS and LAMA respectively), and that LABA/LAMA improves the score compared to LAMA (MD −1.12, 95% CrI −1.88 to −0.37), using the fixed‐treatment‐effect model. Results are more uncertain when considering the random‐treatment‐effects model although there is evidence that LABA/LAMA and LABA/ICS improve the score compared to LABA (MD −2.31, 95% CrI −4.17 to −0.64; MD −1.61, 95% CrI −2.52 to −0.69), and that LABA/LAMA improves the score compared to LAMA (MD −1.49, 95% CrI −3.16 to −0.20). The 95% CI exceeding MCID of 4 suggests a possibility of clinically significant improvement favouring LABA/LAMA over LABA. Table 27 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 2nd).

19. Relative effects: change from baseline in St. George's Respiratory Questionnaire score at 12 months in the high‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	−0.52	−1.42 to 0.36	−0.69	−2.46 to 0.87
LABA/LAMA v LAMA	−1.12	−1.88 to −0.37	−1.49	−3.16 to −0.20
LABA/LAMA v LABA	−2.10	−3.08 to −1.13	−2.31	−4.17 to −0.64
LABA/ICS v LAMA	−0.59	−1.48 to 0.29	−0.79	−2.40 to 0.65
LABA/ICS v LABA	−1.57	−2.23 to −0.92	−1.61	−2.52 to −0.69
LAMA v LABA	−0.98	−1.86 to −0.08	−0.82	−2.29 to 0.84

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

20. Mean and median ranks: change from baseline in St. George's Respiratory Questionnaire score at 12 months in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.1	1	1 to 2
LABA/ICS	2.0	2	1 to 3
LAMA	2.9	3	2 to 3
LABA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.3.3 Pairwise meta‐analyses

There is evidence to suggest that LABA/LAMA improves SGRQ score at 12 months compared to LABA/ICS or LAMA (MD −1.20, 95% CI −2.34 to −0.06 or MD −3.38, 95% CI −5.83 to −0.93), and that LABA/ICS improves the score compared to LABA (MD −1.75, 95% CI −2.61 to −0.89), although the mean differences do not reach the clinical significance of MCID of 4. There is no evidence of significant difference for LABA/ICS versus LAMA and LAMA versus LABA. There was no direct comparison for LABA/LAMA versus LABA. The results were consistent with the fixed‐effect NMA except for LABA/LAMA versus LABA/ICS and LAMA versus LABA. LABA/LAMA significantly improved the score compared to LABA/ICS in the pairwise MA (MD −1.20, 95% CI −2.34 to −0.06), but not in the NMA (MD −0.52, 95% CrI −1.42 to 0.36), and LAMA improved the score compared to LABA in the NMA (MD −0.98, 95% CrI −1.86 to −0.08), but not in the pairwise MA (MD −0.40, 95% CI −1.56 to 0.76; Appendix 6). There is no evidence of clinically significant difference in any comparison except for LABA/LAMA versus LAMA, in which the 95% CI suggested a possibility of clinically significant improvement favouring LABA/LAMA over LAMA (MD −3.38, 95% CI −5.83 to −0.93). The certainty of evidence was high for LABA/LAMA versus LABA/ICS and LAMA versus LABA, moderate for LABA/ICS versus LABA, and low for LABA/LAMA or LABA/ICS versus LAMA. There was no difference between random and fixed analyses.

1.3.4 Rank probabilities for change from baseline in SGRQ score at 3, 6, and 12 months

Figure 10 plots the ranks of SGRQ score at 3, 6, and 12 months for each treatment group. The vertical axis shows the probability of being ranked best, second best, third best, or worst treatment group. LABA/LAMA has a high probability of being ranked first at every time point whereas LABA has a high probability of being ranked worst at 6 and 12 months.

10.

Plot of rank probabilities for each treatment group
 Change from baseline in St George's Respiratory Questionnaire score at 3 (solid line), 6 (dashed line), and 12 months (dotted line), in the high‐risk population

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.4 Outcome: transition dyspnoea index (TDI)

1.4.1 TDI at 3, 6, and 12 months

There were insufficient data to perform a NMA for TDI at 3, 6, and 12 months. The results were based on one trial for the following comparisons: LABA/ICS versus LAMA at 3, 6, and 12 months and LAMA versus LABA at 3, 6, and 12 months. There is no evidence of clinically significant improvement in TDI (MCID of 1), with any treatment group compared to the others although a significant difference was seen for LABA/ICS versus LAMA at three months (MD 0.50, 95% CI 0.18 to 0.82), and LAMA versus LABA at 3, 6, and 12 months (MD −0.14 95% CI −0.15 to −0.13; MD −0.19 95% CI −0.20 to −0.18; and MD −0.26 95% CI −0.27 to −0.25), favouring LABA/ICS over LAMA and LABA over LAMA (Appendix 6). The certainty of evidence was low for LABA/ICS versus LAMA at 12 months and moderate for the rest of the comparisons.

1.5 Outcome: change from baseline in forced expiratory volume in one second (FEV1)

1.5.1 Outcome: change from baseline in FEV1 at three months

We included 11 studies of 12 interventions and four treatment groups for this outcome (Appendix 3; Figure 11 a and b). Note that interventions formoterol 9 μg twice daily, formoterol 12 μg twice daily, formoterol/budesonide 9 μg/320 μg twice daily, and formoterol/beclomethasone 12 μg/200 μg twice daily are disconnected from the main treatment network (Figure 11a), but we included them in a class/group model.

11.

Change from baseline in forced expiratory volume in 1 second at 3 months in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Positive values favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.1.1 Model selection and inconsistency checking

We chose a fixed‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.5.1.2 NMA results

The NMA included a total of 11,668 participants (LABA: 2203, LAMA: 2010, LABA/ICS: 5192, LABA/LAMA: 2263). Figure 11d and Table 28 show the mean difference in change from baseline in FEV1 at three months for each treatment group compared to every other. There is evidence to suggest that all treatment groups improve FEV1 at three months compared to LABA (MD 0.12, 95% CrI 0.10 to 0.15; MD 0.05, 95% CrI 0.04, 0.07; and MD 0.05, 95% CrI 0.02 to 0.07 for LABA/LAMA, LABA/ICS, and LAMA respectively), and that LABA/LAMA improves FEV1 compared to LABA/ICS and LAMA (MD 0.07, 95% CrI 0.05 to 0.09; and MD 0.07, 95% CrI 0.05 to 0.10). The difference for LABA/LAMA versus LABA was of clinical significance favouring LABA/LAMA (MD 0.12, 95% CrI 0.10 to 0.15). The 95% CI reaching MCID of 0.1 L suggests a possibility of clinically significant improvement favouring LABA/LAMA over LAMA. Table 29 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st), whereas LABA was the worst ranked with a median of 4 (95% CrI 4th to 4th).

21. Relative effects: change from baseline in forced expiratory volume in 1 second at 3 months in the high‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	0.07	0.05 to 0.09	0.07	0.03 to 0.10
LABA/LAMA v LAMA	0.07	0.05 to 0.10	0.07	0.04 to 0.11
LABA/LAMA v LABA	0.12	0.10 to 0.15	0.12	0.07 to 0.15
LABA/ICS v LAMA	0	−0.02 to 0.02	0.01	−0.02 to 0.04
LABA/ICS v LABA	0.05	0.04 to 0.07	0.05	0.03 to 0.07
LAMA v LABA	0.05	0.02 to 0.07	0.04	0.00 to 0.08

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

22. Mean and median ranks: change from baseline in forced expiratory volume in 1 second at 3 months in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LABA/ICS	2.4	2	2 to 3
LAMA	2.6	3	2 to 3
LABA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.1.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs. There is no evidence of clinically significant improvement (MCID of 0.1 L or greater), with any treatment group compared to the others except for LABA/LAMA versus LABA/ICS, in which the 95% CI suggested a possibility of clinically significant difference favouring LABA/LAMA over LABA/ICS (MD 0.08, 95% CI 0.06 to 0.10; Appendix 6). There was no direct comparison for LABA/LAMA versus LABA and LAMA versus LABA. The certainty of evidence was high for LABA/LAMA versus LABA/ICS and LABA/ICS versus LAMA and moderate for LABA/LAMA versus LAMA and LABA/ICS versus LABA. There was no difference between random and fixed analyses.

1.5.2 Outcome: change from baseline in FEV1 at six months

Eleven studies of 11 interventions and four treatment groups were available for this outcome (Appendix 3; Figure 12 a and b). Note that interventions formoterol 9 μg twice daily, salmeterol 50 μg twice daily + fluticasone 250 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, and formoterol/budesonide 9 μg/320 μg twice daily are disconnected from the main treatment network (Figure 12a), but we included them were in a class/group model.

12.

Change from baseline in forced expiratory volume in 1 second at 6 months in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Positive values favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.5.2.1 Model selection and inconsistency checking

We chose a fixed‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.5.2.2 NMA results

The NMA included a total of 10,822 participants (LABA: 2111, LAMA: 1700, LABA/ICS: 4263, LABA/LAMA: 2748). Figure 12d and Table 30 show the mean difference in change from baseline in FEV1 at six months for each treatment group compared to every other. There is evidence to suggest that all treatment groups improve FEV1 at six months compared to LABA, (MD 0.13, 95% CrI 0.10 to 0.15; MD 0.04, 95% CrI 0.03 to 0.06; and MD 0.06, 95% CrI 0.03 to 0.08 for LABA/LAMA, LABA/ICS, and LAMA respectively), and that LABA/LAMA improves FEV1 compared to LABA/ICS and LAMA (MD 0.08, 95% CrI 0.06 to 0.10; and MD 0.07, 95% CrI 0.04 to 0.09). The difference was clinically significant (MCID of 0.1 L or greater), for LABA/LAMA versus LABA (MD 0.13, 95% CrI 0.10 to 0.15), favouring LABA/LAMA over LABA with the fixed‐effect model. The 95% CI reaching MCID of 0.1 L suggests a possibility of clinically significant improvement favouring LABA/LAMA over LABA/ICS. Table 31 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st), whereas LABA was the worst ranked with a median of 4 (95% CrI 4th to 4th).

23. Relative effects: change from baseline in forced expiratory volume in 1 second at 6 months in the high‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	0.08	0.06 to 0.10	0.08	0.04 to 0.12
LABA/LAMA v LAMA	0.07	0.04 to 0.09	0.07	0.02 to 0.11
LABA/LAMA v LABA	0.13	0.10 to 0.15	0.13	0.09 to 0.18
LABA/ICS v LAMA	−0.02	−0.04 to 0.01	−0.02	−0.06 to 0.03
LABA/ICS v LABA	0.04	0.03 to 0.06	0.05	0.03 to 0.08
LAMA v LABA	0.06	0.03 to 0.08	0.06	0.02 to 0.11

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

24. Mean and median ranks: change from baseline in forced expiratory volume in 1 second at 6 months in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LAMA	2.1	2	2 to 3
LABA/ICS	2.9	3	2 to 3
LABA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.2.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs. There is no evidence of clinically significant improvement (MCID of 0.1 L or greater), with any treatment group compared to the others except for LABA/LAMA versus LABA/ICS or LAMA, in which the 95% CI suggested a possibility of clinically significant difference favouring LABA/LAMA over LABA/ICS or LAMA (MD 0.09, 95% CI 0.07 to 0.11; or MD 0.06, 95% CI 0.02 to 0.10; Appendix 6). There was no direct comparison for LABA/LAMA versus LABA and LAMA versus LABA. The certainty of evidence was high for LABA/LAMA versus LABA/ICS and moderate for LABA/LAMA versus LAMA and LABA/ICS versus LAMA or LABA. There was no difference between random and fixed analyses.

1.5.3 Outcome: change from baseline in FEV1 at 12 months

We included 13 studies of 13 interventions and four treatment groups for this outcome (Appendix 3; Figure 13a and b). Note that interventions formoterol 9 μg twice daily, formoterol 12 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, formoterol/budesonide 12 μg/400 μg twice daily, and formoterol/beclomethasone 12 μg/200 μg twice daily are disconnected from the main treatment network (Figure 13a), but we included them in a class/group model.

13.

Change from baseline in forced expiratory volume in 1 second at 12 months in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Positive values favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.3.1 Model selection and inconsistency checking

We chose a fixed‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.5.3.2 NMA results

The NMA included a total of 11,171 participants (LABA: 1944, LAMA: 1919, LABA/ICS: 4982, LABA/LAMA: 2326). Figure 13d and Table 32 show the mean difference in change from baseline in FEV1 at 12 months for each treatment group compared to every other. There is evidence to suggest that all treatment groups improve FEV1 at 12 months compared to LABA (MD 0.12, 95% CrI 0.08 to 0.16; MD 0.05, 95% CrI 0.03 to 0.07; and MD 0.08, 95% CrI 0.04 to 0.12 for LABA/LAMA, LABA/ICS, and LAMA respectively), and that LABA/LAMA improves FEV1 compared to LABA/ICS (MD 0.07, 95% CrI 0.04 to 0.1). The 95% CI containing MCID of 0.1 L suggests a possibility of clinically significant improvement favouring LABA/LAMA over LABA/ICS and LABA and favouring LAMA over LABA. Table 33 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st), whereas LABA was the worst ranked with a median of 4 (95% CrI 4th to 4th).

25. Relative effects: change from baseline in forced expiratory volume in 1 second at 12 months in the high‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	0.07	0.05 to 0.09	0.07	0.04 to 0.10
LABA/LAMA v LAMA	0.04	0.01 to 0.07	0.04	0.00 to 0.08
LABA/LAMA v LABA	0.11	0.09 to 0.14	0.12	0.08 to 0.16
LABA/ICS v LAMA	−0.03	−0.06 to 0.00	−0.03	−0.07 to 0.01
LABA/ICS v LABA	0.05	0.03 to 0.06	0.05	0.03 to 0.07
LAMA v LABA	0.07	0.04 to 0.11	0.08	0.04 to 0.12

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

26. Mean and median ranks: change from baseline in forced expiratory volume in 1 second at 12 months in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LAMA	2.0	2	2 to 2
LABA/ICS	3.0	3	3 to 3
LABA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.3.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs except for LABA/LAMA versus LAMA, in which there is evidence of significant improvement favouring LABA/LAMA over LAMA (MD 0.05, 95% CI 0.01 to 0.09). There was no direct comparison for LABA/LAMA versus LABA and LAMA versus LABA. Otherwise there is no evidence of clinically significant improvement (MCID of 0.1 L) with any treatment group compared to the others (Appendix 6). The certainty of evidence was very low for LABA/ICS versus LAMA and moderate for the rest of the available comparisons. There was no difference between random and fixed analyses.

1.5.4 Rank probabilities for change from baseline in FEV1 at 3, 6, and 12 months

Figure 14 plots the ranks of each treatment group for FEV1 at 3, 6 and 12 months. The vertical axis shows the probability of being the best, second best, third best, or worst treatment group. LABA/LAMA has nearly 100% probability of being ranked first at all time points with LABA having a very high probability of being the worst intervention at all time points.

14.

Plot of rank probabilities for each treatment group
 Change from baseline in forced expiratory volume in 1 second at 3 (solid line), 6 months (dashed line) and 12 months in the high‐risk population.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.6 Outcome: mortality

Twenty‐four studies of 18 interventions and four treatment groups were available for this outcome (Appendix 3; Figure 15 a and b). Note that interventions formoterol 9 μg twice daily, formoterol 12 μg twice daily, salmeterol 50 μg twice daily + fluticasone 250 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, formoterol/budesonide 9 μg/320 μg twice daily, formoterol/budesonide 12 μg/400 μg twice daily, and formoterol/beclomethasone 12 μg/200 μg twice daily are disconnected from the main treatment network (Figure 15a), but we included them in a class/group model.

15.

Mortality in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.6.1 Model selection and inconsistency checking

We chose a fixed‐effect model with fixed‐class effects, assuming consistency, although results should be interpreted with caution due to some evidence of inconsistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.6.2 NMA results

The NMA included a total of 31,674 participants (LABA: 11,182, LAMA: 7853, LABA/ICS: 10,084, LABA/LAMA: 2555). The median duration of follow‐up was 52 weeks (range 12 to 156 weeks). Figure 15d and Table 34 show the OR of mortality for each treatment group compared to every other. There was no evidence to suggest that any treatment group increased or decreased the odds of mortality compared to any other. Table 35 shows the rank statistics for the four treatment groups (sorted by mean rank). All treatment groups have high uncertainty in ranks as expected, due to no treatment effect being identified for any treatment group.

27. Relative effects: mortality in the high‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA v LABA/ICS	1.12	0.75 to 1.68	1.15	0.70 to 1.95
LABA/LAMA v LAMA	0.98	0.66 to 1.42	0.99	0.62 to 1.60
LABA/LAMA v LABA	0.97	0.63 to 1.46	1.04	0.63 to 1.86
LABA/ICS v LAMA	0.87	0.65 to 1.16	0.86	0.58 to 1.26
LABA/ICS v LABA	0.86	0.66 to 1.11	0.91	0.68 to 1.23
LAMA v LABA	0.99	0.77 to 1.27	1.05	0.75 to 1.59

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

28. Mean and median ranks: mortality in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/ICS	1.6	1	1 to 4
LABA/LAMA	2.6	3	1 to 4
LAMA	2.8	3	1 to 4
LABA	3.0	3	1 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.6.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs. There was no direct comparison for LABA/LAMA versus LABA (Appendix 6). The certainty of evidence was low for LABA/ICS versus LABA and moderate for the rest of available comparisons. There was no difference between random and fixed analyses.

1.7 Outcome: serious adverse events (SAEs)

1.7.1 Outcome: total SAEs

The analysis for total SAEs included 24 studies of 18 interventions and four treatment groups. We included a total of 31,721 participants (LABA: 10,942, LAMA: 7853, LABA/ICS: 10,371, LABA/LAMA: 2555; Appendix 3; Figure 16 a and b). The median duration of follow‐up was 52 weeks (range 12 to 156 weeks). Note that interventions formoterol 9 μg twice daily, formoterol 12 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, formoterol/budesonide 9 μg/320 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, formoterol/budesonide 12 μg/400 μg twice daily, formoterol/beclomethasone 12 μg/200 μg twice daily and salmeterol 50 μg twice daily + fluticasone 250 μg twice daily are disconnected from the main treatment network (Figure 16a), but we included them in a class/group model.

16.

Total serious adverse events in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.7.1.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.7.2 Outcome: chronic obstructive pulmonary disease (COPD) SAEs

The analysis for COPD SAEs included 20 studies of 14 interventions and four treatment groups. We included a total of 28,614 participants (LABA: 9675, LAMA: 7697, LABA/ICS: 8835, LABA/LAMA: 2407; Appendix 3;Figure 17 a and b). The median duration of follow‐up was 52 weeks (range 12 to 156 weeks). Note that interventions formoterol 9 μg twice daily, salmeterol 50 μg twice daily + fluticasone 250 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily and formoterol/budesonide 9 μg/320 μg twice daily are disconnected from the main treatment network (Figure 17a), but we included them in a class/group model.

17.

Chronic obstructive pulmonary disease serious adverse events in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects.

BUD: budesonide; FM: formoterol; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.7.2.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.7.3 Outcome: cardiac SAEs

The analysis for cardiac SAEs included 19 studies of 16 interventions and four treatment groups (Appendix 3; Figure 18 a and b). We included a total of 29,045 participants (LABA: 10,016, LAMA: 7567, LABA/ICS: 9055, LABA/LAMA: 2407). The median duration of follow‐up was 52 weeks (range 12 to 156 weeks). Note that interventions formoterol 9 μg twice daily, formoterol 12 μg twice daily, salmeterol 50 μg twice daily + fluticasone 250 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, formoterol/budesonide 9 μg/320 μg twice daily, and formoterol/beclomethasone 12 μg/200 μg twice daily are disconnected from the main treatment network (Figure 18a), but we included them in a class/group model.

18.

Cardiac serious adverse events in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects

BDP: beclomethasone; BUD: budesonide; FM: formoterol; FP: fluticasone propionate; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.7.3.1 Model selection and inconsistency checking

We chose a random‐treatment‐effects model with fixed‐class effects, assuming consistency. We also report results based on the fixed‐treatment‐effect model with fixed‐class effects for comparison (Appendix 4).

1.7.4 NMA results

Table 36 shows the OR of each type of adverse event for each treatment group compared to every other. For total SAEs there is evidence to suggest that LABA/ICS increases the odds of SAEs compared to LAMA (OR 1.14, 95% CrI 1.02 to 1.27), and that LAMA decreases the odds of SAEs compared to LABA (OR 0.88, 95% CrI 0.81 to 0.97), although this effect was only seen in the fixed‐effect model. For COPD SAEs there is evidence to suggest that LABA/ICS increases the odds of SAEs compared to LAMA (OR 1.22 95% CrI 1.05 to 1.42), and that LAMA decreases the odds of SAEs compared to LABA (OR 0.77, 95% CrI 0.68 to 0.87), and this was seen in both models. No difference between treatment groups was evident for cardiac SAEs.

29. Relative effects: serious adverse events in the high‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
Total SAEs
LABA/LAMA vs LABA/ICS	0.89	0.77 to 1.02	0.89	0.74 to 1.06
LABA/LAMA vs LAMA	1.01	0.87 to 1.17	1.01	0.83 to 1.21
LABA/LAMA vs LABA	0.89	0.77 to 1.04	0.89	0.73 to 1.08
LABA/ICS vs LAMA	1.14	1.02 to 1.27	1.13	0.99 to 1.31
LABA/ICS vs LABA	1.01	0.92 to 1.10	1.01	0.91 to 1.12
LAMA vs LABA	0.88	0.81 to 0.97	0.89	0.78 to 1.01
COPD SAEs
LABA/LAMA vs LABA/ICS	0.87	0.73 to 1.04	0.87	0.71 to 1.09
LABA/LAMA vs LAMA	1.07	0.89 to 1.28	1.07	0.85 to 1.34
LABA/LAMA vs LABA	0.82	0.68 to 1.00	0.83	0.65 to 1.05
LABA/ICS vs LAMA	1.22	1.05 to 1.42	1.22	1.02 to 1.46
LABA/ICS vs LABA	0.95	0.83 to 1.08	0.94	0.81 to 1.09
LAMA vs LABA	0.77	0.68 to 0.87	0.77	0.66 to 0.91
CARDIAC SAEs
LABA/LAMA vs LABA/ICS	0.91	0.66 to 1.25	0.70	0.03 to 5.88
LABA/LAMA vs LAMA	0.75	0.54 to 1.03	0.69	0.02 to 25.46
LABA/LAMA vs LABA	0.85	0.60 to 1.19	0.83	0.06 to 9.24
LABA/ICS vs LAMA	0.83	0.63 to 1.08	1.08	0.06 to 23.81
LABA/ICS vs LABA	0.93	0.75 to 1.16	1.27	0.37 to 5.97
LAMA vs LABA	1.13	0.89 to 1.42	1.13	0.06 to 21.22

COPD: chronic obstructive pulmonary disease; CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist;SAE: serious adverse event

1.7.5 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs except for LABA/ICS versus LAMA for COPD SAEs in which the NMA suggested LABA/ICS increased the odds of COPD SAEs compared to LAMA (OR 1.22, 95% CrI 1.05 to 1.42), whereas the pairwise MA did not (OR 0.99, 95% CI 0.33 to 2.96). There was no direct comparison for LABA/LAMA versus LABA for total, COPD, and cardiac SAEs. Table 37 shows the certainty of evidence for each treatment group compared to every other. There was no difference between random and fixed analyses (Appendix 6).

30. Certainty of evidence: serious adverse events in the high‐risk population.

Treatment comparison	Total SAEs	COPD SAEs	Cardiac SAEs
LABA/LAMA vs LABA/ICS	Moderate	Moderate	Moderate
LABA/LAMA vs LAMA	Moderate	Moderate	Moderate
LABA/LAMA vs LABA	NA	NA	NA
LABA/ICS vs LAMA	Moderate	Moderate	Moderate
LABA/ICS vs LABA	Moderate	Moderate	Moderate
LAMA vs LABA	High	High	Low

COPD: chronic obstructive pulmonary disease; CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist;NA: not applicable; SAE: serious adverse event

1.8 Outcome: dropouts due to adverse events

We included 25 studies of 18 interventions and four treatment groups for this outcome (Appendix 3; Figure 19 a and b). Note that interventions formoterol 9 μg twice daily, formoterol 12 μg twice daily, salmeterol 50 μg twice daily + fluticasone 250 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, formoterol/budesonide 9 μg/320 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, formoterol/budesonide 12 μg/400 μg twice daily, and formoterol/beclomethasone 12 μg/200 μg twice daily are disconnected from the main treatment network (Figure 19a), but we included them in a class/group model.

19.

Dropouts due to adverse events in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; FP: fluticasone propionate; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.8.1 Model selection and inconsistency checking

We chose a fixed‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

1.8.2 NMA results

The NMA included a total of 32,230 participants (LABA: 11,197, LAMA: 7853, LABA/ICS: 10,625, LABA/LAMA: 2555). The median duration of follow‐up was 52 weeks (range 12 to 156 weeks). Figure 19d and Table 38 show the OR of dropout due to adverse events for each treatment group compared to every other. There was no evidence to suggest that any treatment group increased or decreased the odds of dropout compared to any other. Table 39 shows the rank statistics for the four treatment groups (sorted by mean rank). All treatment groups have high uncertainty in ranks as expected, due to no treatment effect being identified for any treatment group.

31. Relative effects: dropouts due to adverse events in the high‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.93	0.76 to 1.14	0.93	0.73 to 1.19
LABA/LAMA vs LAMA	0.94	0.76 to 1.17	0.95	0.74 to 1.21
LABA/LAMA vs LABA	0.83	0.67 to 1.03	0.83	0.65 to 1.07
LABA/ICS vs LAMA	1.01	0.87 to 1.19	1.02	0.85 to 1.22
LABA/ICS vs LABA	0.89	0.79 to 1.01	0.89	0.79 to 1.01
LAMA vs LABA	0.88	0.77 to 1.01	0.88	0.75 to 1.03

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

32. Mean and median ranks: dropouts due to adverse events in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.6	1	1 to 4
LAMA	2.2	2	1 to 4
LABA/ICS	2.4	2	1 to 4
LABA	3.9	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.8.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs. There was no direct comparison for LABA/LAMA versus LABA (Appendix 6). The certainty of evidence was high for LAMA versus LABA, moderate for LABA/LAMA versus LABA/ICS, LABA/ICS versus LAMA, and low for LABA/LAMA versus LAMA and LABA/ICS versus LABA. There was no difference between random and fixed analyses.

1.9 Outcome: pneumonia

We included 24 studies of 18 interventions and four treatment groups for this outcome (Appendix 3; Figure 20 a and b). Note that interventions formoterol 9 μg twice daily, formoterol 12 μg twice daily, formoterol/budesonide 9 μg/160 μg twice daily, formoterol/budesonide 9 μg/320 μg twice daily, formoterol/budesonide 12 μg/400 μg twice daily, formoterol/beclomethasone 12 μg/200 μg twice daily, indacaterol 150 μg once daily + budesonide 400 μg twice daily, and salmeterol 50 μg twice daily + fluticasone 250 μg twice daily are disconnected from the main treatment network (Figure 20a), but we included them in a class/group model.

20.

Pneumonia in the high‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

BDP: beclomethasone; BUD: budesonide; FM: formoterol; FP: fluticasone propionate; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

1.9.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison. Results should be interpreted with some caution due to poor model fit, which can be attributed to studies with zero cells (Appendix 4).

1.9.2 NMA results

The NMA included a total of 31,812 participants (LABA: 10991, LAMA: 7853, LABA/ICS: 10413, LABA/LAMA: 2555). The median duration of follow‐up was 52 weeks (range 12 to 156 weeks). Figure 20d and Table 40 show the OR of pneumonia for each treatment group compared to every other. There is evidence to suggest that LABA/ICS increases the odds of pneumonia compared to the other treatment groups (OR 1.69, 95% CrI 1.20 to 2.44; OR 1.78, 95% CrI 1.33 to 2.39; OR 1.50, 95% CrI 1.17 to 1.92 for LABA/LAMA, LAMA and LABA respectively), but no evidence of differences across other comparisons (Appendix 6Table 7). Table 41 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LAMA with a median rank of 1st but with wide credible intervals (1st to 3rd), whereas LABA/ICS was ranked the worst (median = 4, 95% CrI 4th to 4th).

33. Relative effects: pneumonia in the high‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.59	0.41 to 0.83	0.59	0.35 to 1.01
LABA/LAMA vs LAMA	1.05	0.72 to 1.5	1.05	0.63 to 1.81
LABA/LAMA vs LABA	0.88	0.60 to 1.29	0.87	0.49 to 1.52
LABA/ICS vs LAMA	1.78	1.33 to 2.39	1.79	1.19 to 2.76
LABA/ICS vs LABA	1.50	1.17 to 1.92	1.48	1.10 to 1.98
LAMA vs LABA	0.84	0.65 to 1.09	0.83	0.54 to 1.21

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

34. Mean and median ranks: pneumonia in the high‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LAMA	1.5	1	1 to 3
LABA/LAMA	1.9	2	1 to 3
LABA	2.6	3	1 to 3
LABA/ICS	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.9.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs. There was no direct comparison for LABA/LAMA versus LABA (Appendix 6). The certainty of evidence was moderate for the all available comparisons (see 'Summary of findings' tables). There was no difference between random and fixed analyses.

2. Results: low‐risk population

2.1 Outcome: exacerbations

2.1.1 Outcome: moderate to severe exacerbations

We included 38 studies of 22 interventions and four treatment groups for this outcome (Appendix 3; Figure 21 a and b). Note that interventions indacaterol 75 μg once daily and indacaterol/glycopyrronium 27.5 μg/15.6 μg twice daily are disconnected from the main treatment network (Figure 21a), but we included them in a class/group model.

21.

Moderate to severe exacerbations in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

ACL: aclidinium; BUD: budesonide; FF: fluticasone furoate; FM: formoterol; FP: fluticasone propionate; Glyco: glycopyrronium; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; MF: mometasone furoate; SAL: salmeterol; Tio: tiotropium; UMEC: umeclidinium; VI: vilanterol

2.1.1.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.1.1.2 NMA results

The NMA included a total of 31,406 participants (LABA: 6845, LAMA: 7364, LABA/ICS: 9592, LABA/LAMA: 7605). The median duration of follow‐up was 24 weeks (range 12 to 156 weeks). Figure 21d and Table 42 show the HR for moderate to severe exacerbations for each treatment group compared to every other. There is evidence that all treatment groups of interventions decrease the rate of moderate to severe exacerbations compared to LABA (HR 0.78, 95% CrI 0.67 to 0.90; HR 0.89, 95% CrI 0.84 to 0.96; HR 0.87, 95% CrI 0.78 to 0.97 for LABA/LAMA, LABA/ICS and LAMA respectively; Appendix 7; Table 7), although there is added uncertainty for the comparison with LAMA in the random‐effects model. Table 43 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 2nd) with LABA the worst ranked treatment group (95% CrI 4th to 4th).

35. Relative effects: moderate to severe exacerbations in the low‐risk population.

Treatment comparison	Hazard ratios: fixed‐effect		Hazard ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.87	0.75 to 1.01	0.89	0.78 to 1.04
LABA/LAMA vs LAMA	0.90	0.76 to 1.06	0.88	0.76 to 1.01
LABA/LAMA vs LABA	0.78	0.67 to 0.90	0.78	0.69 to 0.89
LABA/ICS vs LAMA	1.03	0.91 to 1.17	0.98	0.83 to 1.14
LABA/ICS vs LABA	0.89	0.84 to 0.96	0.88	0.78 to 0.96
LAMA vs LABA	0.87	0.78 to 0.97	0.89	0.78 to 1.01

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

36. Mean and median group ranks: moderate to severe exacerbations in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.1	1	1 to 2
LAMA	2.2	2	1 to 3
LABA/ICS	2.6	3	2 to 3
LABA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.1.1.3 Clinical homogeneity assessment

Table 44 shows the clinical homogeneity assessment across the available comparisons. Bronchial reversibility ranged from 11.1% to 17.5%, which could have introduced a bias favouring an ICS‐containing inhaler in a population with a significant bronchodilator response. The NMA results should be interpreted with caution because of the difference in bronchial reversibility across the pairwise comparisons.

37. Study characteristics of treatment group pair‐wise comparisons and transitivity assessment in moderate to severe exacerbations in the low‐risk population.

Comparison	Comparisons	Number of participants	Mean age (years)	Male (%)	Baseline FEV1 (L) prebronchodilator	Current smoker (%)	Bronchial reversibility (%)
LABA/LAMA vs LABA/ICS	6	4315	63	74	45	1.33	14.9
LABA/LAMA vs LAMA	8	5192	63	71	47	1.32	14.7
LABA/LAMA vs LABA	5	2488	64	68	44	1.36	17.5
LABA/ICS vs LAMA	1	623	63	65	52	1.35	13
LABA/ICS vs LABA	6	6689	64	74	44	1.27	11.1
LAMA vs LABA	5	4567	64	71	39	1.3	17.1

CrI: credible interval; FEV1: forced expiratory volume in 1 second; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.1.1.4 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs except for LAMA versus LABA, in which the 95% CI crossed the line of no difference with the pairwise MA (OR 0.92, 95% CI 0.79 to 1.07; Appendix 7). The certainty of evidence was moderate for the LAMA versus LABA comparison due to a suboptimal information size, which could explain the difference. Otherwise, the certainty of evidence was moderate for LABA/LAMA versus LABA/ICS and LABA/ICS versus LABA, and low for LABA/LAMA versus LAMA and LABA/ICS versus LAMA (see: 'Summary of findings' tables). There was no difference between random and fixed analyses.

2.1.2 Outcome: severe exacerbations

We included 31 studies of 18 interventions and four treatment groups for this outcome (Appendix 3; Figure 22 a and b).

22.

Severe exacerbations in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.1.2.1 Model selection and inconsistency checking

We chose a fixed‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.1.2.2 NMA results

The NMA included a total of 36,285 participants (LABA: 4963, LAMA: 17856, LABA/ICS: 7302, LABA/LAMA: 6164). The median duration of follow‐up was 24 weeks (range 12 to 156 weeks). Figure 22d and Table 45 show the HR for severe exacerbations for each treatment group compared to every other. There is no evidence that any treatment group reduces severe exacerbations compared to the others, although uncertainty is large for some comparisons. HRs for LABA/LAMA versus LABA/ICS, LABA, and LAMA were 0.71 (95% CrI 0.47 to 1.08), 0.90, (95% CrI 0.6 to 1.31), and 0.72 (95% CrI 0.48 to 1.02), respectively (Appendix 7; Table 7). Table 46 shows the rank statistics for the four treatment groups (sorted by mean rank). There is considerable uncertainty in the ranks, which is consistent with there being no evidence of a difference in treatment effects between treatment groups. The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 3rd).

38. Relative effects: severe exacerbations in the low‐risk population.

Treatment comparison	Hazard ratios: fixed‐effect		Hazard ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.71	0.50 to 1.02	0.71	0.47 to 1.08
LABA/LAMA vs LAMA	0.88	0.62 to 1.24	0.90	0.60 to 1.31
LABA/LAMA vs LABA	0.73	0.51 to 1.03	0.72	0.48 to 1.02
LABA/ICS vs LAMA	1.23	0.96 to 1.57	1.25	0.86 to 1.85
LABA/ICS vs LABA	1.02	0.89 to 1.17	1.01	0.72 to 1.28
LAMA vs LABA	0.83	0.67 to 1.03	0.80	0.56 to 1.05

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

39. Mean and median ranks: severe exacerbations in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.3	1	1 to 3
LAMA	1.9	2	1 to 3
LABA	3.3	3	2 to 4
LABA/ICS	3.5	4	2 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.1.2.3 Clinical homogeneity assessment

Table 12 shows the clinical homogeneity assessment across the available comparisons. Bronchial reversibility ranged from 11.1% to 18.3%. The average bronchial reversibility for LABA/ICS versus LAMA was 11.1% which could have underestimated the effects of LABA/ICS. The NMA results should be interpreted with caution because of the difference in bronchial reversibility across the pairwise comparisons.

2.1.2.4 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs and showed no evidence that any treatment group reduced severe exacerbations compared to the others (Appendix 7). ORs for LABA/LAMA versus LABA/ICS, LAMA, and LABA were 0.66 (95% CI 0.27 to 1.63), 0.99 ( 95% CI 0.57 to 1.72), and 0.78 (95% CI 0.55 to 1.12). The certainty of evidence was high for LABA/ICS versus LABA, moderate for LABA/LAMA versus LABA/ICS, LABA/LAMA versus LAMA, and LABA/LAMA versus LABA, and low for LABA/ICS versus LAMA and LAMA versus LABA (see 'Summary of findings' tables). There was no difference between random and fixed analyses.

2.1.3 Rank probabilities for exacerbations

Figure 23 plots the ranks of each treatment group for severe exacerbations and moderate to severe exacerbations. The vertical axis shows the probability of being ranked best, second best, third best, or worst treatment group. LABA/LAMA has a high probability of being the best intervention for both severe and moderate to severe exacerbations in the low‐risk population with a probability of about 90% of being the best treatment group to reduce moderate to severe exacerbations. LABA has a high probability of being the worst treatment group for reducing moderate to severe exacerbations and has a very small probability of ranking among the best treatment groups for reducing both severe and moderate to severe exacerbations.

23.

Plot of rank probabilities for each treatment group for chronic obstructive pulmonary disease exacerbations in the low‐risk population
 Severe exacerbations (solid line), and moderate/severe exacerbations (dashed line), in the low‐risk population

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.2 Outcome: St George's Respiratory Questionnaire (SGRQ) responders

2.2.1 Outcome: SGRQ responders at three months

We included 22 studies of 17 interventions and four treatment groups for this outcome (Appendix 3; Figure 24 a and b). Note that interventions formoterol 4.5 μg twice daily, formoterol 9 μg twice daily, glycopyrronium 15.6 μg twice daily, tiotropium 5 μg once daily, indacaterol/glycopyrronium 27.5 μg/15.6 μg twice daily and olodaterol/tiotropium 5 μg/5 μg once daily are disconnected from the main treatment network (Figure 24a), but we included them in a class/group model.

24.

St George's Respiratory Questionnaire score responders at 3 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values greater than 1 favour the first named treatment group.

FM: formoterol; Glyco: glycopyrronium; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; Olo: olodaterol; Tio: tiotropium

2.2.1.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.2.1.2 NMA results

The NMA included a total of 14,351 participants (LABA: 2371, LAMA: 5356, LABA/ICS: 2213, LABA/LAMA: 4411). Figure 24d and Table 47 show the OR of SGRQ responders at three months for each treatment group compared to every other. There is evidence to suggest that LABA/LAMA, LABA/ICS, and LABA increase the odds of SGRQ response at three months compared to LAMA (OR 1.33, 95% CrI 1.19 to 1.48; OR 1.24, 95% CrI 1.07 to 1.43; OR 1.37, 95% CrI 1.18 to 1.61)). Table 48 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LABA with a median rank of 1 although with large uncertainty (95% CrI 1st to 3rd), whereas LAMA was ranked the worst (median = 4, 95% CrI 4th to 4th).

40. Relative effects: St. George's Respiratory Questionnaire responders at 3 months in the low‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	1.07	0.94 to 1.23	1.07	0.93 to 1.23
LABA/LAMA vs LAMA	1.33	1.19 to 1.48	1.32	1.18 to 1.49
LABA/LAMA vs LABA	0.96	0.81 to 1.15	0.96	0.79 to 1.17
LABA/ICS vs LAMA	1.24	1.07 to 1.43	1.24	1.06 to 1.45
LABA/ICS vs LABA	0.9	0.76 to 1.06	0.9	0.75 to 1.08
LAMA vs LABA	0.73	0.62 to 0.85	0.72	0.60 to 0.87

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

41. Mean and median ranks: St. George's Respiratory Questionnaire responders at 3 months in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA	1.4	1	1 to 3
LABA/LAMA	1.8	2	1 to 3
LABA/ICS	2.8	3	1 to 3
LAMA	4.0	4	4 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.2.1.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs except for LABA/ICS versus LAMA (Appendix 7), in which the 95% CI crossed the line of no difference with the pairwise MA (OR 1.26 (95% CI 0.92 to 1.74), low confidence due to a wide 95% CI and a small sample size). There was no direct comparison for LABA/LAMA versus LABA. Otherwise, the certainty of evidence was high for LAMA/LABA versus LAMA, and LAMA versus LABA, and moderate for LABA/LAMA versus LABA/ICS, and low for LABA/ICS versus LABA. There was no difference between random and fixed analyses.

2.2.2 Outcome: SGRQ responders at six months

We included 18 studies of 19 interventions and four treatment groups for this outcome (Appendix 3; Figure 25 a and b).

25.

St George's Respiratory Questionnaire score responders at 6 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values greater than 1 favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.2.2.1 Model selection and inconsistency checking

We chose a random‐treatment‐effects model with a fixed‐class effect, assuming consistency (Appendix 4).

2.2.2.2 NMA results

The NMA included a total of 20,385 participants (LABA: 8259, LAMA: 5164, LABA/ICS: 2721, LABA/LAMA: 4241). Figure 25d and Table 49 show the OR of SGRQ responders at six months for each treatment group compared to every other. There is evidence to suggest that LABA/LAMA increases SGRQ responders at six months compared to both LAMA and LABA monotherapies (OR 1.26, 95% CrI 1.10 to 1.42; OR 1.28, 95% CrI 1.11 to 1.47). Table 50 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st – 2nd), with LAMA and LABA the worst ranked treatment groups.

42. Relative effects: SGRQ responders at 6 months in the low‐risk population.

Treatment comparison	Odds ratios: random‐effects
	Median	95% CrI
LABA/LAMA vs LABA/ICS	1.22	0.99 to 1.51
LABA/LAMA vs LAMA	1.26	1.10 to 1.42
LABA/LAMA vs LABA	1.28	1.11 to 1.47
LABA/ICS vs LAMA	1.03	0.83 to 1.27
LABA/ICS vs LABA	1.05	0.87 to 1.25
LAMA vs LABA	1.02	0.90 to 1.16

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

43. Mean and median ranks: St. George's Respiratory Questionnaire responders at 6 months in the low‐risk population.

Treatment group	Rank (from random‐effects model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 2
LABA/ICS	2.7	2	1 to 4
LAMA	3.0	3	2 to 4
LABA	3.3	3	2 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.2.2.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs across all comparisons for SGRQ responders at six months (Appendix 7). There is evidence to suggest that LABA/LAMA increases SGRQ responders at six months compared to both LAMA and LABA monotherapies (OR 1.26, 95% CI 1.15 to 1.37; OR 1.20, 95% CI 1.06 to 1.37). The certainty of evidence was moderate for LABA/LAMA versus LAMA and LABA/ICS versus LABA and low for LABA/LAMA versus LABA/ICS, LABA/LAMA versus LABA, and LAMA versus LABA. There was no direct comparison for LABA/ICS versus LAMA. There was no difference between random and fixed analyses.

2.2.3 Rank probabilities for SGRQ responders at three and six months

Figure 26 plots the ranks of SGRQ responders at three and six months for each treatment group. The vertical axis shows the probability of being ranked best, second best, third best, or worst treatment group. There is uncertainty as to the ranking of treatment groups at three months but LAMA is clearly ranked worst. LABA has the highest probability of being ranked first at three months but there is also a small probability that it is ranked third or last. At six months, LABA/LAMA has nearly 100% probability of being the best.

26.

Plot of rank probabilities for each treatment group for St George's Respiratory Questionnaire responders in the low‐risk population
 St George's Respiratory Questionnaire responders at 3 (solid line), and 6 months (dashed line), in the low‐risk population

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.2.4 Outcome: SGRQ responders at 12 months

2.2.4.1 Pairwise meta‐analyses

There is evidence to suggest LABA/ICS is associated with a significantly higher proportion in SGRQ responders at 12 months compared to LABA (OR 1.42, 95% CI 1.18 to 1.70; moderate‐certainty evidence). There was no direct comparison for LABA/LAMA versus LABA/ICS and LABA/ICS versus LAMA. There is no evidence of significant differences for LABA/LAMA versus LAMA or LABA (moderate‐certainty evidence), and LAMA versus LABA (low‐certainty evidence; Appendix 7).

2.3 Outcome: change from baseline in SGRQ score

2.3.1 Outcome: change from baseline in SGRQ score at three months

We included 28 studies of 19 interventions and four treatment groups for this outcome (Appendix 3; Figure 27 a and b). Note that interventions formoterol 4.5 μg twice daily, formoterol 9 μg twice daily, glycopyrronium 15.6 μg twice daily, tiotropium 5 μg once daily, indacaterol/glycopyrronium 27.5 μg/15.6 μg twice daily, and olodaterol/tiotropium 5 μg/5 μg once daily are disconnected from the main treatment network (Figure 27a), but we included them in a class/group.

27.

Change from baseline in SGRQ score at 3 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 0 favour the first named treatment group.

FM: formoterol; Glyco: glycopyrronium; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; Olo: olodaterol; Tio: tiotropium

2.3.1.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.3.1.2 NMA results

The NMA included a total of 20,594 participants (LABA: 3933, LAMA: 7849, LABA/ICS: 2396, LABA/LAMA: 6416). Figure 27d and Table 51 show the mean difference in change from baseline in SGRQ score at three months for each treatment group compared to every other. There is evidence to suggest that both LABA/LAMA and LABA/ICS improve SGRQ score at three months compared to LAMA (MD −1.64, 95% CrI −2.2 to −1.08; MD −1.68, 95% CrI −2.59 to −0.78), although the MDs do not reach the clinical significance of MCID of 4. There is no evidence of differences across the other comparisons. Table 52 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment groups are LABA/ICS and LABA/LAMA, both with a median rank of 2 (95% CrI 1st to 3rd).

44. Change from baseline in St. George's Respiratory Questionnaire score at 3 months in the low‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.04	−0.79 to 0.88	0.04	−0.84 to 0.88
LABA/LAMA vs LAMA	−1.64	−2.2 to −1.08	−1.64	−2.25 to −1.05
LABA/LAMA vs LABA	−0.63	−1.86 to 0.6	−0.62	−1.95 to 0.65
LABA/ICS vs LAMA	−1.68	−2.59 to −0.78	−1.68	−2.6 to −0.74
LABA/ICS vs LABA	−0.67	−1.88 to 0.54	−0.67	−1.92 to 0.57
LAMA vs LABA	1.01	−0.2 to 2.22	1.02	−0.26 to 2.27

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

45. Mean and median ranks: change from baseline in St. George's Respiratory Questionnaire score at 3 months in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/ICS	1.6	2	1 to 3
LABA/LAMA	1.7	2	1 to 3
LABA	2.8	3	1 to 4
LAMA	3.9	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.3.1.3 Pairwise meta‐analyses

There is evidence to suggest that LABA/LAMA improves SGRQ score at three months compared to LAMA (MD −1.60, 95% CI −2.19 to −1.01), and that LAMA improves the score compared to LABA (MD 1.84, 95% CI 0.87 to 2.80), but the mean differences do not reach the clinical significance of MCID of 4. There is no evidence of differences across the other comparisons, however, a clinically significant difference cannot be excluded favouring LABA/LAMA over LABA given its 95% CI crossing the line of MCID of 4 (MD −1.29, 95% CI −4.29, 1.71; Appendix 7). The certainty of evidence for LABA/ICS versus LAMA and LAMA versus LABA was moderate due to a suboptimal information size, which could explain discrepancies with the NMA results. Otherwise all other results were consistent with the NMAs. The certainty of evidence was moderate for LABA/LAMA versus LAMA or LABA and high for LABA/LAMA versus LABA/ICS and LABA/ICS versus LABA. There was no difference between random and fixed analyses.

2.3.2 Outcome: change from baseline in SGRQ score at six months

We included 20 studies of 17 interventions and four treatment groups for this outcome (Appendix 3; Figure 28 a and b).

28.

Change from baseline in St George's Respiratory Questionnaire score at 6 months in the low‐risk population.
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 0 favour the first named treatment group.

FM: formoterol; Glyco: glycopyrronium; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; Olo: olodaterol; Tio: tiotropium

2.3.2.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.3.2.2 NMA results

The NMA included a total of 16,508 participants (LABA: 4351, LAMA: 4454, LABA/ICS: 2880, LABA/LAMA: 4823). Figure 28d and Table 53 show the mean difference in change from baseline in SGRQ score at six months for each treatment group compared to every other. There is evidence to suggest that both LABA/LAMA and LABA/ICS reduce SGRQ score compared to LABA at six months (MD −1.36, 95% CrI −2.12 to −0.60; MD −1.14, 95% CrI −1.90 to −0.37), and that LABA/LAMA reduces SGRQ score compared to LAMA (MD −1.18, 95% CrI −1.80 to ‐0.56), although the differences do not reach the clinical significance of MCID of 4. Table 54 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 2nd).

46. Relative effects: change from baseline in SGRQ score at 6 months in the low‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	−0.22	−1.28 to 0.82	−0.3	−1.50 to 0.93
LABA/LAMA vs LAMA	−1.18	−1.80 to −0.56	−1.17	−1.91 to −0.48
LABA/LAMA vs LABA	−1.36	−2.12 to −0.6	−1.4	−2.24 to −0.51
LABA/ICS vs LAMA	−0.96	−1.98 to 0.09	−0.89	−2.08 to 0.33
LABA/ICS vs LABA	−1.14	−1.90 to −0.37	−1.11	−2.01 to −0.16
LAMA vs LABA	−0.18	−0.91 to 0.55	−0.21	−1.05 to 0.61

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

47. Mean and median ranks: St. George's Respiratory Questionnaire at 6 months in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.3	1	1 to 2
LABA/ICS	1.7	2	1 to 3
LAMA	3.3	3	2 to 4
LABA	3.7	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.3.2.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs and there is no evidence of clinically significant improvement in SGRQ score at six months (MCID of 4 or greater), with any treatment group compared to the others (Appendix 7). There were no data available for LABA/ICS versus LAMA. The certainty of evidence was high for LAMA versus LABA, moderate for LABA/LAMA versus LAMA or LABA and LABA/ICS versus LABA, and low for LABA/LAMA versus LABA/ICS. There was no difference between random and fixed analyses.

2.3.3 Outcome: change from baseline in SGRQ score at 12 months

We included six studies of 10 interventions and four treatment groups for this outcome (Appendix 3; Figure 29 a and b). Note that interventions salmeterol 50 μg twice daily and salmeterol/fluticasone 50 μg/500 μg twice daily are disconnected from the main treatment network (Figure 29a), but we included them in a class/group model.

29.

Change from baseline in SGRQ score at 12 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 0 favour the first named treatment group.

FP: fluticasone propionate; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAL: salmeterol

2.3.3.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.3.3.2 NMA results

The NMA included a total of 6849 participants (LABA: 2021, LAMA: 2163, LABA/ICS: 873, LABA/LAMA: 1792). Figure 29d and Table 55 show the mean difference in change from baseline in SGRQ score at 12 months for each treatment group compared to every other. There is some evidence to suggest that LABA/ICS improves SGRQ score at 12 months compared to LABA using the fixed‐effect model (MD −1.69, 95% CrI −2.81 to −0.57). Both LABA/LAMA and LABA/ICS showed a reduction in SGRQ score compared to LAMA when using the fixed effect model (MD −0.89, 95% CrI −1.66 to −0.11) and MD −1.85, 95% CrI −3.28 to −0.43). Increased uncertainty in the random‐effects model leads to inconclusive results and the mean differences do not reach the clinical significance of MCID of 4. Table 56 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/ICS with a median rank of 1 (95% CrI 1st to 2nd).

48. Relative effects: change from baseline in St. George's Respiratory Questionnaire score at 12 months in the low‐risk population.

Treatment comparison	Mean differences: fixed‐effect		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.97	−0.48 to 2.42	1.05	−1.78 to 3.98
LABA/LAMA vs LAMA	−0.89	−1.66 to −0.11	−0.8	−2.05 to 0.62
LABA/LAMA vs LABA	−0.72	−1.64 to 0.20	−0.65	−2.29 to 1.11
LABA/ICS vs LAMA	−1.85	−3.28 to −0.43	−1.86	−4.63 to 1.02
LABA/ICS vs LABA	−1.69	−2.81 to −0.57	−1.71	−4.02 to 0.65
LAMA vs LABA	0.16	−0.72 to 1.04	0.13	−1.48 to 1.74

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

49. Mean and median ranks: change from baseline in St. George's Respiratory Questionnaire score at 12 months in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/ICS	1.1	1	1 to 2
LABA/LAMA	2.0	2	1 to 3
LABA	3.3	3	2 to 4
LAMA	3.6	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.3.3.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs and there is no evidence that any treatment group is associated with clinically significant improvement in SGRQ score at 12 months compared to the others (Appendix 7). The certainty of evidence was high for LABA/LAMA versus LABA and LAMA versus LABA, moderate for LABA/ICS versus LABA, and very low for LABA/LAMA versus LAMA. There was no direct comparison for LABA/LAMA versus LABA/ICS and LABA/ICS versus LAMA. There was no difference between random and fixed analyses.

2.3.4 Rank probabilities for change from baseline in SGRQ score

Figure 30 plots the ranks of SGRQ score at 3, 6 and 12 months for each treatment group. The vertical axis shows the probability of being ranked best, second best, third best, or worst treatment group. LABA and LAMA have a high probability of ranking 3rd or 4th at all time points whereas LABA/ICS has a high probability of being the best at 12 months.

30.

Plot of rank probabilities for each treatment group
 Change from baseline in St George's Respiratory Questionnaire score at 3 (solid line), 6 (dashed line), and 12 months (dotted line), in the low‐risk population

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4 Outcome: transitional dyspnoea index (TDI)

2.4.1 Outcome: TDI at three months

We included 30 studies of 19 interventions and four treatment groups for this outcome (Appendix 3; Figure 31 a and b). Note that interventions glycopyrronium 15.6 μg twice daily and indacaterol/glycopyrronium 27.5 μg/15.6 μg twice daily are disconnected from the main treatment network (Figure 31a), but we included them in a class/group model.

31.

Transition Dyspnea Index at 3 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Positive values favour the first named treatment group.

FM: formoterol; Glyco: glycopyrronium; ICS: inhaled corticosteroid; IND: indacaterol; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; Olo: olodaterol; Tio: tiotropium

2.4.1.1 Model selection and inconsistency checking

We chose a random‐treatment‐effects model with fixed‐class effects, assuming consistency. We also report results for a fixed‐treatment‐effect model with random‐class effects for comparison (Appendix 4).

2.4.1.2 NMA results

The NMA included a total of 21,750 participants (LABA: 5113, LAMA: 7046, LABA/ICS: 2838, LABA/LAMA: 6753). Figure 31d and Table 57 show the mean difference in TDI score at three months for each treatment group compared to every other, using the two models. There is evidence to suggest that LABA/LAMA increases TDI at three months compared to all other treatment groups (MD 0.35, 95% CrI 0.12 to 0.56; MD 0.54, 95% CrI 0.36 to 0.73; MD 0.44, 95% CrI 0.20 to 0.67 against LABA/ICS, LAMA and LABA), although the MDs do not reach the clinical significance of MCID of 1. There is no evidence of differences across the other treatment groups using the model with random‐treatment and fixed‐class effects. Table 58 shows the rank statistics for the four treatment groups (sorted by mean rank) for the preferred model. The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st).

50. Relative effects: Transition Dyspnea Index at 3 months in the low‐risk population.

Treatment comparison	Mean differences: random‐effects (fixed‐class)		Mean differences: fixed‐effect (random‐class)
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.35	0.12 to 0.56	0.48	0.09 to 0.99
LABA/LAMA vs LAMA	0.54	0.36 to 0.73	0.55	0.22 to 0.90
LABA/LAMA vs LABA	0.44	0.20 to 0.67	0.47	0.09 to 0.85
LABA/ICS vs LAMA	0.19	−0.07 to 0.47	0.06	−0.43 to 0.48
LABA/ICS vs LABA	0.09	−0.18 to 0.36	−0.02	−0.48 to 0.37
LAMA vs LABA	−0.1	−0.36 to 0.14	−0.08	−0.46 to 0.28

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

51. Median and mean ranks: Transition Dyspnea Index at 3 months in the low‐risk population.

Treatment group	Rank (from random‐effects, fixed‐class)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LABA/ICS	2.3	2	2 to 4
LABA	3.0	3	2 to 4
LAMA	3.7	4	2 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.1.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs and there is no evidence that any treatment group is associated with clinically significant improvement in TDI at three months (MCID of 1), compared to the others, despite a significant difference in some comparisons (Appendix 7). The certainty of evidence was high for LABA/ICS versus LABA, moderate for LABA/LAMA versus LAMA, low for LABA/LAMA versus LABA/ICS or LABA, and very low for LABA/ICS versus LAMA. There was no difference between random and fixed analyses.

2.4.2 Outcome: TDI at six months

We included 18 studies of 16 interventions and four treatment groups for this outcome (Appendix 3; Figure 32 a and b).

32.

Transition Dyspnea Index at 6 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Positive values favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.2.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.4.2.2 NMA results

The NMA included a total of 14,315 participants (LABA: 3878, LAMA: 3977, LABA/ICS: 1825, LABA/LAMA: 4635). Figure 32d and Table 59 show the mean difference in TDI score at six months for each treatment group compared to every other. There is evidence to suggest that LABA/LAMA increases TDI at six months compared to LAMA and LABA monotherapies (MD 0.33, 95% CrI 0.18 to 0.47; MD 0.37, 95% CrI 0.21, 0.52), although the MDs do not reach the clinical significance of MCID of 1. There is no evidence of differences across the other comparisons. Table 60 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group is LABA/LAMA with a median rank of 1 (95% CrI 1st to 2nd).

52. Relative effects: Transition Dyspnea Index at 6 months in the low‐risk population.

Treatment comparison	Mean differences: random‐effects (fixed‐class)		Mean differences: fixed‐effect (random‐class)
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.15	−0.10 to 0.4	0.14	−0.14 to 0.41
LABA/LAMA vs LAMA	0.33	0.18 to 0.47	0.32	0.15 to 0.48
LABA/LAMA vs LABA	0.37	0.21 to 0.52	0.36	0.18 to 0.55
LABA/ICS vs LAMA	0.18	−0.09 to 0.45	0.18	−0.12 to 0.50
LABA/ICS vs LABA	0.22	−0.02 to 0.46	0.22	−0.04 to 0.50
LAMA vs LABA	0.04	−0.12 to 0.21	0.04	−0.15 to 0.24

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

53. Mean and median ranks: Transition Dyspnea Index at 6 months in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.1	1	1 to 2
LABA/ICS	2.0	2	1 to 4
LAMA	3.2	3	2 to 4
LABA	3.6	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.2.3 Pairwise meta‐analyses

There was no direct comparison for LABA/ICS versus LAMA. Otherwise, the results from pairwise MAs were consistent with the NMAs and there is no evidence that any treatment group is associated with clinically significant improvement in TDI at six months (MCID of 1), compared to the others (Appendix 7). The certainty of evidence was high for LABA/LAMA versus LABA/ICS and LABA/ICS versus LABA, moderate for LABA/LAMA versus LAMA or LABA, and low for LAMA versus LABA. There was no difference between random and fixed analyses .

2.4.3 Outcome: TDI at 12 months

We included six studies of 10 interventions and three treatment groups for this outcome (Appendix 3; Figure 33 a and b).

33.

Transition Dyspnea Index at 12 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Positive values favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.3.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.4.3.2 NMA results

The NMA included a total of 38,861 participants (LABA: 3908, LAMA: 32,624, LABA/ICS: 0, LABA/LAMA: 2329). Figure 33d and Table 61 show the mean difference in TDI score at 12 months for each treatment group compared to every other. There is evidence to suggest that LABA/LAMA increases TDI at 12 months compared to LAMA and LABA monotherapies (MD 0.20, 95% CrI 0.09 to 0.32; MD 0.30, 95% CrI 0.17 to 0.42). There is no evidence of differences across other comparisons. Table 62 shows the rank statistics for the three treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st).

54. Relative effects: Transition Dyspnea Index at 12 months in the low‐risk population.

Treatment comparison	Mean differences: random‐effects (fixed‐class)		Mean differences: fixed‐effect (random‐class)
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LAMA	0.20	0.09 to 0.32	0.22	−0.05 to 0.51
LABA/LAMA vs LABA	0.30	0.17 to 0.42	0.37	0.11 to 0.71
LAMA vs LABA	0.09	−0.02 to 0.21	0.15	−0.10 to 0.46

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

55. Mean and median ranks: Transition Dyspnea Index at 12 months in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.00	1	1 to 1
LAMA	2.06	2	2 to 3
LABA	2.94	3	2 to 3

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.3.3 Pairwise meta‐analyses

There was no direct comparison for LABA/LAMA versus LABA/ICS and LABA/ICS versus LAMA or LABA. Otherwise, the results from pairwise MAs were consistent with the NMAs and there is no evidence that any treatment group is associated with clinically significant improvement in TDI at 12 months (MCID of 1), compared to the others (Appendix 7). The certainty of evidence was high for LAMA versus LAMA, moderate for LABA/LAMA versus LAMA, and very low for LABA/LAMA versus LABA. There was no difference between random and fixed analyses.

2.4.4 Rank probabilities for TDI

Figure 34 plots the ranks of TDI score for each treatment group at three and six months only. Ranks at 12 months are not plotted as only three treatment groups were available for comparison. The vertical axis shows the probability of being ranked best, second best, third best, or worst treatment group. LABA/LAMA has the highest probability of being ranked first at six months and nearly 100% probability of being the best at three months. There is uncertainty in the ranking of the other interventions.

34.

Plot of rank probabilities for each treatment group for Transition Dyspnea Index
 Transition Dyspnea Index score at 3 and 6 months in the low‐risk population.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5 Outcome: change from baseline in forced expiratory volume in one second (FEV1)

2.5.1 Outcome: change from baseline in FEV1 at three months

We included 50 studies of 23 interventions and four treatment groups for this outcome (Appendix 3; Figure 35 a and b). Note that interventions indacaterol 75 μg once daily, glycopyrronium 15.6 μg twice daily and indacaterol/glycopyrronium 27.5/12.5 μg twice daily are disconnected from the main treatment network (Figure 35a), but we included them in a class/group model.

35.

Change from baseline in forced expiratory volume in 1 second at 3 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Positive values favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.1.1 Model selection and inconsistency checking

We chose a random‐treatment‐effects model with fixed‐class effects, assuming consistency (Appendix 4).

2.5.1.2 NMA results

The NMA included a total of 30,962 participants (LABA: 6725, LAMA: 9977, LABA/ICS: 6126, LABA/LAMA: 8134) Figure 35d and Table 63 show the mean difference in change from baseline in FEV1 at three months for each treatment group compared to every other. There is evidence to suggest that LABA/LAMA and LABA/ICS increase FEV1 at three months compared to LAMA (MD 0.08, 95% CrI 0.06 to 0.09; MD 0.02, 95% CrI 0 to 0.04), and LABA (MD 0.09, 95% CrI 0.07 to 0.11; 0.03 95% CrI 0.01 to 0.05), monotherapies and that LABA/LAMA improves FEV1 compared to LABA/ICS (MD 0.05, 95% CrI 0.03 to 0.07). The 95% CI exceeding MCID of 0.1 L suggests a possibility of clinically significant improvement favouring LABA/LAMA over LABA. Table 64 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st).

56. Relative effects: change from baseline in forced expiratory volume in 1 second at 3 months in the low‐risk population.

Treatment comparison	Mean differences: random‐effects
	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.05	0.03 to 0.07
LABA/LAMA vs LAMA	0.08	0.06 to 0.09
LABA/LAMA vs LABA	0.09	0.07 to 0.11
LABA/ICS vs LAMA	0.02	0.00 to 0.04
LABA/ICS vs LABA	0.03	0.01 to 0.05
LAMA vs LABA	0.01	−0.01 to 0.03

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

57. Mean and median ranks: change from baseline in forced expiratory volume in 1 second at 3 months in the low‐risk population.

Treatment group	Rank (from random‐effects model)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LABA/ICS	2.0	2	2 to 2
LAMA	3.2	3	3 to 4
LABA	3.8	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.1.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs and there is no evidence that any treatment group is associated with clinically significant improvement (MCID of 0.1 L) in change from baseline in FEV1 at three months compared to the others (Appendix 7). However, a clinically significant improvement in change from baseline in FEV1 at three months cannot be excluded favouring LABA/LAMA over LABA/ICS (MD 0.08, 95% CI 0.03 to 0.12; low‐certainty evidence), and LABA (MD 0.07, 95% CI 0.03 to 0.12; very low‐certainty evidence), given the 95% CI crossing the line of MCID of 0.1 L. Otherwise, the certainty of evidence was moderate for LABA/ICS versus LABA, low for LABA/LAMA versus LABA/ICS or LAMA, LABA/ICS versus LAMA, and LAMA versus LABA. There was no difference between random and fixed analyses except for LABA/ICS versus LAMA, in which the random‐effects model had a wider 95% CI containing the line of no difference (MD 0.02, 95% CI −0.02 to 0.06).

2.5.2 Outcome: change from baseline in FEV1 at six months

We included 30 studies of 21 interventions and four treatment groups for this outcome (Appendix 3; Figure 36 a and b).

36.

Change from baseline in forced expiratory volume in 1 second at 6 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot (deviance points from the fixed‐effect model with random‐treatment‐group effect on the x‐axis and from the fixed‐effect inconsistency model with random‐class effect on the y‐axis); d. plot of relative effects. Positive values favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.2.1 Model selection and inconsistency checking

We chose a random‐treatment‐effects model with fixed‐class effects, assuming consistency. We also report results for a fixed‐treatment‐effect model with random‐class effects for comparison. However, there is weak evidence of potential inconsistency in this network and results should be interpreted with some caution (Appendix 4).

2.5.2.2 NMA results

The NMA included a total of 21,224 participants (LABA: 5959, LAMA: 6360, LABA/ICS: 2155, LABA/LAMA: 6750). Figure 36d and Table 65 show the mean difference in change from baseline in FEV1 at six months for each treatment group compared to every other. There is evidence to suggest that LABA/LAMA increases FEV1 at six months compared to all other treatment groups (MD 0.05, 95% CrI 0.03 to 0.08; MD 0.06, 95% CrI 0.05 to 0.08; MD 0.08, 95% CrI 0.06 to 0.09 against LABA/ICS, LAMA, and LABA respectively), and that LAMA slightly increases FEV1 compared to LABA (MD 0.01, 95% CrI 0.00 to 0.03), in the random‐effects‐model with fixed‐class effects although the mean differences do not reach the clinical significance of MCID of 0.1 L. Table 66 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 1st). Results are more uncertain when considering the fixed‐treatment‐effect model with random‐class effects.

58. Relative effects: change from baseline in forced expiratory volume in 1 second at 6 months in the low‐risk population.

Treatment comparison	Mean differences: random‐effects		Mean differences: fixed‐effect (random‐class)
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.05	0.03 to 0.08	0.05	−0.01 to 0.11
LABA/LAMA vs LAMA	0.06	0.05 to 0.08	0.06	0.02 to 0.09
LABA/LAMA vs LABA	0.08	0.06 to 0.09	0.08	0.04 to 0.11
LABA/ICS vs LAMA	0.01	−0.02 to 0.04	0.01	−0.05 to 0.07
LABA/ICS vs LABA	0.02	−0.01 to 0.05	0.03	−0.02 to 0.08
LAMA vs LABA	0.01	0.00 to 0.03	0.02	−0.01 to 0.05

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

59. Mean and median ranks: change from baseline in forced expiratory volume in 1 second at 6 months in the low‐risk population.

Treatment group	Rank (from random‐effects to fixed‐class)
	Mean	Median	95% CrI
LABA/LAMA	1.0	1	1 to 1
LABA/ICS	2.3	2	2 to 4
LAMA	2.7	3	2 to 4
LABA	3.9	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.2.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs except for LABA/ICS versus LABA in which LABA/ICS significantly increased FEV1 at six months compared to LABA (MD 0.04, 95% CI 0.01 to 0.07). There is no evidence of clinically significant improvement (MCID of 0.1 L or greater) with any treatment group compared to the others, except for LABA/LAMA versus LABA/ICS in which its 95% CI suggested a possibility of clinically significant difference favouring LABA/LAMA over LABA/ICS (MD 0.10, 95% CI 0.05 to 0.15; Appendix 7). The certainty of evidence was high for LABA/LAMA versus LABA/ICS and LABA/ICS versus LAMA, and moderate for LABA/LAMA versus LAMA and LABA/ICS versus LABA. There was no difference between random and fixed analyses.

2.5.3 Outcome: change from baseline in FEV1 at 12 months

We included 13 studies of 13 interventions and three treatment groups for this outcome (Appendix 3; Figure 37 a and b).

37.

Change from baseline in forced expiratory volume in 1 second at 12 months in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot (deviance points from the fixed‐effect model with random‐class effect on the x‐axis and from the fixed‐effect inconsistency model with random‐class effect on the y‐axis); d. plot of relative effects. Positive values favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.3.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with random‐class effects, assuming consistency. We also reported results for a random‐treatment‐effects model with fixed‐class effects for comparison. However, there is weak evidence of potential inconsistency in the latter model so results should be interpreted with caution (Appendix 4).

2.5.3.2 NMA results

The NMA included a total of 10,676 participants (LABA: 3577, LAMA: 4057, LABA/ICS: 0, LABA/LAMA: 3042). Figure 37d and Table 67 show the mean difference in change from baseline in FEV1 at 12 months for each treatment group compared to every other. There is evidence to suggest that LABA/LAMA increases FEV1 at 12 months compared to LABA (MD 0.08, 95% CrI 0.02 to 0.14). However there is high uncertainty in the results. Comparisons based on the random‐treatment‐effects model with fixed class are more precise with similar MDs. The 95% CI containing MCID of 0.1 L in both models (MD 0.08, 95% CrI 0.02 to 0.14 and MD 0.08, 95% CrI 0.06 to 0.1), suggests a possibility of clinically significant improvement favouring LABA/LAMA over LABA. Table 68 shows the rank statistics for the three treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/LAMA with a median rank of 1 (95% CrI 1st to 2nd).

60. Relative effects: change from baseline in forced expiratory volume in 1 second at 12 months in the low‐risk population.

Treatment comparison	Mean differences− fixed effects		Mean differences: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LAMA	0.06	−0.01 to 0.12	0.06	0.04 to 0.08
LABA/LAMA vs LABA	0.08	0.02 to 0.14	0.08	0.06 to 0.10
LAMA vs LABA	0.02	0.00 to 0.06	0.02	0.00 to 0.04

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

61. Mean and median ranks: change from baseline in forced expiratory volume in 1 second at 12 months in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/LAMA	1.1	1	1 to 2
LAMA	2.0	2	1 to 3
LABA	3.0	3	2 to 3

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

The random‐class effects model assumes that treatment effects within a class or group can vary. Table 69 reports the mean difference of each individual intervention compared to formoterol 9 to 12 μg twice daily. Tiotropium 18 μg once daily, tiotropium 5 μg once daily, and all the interventions in the LABA/LAMA group (formoterol/glycopyrronium 9.6 μg/18 μg twice daily, indacaterol/glycopyrronium 27.5 μg/15.6 μg twice daily, indacaterol/glycopyrronium 110 μg/50 μg once daily, olodaterol/tiotropium 5 μg/5 μg once daily and formoterol/aclidinium 12 μg/400 μg twice daily) showed an increase in FEV1 at 12 months compared to formoterol 9 to 12 μg twice daily.

62. Intervention effects: change from baseline in forced expiratory volume in 1 second at 12 months in the low‐risk population.

Intervention	Median	95% CrI
Formoterol 9−12 twice daily	Reference
Indacaterol 75 once daily	0.002	−0.029 to 0.048
Olodaterol 5 once daily	0.001	−0.018 to 0.022
Tiotripium 18 once daily	0.034	0.016 to 0.054
Tiotripium 5 once daily	0.031	0.009 to 0.056
Aclidinium 400 twice daily	0.027	−0.002 to 0.060
Glycopyrronium 15.6 twice daily	0.010	−0.006 to 0.027
Glycopyrronium 50 once daily	0.022	−0.022 to 0.062
Formoterol/glycopyrronium 9.6/18 twice daily	0.066	0.050 to 0.081
Indacaterol/glycopyrronium 27.5/15.6 twice daily	0.083	0.034 to 0.137
Indacaterol/glycopyrronium 110/50 once daily	0.128	0.091 to 0.165
Olodaterol/tiotropium 5/5 once daily	0.089	0.066 to 0.114
Formterol/aclidinium 12/400 twice daily	0.044	0.005 to 0.081

CrI: credible interval

2.5.3.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMA (the random‐treatment‐effects model with fixed classes), except for LAMA versus LABA, in which there was a significant improvement with LAMA compared to LABA (MD 0.02, 95% CI 0.01 to 0.03; Appendix 7). However, there is no evidence that any treatment group is associated with clinically significant improvement (MCID of 0.1 L), compared to the others (very low‐certainty evidence). Appendix 7 shows the certainty of evidence for the rest of the comparisons. There was no difference between random and fixed analyses.

2.5.4 Rank probabilities for change from baseline in FEV1

Figure 38 plots the ranks of each treatment group for FEV1 at three and six months only. We have not plotted ranks at 12 months, as only three treatment groups were available for comparison. The vertical axis shows the probability of being the best, second best, third best, or worst treatment group. LABA/LAMA has nearly 100% probability of being ranked first at three and six months, with LABA having a very high probability of being the worst intervention at three and six months.

38.

Plot of rank probabilities for each treatment group in change in forced expiratory volume in 1 second in the low‐risk population
 Change from baseline in forced expiratory volume in 1 second at 3 (solid line), and 6 months (dashed line).

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.6 Outcome: mortality

We included 51 studies of 27 interventions and four treatment groups for this outcome (Appendix 3; Figure 39 a and b).

39.

Mortality in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.6.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects for comparison. Results should be interpreted with some caution due to poor model fit, which can be attributed to studies with zero cells (Appendix 4).

2.6.2 NMA results

The NMA included a total of 56,493 participants (LABA: 11,488, LAMA: 25,324, LABA/ICS: 7586, LABA/LAMA: 12,095). The median duration of follow‐up was 24 weeks (range 12 to 156 weeks). Figure 39d and Table 70 show the OR of mortality for each treatment group compared to every other. There was no evidence to suggest that any treatment group increased or decreased the odds of mortality compared to any other.

63. Relative effects: mortality in the low‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	1.25	0.79 to 2.00	1.27	0.69 to 2.30
LABA/LAMA vs LAMA	0.91	0.63 to 1.32	0.90	0.59 to 1.34
LABA/LAMA vs LABA	1.16	0.75 to 1.81	1.19	0.73 to 1.98
LABA/ICS vs LAMA	0.73	0.45 to 1.16	0.72	0.37 to 1.30
LABA/ICS vs LABA	0.93	0.76 to 1.14	0.94	0.59 to 1.52
LAMA vs LABA	1.28	0.83 to 1.98	1.31	0.82 to 2.22

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

Table 71 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LABA/ICS with a median rank of 1 (95% CrI 1st to 4th), although the wide CrIs around the mean highlight the uncertainty in the results.

64. Mean and median ranks: mortality in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LABA/ICS	1.5	1	1 to 4
LABA	2.1	2	1 to 4
LABA/LAMA	3.0	3	1 to 4
LAMA	3.5	4	1 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.6.3 Pairwise meta‐analyses

The results from pairwise MAs were consistent with the NMAs and there is no evidence to suggest that any treatment group increased or decreased the odds of mortality compared to any other (Appendix 7). The certainty of evidence was moderate for all comparisons. There was no difference between random and fixed analyses.

2.7 Outcome: serious adverse events (SAEs)

SAEs were separated into total SAEs, COPD SAEs and cardiac SAEs.

2.7.1 Outcome: total SAEs

The analysis for total SAEs included 67 studies of 30 interventions and four treatment groups. We included a total of 64,855 participants (LABA: 13,703, LAMA: 27,712, LABA/ICS: 8609, LABA/LAMA: 14,831; Appendix 3, Figure 40 a and b). The median duration of follow‐up was 24 weeks (range 12 to 156 weeks).

40.

Total serious adverse events in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.7.1.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison (Appendix 4).

2.7.2 Outcome: COPD SAEs

The analysis for COPD SAEs included 63 studies of 30 interventions and four treatment groups (Appendix 3; Figure 41 a and b). We included a total of 61,759 participants (LABA: 12,981, LAMA: 27,819, LABA/ICS: 7971, LABA/LAMA: 12,988)

41.

Chronic obstructive pulmonary disease serious adverse events in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.7.2.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison. Results should be interpreted with some caution due to poor model fit, which can be attributed to studies with zero cells (Appendix 4).

2.7.3 Outcome: cardiac SAEs

The analysis for cardiac SAEs included 58 studies of 29 interventions and four treatment groups (Appendix 3; Figure 42 a and b). We included a total of 62,007 participants (LABA: 12,581, LAMA: 24,747, LABA/ICS: 10,303, LABA/LAMA: 14,376).

42.

Cardiac serious adverse events in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.7.3.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison. Results should be interpreted with some caution due to poor model fit, which can be attributed to studies with zero cells.

2.7.4 NMA results

Table 72 shows the OR of each type of adverse event for each treatment group compared to every other. For total SAEs there was evidence of an increase in the odds of an event for LABA/ICS compared to LABA (OR 1.13, 95% CrI 1.01 to 1.27), although only if we used the fixed‐effect model. For cardiac and COPD SAEs, there was no evidence that any treatment group increases or decreases the odds of an event compared to any other.

65. Relative effects: serious adverse events in the low‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
Total SAEs
LABA/LAMA vs LABA/ICS	0.91	0.78 to 1.05	0.91	0.77 to 1.06
LABA/LAMA vs LAMA	1.03	0.93 to 1.15	1.03	0.92 to 1.16
LABA/LAMA vs LABA	1.02	0.91 to 1.15	1.02	0.90 to 1.16
LABA/ICS vs LAMA	1.14	0.98 to 1.32	1.14	0.97 to 1.35
LABA/ICS vs LABA	1.13	1.01 to 1.27	1.13	0.99 to 1.29
LAMA vs LABA	0.99	0.88 to 1.11	0.99	0.87 to 1.12
COPD SAEs
LABA/LAMA vs LABA/ICS	0.96	0.75 to 1.22	0.92	0.67 to 1.26
LABA/LAMA vs LAMA	0.99	0.82 to 1.19	0.98	0.78 to 1.21
LABA/LAMA vs LABA	0.92	0.75 to 1.13	0.89	0.68 to 1.13
LABA/ICS vs LAMA	1.04	0.81 to 1.32	1.06	0.77 to 1.48
LABA/ICS vs LABA	0.96	0.82 to 1.13	0.96	0.73 to 1.25
LAMA vs LABA	0.93	0.76 to 1.14	0.9	0.71 to 1.14
Cardiac SAEs
LABA/LAMA vs LABA/ICS	1.28	0.91 to 1.81	1.24	0.81 to 1.83
LABA/LAMA vs LAMA	1.05	0.80 to 1.36	1.04	0.77 to 1.37
LABA/LAMA vs LABA	1.24	0.92 to 1.68	1.24	0.89 to 1.71
LABA/ICS vs LAMA	0.82	0.58 to 1.15	0.84	0.56 to 1.27
LABA/ICS vs LABA	0.97	0.79 to 1.19	0.99	0.74 to 1.41
LAMA vs LABA	1.19	0.89 to 1.59	1.19	0.88 to 1.64

COPD: chronic obstructive pulmonary disease; CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAE: serious adverse event

2.7.5 Pairwise meta‐analyses

There is no evidence to suggest that any treatment group increases or decreases the odds of an event compared to the others with pairwise MAs. The results were consistent with the NMAs except for LABA/ICS versus LABA, in which LABA/ICS was associated with a significant increase in total SAEs compared to LABA with the fixed‐effect NMA but not with the pairwise MAs or random‐effects NMA (Appendix 7; Table 72). Table 73 shows the certainty of evidence for each treatment group compared to every other. There was no difference between random and fixed analyses.

66. Certainty of evidence: serious adverse events in the low‐risk population.

Treatment comparison	Total SAEs	COPD SAEs	Cardiac SAEs
LABA/LAMA vs LABA/ICS	Moderate	Low	Moderate
LABA/LAMA vs LAMA	High	High	Moderate
LABA/LAMA vs LABA	High	Moderate	Moderate
LABA/ICS vs LAMA	Moderate	Moderate	Moderate
LABA/ICS vs LABA	Low	High	High
LAMA vs LABA	High	Low	Moderate

COPD: chronic obstructive pulmonary disease; CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; SAE: serious adverse event

2.8 Outcome: dropouts due to serious adverse events (SAEs)

We included 65 studies of 29 interventions and four treatment groups for this outcome (Appendix 3; Figure 43 a and b).

43.

Dropouts due to adverse events in the low‐risk population.
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot; d: plot of relative effects. Values less than 1 favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.8.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on the random‐treatment‐effects model with fixed‐class effects for comparison. Results should be interpreted with some caution due to poor model fit (Appendix 4).

2.8.2 NMA results

The NMA included a total of 62,831 participants (LABA: 13,074, LAMA: 27,155, LABA/ICS: 8394, LABA/LAMA: 14,208). The median duration of follow‐up was 24 weeks (range 12 to 156 weeks). Figure 43d and Table 74 show the OR of dropouts due to adverse events for each treatment group compared to every other. There was no evidence to suggest that any treatment group increased or decreased the odds of dropout compared to any other except for LAMA versus LABA (OR 0.84, 95% CrI 0.72 to 0.97). Table 75 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LAMA with a median rank of 1 (95% CrIs 1st to 3rd), although the wide CrIs around the mean highlight the uncertainty in the results.

67. Relative effects: dropouts due to adverse events in the low‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.99	0.83 to 1.18	0.99	0.82 to 1.2
LABA/LAMA vs LAMA	1.09	0.95 to 1.26	1.09	0.94 to 1.28
LABA/LAMA vs LABA	0.91	0.78 to 1.06	0.91	0.77 to 1.07
LABA/ICS vs LAMA	1.11	0.92 to 1.33	1.11	0.89 to 1.37
LABA/ICS vs LABA	0.92	0.8 to 1.06	0.92	0.77 to 1.09
LAMA vs LABA	0.84	0.72 to 0.97	0.83	0.7 to 0.98

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

68. Mean and median ranks: dropouts due to adverse events in the low‐risk population.

Treatment group	Rank (from fixed‐effect model)
	Mean	Median	95% CrI
LAMA	1.3	1	1 to 3
LABA/ICS	2.5	3	1 to 4
LABA/LAMA	2.5	2	1 to 4
LABA	3.7	4	2 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.8.3 Pairwise meta‐analyses

There is no evidence to suggest that any treatment group increases or decreases the odds of an event compared to the others with pairwise MAs. The results were consistent with the NMAs except for LAMA versus LABA, in which LAMA was associated with a significant decrease in dropouts due to adverse events compared to LABA in the NMA (OR 0.84, 95% CrI 0.72 to 0.97), but not in the pairwise MA (OR 0.90, 95% CI 0.73 to 1.10; Appendix 7). The certainty of evidence was moderate for LABA/ICS or LAMA versus LABA, low for LABA/LAMA versus LABA/ICS or LAMA and LABA/ICS versus LAMA, and very low for LABA/LAMA versus LABA. There was no difference between random and fixed analyses.

2.9 Outcome: pneumonia

We included 61 studies of 29 interventions and four treatment groups for this outcome (Appendix 3; Figure 44 a and b).

44.

Pneumonia in the low‐risk population
 a: network diagram of interventions; b: network diagram of treatment groups; c: deviance plot (deviance points from the fixed‐effect model with fixed‐class effect and from the fixed‐effect inconsistency model with fixed‐class effect); d: plot of relative effects. Values less than 1 favour the first named treatment group.

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.9.1 Model selection and inconsistency checking

We chose a fixed‐treatment‐effect model with fixed‐class effects, assuming consistency. We also report results based on a random‐treatment‐effects model with fixed‐class effects and informative prior distribution on the heterogeneity parameter for comparison. Results should be interpreted with caution due to potential inconsistency in the data (Appendix 4).

2.9.2 NMA results

The NMA included a total of 61,157 participants (LABA: 12,640, LAMA: 26,596, LABA/ICS: 7518, LABA/LAMA: 14,403). The median duration of follow‐up was 24 weeks (range 12 to 156 weeks). Figure 44d and Table 76 show the OR of pneumonia for each treatment group compared to every other. There is evidence to suggest that LABA/ICS increases the odds of pneumonia compared to LAMA and LABA (OR 2.02, 95% CrI 1.16 to 3.72; OR 1.93, 95% CrI 1.29 to 3.22), but no evidence of differences across other comparisons (Appendix 7; Table 7). Table 77 shows the rank statistics for the four treatment groups (sorted by mean rank). The highest ranked treatment group was LAMA with a median rank of 1 (95% CrI 1st to 3rd), although note the uncertainty in all the rankings.

69. Relative effects: pneumonia in the low‐risk population.

Treatment comparison	Odds ratios: fixed‐effect		Odds ratios: random‐effects
	Median	95% CrI	Median	95% CrI
LABA/LAMA vs LABA/ICS	0.67	0.44 to 1.01	0.61	0.34 to 1.01
LABA/LAMA vs LAMA	1.24	0.87 to 1.77	1.23	0.82 to 1.84
LABA/LAMA vs LABA	1.21	0.83 to 1.77	1.18	0.75 to 1.81
LABA/ICS vs LAMA	1.87	1.21 to 2.91	2.02	1.16 to 3.72
LABA/ICS vs LABA	1.82	1.41 to 2.36	1.93	1.29 to 3.22
LAMA vs LABA	0.97	0.66 to 1.44	0.96	0.62 to 1.49

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

70. Mean and median ranks: pneumonia in the low‐risk population.

Treatment group	Rank (from random‐effects model)
	Mean	Median	95% CrI
LAMA	1.6	1	1 to 3
LABA	1.8	2	1 to 3
LABA/LAMA	2.7	3	1 to 4
LABA/ICS	4.0	4	3 to 4

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.9.3 Clinical homogeneity assessment

Table 13 shows the clinical homogeneity assessment across the available comparisons. Pre‐bronchodilator baseline FEV1 ranged from 1.14 L to 1.34 L. The comparisons of LABA/ICS versus monotherapies had a lower baseline FEV1 compared with those of LABA/LAMA versus monotherapies, which could have introduced a bias against LABA/ICS. The NMA results should be interpreted with caution because of the difference in the baseline FEV1 across the pairwise comparisons.

2.9.4 Pairwise meta‐analyses

The results from pairwise MAs suggest that LABA/ICS increases the odds of pneumonia compared to LABA/LAMA and LABA (OR 2.33, 95% CI 1.03 to 5.26; OR 1.64, 95% CI 1.25 to 2.14). The difference was significant for LABA/LAMA versus LABA/ICS with the pairwise MAs (moderate‐certainty evidence), but not with the NMAs, and significant for LABA/ICS versus LAMA (OR 2.02, 95% CrI 1.16 to 3.72), with the NMA but not with the pairwise MA (OR 5.82, 95% CI 0.70 to 48.80; low‐certainty evidence; Appendix 7). The certainty of evidence was high for LABA/ICS versus LABA, moderate for LABA/LAMA versus LAMA or LABA, and LAMA versus LABA (see 'Summary of findings' tables). The aforementioned difference in the baseline FEV1 across the pairwise comparisons may have affected the NMA results. There was no difference between random and fixed analyses.

Discussion

Summary of main results

We assumed a class/group effect in all treatment groups because the random‐class‐effects model did not significantly improve model fit compared to the fixed‐class‐effects model except for change from baseline in FEV1 at 12 months in the low‐risk population, which argues against intraclass/group differences in any of the treatment groups we analysed. We have summarised the results in Appendix 6, Appendix 7, and Appendix 5.

The NMAs suggested that LABA/LAMA combination was the highest ranked treatment group to reduce moderate to severe and severe exacerbations, followed by LAMA. There is evidence that LABA/LAMA significantly reduces moderate to severe exacerbations compared to all others, and severe exacerbations compared to LABA/ICS and LABA in the high‐risk population.

The LABA/ICS combination was ranked third for moderate to severe exacerbations and severe exacerbations in the high‐risk population and ranked fourth for the severe exacerbations in the low‐risk population. LABA was the worst ranked, except for severe exacerbations in the low‐risk population, for which they were ranked third.

In the pairwise MAs, there was no definite evidence that LABA/LAMA or LAMA reduced moderate to severe or severe exacerbations compared to LABA/ICS in both populations, although a clinically meaningful reduction could not be excluded due to a wide 95% CI.

With regard to symptom and quality‐of‐life scores, the combination therapies, LABA/LAMA and LABA/ICS were generally ranked higher than monotherapies in both populations. LAMA/LABA was ranked higher than LABA/ICS in the high‐risk population. There were significant overlaps in the rank statistics between LABA/LAMA and LABA/ICS as well as between LAMA and LABA in the low‐risk population.

In the high‐risk population of pairwise MAs, the LABA/LAMA combination significantly increased SGRQ responders compared to LAMA at six months, LABA/ICS at 12 months, and LAMA at 12 months (Appendix 6).

In the low‐risk population of pairwise MAs, the LABA/LAMA combination significantly increased SGRQ responders compared to LAMA at three and six months and LABA at six months (Appendix 7).

The LABA/ICS combination significantly increased SGRQ responders compared to LABA at 12 months and the odds ratio of SGRQ response was significantly lower with LAMA compared to LABA at three months. Otherwise, none of the differences in symptom and quality‐of‐life scores met the MCID criteria of clinical significance in either high‐ or low‐risk populations.

The LABA/ICS combination was the lowest ranked in pneumonia SAEs in the high‐ and low‐risk populations. In the high‐risk population, LABA/ICS significantly increased the odds of pneumonia compared to LAMA/LABA, LAMA, and LABA both in the NMA and pairwise MAs. In the low‐risk population, LABA/ICS increased the odds of pneumonia compared to LAMA and LABA in the NMA and compared to LABA/LAMA and LABA in the pairwise MAs.

There were significant overlaps in the rank statistics in the other safety outcomes. LABA/ICS significantly increased total SAEs compared to LABA, and LAMA significantly reduced COPD SAEs compared to LABA, both in the NMAs and pairwise MAs. In the low‐risk population, LABA/ICS significantly increased total SAEs and LAMA significantly reduced dropouts due to adverse events compared to LABA in the NMAs but not in the pairwise MAs. Otherwise, there was no evidence to suggest that any treatment group increased the odds of SAEs or dropout compared to the others.

With regard to pre‐bronchodilator FEV1, the highest ranked treatment group was LABA/LAMA with a median rank of 1 whereas LABA was the worst ranked with a median of 4 at all time points. LABA/ICS and LAMA were ranked second or third. In the pairwise MAs, a significant difference was seen in some comparisons but the 95% CIs crossed the line of MCID of 0.1 L, suggesting none of the differences was clinically meaningful.

Overall completeness and applicability of evidence

The study results are not applicable to those with a milder form of COPD because people with mild COPD do not usually require a maintenance inhaler therapy and we did not include them in our analysis.

We also excluded people with asthma, although the baseline bronchodilator response was quite significant in some studies despite the exclusion (Table 8). It is unclear whether efficacies of ICS/LABA would be different in people without a history of asthma but with a significant bronchodilator response, which is usually seen in a more severe form of the disease. Cardiac SAEs could have been underestimated due to the exclusion of people with a significant cardiovascular comorbidity in a majority of included studies.

We excluded drug formulations or doses that were not approved or available for clinical use, as well as nebulised medications. Therefore, the results are not applicable for nebulised or off‐label use of available medications.

Otherwise, we included a total of 101,311 participants from 99 studies from across the world to be as comprehensive as possible. We used a Bayesian shared parameter model for COPD exacerbations and were able to avoid losing a substantial amount of relevant data (e.g. 6 out of 13 studies in severe exacerbations in the high‐risk population). We were able to collect a substantial amount of data from manufacturers' websites and ClinicalTrials.gov due to greater transparency from pharmaceutical companies.

Quality of the evidence

All included studies were RCTs, and the quality of included RCTs was generally good (Figure 2). Nineteen studies had an open tiotropium arm and 16 studies had relatively uneven dropouts. The results were unchanged in most of comparisons when we excluded those studies one by one or all together in the pairwise analyses. Otherwise, we downgraded the certainty rating by one or even two levels in some comparisons.

We had a total of 189 head‐to‐head comparisons in the pairwise MAs and the certainty of evidence was high, moderate, low and very low in 40, 99, 39, and 11 comparisons respectively. The primary reason for downgrading was a suboptimal information size or a wide 95% CI. Our confidence in the findings increased when the NMAs supported the pairwise results with a much greater information size. The results should be interpreted with caution for those derived from a small sample size or with low or very low certainty of evidence, or both (see 'Summary of findings' tables; Appendix 6; Appendix 7).

We found no evidence of inconsistency or effect modifiers when we compared the model fit and between‐study heterogeneity from NMA models with those from an unrelated effects (inconsistency) model except for mortality in the high‐risk population, as well as in change from baseline in FEV1 at six months, cardiac SAEs, and pneumonia in the low‐risk population.

The results from the NMAs and pairwise MAs were consistent, which would make significant inconsistency less likely except for pneumonia in the low‐risk population (Appendix 6; Appendix 7).

The mean baseline FEV1 of between‐treatment group comparisons for pneumonia in the low‐risk population, ranged from 1.14 L to 1.34 L (Table 13), which could be a potential effect modifier and possibly explain the inconsistency in this outcome. Therefore the NMA results of this outcome should be interpreted cautiously and in relation to the results from direct comparisons.

Potential biases in the review process

Incorporating indirect comparisons increases information size and statistical power. However it could introduce bias if there is a difference in participants, co‐interventions, or trial methodology between contrasts in a network (intransitivity), which is an inherent issue to a NMA. We took several measures to assess and minimise intransitivity.

1. We reviewed the study population after the first draft of our protocol and divided the entire population into high‐ and low‐risk populations because we thought such differences in the study population could introduce intransitivity. We acknowledge that blood eosinophil counts could be an effect modifier for LABA/ICS but available data were insufficient to include them as a covariate as a way of exploring subgroup effects.

2. We constructed summary tables organised by treatment group pair‐wise comparisons (Table 9; Table 10; Table 11; Table 12; Table 13), for the primary outcomes in both populations and also in pneumonia in the low‐risk population to assess clinical and methodological similarities/dissimilarities of the studies.

3. We performed NMAs and pairwise MAs to address possible intransitivity when there was a discrepancy between them (Appendix 6; Appendix 7).

4. We analysed several outcomes at different time points (e.g. 3, 6, and 12 months), when feasible.

5. We assessed consistency using the inconsistency models, acknowledged a possibility of intransitivity when suspected, and interpreted the results accordingly.

Agreements and disagreements with other studies or reviews

There are an increasing number of systematic reviews comparing LAMA/LABA with existing maintenance inhalers (Farne 2015; Oba 2016a; Oba 2016b). Our results are essentially similar to the existing reports but there are some differences in data collection and interpretations of the results.

Chen 2017 concluded that, “LAMA were associated with a greater reduction in acute exacerbations and fewer adverse effects compared with LABA.” They analysed all severities of exacerbation (mild, moderate, and severe), and adverse event (serious and non‐serious), including vilanterol, which was not approved or available for clinical use whereas our study analysed moderate to severe and severe exacerbations and SAEs (i.e. serious only), excluding vilanterol, which would be of greater clinical relevance in our opinion.

Horita 2017 reported “LAMA+LABA has fewer exacerbations… And more frequent improvement in quality of life as measured by an increase over 4 units or more of the SGRQ” compared to LABA/ICS. They included all severities of COPD exacerbation and analysed SGRQ responders at all time points combined together whereas we separated out moderate to severe and severe exacerbations and assessed SGRQ responders at different time points because previous reports suggested that a proportion of SGRQ responders changed over time after study entry.

Kew 2014 compared LABA/ICS, LAMA, LABA, and placebo, and concluded, “Quality of life and lung function were improved most on combination inhalers (LABA and ICS) and least on ICS alone at 6 and at 12 months.” We did not include ICS because it is now not commonly used as monotherapy in COPD and emphasised clinical significance/insignificance of the reported differences based on the recommended MCIDs.

Rodrigo 2017 concluded “The greater efficacy and comparable safety profiles observed with LABA/LAMA combinations versus LAMA or LABA/ICS” and “LABA/LAMA significantly reduced moderate/severe exacerbation rate compared with LABA/ICS”, which was based on two studies. Our pairwise analyses included seven studies for moderate to severe exacerbations (one in the high‐risk and six in the low‐risk populations) and five studies for severe exacerbations (one in the high‐risk and four in the low‐risk populations). In addition, we performed NMAs with much greater statistical power and addressed uncertainty surrounding these outcomes, taking effect modifiers into consideration.

Schlueter 2016 concluded “All LAMA/LABA FDCs were found to have similar efficacy and safety”, which agrees with our results. We examined a class/group effect not only in LABA/LAMA combinations but also in LABA/ICS combinations, LAMAs, and LABAs.

Welsh 2013 compared LABA/ICS versus tiotropium (LAMA), and concluded, “The relative efficacy and safety of combined inhalers and tiotropium remains uncertain” because of missing outcome data. We examined the proportion of missing data in each outcome, which varied widely, and downgraded the certainty of evidence accordingly.

Authors' conclusions

Implications for practice.

In conclusion, long‐acting β‐agonist/long‐acting muscarinic antagonist (LABA/LAMA), may have an advantage over LABA/inhaled corticosteroid (ICS), to reduce chronic obstructive pulmonary disease (COPD), exacerbations in the high‐risk population and over monotherapies to improve participant‐reported outcomes, such as symptoms and perceived health status, in people with or without a history of COPD exacerbations. LAMA may be preferred over LABA to reduce COPD exacerbations, especially in the high‐risk population. ICS‐containing inhalers are associated with an increased risk of pneumonia.

Implications for research.

The efficacy of maintenance inhaler therapies appears modest at best. Research and development of a new therapy, such as triple combination therapy, which would have a greater impact on controlling symptoms and preventing exacerbations, are much desired. Meanwhile further investigation on how best to use the existing inhaler therapies in subgroups of patients, such as in those with blood eosinophilia and varying degrees of bronchial reactivity would be helpful. There is a need for more studies evaluating COPD subpopulations or phenotypes.

Appendices

Appendix 1. Sources and search methods for the Cochrane Airways Trials Register

Electronic searches: core databases

Database	Dates searched	Frequency of search
CENTRAL (via the Cochrane Register of Studies (CRS))	From inception	Monthly
MEDLINE (Ovid)	1946 onwards	Weekly
Embase (Ovid)	1974 onwards	Weekly
PsycINFO (Ovid)	1967 onwards	Monthly
CINAHL (EBSCO)	1937 onwards	Monthly
AMED (EBSCO)	From inception	Monthly

Handsearches: core respiratory conference abstracts

Conference	Years searched
American Academy of Allergy, Asthma and Immunology (AAAAI)	2001 onwards
American Thoracic Society (ATS)	2001 onwards
Asia Pacific Society of Respirology (APSR)	2004 onwards
British Thoracic Society Winter Meeting (BTS)	2000 onwards
Chest Meeting	2003 onwards
European Respiratory Society (ERS)	1992, 1994, 2000 onwards
International Primary Care Respiratory Group Congress (IPCRG)	2002 onwards
Thoracic Society of Australia and New Zealand (TSANZ)	1999 onwards

Chronic obstructive pulmonary disease (COPD) search

1. Lung Diseases, Obstructive/

2. exp Pulmonary Disease, Chronic Obstructive/

3. emphysema$.mp.

4. (chronic$ adj3 bronchiti$).mp.

5. (obstruct$ adj3 (pulmonary or lung$ or airway$ or airflow$ or bronch$ or respirat$)).mp.

6. COPD.mp.

7. COAD.mp.

8. COBD.mp.

9. AECB.mp.

10. or/1‐9

Filter to identify randomised controlled trials (RCTs)

1. exp "clinical trial [publication type)"/

2. (randomized or randomised).ab,ti.

3. placebo.ab,ti.

4. dt.fs.

5. randomly.ab,ti.

6. trial.ab,ti.

7. groups.ab,ti.

8. or/1‐7

9. Animals/

10. Humans/

11. 9 not (9 and 10)

12. 8 not 11

The MEDLINE strategy and RCT filter are adapted to identify trials in other electronic databases

Appendix 2. Search strategy to identify relevant trials from the Cochrane Airways Trials Register

#1 MeSH DESCRIPTOR Pulmonary Disease, Chronic Obstructive Explode All

#2 MeSH DESCRIPTOR Bronchitis, Chronic

#3 (obstruct*) near3 (pulmonary or lung* or airway* or airflow* or bronch* or respirat*)

#4 COPD:MISC1

#5 (COPD OR COAD OR COBD OR AECOPD):TI,AB,KW

#6 #1 OR #2 OR #3 OR #4 OR #5

#7 mometasone* AND formoterol*

#8 fluticasone* AND salmeterol*

#9 budesonide* AND formoterol*

#10 beclomethasone* AND formoterol*

#11 fluticasone* AND formoterol*

#12 Flutiform or Fostair or Simplyone

#13 fluticasone* AND vilanterol*

#14 mometasone* AND indacaterol*

#15 formoterol* and ciclesonide*

#16 QMF149

#17 GW685698 AND GW642444

#18 steroid* OR corticosteroid* or ICS

#19 (long‐acting* or long NEXT acting*) NEAR beta*

#20 #18 AND #19

#21 #7 or #8 or #9 or #10 or #11 or #12 or #13 or #14 or #15 or #16 or #17 or #20

#21 formoterol* AND aclidinium*

#22 indacaterol* AND glycopyrronium*

#23 indacaterol* AND tiotropium*

#24 olodaterol* AND tiotropium*

#25 vilanterol* AND umeclidinium*

#26 QVA149

#27 Ultibro or Stiolto or Duaklir Genuair

#28 Muscarinic* Next Antagonist*

#29 #19 AND #28

#30 #21 or # 22 or #23 or #24 or #25 or #26 or #27 or # 29

#31 combin* NEAR inhaler*

#32 FDC:ti,ab

#33 #21 or #30 or #31 or #32

#34 #6 AND #33

(In search line #4, MISC1 denotes the field in which the reference has been coded for condition, in this case, COPD)

Appendix 3. Tables of interventions and treatment groups in the NMAs

1. Population: high‐risk

1.1.1 Moderate to severe exacerbations

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Glycopyrronium 50 μg once daily	LAMA
6	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
7	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
8	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
9	Salmeterol 50 twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Formoterol/budesonide 9/160 μg twice daily	LABA/ICS
11	Formoterol/budesonide 9/320 μg twice daily	LABA/ICS
12	Formoterol/beclomethasone 12/200 μg twice daily	LABA/ICS
13	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
14	Salmeterol 50 twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.1.2 Severe exacerbations

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Tiotropium 18 μg once daily	LAMA
4	Glycopyrronium 50 μg once daily	LAMA
5	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
6	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
7	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
8	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
9	Salmeterol 50 twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.2.2 St George's Respiratory Questionnaire responders at 12 months

	Intervention	Treatment group
1	Salmeterol 50 twice daily	LABA
2	Indacaterol 150 once daily	LABA
3	Formoterol 9‐12 twice daily	LABA
4	Tiotropium 18 once daily	LAMA
5	Glycopyrronium 50 once daily	LAMA
6	Salmeterol/fluticasone 50/250 twice daily	LABA/ICS
7	Salmeterol/fluticasone 50/500 twice daily	LABA/ICS
8	Formoterol/budesonide 12/400 twice daily DPI	LABA/ICS
9	Formoterol/beclomethasone 12/200 twice daily	LABA/ICS
10	Indacaterol/glycopyrronium 110/50 once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.1 Change from baseline in St George's Respiratory Questionnaire score at 3 months

	Intervention	Treatment group
1	Indacaterol 150 μg once daily	LABA
2	Salmeterol 50 μg twice daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Glycopyrronium 50 μg once daily	LAMA
6	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
7	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
8	Salmeterol 50 μg twice daily + fluticasone 250 μg twice daily	LABA/ICS
9	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
10	Indacaterol 150 μg once daily + budesonide 400 μg twice daily	LABA/ICS
11	Formoterol/budesonide 9/320 μg twice daily	LABA/ICS
12	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.2 Change from baseline in St George's Respiratory Questionnaire score at 6 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Glycopyrronium 50 μg once daily	LAMA
6	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
7	Salmeterol/fluticasone 50/50 μg twice daily	LABA/ICS
8	Salmeterol 50 μg twice daily + fluticasone 250 μg twice daily	LABA/ICS
9	Indacaterol 150 μg once daily + budesonide 400 μg twice daily	LABA/ICS
10	budesonide/formoterol 160/9 μg twice daily	LABA/ICS
11	budesonide/formoterol 320/9 μg twice daily	LABA/ICS
12	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.3.3 Change from baseline in St George's Respiratory Questionnaire score at 12 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9 μg twice daily	LABA
4	Formoterol 12 μg twice daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Glycopyrronium 50 μg once daily	LAMA
7	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
8	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
9	Salmeterol 50 μg twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
11	Budesonide/formoterol 400/12 μg twice daily	LABA/ICS
12	Beclomethasone/formoterol 200/12 μg twice daily	LABA/ICS
13	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
14	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
15	Salmeterol 50 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.1 Change from baseline in forced expiratory volume in 1 second at 3 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9 μg twice daily	LABA
3	Formoterol 12 μg twice daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Glycopyrronium 50 μg once daily	LAMA
6	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
7	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
8	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
9	Budesonide + indacaterol 400/150 μg twice daily	LABA/ICS
10	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
11	Beclomethasone/formoterol 200/12 μg twice daily	LABA/ICS
12	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.2 Change from baseline in forced expiratory volume in 1 second at 6 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9 μg twice daily	LABA
3	Tiotropium 18 μg once daily	LAMA
4	Glycopyrronium 50 μg once daily	LAMA
5	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
6	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
7	Salmeterol 50 twice daily + fluticasone 250 μg twice daily	LABA/ICS
8	Budesonide + indacaterol 400/150 μg twice daily	LABA/ICS
9	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
10	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
11	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.5.3 Change from baseline in forced expiratory volume in 1 second at 12 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9 μg twice daily	LABA
3	Formoterol 12 μg twice daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Glycopyrronium 50 μg once daily	LAMA
6	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
7	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
8	Budesonide/formoterol 400/12 μg twice daily	LABA/ICS
9	Beclomethasone/formoterol 200/12 μg twice daily	LABA/ICS
10	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
11	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
12	Salmeterol 50 twice daily + fluticasone 500 μg twice daily	LABA/ICS
13	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.6 Mortality

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9 μg twice daily	LABA
4	Formoterol 12 μg twice daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Glycopyrronium 50 μg once daily	LAMA
7	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
8	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
9	Salmeterol 50 μg twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
11	Salmeterol 50 twice daily + fluticasone 250 μg twice daily	LABA/ICS
12	Budesonide 400 μg twice daily + indacaterol 150 μg once daily	LABA/ICS
13	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
14	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
15	Budesonide/formoterol 400/12 μg	LABA/ICS
16	Beclomethasone/formoterol 200/12 μg	LABA/ICS
17	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
18	Salmeterol 50 twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.7.1 Total serious adverse events

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9 μg twice daily	LABA
4	Formoterol 12 μg twice daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Glycopyrronium 50 μg once daily	LAMA
7	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
8	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
9	Salmeterol 50 μg twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
11	Budesonide 400 μg twice daily + indacaterol 150 μg once daily	LABA/ICS
12	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
13	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
14	Budesonide/formoterol 400/12 μg	LABA/ICS
15	Beclomethasone/formoterol 200/12 μg	LABA/ICS
16	Salmeterol 50 μg twice daily + fluticasone 250 μg twice daily	LABA/ICS
17	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
18	Salmeterol 50 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.7.2 Chronic obstructive pulmonary disease serious adverse events

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9 μg twice daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Glycopyrronium 50 μg once daily	LAMA
6	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
7	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
8	Salmeterol 50 μg twice daily + fluticasone 250 μg twice daily	LABA/ICS
9	Salmeterol 50 μg twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
11	Indacaterol 150 μg once daily + budesonide 400 μg twice daily	LABA/ICS
12	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
13	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
14	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.7.3 Cardiac serious adverse events

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9 μg twice daily	LABA
4	Formoterol 12 μg twice daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Glycopyrronium 50 μg once daily	LAMA
7	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
8	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
9	Salmeterol 50 μg twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
11	Fluticasone 250 μg + salmeterol 50 μg twice daily	LABA/ICS
12	Budesonide 400 μg twice daily + indacaterol 150 μg once daily	LABA/ICS
13	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
14	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
15	Beclomethasone/formoterol 200/12 μg	LABA/ICS
16	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.8 Dropouts due to adverse events

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9 μg twice daily	LABA
4	Formoterol 12 μg twice daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Glycopyrronium 50 μg once daily	LAMA
7	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
8	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
9	Salmeterol 50 μg twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
11	Fluticasone 250 μg + salmeterol 50 μg twice daily	LABA/ICS
12	Budesonide 400 μg twice daily + indacaterol 150 μg once daily	LABA/ICS
13	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
14	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
15	Budesonide/formoterol 400/12 μg	LABA/ICS
16	Beclomethasone/formoterol 200/12	LABA/ICS
17	Indacaterol/glycopyrronium 110/50 once daily	LABA/LAMA
18	Salmeterol 50 twice daily + tiotropium 18 once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

1.9 Pneumonia

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 9 μg twice daily	LABA
4	Formoterol 12 μg twice daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Glycopyrronium 50 μg once daily	LAMA
7	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
8	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
9	Salmeterol 50 twice daily + fluticasone 500 μg twice daily	LABA/ICS
10	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
11	Budesonide/formoterol 160/9 μg twice daily	LABA/ICS
12	Budesonide/formoterol 320/9 μg twice daily	LABA/ICS
13	Budesonide/formoterol 400/12 μg	LABA/ICS
14	Beclomethasone/formoterol 200/12 μg	LABA/ICS
15	Budesonide 400 μg twice daily + indacaterol 150 μg once daily	LABA/ICS
16	Fluticasone 250 μg + salmeterol 50 μg twice daily	LABA/ICS
17	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
18	Salmeterol 50 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2 Population: low‐risk

2.1.1 Moderate to severe exacerbations

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9‐12 μg twice daily	LABA
3	Indacaterol 75 μg once daily	LABA
4	Indacaterol 150 μg once daily	LABA
5	Indacaterol 300 μg once daily	LABA
6	Tiotropium 18 μg once daily	LAMA
7	Tiotropium 5 μg once daily	LAMA
8	Aclidinium 400 μg twice daily	LAMA
9	Umeclidinium 62.5 μg once daily	LAMA
10	Glycopyrronium 50 μg once daily	LAMA
11	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
12	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
13	Salmeterol/fluticasone 42/230 μg (HFA) twice daily	LABA/ICS
14	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
15	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
16	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
17	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
18	Indacaterol/glycopyrronium 27.5/12.5 μg twice daily	LABA/LAMA
19	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
20	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
21	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
22	Tiotropium 18 μg once daily + formoterol 10 μg twice daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.1.2 Severe exacerbations

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9‐12 μg twice daily	LABA
3	Indacaterol 150 μg once daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Tiotropium 5 μg once daily	LAMA
6	Umeclidinium 62.5 μg once daily	LAMA
7	Glycopyrronium 50 μg once daily	LAMA
8	Salmetrol/fluticasone 50/250 μg twice daily	LABA/ICS
9	Salmetrol/fluticasone 50/500 μg twice daily	LABA/ICS
10	Salmetrol/fluticasone 42/230 μg (HFA) twice daily	LABA/ICS
11	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
12	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
13	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
14	Vilaterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
15	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
16	Formterol/aclidinium 12/400 μg twice daily	LABA/LAMA
17	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
18	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.2.1 St George's Respiratory Questionnaire responders at 3 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Formoterol 4.5 μg twice daily	LABA
4	Formoterol 9‐12 μg twice daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Umeclidinium 62.5 μg once daily	LAMA
7	Glycopyrronium 50 μg once daily	LAMA
8	Glycopyrronium 15.6 μg twice daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
11	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
12	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
13	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
14	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
15	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
16	Indacaterol/glycopyrronium 27.5/12.5 μg	LABA/LAMA
17	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.2.2 St George's Respiratory Questionnaire responders at 6 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9‐12 μg twice daily	LABA
3	Indacaterol 150 μg once daily	LABA
4	Indacaterol 300 μg once daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Aclidinium 400 μg twice daily	LAMA
7	Umeclidinium 62.5 μg once daily	LAMA
8	Glycopyrronium 15.6 μg twice daily	LAMA
9	Glycopyrronium 50 μg once daily	LAMA
10	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
11	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
12	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
13	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
14	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
15	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
16	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
17	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.3.1 Change from baseline in St George's Respiratory Questionnaire score at 3 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Indacaterol 300 μg once daily	LABA
4	Formoterol 4.5 μg twice daily	LABA
5	Formoterol 9‐12 μg twice daily	LABA
6	Tiotropium 18 μg once daily	LAMA
7	Umeclidinium 62.5 μg once daily	LAMA
8	Glycopyrronium 50 μg once daily	LAMA
9	Glycopyrronium 15.6 μg twice daily	LAMA
10	Tiotropium 5 μg once daily	LAMA
11	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
12	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
13	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
14	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
15	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
16	Indacaterol/ glycopyrronium 27.5/12.5 μg twice daily	LABA/LAMA
17	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
18	Olodaterol 5 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
19	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.3.2 Change from baseline in St George's Respiratory Questionnaire score at 6 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9‐12 μg twice daily	LABA
3	Indacaterol 150 μg once daily	LABA
4	Indacaterol 300 μg once daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Aclidinium 400 μg twice daily	LAMA
7	Umeclidinium 62.5 μg once daily	LAMA
8	Glycopyrronium 15.6 μg twice daily	LAMA
9	Glycopyrronium 50 μg once daily	LAMA
10	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
11	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
12	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
13	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
14	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
15	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
16	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
17	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.3.3 Change from baseline in St George's Respiratory Questionnaire score at 12 months

	Intervention	Treatment group
1	Formoterol 9‐12 μg twice daily	LABA
2	Salmeterol 50 μg twice daily	LABA
3	Tiotropium 18 μg once daily	LAMA
4	Aclidinium 400 μg twice daily	LAMA
5	Glycopyrronium 15.6 μg twice daily	LAMA
6	Glycopyrronium 50 μg once daily	LAMA
7	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
8	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
9	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
10	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.1 Transition Dyspnea Index at 3 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Indacaterol 150 μg once daily	LABA
3	Indacaterol 300 μg once daily	LABA
4	Olodaterol 5 μg once daily	LABA
5	Formoterol 9‐12 μg twice daily	LABA
6	Tiotropium 18 μg once daily	LAMA
7	Umeclidinium 62.5 μg once daily	LAMA
8	Glycopyrronium 50 μg once daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Glycopyrronium 15.6 μg twice daily	LAMA
11	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
12	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
13	ICS/LABA free or fixed combination	LABA/ICS
14	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
15	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
16	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
17	Indacaterol 110 μg once daily + glycopyrronium 50 μg once daily	LABA/LAMA
18	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
19	Indacaterol/glycopyrronium 27.5/12.5 μg twice daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.2 Transition Dyspnea Index at 6 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9‐12 μg twice daily	LABA
3	Indacaterol 150 μg once daily	LABA
4	Olodaterol 5 μg once daily	LABA
5	Tiotropium 18 μg once daily	LAMA
6	Tiotropium 5 μg once daily	LAMA
7	Aclidinium 400 μg twice daily	LAMA
8	Umeclidinium 62.5 μg once daily	LAMA
9	Glycopyrronium 50 μg once daily	LAMA
10	Salmeterol/fluticasone 250/50 μg twice daily	LABA/ICS
11	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
12	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
13	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
14	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
15	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
16	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.4.3 Transition Dyspnea Index at 12 months

	Intervention	Treatment group
1	Formoterol 9‐12 μg twice daily	LABA
2	Indacaterol 300 μg once daily	LABA
3	Olodaterol 5 μg once daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Tiotropium 5 μg once daily	LAMA
6	Aclidinium 400 μg twice daily	LAMA
7	Glycopyrronium 15.6 μg twice daily	LAMA
8	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
9	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
10	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.1 Change from baseline in forced expiratory volume in 1 second at 3 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9‐12 μg twice daily	LABA
3	Indacaterol 75 μg once daily	LABA
4	Indacaterol 150 μg once daily	LABA
5	Indacaterol 300 μg once daily	LABA
6	Olodaterol 5 μg once daily	LABA
7	Tiotropium 18 once daily	LAMA
8	Tiotropium 5 once daily	LAMA
9	Umeclidinium 62.5 μg once daily	LAMA
10	Glycopyrronium 15.6 μg twice daily	LAMA
11	Glycopyrronium 50 μg once daily	LAMA
12	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
13	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
14	Salmeterol/fluticasone 42/230 μg (HFA) twice daily	LABA/ICS
15	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
16	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
17	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
18	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
19	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
20	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
21	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
22	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
23	Olodaterol 5 μg once daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.2 Change from baseline in forced expiratory volume in 1 second at 6 months

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 9‐12 μg twice daily	LABA
3	Indacaterol 75 μg once daily	LABA
4	Indacaterol 150 μg once daily	LABA
5	Indacaterol 300 μg once daily	LABA
6	Olodaterol 5 μg once daily	LABA
7	Tiotropium 18 μg once daily	LAMA
8	Tiotropium 5 μg once daily	LAMA
9	Aclidinium 400 μg twice daily	LAMA
10	Umeclidinium 62.5 μg once daily	LAMA
11	Glycopyrronium 15.6 μg twice daily	LAMA
12	Glycopyrronium 50 μg once daily	LAMA
13	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
14	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
15	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
16	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
17	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
18	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
19	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
20	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
21	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.5.3 Change from baseline in forced expiratory volume in 1 second at 12 months

	Intervention	Treatment group
1	Formoterol 9‐12 μg twice daily	LABA
2	Indacaterol 75 μg once daily	LABA
3	Olodaterol 5 μg once daily	LABA
4	Tiotropium 18 μg once daily	LAMA
5	Tiotropium 5 μg once daily	LAMA
6	Aclidinium 400 μg twice daily	LAMA
7	Glycopyrronium 15.6 μg twice daily	LAMA
8	Glycopyrronium 50 μg once daily	LAMA
9	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
10	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
11	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
12	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
13	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.6 Mortality

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 4.5 μg twice daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Indacaterol 75 μg once daily	LABA
5	Indacaterol 150 μg once daily	LABA
6	Indacaterol 300 μg once daily	LABA
7	Olodaterol 5 μg once daily	LABA
8	Tiotropium 18 μg once daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Aclidinium 400 μg twice daily	LAMA
11	Umeclidinium 62.5 μg once daily	LAMA
12	Glycopyrronium 15.6 μg twice daily	LAMA
13	Glycopyrronium 50 μg once daily	LAMA
14	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
15	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
16	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
17	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
18	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
19	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
20	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
21	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
22	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
23	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
24	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
25	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
26	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA
27	Olodaterol 5 μg once daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.7.1 Total serious adverse events

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 4.5 μg twice daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Indacaterol 75 μg once daily	LABA
5	Indacaterol 150 μg once daily	LABA
6	Indacaterol 300 μg once daily	LABA
7	Olodaterol 5 μg once daily	LABA
8	Tiotropium 18 μg once daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Aclidinium 400 μg twice daily	LAMA
11	Umeclidinium 62.5 μg once daily	LAMA
12	Glycopyrronium 15.6 μg twice daily	LAMA
13	Glycopyrronium 50 μg once daily	LAMA
14	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
15	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
16	Salmeterol/fluticasone 42/230 μg (HFA) twice daily	LABA/ICS
17	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
18	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
19	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
20	ICS/LABA free or fixed combination	LABA/ICS
21	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
22	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
23	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
24	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
25	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
26	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
27	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
28	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA
29	Olodaterol 5 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
30	Indacaterol 110 μg once daily + glycopyrronium 50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.7.2 Chronic obstructive pulmonary disease serious adverse events

1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 4.5 μg twice daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Indacaterol 75 μg once daily	LABA
5	Indacaterol 150 μg once daily	LABA
6	Indacaterol 300 μg once daily	LABA
7	Olodaterol 5 μg once daily	LABA
8	Tiotropium 18 μg once daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Aclidinium 400 μg twice daily	LAMA
11	Umeclidinium 62.5 μg once daily	LAMA
12	Glycopyrronium 15.6 μg twice daily	LAMA
13	Glycopyrronium 50 μg once daily	LAMA
14	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
15	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
16	Salmeterol/fluticasone 42/230 μg (HFA) twice daily	LABA/ICS
17	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
18	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
19	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
20	ICS/LABA free or fixed combination	LABA/ICS
21	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
22	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
23	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
24	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
25	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
26	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
27	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
28	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA
29	Olodaterol 5 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
30	Indacaterol 110 μg once daily + glycopyrronium 50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.7.3 Cardiac serious adverse events

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 4.5 μg twice daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Indacaterol 75 μg once daily	LABA
5	Indacaterol 150 μg once daily	LABA
6	Indacaterol 300 μg once daily	LABA
7	Olodaterol 5 μg once daily	LABA
8	Tiotropium 18 μg once daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Aclidinium 400 μg twice daily	LAMA
11	Umeclidinium 62.5 μg once daily	LAMA
12	Glycopyrronium 15.6 μg twice daily	LAMA
13	Glycopyrronium 50 μg once daily	LAMA
14	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
15	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
16	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
17	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
18	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
19	ICS/LABA free or fixed combination	LABA/ICS
20	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
21	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
22	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
23	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
24	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
25	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
26	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
27	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA
28	Olodaterol 5 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
29	Indacaterol 110 μg once daily + glycopyrronium 50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.8 Dropouts

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 4.5 μg twice daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Indacaterol 75 μg once daily	LABA
5	Indacaterol 150 μg once daily	LABA
6	Indacaterol 300 μg once daily	LABA
7	Olodaterol 5 μg once daily	LABA
8	Tiotropium 18 μg once daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Aclidinium 400 μg twice daily	LAMA
11	Umeclidinium 62.5 μg once daily	LAMA
12	Glycopyrronium 15.6 μg twice daily	LAMA
13	Glycopyrronium 50 μg once daily	LAMA
14	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
15	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
16	Salmeterol/fluticasone 42/230 μg twice daily	LABA/ICS
17	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
18	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
19	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
20	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
21	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
22	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
23	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
24	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
25	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
26	Indacaterol 150 once daily + tiotropium 18 μg once daily	LABA/LAMA
27	Formoterol 10‐12 twice daily + tiotropium 18 μg once daily	LABA/LAMA
28	Olodaterol 5 once daily + tiotropium 18 μg once daily	LABA/LAMA
29	Indacaterol 110 μg once daily + glycopyrronium 50 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

2.9 Pneumonia

	Intervention	Treatment group
1	Salmeterol 50 μg twice daily	LABA
2	Formoterol 4.5 μg twice daily	LABA
3	Formoterol 9‐12 μg twice daily	LABA
4	Indacaterol 75 μg once daily	LABA
5	Indacaterol 150 μg once daily	LABA
6	Indacaterol 300 μg once daily	LABA
7	Olodaterol 5 μg once daily	LABA
8	Tiotropium 18 μg once daily	LAMA
9	Tiotropium 5 μg once daily	LAMA
10	Aclidinium 400 μg twice daily	LAMA
11	Umeclidinium 62.5 μg once daily	LAMA
12	Glycopyrronium 15.6 μg twice daily	LAMA
13	Glycopyrronium 50 μg once daily	LAMA
14	Salmeterol/fluticasone 50/250 μg twice daily	LABA/ICS
15	Salmeterol/fluticasone 50/500 μg twice daily	LABA/ICS
16	Salmeterol/fluticasone 42/230 μg twice daily	LABA/ICS
17	Formoterol/mometasone 200/10 μg twice daily	LABA/ICS
18	Formoterol/mometasone 400/10 μg twice daily	LABA/ICS
19	Vilanterol/fluticasone 25/100 μg once daily	LABA/ICS
20	ICS/LABA free or fixed combination	LABA/ICS
21	Vilanterol/umeclidinium 25/62.5 μg once daily	LABA/LAMA
22	Formoterol/glycopyrronium 9.6/18 μg twice daily	LABA/LAMA
23	Indacaterol/glycopyrronium 27.5/15.6 μg twice daily	LABA/LAMA
24	Indacaterol/glycopyrronium 110/50 μg once daily	LABA/LAMA
25	Olodaterol/tiotropium 5/5 μg once daily	LABA/LAMA
26	Formoterol/aclidinium 12/400 μg twice daily	LABA/LAMA
27	Indacaterol 150 μg once daily + tiotropium 18 μg once daily	LABA/LAMA
28	Formoterol 10‐12 μg twice daily + tiotropium 18 μg once daily	LABA/LAMA
29	Olodaterol 5 μg once daily + tiotropium 18 μg once daily	LABA/LAMA

ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

Appendix 4. Model fit description and statistics

Population: high‐risk

Outcome: moderate to severe exacerbations

We fitted random‐ and fixed‐treatment‐effects network meta‐analysis (NMA) models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model with lower deviance information criterion (DIC) and between‐study heterogeneity was low (standard deviation (SD) 0.07, 95% credible interval (CrI) 0.008 to 0.14). We considered a random‐class model with fixed‐treatment effects, which only slightly improved fit compared to the fixed‐treatment‐effect model with fixed‐class. We chose the random‐treatment‐effects model with fixed‐class effects as it had the lowest DIC.

The inconsistency model with random treatment effects (and fixed‐class effects), did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA (consistency), and inconsistency models showed small improvements for two data points in the inconsistency model with other points fitting worse (Figure 3c). Reported results are therefore based on the random‐treatment‐effects NMA model with fixed‐class effects assuming consistency.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	42.65	0.07 (0.008 to 0.14)	24.52
Fixed‐effect model	48.22		36.45
Random‐effects inconsistency model	42.04	0.05 (0.003 to 0.13)	24.31
Random‐class‐effects models
Fixed‐effect model	49.36		33.33

^acompare to 27 data points

Outcome: severe exacerbations

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was low (SD 0.07, 95% CrI 0.003 to 0.26). We chose the fixed‐effect model as it had the lowest DIC.
 The inconsistency model with fixed‐treatment effects (and fixed‐class effects) did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, which showed no substantial improvement in fit for any data point (Figure 4). Reported results are therefore based on the fixed‐effect NMA model, assuming consistency with results based on the random‐effects model also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	71.89	0.07 (0.003 to 0.26)	16.64
Fixed‐effect model	70.30		17.44
Fixed‐effect inconsistency model	73.68		18.84

^acompare to 19 data points

Outcome: St George's Respiratory Questionnaire responders at 12 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model although their DIC were comparable and between‐study heterogeneity was moderate (SD 0.26, 95% CrI 0.03 to 1.01). We considered a random‐class model with fixed‐treatment effects but this did not meaningfully improve fit. As there were not enough data to estimate the within‐class variance for the LAMA and LABA/LAMA groups, we assumed that these were equal to the variance in the other monotherapy and combination class respectively. We chose the fixed‐treatment‐effect model with fixed‐class effects as it is the simplest and had comparable DIC to the other models.

The inconsistency model with fixed‐treatment effects (and fixed‐class effects) did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models showed some improvement in fit for data points from one study (Figure 6c). Reported results are based on the fixed‐treatment‐effect NMA model with fixed‐class effects assuming consistency. Results based on the random‐treatment‐effects model with fixed‐classes are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	137.86	0.16 (0.01 to 0.48)	16.91
Fixed‐effect model	139.08		22.01
Fixed‐effect inconsistency model	141.81		22.78
Random‐class‐effects models: class 2 uses variance from class 1, class 4 from class 3
Fixed‐effect model	144.12		22.17

^acompare to 16 data points

Outcome: change from baseline in St George's Respiratory Questionnaire score at 3 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was moderate (SD 0.66, 95% CrI 0.03 to 2.93). We chose the fixed‐treatment‐effect model as it had the lowest DIC. The inconsistency model with fixed‐treatment effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency), and inconsistency models, which showed an equal or better fit of points in the consistency model compared to the inconsistency model (Figure 7c). Reported results are therefore based on the fixed‐treatment‐effects NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	60.89	0.66 (0.03 to 2.93)	20.39
Fixed‐effect model	59.35		21.26
Fixed‐effect inconsistency model	62.90		22.84

^acompare to 19 data points

Outcome: change from baseline in St George's Respiratory Questionnaire score at 6 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was moderate (SD 0.61, 95% CrI 0.31 to 2.03). We chose the fixed‐treatment‐effect model as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models, which showed an equal or better fit of points in the consistency model compared to the inconsistency model (Figure 8c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	65.03	0.61 (0.31 to 2.03)	22.94
Fixed‐effect model	64.00		25.08
Fixed‐effect inconsistency model	66.70		25.79

^acompare to 22 data points

Outcome: change from baseline in St George's Respiratory Questionnaire score at 12 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model but comparable DIC and between‐study heterogeneity was moderate (SD 0.81, 95% CrI 0.12 to 1.75). We considered a random‐class model with fixed‐treatment effects which only slightly improved fit compared to the fixed‐treatment‐effect model with fixed‐class. As there were not enough data to estimate the within‐class variance for the LAMA and LABA/LAMA groups, we assumed that these were equal to the variance in the other monotherapy and combination group respectively. We chose the fixed‐treatment‐effect model with fixed‐class effects as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects (and fixed‐class effects) did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models showed a small improvement for data points from one study in the inconsistency model with other points fitting worse (Figure 9c).

Reported results are therefore based on the fixed‐effect NMA model, assuming consistency with results based on the random‐effects model also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	94.26	0.81 (0.12 to 1.75)	31.42
Fixed‐effect model	96.60		39.8
Fixed‐effect inconsistency model	96.96		38.2
Random‐class‐effects models
Fixed‐effect model	98.69		37.05

^acompare to 32 data points

Outcome: change from baseline in forced expiratory volume in 1 second at 3 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well with equivalent DIC and low between‐study heterogeneity (SD 0.01, 95% CrI 0.00 to 0.04). The fixed‐effect model with fixed‐class effects was chosen as it is the simplest.

The inconsistency model with fixed‐treatment effects and fixed‐class effects showed a very small improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models, which showed no substantial improvement in fit for any data point (Figure 11c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	−114.44	0.01 (0 to 0.04)	22.9
Fixed‐effect model	−114.95		26.0
Fixed‐effect inconsistency model	−115.14		24.8

^acompare to 23 data points

Outcome: change from baseline in forced expiratory volume in 1 second at 6 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was low (SD = 0.02, 95% CrI 0 to 0.05). The fixed‐effect model with fixed‐class effects was chosen as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects and fixed‐class effects did not show improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models, which showed no substantial improvement in fit for any data point (Figure 12c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	−103.62	0.02 (0.00 to 0.05)	22.70
Fixed‐effect model	−103.97		25.87
Fixed‐effect inconsistency model	−102.38		26.47

^acompare to 24 data points

Outcome: change from baseline in forced expiratory volume in 1 second at 12 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was low (SD 0.01, 95% CrI 0.00 to 0.03). The fixed‐effect model with fixed‐class effects was chosen as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects and fixed‐class effects did not show improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models, which showed no improvement in fit for any data point (Figure 13c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	−128.14	0.01 (0.00 to 0.03)	26.19
Fixed‐effect model	−129.43		28.16
Fixed‐effect inconsistency model	−128.31		28.28

^acompare to 29 data points

Outcome: mortality

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was moderate (SD 0.17, 95% CrI 0.01 to 0.49). The fixed‐effect model with fixed‐class effects was chosen as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects and fixed‐class effects showed a small improvement in fit compared to the NMA model assuming consistency. Plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models, which showed some improvement in fit for data points from one study suggesting a possibility of inconsistency (Figure 15c).

Reported results are based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency although results should be interpreted with caution due to some evidence of inconsistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	271.00	0.17 (0.009 to 0.49)	51.45
Fixed‐effect model	269.87		53.87
Fixed‐effect inconsistency model	268.35		50.36

^acompare to 53 data points

Outcome: total serious adverse events

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was very low (SD 0.05, 95% CrI 0.00 to 0.17). The fixed‐effect model with fixed‐class effects was chosen as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects and fixed‐class effects showed no improvement in fit compared to the NMA model assuming consistency. Plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models confirmed this as there was no improvement in fit for any data points in the inconsistency model (Figure 16c).

Reported results are based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	378.46	0.06 (0.002 to 0.17)	49.12
Fixed‐effect model	376.7		50.94
Fixed‐effect inconsistency model	379.24		51.44

^acompare to 53 data points

Outcome: COPD serious adverse events

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was very low (SD 0.06, 95% CrI 0.00 to 0.21). The fixed‐effect model with fixed‐class effects was chosen as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects and fixed‐class effects showed no improvement in fit compared to the NMA model assuming consistency. Plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models confirmed this as there was no improvement in fit for any data points in the inconsistency model (Figure 17c).

Reported results are based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	283.74	0.06 (0.002 to 0.21)	42.55
Fixed‐effect model	282.07		43.21
Fixed‐effect inconsistency model	285.67		44.73

^acompare to 44 data points

Outcome: cardiac serious adverse events

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model with a slightly lower DIC although the posterior mean of the residual deviance was still considerably larger than the number of data points, and the between‐study heterogeneity was moderate (SD 0.28 to 95% CrI 0.02 to 0.67). Random‐class models with fixed‐ and random‐treatment effects were fitted, which improved fit compared to the fixed‐class models. As there were not enough data to estimate the within‐class variance for the LABA/LAMA group, we assumed that this was equal to the variance in the other combination group (LABA/ICS). DIC was lowest for the random‐treatment‐effects model with a fixed‐class so we chose this model. However, note that this DIC differed by only 1 point from the DIC for the fixed‐treatment‐effect model with a fixed‐class.

The inconsistency models with random‐treatment effects (and fixed‐class), showed no improvement in fit and DIC compared to the NMA model assuming consistency to suggesting no evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA and inconsistency models confirmed this as there was no improvement in fit for any points in the inconsistency model (Figure 18c).

Reported results are therefore based on the random‐treatment‐effects NMA model with fixed‐class effects to assuming consistency. Results based on the fixed‐treatment‐effect model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	256.42	0.28 (0.02, 0.67)	51.51
Fixed‐effect model	257.45		59.83
Fixed‐effect inconsistency model	260.69		61.06
Random‐class‐effects models
Random‐effects model	253.42	0.23 (0.01, 0.65)	44.88
Fixed‐effect model	253.13		48.23

^acompare to 42 data points

Outcome: dropouts due to adverse events

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was very low (SD 0.05 to 95% CrI 0.00 to 0.18). The fixed‐effect model with fixed‐class effects was chosen as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects and fixed‐class effects showed no improvement in fit compared to the NMA model assuming consistency. Plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models confirmed this as there was no improvement in fit for any data points in the inconsistency model (Figure 19c).

Reported results are based on the fixed‐treatment‐effect NMA model with fixed‐class effects to assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	344.54	0.05 (0.002 to 0.18)	45.35
Fixed‐effect model	342.43		45.35
Fixed‐effect inconsistency model	345.77		46.7

^acompare to 55 data points

Outcome: pneumonia

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The posterior mean of the residual deviance was substantially larger than the number of data points for both models and the between‐study heterogeneity was moderate (SD 0.18, 95% CrI 0.01 to 0.61). Random‐class models with fixed‐ and random‐treatment‐effects were fitted and although model fit was improved, the DIC was comparable to the fixed‐class models. As there were not enough data to estimate the within‐class variance for the LAMA and LABA/LAMA groups, we assumed that these were equal to the variance in the other monotherapy and combination groups respectively. The fixed‐treatment‐effect model with fixed‐class had the lowest DIC so we chose this model.

The inconsistency model with fixed‐treatment effects (and fixed‐class), showed no improvement in fit or DIC compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, where fit was the same or better for the consistency model for most data points (Figure 20c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison. Results should be interpreted with some caution due to poor model fit, which can be attributed to studies with zero cells.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	280.12	0.18 (0.01 to 0.61)	60.01
Fixed‐effect model	278.71		63.19
Fixed‐effect inconsistency model	282.65		65.11
Random‐class‐effects models
Fixed‐effect model	281.64		60.95
Random‐effects model	281.35	0.24 (0.01 to 0.71)	56.87

^acompare to 53 data points

Population: low‐risk

Outcome: moderate to severe chronic obstructive pulmonary disease exacerbations

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model although their DIC were comparable and between‐study heterogeneity was low (SD 0.054, 95% CrI 0.002 to 0.14). We considered a random‐class model with fixed‐treatment effects but this did not meaningfully improve fit. We chose the fixed‐treatment‐effect model with fixed‐class effects as it is the simplest and had comparable DIC to the other models.

The inconsistency model with fixed‐treatment effects (and fixed‐class effects) did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models, which showed no substantial improvement in fit for any data point (Figure 21c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects assuming consistency. Results based on the random‐treatment‐effects model with fixed‐classes are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	386.49	0.05 (0.002 to 0.14)	76.97
Fixed‐effect model	387.13		81.9
Fixed‐effect inconsistency model	390.02		81.8
Random‐class‐effects models
Fixed‐effect model	392.54		79.89

^acompare to 72 data points

Outcome: severe chronic obstructive pulmonary disease exacerbations

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model although the latter had lower DIC and between‐study heterogeneity was low (SD 0.10, 95% CrI 0.006 to 0.43). A random‐class model with fixed‐treatment effect was considered but this did not improve fit so we chose the fixed‐effect model with fixed‐class effects as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects and fixed‐class effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency), and inconsistency models, which showed no substantial improvement in fit for any data point (Figure 22c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	270.29	0.10 (0.006 to 0.43)	64.82
Fixed‐effect model	268.61		66.19
Fixed‐effect inconsistency model	273.57		68.36
Random‐class‐effects models
Fixed‐effect model	275.61		68.46

^acompare to 60 data points

Outcome: St George's Respiratory Questionnaire responders at 3 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was low (SD 0.04, 95% CrI 0.002 to 0.15). We chose the fixed‐treatment‐effect model as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency) and inconsistency models, which showed an equal or better fit of points in the consistency model compared to the inconsistency model (Figure 24c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	337.64	0.04 (0.002 to 0.15)	39.84
Fixed‐effect model	335.70		40.29
Fixed‐effect inconsistency model	339.79		42.32

^acompare to 44 data points

Outcome: St George's Respiratory Questionnaire responders at 6 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model with a lower DIC and the between‐study heterogeneity estimated was low (SD 0.14, 95% CrI 0.06 to 0.23). A random‐class model with fixed‐treatment effects was fitted, which improved fit compared to the fixed treatment with fixed‐class effects model. However, we selected the random‐treatment‐effects model with a fixed‐class as it had the lowest DIC.

The inconsistency model with random‐treatment effects and fixed‐class effects did not show an improvement in fit or a reduction in the between‐study heterogeneity compared to the selected NMA model assuming consistency, suggesting no evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA and inconsistency models did not show substantial improvement in fit for any data points (Figure 25c). Reported results are therefore based on the random‐treatment‐effects NMA model with fixed‐class effects (assuming consistency).

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	380.57	0.14 (0.06 to 0.23)	46.38
Fixed‐effect model	391.67		70.62
Random‐effects inconsistency model	383.65	0.13 (0.05 to 0.22)	47.95
Random‐class‐effects models
Fixed‐effect model	385.45		53.20

^acompare to 47 data points

Outcome: change from baseline in St George's Respiratory Questionnaire score at 3 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was low (SD 0.19, 95% CrI 0.006 to 0.67). We chose the fixed‐treatment‐effect model as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency), and inconsistency models, which showed an equal or better fit of points in the consistency model compared to the inconsistency model (Figure 27c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	170.91	0.19 (0.006 to 0.67)	43.82
Fixed‐effect model	169.00		43.55
Fixed‐effect inconsistency model	174.43		45.99

^acompare to 59 data points

Outcome: change from baseline in St George's Respiratory Questionnaire score at 6 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was moderate to low (SD 0.36, 95% CrI 0.17 to 1.08). We chose the fixed‐treatment‐effect model as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA (consistency), and inconsistency models, which showed no improvement in fit for any points in the inconsistency model (Figure 28c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	149.50	0.36 (0.17 to 1.08)	45.83
Fixed‐effect model	148.02		48.20
Fixed‐effect inconsistency model	151.37		49.56

^acompare to 47 data points

Outcome: change from baseline in St George's Respiratory Questionnaire score at 12 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was moderate (SD 0.61, 95% CrI 0.29 to 2.51). We chose the fixed‐treatment‐effect model as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, which showed an equal or better fit of points in the consistency model compared to the inconsistency model (Figure 29c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	42.48	0.61 (0.29 to 2.51)	14.22
Fixed‐effect model	41.25		15.09
Fixed‐effect inconsistency model	43.24		16.07

^acompare to 15 data points

Outcome: Transition Dyspnoea Index at 3 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model with a lower DIC and the between‐study heterogeneity was moderate (SD 0.17, 95% CrI 0.02 to 0.32). We fitted a random‐class model with fixed‐treatment effects, which improved fit substantially compared to the fixed‐treatment‐effect models with a fixed‐class but only slightly compared to the random‐treatment‐effects model with a fixed‐class. As there were not enough data to estimate the within‐class variance for the LABA/ICS group, we assumed that this was equal to the variance in the other combination therapy group (LABA/LAMA).

DIC slightly favoured the fixed‐treatment‐effect model with a random‐class over the random‐treatment‐effects model with a fixed‐class (difference of 3.6 points, which is close to the value for no meaningful difference). Within‐class variability in the fixed‐treatment‐effect model with random‐class was moderate (Table 84). We chose the random‐treatment‐effects model with a fixed‐class as it is more interpretable. However, there is statistical uncertainty as to whether the variability observed across treatment effects is due to between‐study or within‐class/group differences.

71. Within‐class/group standard deviation for forced expiratory volume in 1 second at 12 months in the low‐risk population: fixed‐treatment‐effect model with random‐class.

Treatment group	Median	95% CrI
LABA	0.273	0.022 to 1.190
LAMA	0.109	0.005 to 0.589
LABA/ICS	0.181	0.036 to 0.612
LABA/LAMA	0.181	0.036 to 0.612

CrI: credible interval; ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist

The inconsistency model with random‐treatment effects and fixed‐class did not show an improvement in fit or reduction in heterogeneity compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, which showed no substantial improvement in fit of any points in the inconsistency model (Figure 31c).

Reported results are based on the random‐treatment‐effects model with fixed‐class NMA model (assuming consistency), with the results for the fixed‐treatment‐effect model with random‐class also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	14.34	0.17 (0.02 to 0.32)	61.72
Fixed‐effect model	17.97		75.50
Random‐effects inconsistency model	18.29	0.19 (0.04 to 0.35)	62.33
Random‐class‐effects models
Fixed‐effect model	10.71		59.48

^acompare to 63 data points

Outcome: Transition Dyspnoea Index at 6 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was low (SD 0.09, 95% CrI 0.004 0 0.24). We chose the fixed‐treatment‐effect model as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, showed only a small improvement in fit for some points in the inconsistency model compared to the consistency model (Figure 32c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐classes are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	2.31	0.09 (0.004 to 0.24)	36.56
Fixed‐effect model	0.59		37.73
Fixed‐effect inconsistency model	2.08		37.24

^acompare to 41 data points

Outcome: Transition Dyspnoea Index at 12 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model although their DIC was comparable and between‐study heterogeneity was moderate (SD 0.16, 95% CrI 0.02 to 0.43). We fitted a random‐class model with fixed‐treatment effects, which improved fit compared to the fixed‐treatment‐effect model with a fixed‐class although with a similar DIC. Since all models had similar DIC, we chose the fixed‐treatment‐effect model with a fixed‐class, as it is the simplest.

The inconsistency model with fixed‐treatment effects (and fixed‐class), did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, which showed an equal or better fit of points in the consistency model compared to the inconsistency model (Figure 33c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects assuming consistency. Results based on the random‐treatment‐effects model with fixed‐classes are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	‐6.91	0.16 (0.01 to 0.43)	14.19
Fixed‐effect model	‐5.15		19.59
Fixed‐effect inconsistency model	‐5.15		19.59
Random‐class‐effects models
Fixed‐effect model	‐5.04		15.06

^acompare to 16 data points

Outcome: change from baseline in forced expiratory volume in 1 second at 3 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model with a lower DIC and the between‐study heterogeneity was moderate (SD 0.03, 95% CrI 0.02 to 0.03). A random‐class model with fixed‐treatment effects was fitted which improved fit compared to the fixed‐treatment‐effect model with a fixed‐class. However, the random‐treatment‐effects model with a fixed‐class was selected as it had the lowest DIC.

The inconsistency model with random‐treatment effects (and fixed‐class) did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, which showed no substantial improvement in the fit of points in the inconsistency model (Figure 35c).

Reported results are therefore based on the random‐effects NMA model with fixed‐classes (assuming consistency).

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	‐513.575	0.03 (0.02 to 0.03)	105.6
Fixed‐effect model	‐421.49		229.0
Random‐effects inconsistency model	‐514.67	0.02 (0.02 to 0.03)	104.4
Random‐class‐effects models
Fixed‐effect model	‐481.10		155.2

^acompare to 107 data points

Outcome: change from baseline in forced expiratory volume in 1 second at 6 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model with a lower DIC and the between‐study heterogeneity was moderate (SD 0.02, 95% CrI 0.007 to 0.03). We fitted a random‐class model with fixed‐treatment effects, which improved fit substantially compared to the fixed‐treatment‐effect models with a fixed‐class but not compared to the random‐treatment‐effects model with a fixed‐class. As there were not enough data to estimate the within‐class variance for the LABA/ICS group, we assumed that this was equal to the variance in the other combination therapy group (LABA/LAMA).

The difference in DIC between the fixed‐treatment‐effect model with a random‐class and the random‐treatment‐effects model with a fixed‐class was less than 3 points. Within‐class variability in the fixed‐treatment‐effect model with random‐class was moderate. We chose the random‐treatment‐effects model with a fixed‐class as it is more interpretable. However, there is statistical uncertainty as to whether the variability observed across treatment effects is due to between‐study or within‐class differences.

The inconsistency model with random‐treatment effects (and fixed‐class) showed some improvement in fit compared to the NMA model assuming consistency and had lower between‐study heterogeneity and DIC, suggesting some evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA and inconsistency models showed that fit improved for some studies in the inconsistency model compared to the consistency models, although for other studies fit was worse (Figure 36c).

Reported results are based on the random‐treatment‐effects model with fixed‐class NMA model (assuming consistency) with the results for the fixed‐treatment‐effect model with random‐class also reported for comparison. However, there is weak evidence of potential inconsistency in this network and results should be interpreted with some caution.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	‐324.38	0.02 (0.007 to 0.03)	68.26
Fixed‐effect model	‐315.31		91.40
Random‐effects inconsistency model	‐328.14	0.01 (0.000 to 0.02)	66.91
Random‐class‐effects models
Fixed‐effect model	‐326.62		68.99

^acompare to 69 data points

Within class/group standard deviation for change from baseline in FEV1 at 6 months in the low‐risk population

Fixed‐treatment‐effect model with random‐class

	Median	95% CrI
LABA	0.010	(0.000 to 0.052)
LAMA	0.020	(0.003 to 0.064)
LABA/ICS	0.025	(0.009 to 0.068)
LABA/LAMA	0.025	(0.009 to 0.068)

Outcome: change from baseline in forced expiratory volume in 1 second at 12 months

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The random‐effects model had a better fit than the fixed‐effect model with a lower DIC and the between‐study heterogeneity was moderate (SD 0.02, 95% CrI 0.01 to 0.04). We fitted a random‐class model with fixed‐treatment effects, which improved fit compared to the fixed‐treatment‐effect model with a fixed‐class. DIC was lower in the model with fixed‐treatment and random‐class effects, although there was evidence of overfitting. We therefore report results for both the random‐treatment‐effects model with a fixed‐class and the fixed‐treatment‐effect model with a random‐class (Table 67). Within‐class variability in the fixed‐treatment‐effect model with random‐class was moderate. There is some evidence that the variability observed across treatment effects may be due to within‐class/group differences rather than between‐study heterogeneity.

The inconsistency model with random‐treatment effects and fixed‐class had an improved model fit and lower between‐study heterogeneity and DIC when compared to the equivalent consistency model.

The inconsistency model with fixed‐treatment effects with random‐class did not show an improvement in fit or DIC when compared to the equivalent consistency model therefore suggesting no evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA and inconsistency models confirmed this (Figure 37c).

Reported results are based on the fixed‐treatment‐effect NMA model with random‐classes (assuming consistency), with the results for the random‐treatment‐effects model with fixed‐classes also reported for comparison. However, there is weak evidence of potential inconsistency in the latter model so results should be interpreted with caution.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	−150.21	0.02 (0.01 to 0.04)	32.70
Fixed‐effect model	−142.19		49.03
Random‐effects inconsistency model	−154.87	0.01 (0.00 to 0.03)	29.46
Random‐class‐effects models
Fixed‐effect model	−155.96		27.93
Fixed‐effect inconsistency model	−154.3		28.87

^acompare to 31 data points

Within class/group standard deviation for change from baseline in FEV1 at 12 months in the low‐risk population

Fixed‐treatment‐effect model with random‐class

	Median	95% CrI
LABA	0.019	(0.001 to 0.422)
LAMA	0.018	(0.004 to 0.073)
LABA/LAMA	0.045	(0.016 to 0.158)

Outcome: mortality

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The posterior mean of the residual deviance was substantially larger than the number of data points for both models and the between‐study heterogeneity was moderate (SD 0.15, 95% CrI 0.007 to 0.70). We considered random‐class models with fixed‐ and random‐treatment effects but this only slightly improved fit compared to the fixed‐class models. The fixed‐treatment‐effect model with fixed‐class had the lowest DIC so we chose this model.

The inconsistency model with fixed‐treatment effects (and fixed‐class) showed no improvement in fit or DIC compared to the NMA model assuming consistency, suggesting no evidence of inconsistency (Figure 39c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison. Results should be interpreted with some caution due to poor model fit which can be attributed to studies with zero cells.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	432.52	0.15 (0.007 to 0.70)	129.4
Fixed‐effect model	430.85		131.9
Fixed‐effect inconsistency model	430.73		132.4
Random‐class‐effects models
Fixed‐effect model	435.98		134.5

^acompare to 110 data points

Outcome: total serious adverse events

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. Both models fitted the data well and between‐study heterogeneity was low (SD 0.04, 95% CrI 0.00 to 0.15). We chose the fixed‐effect model as it had the lowest DIC.

The inconsistency model with fixed‐treatment effects (and fixed‐class effects) did not show an improvement in fit compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, which showed no improvement in fit for any data point (Figure 40c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	891.21	0.04 (0 to 0.15)	145.8
Fixed‐effect model	889.36		147.7
Fixed‐effect inconsistency	894.82		150.2

^acompare to 145 data points

Outcome: chronic obstructive pulmonary disease serious adverse events

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The posterior mean of the residual deviance was substantially larger than the number of data points for both models and the between‐study heterogeneity was moderate (SD 0.16, 95% CrI 0.002 to 0.38). Random‐class models with fixed‐ and random‐treatment effects were fitted and although model fit was improved the fixed‐class models had lower DIC. The fixed‐treatment‐effect model with fixed‐class had the lowest DIC so we chose this model.

The inconsistency model with fixed‐treatment effects (and fixed‐class) showed no improvement in fit or DIC compared to the NMA model assuming consistency, suggesting no evidence of inconsistency (Figure 41c). However, plotting each data point's contribution to the residual deviance in the NMA and inconsistency models there were a few studies with slightly improved fit in the inconsistency, compared to the consistency model, suggesting some evidence of inconsistency (Figure 41c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison. Results should be interpreted with some caution due to poor model fit, which can be attributed to studies with zero cells.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	662.62	0.16 (0.002 to 0.38)	144.2
Fixed‐effect model	661.91		151.0
Fixed‐effect inconsistency	666.00		152.4
Random‐class‐effects models
Random‐effects model	665.07	0.13 (0.006 to 0.37)	140.1
Fixed‐effect model	664.86		143.9

^acompare to 135 data points

Outcome: cardiac serious adverse events

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. The posterior mean of the residual deviance was substantially larger than the number of data points for both models and the between‐study heterogeneity was moderate (SD 0.16, 95% CrI 0.006 to 0.48). We fitted random‐class models with fixed‐ and random‐treatment effects and although model fit was improved the fixed‐class models had lower DIC. The fixed‐treatment‐effect model with fixed‐class had the lowest DIC so we chose this model.

The inconsistency model with fixed‐treatment effects (and fixed‐class) showed some improvement in fit or DIC compared to the NMA model assuming consistency, suggesting evidence of inconsistency. Plotting each data point's contribution to the residual deviance in the NMA and inconsistency models showed improved fit for one study in the inconsistency model, suggesting some evidence of inconsistency (Figure 42c). Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison. Results should be interpreted with some caution due to poor model fit, which can be attributed to studies with zero cells.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	578.42	0.17 (0.006 to 0.48)	151.2
Fixed‐effect model	577.25		155.8
Fixed‐effect inconsistency	572.69		149.3
Random‐class‐effects models
Random‐effects model	581.73	0.16 (0.008 to 0.49)	147.0
Fixed‐effect model	581.40		150.5

^acompare to 127 data points

Outcome: dropouts due to adverse events

We fitted random‐ and fixed‐treatment‐effect NMA models with fixed‐class effects. The posterior mean of the residual deviance was substantially larger than the number of data points for both models and the between‐study heterogeneity was low (SD 0.09, 95% CrI 0.004 to 0.24). Random‐class models with fixed‐ and random‐treatment effects were fitted and although model fit was improved the DIC was comparable to the fixed‐class models. The fixed‐treatment‐effect model with fixed‐class had the lowest DIC so we chose this model.

The inconsistency model with fixed‐treatment effects (and fixed‐class) showed no improvement in fit or DIC compared to the NMA model assuming consistency, suggesting no evidence of inconsistency. We confirmed this by plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, where fit was the same or better for the consistency model for most data points (Figure 43c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class are also reported for comparison. Results should be interpreted with some caution due to poor model fit.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	848.0	0.09 (0.004 to 0.24)	155.6
Fixed‐effect model	846.7		160.5
Fixed‐effect inconsistency	849.3		160.2
Random‐class‐effects models
Random‐effects model	847.3	0.09 (0.003 to 0.23)	144.8
Fixed‐effect model	846.9		148.6

^acompare to 146 data points

Outcome: pneumonia

We fitted random‐ and fixed‐treatment‐effects NMA models with fixed‐class effects. There was some evidence that the posterior distribution of the between‐study heterogeneity was poorly estimated so we used an informative prior distribution, based on Turner 2012. We selected the prior distribution suggested for the between‐study variance of a subjective outcome (infection, new disease), for comparisons of pharmacological interventions.

The random‐effects model had a better fit than the fixed‐effect model with a lower DIC although the posterior mean of the residual deviance was still considerably larger than the number of data points and the between‐study heterogeneity was moderate (SD 0.23, 95% CrI 0.05 to 0.65). We fitted random‐class models with fixed‐ and random‐treatment effects, which improved fit slightly compared to the fixed‐class model. However, DIC was lowest for the fixed‐treatment‐effect model with a fixed‐class so we chose this model.

The inconsistency models with fixed‐treatment effects (and fixed‐class) showed an improvement in fit and DIC compared to the NMA model assuming consistency, suggesting some evidence of inconsistency.

Plotting each data point's contribution to the residual deviance in the NMA and inconsistency models, there was some improvement in fit for a few studies in the inconsistency model although most of the studies with high residual deviance contained zero‐event arms, of which there were many in the dataset (Figure 44c).

Reported results are therefore based on the fixed‐treatment‐effect NMA model with fixed‐class effects, assuming consistency. Results based on the random‐treatment‐effects model with fixed‐class and informative prior distribution on the heterogeneity parameter are also reported for comparison. Results should be interpreted with caution due to potential inconsistency in the data.

	DIC	SD (95% CrI)	Total residual deviancea
Fixed‐class‐effect models
Random‐effects model	531.76	0.23 (0.05 to 0.65)	167.3
Fixed‐effect model	532.14		174.3
Fixed‐effect inconsistency model	525.77		166.0
Random‐class‐effects models
Random‐effects model	531.13	0.22 (0.05 to 0.61)	158.4
Fixed‐effect model	531.66		162.0

^acompare to 133 data points

DIC: deviance information criterion; SD: standard deviation

Appendix 5. Ranking summary

Outcome	Treatment group	High‐risk population			Low‐risk population
		Mean	Median	95% CrI	Mean	Median	95% CrI
Moderate to severe exacerbations	LABA/LAMA	1	1	(1 to 2)	1.1	1	(1 to 2)
	LAMA	2.4	2	(2 to 3)	2.2	2	(1 to 3)
	LABA/ICS	2.6	3	(2 to 3)	2.6	3	(2 to 3)
	LABA	4	4	(4 to 4)	4	4	(4 to 4)
Severe exacerbations	LABA/LAMA	1.2	1	(1 to 2)	1.3	1	(1 to 3)
	LAMA	1.9	2	(1 to 3)	1.9	2	(1 to 3)
	LABA/ICS	3	3	(2 to 3)	3.3	3	(2 to 4)
	LABA	4	4	(4 to 4)	3.5	4	(2 to 4)
SGRQ responders at 3 months	LABA	NA	NA	NA	1.4	1	(1 to 3)
	LABA/LAMA	NA	NA	NA	1.8	2	(1 to 3)
	LABA/ICS	NA	NA	NA	2.8	3	(1 to 3)
	LAMA	NA	NA	NA	4	4	(4 to 4)
SGRQ responders at 6 months	LABA/LAMA	NA	NA	NA	1	1	(1 to 2)
	LABA/ICS	NA	NA	NA	2.7	2	(1 to 4)
	LAMA	NA	NA	NA	3	3	(2 to 4)
	LABA	NA	NA	NA	3.3	3	(2 to 4)
SGRQ score at 3 months	LABA/LAMA	1	1	(1 to 1)	1.7	2	(1 to 3)
	LABA/ICS	2	2	(2 to 2)	1.6	2	(1 to 3)
	LABA	3.4	3	(3 to 4)	2.8	3	(1 to 4)
	LAMA	3.6	4	(3 to 4)	3.9	4	(3 to 4)
SGRQ score at 6 months	LABA/LAMA	1	1	(1 to 1)	1.3	1	(1 to 2)
	LABA/ICS	2	2	(2 to 2)	1.7	2	(1 to 3)
	LAMA	3.2	3	(3 to 4)	3.3	3	(2 to 4)
	LABA	3.8	4	(3 to 4)	3.7	4	(3 to 4)
SGRQ score at 12 months	LABA/LAMA	1.1	1	(1 to 2)	2	2	(1 to 3)
	LABA/ICS	2	2	(1 to 3)	1.1	1	(1 to 2)
	LAMA	2.9	3	(2 to 3)	3.3	3	(2 to 4)
	LABA	4	4	(4 to 4)	3.6	4	(3 to 4)
TDI at 3 months	LABA/LAMA	NA	NA	NA	1	1	(1 to 1)
	LABA/ICS	NA	NA	NA	2.3	2	(2 to 4)
	LABA	NA	NA	NA	3	3	(2 to 4)
	LAMA	NA	NA	NA	3.7	4	(2 to 4)
TDI at 6 months	LABA/LAMA	NA	NA	NA	1.1	1	(1 to 2)
	LABA/ICS	NA	NA	NA	2	2	(1 to 4)
	LAMA	NA	NA	NA	3.2	3	(2 to 4)
	LABA	NA	NA	NA	3.6	4	(3 to 4)
TDI at 12 months	LABA/LAMA	NA	NA	NA	1	1	(1 to 1)
	LAMA	NA	NA	NA	2.06	2	(2 to 3)
	LABA	NA	NA	NA	2.94	3	(2 to 3)
	LABA/ICS	NA	NA	NA	NA	NA	NA
FEV1 at 3 months	LABA/LAMA	1	1	(1 to 1)	1	1	(1 to 1)
	LABA/ICS	2.4	2	(2 to 3)	2	2	(2 to 2)
	LAMA	2.6	3	(2 to 3)	3.2	3	(3 to 4)
	LABA	4	4	(4 to 4)	3.8	4	(3 to 4)
FEV1 at 6 months	LABA/LAMA	1	1	(1 to 1)	1	1	(1 to 1)
	LAMA	2.1	2	(2 to 3)	2.7	3	(2 to 4)
	LABA/ICS	2.9	3	(2 to 3)	2.3	2	(2 to 4)
	LABA	4	4	(4 to 4)	3.9	4	(3 to 4)
FEV1 at 12 months	LABA/LAMA	1	1	(1 to 1)	1.1	1	(1 to 2)
	LAMA	2	2	(2 to 2)	2	2	(1 to 3)
	LABA/ICS	3	3	(3 to 3)	NA	NA	NA
	LABA	4	4	(4 to 4)	3	3	(2 to 3)
Mortality	LABA/ICS	1.6	1	(1 to 4)	1.5	1	(1 to 4)
	LABA/LAMA	2.6	3	(1 to 4)	3	3	(1 to 4)
	LAMA	2.8	3	(1 to 4)	3.5	4	(1 to 4)
	LABA	3	3	(1 to 4)	2.1	2	(1 to 4)
Dropouts due to adverse event	LABA/LAMA	1.6	1	(1 to 4)	2.5	2	(1 to 4)
	LAMA	2.2	2	(1 to 4)	1.3	1	(1 to 3)
	LABA/ICS	2.4	2	(1 to 4)	2.5	3	(1 to 4)
	LABA	3.9	4	(3 to 4)	3.7	4	(2 to 4)
Pneumonia	LAMA	1.5	1	(1 to 3)	1.6	1	(1 to 3)
	LABA/LAMA	1.9	2	(1 to 3)	2.7	3	(1 to 4)
	LABA	2.6	3	(1 to 3)	1.8	2	(1 to 3)
	LABA/ICS	4	4	(4 to 4)	4	4	(3 to 4)
FEV1: forced expiratory volume in one second;ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; NA: not applicable; SGRQ: St George's Respiratory Questionnaire; TDI: Transition Dyspnoea Index

Appendix 6. Summary of results for pairwise and network meta‐analyses in the high‐risk population

Moderate to severe exacerbations, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA (random‐effects/fixed‐class) HR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.87 (0.76 to 1.00)	0.87 (0.76 to 1.00)	0.86 (0.76 to 0.99)
LABA/LAMA vs LAMA	Moderate	1.06 (0.89 to 1.27)	1.06 (0.89 to 1.27)	0.87 (0.78 to 0.99)
LABA/LAMA vs LABA	NA	NA	NA	0.70 (0.61 to 0.80)
LABA/ICS vs LAMA	Moderate	1.12 (0.90 to 1.39)	1.12 (0.90 to 1.39)	1.01 (0.91 to 1.13)
LABA/ICS vs LABA	High	0.81 (0.75 to 0.89)	0.81 (0.75 to 0.89)	0.80 (0.75 to 0.86)
LAMA vs LABA	High	0.84 (0.76 to 0.92)	0.84 (0.76 to 0.92)	0.80 (0.71 to 0.88)
Severe exacerbations, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) HR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.88 (0.74 to 1.06)	0.88 (0.74 to 1.06)	0.78 (0.64 to 0.93)
LABA/LAMA vs LAMA	Moderate	0.73 (0.45 to 1.16)	0.73 (0.45 to 1.16)	0.89 (0.71 to 1.11)
LABA/LAMA vs LABA	NA	NA	NA	0.64 (0.51 to 0.81)
LABA/ICS vs LAMA	Moderate	1.28 (0.95 to 1.73)	1.28 (0.95 to 1.73)	1.15 (0.97 to 1.36)
LABA/ICS vs LABA	Moderate	0.91 (0.74 to 1.13)	0.91 (0.74 to 1.12)	0.83 (0.71 to 0.97)
LAMA vs LABA	Moderate	0.88 (0.78 to 1.01)	0.88 (0.78 to 1.01)	0.72 (0.63 to 0.82)
SGRQ responders at 3 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	NA	NA	NA	NA
LABA/LAMA vs LAMA	NA	NA	NA	NA
LABA/LAMA vs LABA	NA	NA	NA	NA
LABA/ICS vs LAMA	Low	0.96 (0.56 to 1.65)	0.96 (0.56 to 1.65)	NA
LABA/ICS vs LABA	NA	NA	NA	NA
LAMA vs LABA	Moderate	0.97 (0.84 to 1.12)	0.97 (0.84 to 1.12)	NA
SGRQ responders at 6 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(random‐effects/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	NA	NA	NA	NA
LABA/LAMA vs LAMA	Moderate	1.30 (1.08 to 1.56)	1.30 (1.08 to 1.56)	NA
LABA/LAMA vs LABA	NA	NA	NA	NA
LABA/ICS vs LAMA	Moderate	1.26 (0.99 to 1.59)	1.26 (0.99 to 1.59)	NA
LABA/ICS vs LABA	NA	NA	NA	NA
LAMA vs LABA	Low	1.08 (0.93 to 1.25)	1.08 (0.93 to 1.25)	NA
SGRQ responders at 12 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) OR (95% CrI)
LABA/LAMA vs LABA/ICS	High	1.25 (1.09 to 1.43)	1.25 (1.09 to 1.43)	1.21 (1.07 to 1.36)
LABA/LAMA vs LAMA	Low	1.27 (1.04 to 1.55)	1.27 (1.04 to 1.55)	1.36 (1.18 to 1.58)
LABA/LAMA vs LABA	NA	NA	NA	1.41 (1.2 to 1.66)
LABA/ICS vs LAMA	Moderate	1.15 (0.90 to 1.47)	1.15 (0.90 to 1.47)	1.13 (0.98 to 1.3)
LABA/ICS vs LABA	Moderate	1.15 (0.78 to 1.72)	1.22 (1.03 to 1.46)	1.17 (1.02 to 1.34)
LAMA vs LABA	Moderate	1.00 (0.86 to 1.17)	1.00 (0.86 to 1.17)	1.03 (0.91 to 1.18)
CFB in SGRQ at 3 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	High	−1.30 (−2.35 to −0.25)	−1.30 (−2.35 to −0.25)	−1.39 (−2.37 to −0.42)
LABA/LAMA vs LAMA	Moderate	‐3.68 (‐5.84 to −1.52)	‐3.68 (‐5.84 to −1.52)	‐3.31 (‐4.67 to −1.97)
LABA/LAMA vs LABA	NA	NA	NA	‐3.21 (‐4.52 to −1.92)
LABA/ICS vs LAMA	Low	−1.06 (‐4.39 to 2.27)	−1.06 (‐4.39 to 2.27)	−1.92 (‐3.11 to −0.74)
LABA/ICS vs LABA	Low	−1.81 (−2.99 to −0.64)	−1.81 (−2.99 to −0.64)	−1.82 (−2.86 to −0.78)
LAMA vs LABA	High	0.10 (−0.82 to 1.02)	0.10 (−0.82 to 1.02)	0.10 (−0.76 to 0.96)
CFB in SGRQ at 6 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	High	−1.20 (−2.28 to −0.12)	−1.20 (−2.28 to −0.12)	−1.27 (−2.26 to −0.29)
LABA/LAMA vs LAMA	Moderate	−2.79 (‐5.02 to −0.56)	−2.79 (‐5.02 to −0.56)	−2.48 (‐3.72 to −1.24)
LABA/LAMA vs LABA	NA	NA	NA	−2.88 (‐4.03 to −1.73)
LABA/ICS vs LAMA	Low	−1.97 (‐3.79 to −0.15)	−1.97 (‐3.79 to −0.15)	−1.21 (−2.16 to −0.25)
LABA/ICS vs LABA	Very low	−1.40 (−2.53 to −0.26)	−1.45 (−2.17 to −0.73)	−1.6 (−2.27 to −0.93)
LAMA vs LABA	High	−0.70 (−1.74 to 0.34)	−0.70 (−1.74 to 0.34)	−0.39 (−1.27 to 0.47)
CFB in SGRQ at 12 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	High	−1.20 (−2.34 to −0.06)	−1.20 (−2.34 to −0.06)	−0.52 (−1.42 to 0.36)
LABA/LAMA vs LAMA	Low	‐3.38 (‐5.83 to −0.93)	‐3.38 (‐5.83 to −0.93)	−1.12 (−1.88 to −0.37)
LABA/LAMA vs LABA	NA	NA	NA	−2.1 (‐3.08 to −1.13)
LABA/ICS vs LAMA	Low	−0.99 (−2.98 to 1.00)	−0.99 (−2.98 to 1.00)	−0.59 (−1.48 to 0.29)
LABA/ICS vs LABA	Moderate	−1.75 (−2.61 to −0.89)	−1.78 (−2.49 to −1.07)	−1.57 (−2.23 to −0.92)
LAMA vs LABA	High	−0.40 (−1.56 to 0.76)	−0.40 (−1.56 to 0.76)	−0.98 (−1.86 to −0.08)
TDI at 3 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA
LABA/LAMA vs LABA/ICS	NA	NA	NA	NA
LABA/LAMA vs LAMA	NA	NA	NA	NA
LABA/LAMA vs LABA	NA	NA	NA	NA
LABA/ICS vs LAMA	Moderate	0.50 (0.18 to 0.82)	0.50 (0.18 to 0.82)	NA
LABA/ICS vs LABA	NA	NA	NA	NA
LAMA vs LABA	Moderate	−0.14 (−0.15 to −0.13)	−0.14 (−0.15 to −0.13)	NA
TDI at 6 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA
LABA/LAMA vs LABA/ICS	NA	NA	NA	NA
LABA/LAMA vs LAMA	NA	NA	NA	NA
LABA/LAMA vs LABA	NA	NA	NA	NA
LABA/ICS vs LAMA	Moderate	0.30 (−0.06 to 0.66)	0.30 (−0.06 to 0.66)	NA
LABA/ICS vs LABA	NA	NA	NA	NA
LAMA vs LABA	Moderate	−0.19 (−0.20 to −0.18)	−0.19 (−0.20 to −0.18)	NA
TDI at 12 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA
LABA/LAMA vs LABA/ICS	NA	NA	NA	NA
LABA/LAMA vs LAMA	Moderate	−0.38 (−1.28 to 0.52)	−0.38 (−1.28 to 0.52)	NA
LABA/LAMA vs LABA	NA	NA	NA	NA
LABA/ICS vs LAMA	Low	0.00 (−0.40 to 0.40)	0.00 (−0.40 to 0.40)	NA
LABA/ICS vs LABA	NA	NA	NA	NA
LAMA vs LABA	Moderate	−0.26 (−0.27 to −0.25)	−0.26 (−0.27 to −0.25)	NA
CFB in FEV1 at 3 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	High	0.08 (0.06 to 0.10)	0.08 (0.06 to 0.10)	0.07 (0.05 to 0.09)
LABA/LAMA vs LAMA	Moderate	0.06 (0.02 to 0.09)	0.06 (0.02 to 0.09)	0.07 (0.05 to 0.10)
LABA/LAMA vs LABA	NA	NA	NA	0.12 (0.10 to 0.15)
LABA/ICS vs LAMA	High	0.01 (−0.02 to 0.04)	0.01 (−0.02 to 0.03)	0.00 (−0.02 to 0.02)
LABA/ICS vs LABA	Moderate	0.05 (0.03 to 0.07)	0.05 (0.04 to 0.07)	0.05 (0.04 to 0.07)
LAMA vs LABA	NA	NA	NA	0.05 (0.02 to 0.07)
CFB in FEV1 at 6 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	High	0.09 (0.07 to 0.11)	0.09 (0.07 to 0.11)	0.08 (0.06 to 0.10)
LABA/LAMA vs LAMA	Moderate	0.06 (0.02 to 0.10)	0.06 (0.02 to 0.10)	0.07 (0.04 to 0.09)
LABA/LAMA vs LABA	NA	NA	NA	0.13 (0.10 to 0.15)
LABA/ICS vs LAMA	Moderate	−0.01 (−0.04 to 0.02)	−0.01 (−0.04 to 0.02)	−0.02 (−0.04 to 0.01)
LABA/ICS vs LABA	Moderate	0.05 (0.03 to 0.07)	0.04 (0.03 to 0.06)	0.04 (0.03 to 0.06)
LAMA vs LABA	NA	NA	NA	0.06 (0.03 to 0.08)
CFB in FEV1 at 12 months, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD(95% CI)	Pairwise, fixed‐effect MD(95% CI)	NMA (random‐effects/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.06 (0.04 to 0.08)	0.06 (0.04 to 0.08)	0.07 (0.04 to 0.1)
LABA/LAMA vs LAMA	Moderate	0.05 (0.01 to 0.09)	0.05 (0.01 to 0.09)	0.04 (0 to 0.08)
LABA/LAMA vs LABA	NA	NA	NA	0.12 (0.08 to 0.16)
LABA/ICS vs LAMA	Very low	−0.01 (−0.08 to 0.05)	−0.03 (−0.06 to 0.00)	−0.03 (−0.07 to 0.01)
LABA/ICS vs LABA	Moderate	0.05 (0.03 to 0.07)	0.04 (0.03 to 0.06)	0.05 (0.03 to 0.07)
LAMA vs LABA	NA	NA	NA	0.08 (0.04 to 0.12)
Mortality, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) ORa (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	1.00 (0.57 to 1.77)	1.00 (0.57 to 1.77)	1.12 (0.75 to 1.68)
LABA/LAMA vs LAMA	Moderate	1.06 (0.66 to 1.69)	1.06 (0.66 to 1.69)	0.98 (0.66 to 1.42)
LABA/LAMA vs LABA	NA	NA	NA	0.97 (0.63 to 1.46)
LABA/ICS vs LAMA	Moderate	0.53 (0.31 to 0.90)	0.52 (0.31 to 0.89)	0.87 (0.65 to 1.16)
LABA/ICS vs LABA	Low	0.95 (0.69 to 1.30)	0.98 (0.73 to 1.33)	0.86 (0.66 to 1.11)
LAMA vs LABA	Moderate	0.87 (0.66 to 1.16)	0.87 (0.66 to 1.16)	0.99 (0.77 to 1.27)
Total SAEs, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) OR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.91 (0.76 to 1.08)	0.91 (0.76 to 1.08)	0.89 (0.77 to 1.02)
LABA/LAMA vs LAMA	Moderate	0.98 (0.80 to 1.20)	0.98 (0.80 to 1.20)	1.01 (0.87 to 1.17)
LABA/LAMA vs LABA	NA	NA	NA	0.89 (0.77 to 1.04)
LABA/ICS vs LAMA	Moderate	1.29 (1.03 to 1.63)	1.29 (1.03 to 1.63)	1.14 (1.02 to 1.27)
LABA/ICS vs LABA	High	0.99 (0.89 to 1.09)	0.99 (0.89 to 1.09)	1.01 (0.92 to 1.10)
LAMA vs LABA	Moderate	0.90 (0.81 to 1.00)	0.90 (0.81 to 1.00)	0.88 (0.81 to 0.97)
COPD SAEs high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) OR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.87 (0.70 to 1.07)	0.87 (0.70 to 1.07)	0.87 (0.73 to 1.04)
LABA/LAMA vs LAMA	Moderate	1.08 (0.84 to 1.39)	1.08 (0.84 to 1.39)	1.07 (0.89 to 1.28)
LABA/LAMA vs LABA	NA	NA	NA	0.82 (0.68 to 1.00)
LABA/ICS vs LAMA	Low	0.99 (0.33 to 2.96)	1.33 (0.99 to 1.79)	1.22 (1.05 to 1.42)
LABA/ICS vs LABA	Moderate	0.92 (0.78 to 1.07)	0.92 (0.79 to 1.07)	0.95 (0.83 to 1.08)
LAMA vs LABA	High	0.79 (0.69 to 0.91)	0.79 (0.69 to 0.91)	0.77 (0.68 to 0.87)
Cardiac SAEs, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(random‐effects/fixed‐class) OR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.86 (0.58 to 1.29)	0.86 (0.58 to 1.29)	0.7 (0.03 to 5.88)
LABA/LAMA vs LAMA	Low	0.80 (0.53 to 1.20)	0.80 (0.53 to 1.20)	0.69 (0.02 to 25.46)
LABA/LAMA vs LABA	NA	NA	NA	0.83 (0.06 to 9.24)
LABA/ICS vs LAMA	Moderate	0.67 (0.39 to 1.15)	0.67 (0.39 to 1.15)	1.08 (0.06 to 23.81)
LABA/ICS vs LABA	Very low	0.97 (0.68 to 1.38)	0.96 (0.75 to 1.22)	1.27 (0.37 to 5.97)
LAMA vs LABA	Low	1.09 (0.83 to 1.44)	1.09 (0.84 to 1.43)	1.13 (0.06 to 21.22)
Dropouts due to AEs, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(random‐effects/fixed‐class) OR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.88 (0.69 to 1.13)	0.88 (0.69 to 1.13)	0.93 (0.73 to 1.19)
LABA/LAMA vs LAMA	Low	1.03 (0.75 to 1.41)	1.03 (0.75 to 1.40)	0.95 (0.74 to 1.21)
LABA/LAMA vs LABA	NA	NA	NA	0.83 (0.65 to 1.07)
LABA/ICS vs LAMA	Moderate	1.04 (0.74 to 1.47)	1.04 (0.74 to 1.47)	1.02 (0.85 to 1.22)
LABA/ICS vs LABA	Low	0.88 (0.77 to 1.00)	0.88 (0.77 to 1.00)	0.89 (0.79 to 1.01)
LAMA vs LABA	High	0.91 (0.79 to 1.04)	0.91 (0.79 to 1.04)	0.88 (0.75 to 1.03)
Pneumonia, high‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class)OR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.62 (0.40 to 0.96)	0.62 (0.40 to 0.96)	0.59 (0.41 to 0.83)
LABA/LAMA vs LAMA	Moderate	0.98 (0.59 to 1.61)	0.98 (0.60 to 1.61)	1.05 (0.72 to 1.5)
LABA/LAMA vs LABA	NA	NA	NA	0.88 (0.6 to 1.29)
LABA/ICS vs LAMA	Moderate	1.80 (1.06 to 3.06)	1.82 (1.07 to 3.09)	1.78 (1.33 to 2.39)
LABA/ICS vs LABA	Moderate	1.46 (1.03 to 2.08)	1.51 (1.14 to 1.99)	1.50 (1.17 to 1.92)
LAMA vs LABA	Moderate	0.83 (0.61 to 1.13)	0.83 (0.62 to 1.12)	0.84 (0.65 to 1.09)
aPotential inconsistency in the date. Results should be interpreted with caution. AE: adverse event; CFB: change from baseline; HR: hazard ratio; FEV1: forced expiratory volume in one second;ICS: inhaled corticosteroid; LABA: long‐acting beta2‐agonist; LAMA: long‐acting muscarinic antagonist; MA: meta‐analysis; MD: mean difference; NA: not applicable; NMA: network meta‐analysis; OR: odds ratio; SAE: serious adverse event; SGRQ: St George's Respiratory Questionnaire; TDI: Transition Dyspnoea Index

Appendix 7. Summary of results for pairwise and network meta‐analyses in the low‐risk population

Moderate to severe exacerbations, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) HR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.86 (0.65 to 1.14)	0.84 (0.68 to 1.06)	0.87 (0.75 to 1.01)
LABA/LAMA vs LAMA	Low	0.93 (0.66 to 1.30)	0.94 (0.78 to 1.14)	0.90 (0.76 to 1.06)
LABA/LAMA vs LABA	Moderate	0.77 (0.62 to 0.97)	0.77 (0.62 to 0.96)	0.78 (0.67 to 0.90)
LABA/ICS vs LAMA	Low	0.63 (0.24 to 1.66)	0.63 (0.24 to 1.66)	1.03 (0.91 to 1.17)
LABA/ICS vs LABA	Moderate	0.83 (0.70 to 0.98)	0.85 (0.76 to 0.95)	0.89 (0.84 to 0.96)
LAMA vs LABA	Moderate	0.92 (0.79 to 1.07)	0.92 (0.79 to 1.07)	0.87 (0.78 to 0.97)
Severe exacerbations, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(random‐effects/fixed‐class) HR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	0.66 (0.27 to 1.63)	0.62 (0.33 to 1.19)	0.71 (0.47 to 1.08)
LABA/LAMA vs LAMA	Moderate	0.99 (0.57 to 1.72)	1.01 (0.65 to 1.55)	0.90 (0.6 to 1.31)
LABA/LAMA vs LABA	Moderate	0.78 (0.55 to 1.12)	0.78 (0.55 to 1.11)	0.72 (0.48 to 1.02)
LABA/ICS vs LAMA	Low	3.05 (0.32 to 29.47)	3.05 (0.32 to 29.47)	1.25 (0.86 to 1.85)
LABA/ICS vs LABA	High	1.06 (0.90 to 1.24)	1.06 (0.90 to 1.24)	1.01 (0.72 to 1.28)
LAMA vs LABA	Low	0.64 (0.36 to 1.13)	0.65 (0.41 to 1.03)	0.80 (0.56 to 1.05)
SGRQ responders at 3 months, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(fixed‐effect/fixed‐class) OR (95% CrI)
LABA/LAMA vs LABA/ICS	Moderate	1.08 (0.92 to 1.27)	1.08 (0.92 to 1.27)	1.07 (0.94 to 1.23)
LABA/LAMA vs LAMA	High	1.32 (1.16 to 1.51)	1.32 (1.17 to 1.49)	1.33 (1.19 to 1.48)
LABA/LAMA vs LABA	NA	NA	NA	0.96 (0.81 to 1.15)
LABA/ICS vs LAMA	Low	1.26 (0.92 to 1.74)	1.26 (0.92 to 1.74)	1.24 (1.07 to 1.43)
LABA/ICS vs LABA	Low	0.90 (0.73 to 1.11)	0.90 (0.73 to 1.11)	0.9 (0.76 to 1.06)
LAMA vs LABA	High	0.73 (0.59 to 0.89)	0.73 (0.59 to 0.89)	0.73 (0.62 to 0.85)
SGRQ responders at 6 months, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA(random‐effects/fixed‐class) OR (95% CrI)
LABA/LAMA vs LABA/ICS	Low	1.29 (0.88 to 1.89)	1.29 (0.88 to 1.89)	1.22 (0.99 to 1.51)
LABA/LAMA vs LAMA	Moderate	1.26 (1.15 to 1.37)	1.26 (1.15 to 1.37)	1.26 (1.1 to 1.42)
LABA/LAMA vs LABA	Low	1.20 (1.06 to 1.37)	1.20 (1.06 to 1.37)	1.28 (1.11 to 1.47)
LABA/ICS vs LAMA	NA	NA	NA	1.03 (0.83 to 1.27)
LABA/ICS vs LABA	Moderate	1.08 (0.96 to 1.22)	1.08 (0.96 to 1.22)	1.05 (0.87 to 1.25)
LAMA vs LABA	Low	1.02 (0.89 to 1.16)	1.02 (0.93 to 1.11)	1.02 (0.9 to 1.16)
SGRQ responders at 12 months, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects OR (95% CI)	Pairwise, fixed‐effect OR (95% CI)	NMA
LABA/LAMA vs LABA/ICS	NA	NA	NA	NA
LABA/LAMA vs LAMA	Moderate	1.13 (0.95 to 1.34)	1.13 (0.95 to 1.34)	NA
LABA/LAMA vs LABA	Moderate	1.19 (0.99 to 1.44)	1.19 (0.99 to 1.44)	NA
LABA/ICS vs LAMA	NA	NA	NA	NA
LABA/ICS vs LABA	Moderate	1.42 (1.18 to 1.70)	1.42 (1.18 to 1.70)	NA
LAMA vs LABA	Low	1.05 (0.88 to 1.26)	1.05 (0.88 to 1.26)	NA
CFB in SGRQ at 3 months, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD (95% CI)	Pairwise, fixed‐effect MD (95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	High	−0.03 (−1.02 to 0.96)	−0.03 (−1.02 to 0.96)	0.04 (−0.79 to 0.88)
LABA/LAMA vs LAMA	Moderate	−1.60 (−2.19 to −1.01)	−1.60 (−2.19 to −1.01)	−1.64 (−2.2 to −1.08)
LABA/LAMA vs LABA	Moderate	−1.29 (‐4.29 to 1.71)	−1.29 (‐4.29 to 1.71)	−0.63 (−1.86 to 0.6)
LABA/ICS vs LAMA	Moderate	−1.48 (‐3.41 to 0.45)	−1.48 (‐3.41 to 0.45)	−1.68 (−2.59 to −0.78)
LABA/ICS vs LABA	High	−1.00 (−2.61 to 0.61)	−1.00 (−2.61 to 0.61)	−0.67 (−1.88 to 0.54)
LAMA vs LABA	Moderate	1.84 (0.87 to 2.80)	1.84 (0.87 to 2.80)	1.01 (−0.2 to 2.22)
CFB in SGRQ at 6 months, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD (95% CI)	Pairwise, fixed‐effect MD (95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	Low	−0.99 (‐4.12 to 2.14)	−0.99 (‐4.12 to 2.14)	−0.22 (−1.28 to 0.82)
LABA/LAMA vs LAMA	Moderate	−1.20 (−1.83 to −0.57)	−1.20 (−1.83 to −0.57)	−1.18 (−1.8 to −0.56)
LABA/LAMA vs LABA	Moderate	−1.09 (−1.96 to −0.22)	−1.09 (−1.96 to −0.22)	−1.36 (−2.12 to −0.60)
LABA/ICS vs LAMA	NA	NA	NA	−0.96 (−1.98 to 0.09)
LABA/ICS vs LABA	Moderate	−1.18 (−1.97 to −0.40)	−1.18 (−1.97 to −0.40)	−1.14 (−1.90 to −0.37)
LAMA vs LABA	High	−0.25 (−1.09 to 0.58)	−0.23 (−0.99 to 0.54)	−0.18 (−0.91 to 0.55)
CFB in SGRQ at 12 months, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD (95% CI)	Pairwise, fixed‐effect MD (95% CI)	NMA(fixed‐effect/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	NA	NA	NA	0.97 (0.48 to 2.42)
LABA/LAMA vs LAMA	Very low	−0.87 (−1.64 to −0.10)	−0.87 (−1.64 to −0.10)	−0.89 (−1.66 to −0.11)
LABA/LAMA vs LABA	High	−0.69 (−1.64 to 0.25)	−0.69 (−1.64 to 0.25)	−0.72 (−1.64 to 0.20)
LABA/ICS vs LAMA	NA	NA	NA	−1.85 (−3.28 to −0.43)
LABA/ICS vs LABA	Moderate	−1.70 (−2.82 to −0.58)	−1.70 (−2.82 to −0.58)	−1.69 (−2.81 to −0.57)
LAMA vs LABA	High	0.10 (−0.79 to 0.99)	0.10 (−0.79 to 0.99)	0.16 (−0.72 to 1.04)
TDI at 3 months, low‐risk	Certainty of evidence in the pairwise MA	Pairwise, random‐effects MD (95% CI)	Pairwise, fixed‐effect MD (95% CI)	NMA(random‐effects/fixed‐class) MD (95% CrI)
LABA/LAMA vs LABA/ICS	Low	0.40 (0.02 to 0.78)	0.36 (0.16 to 0.56)	0.35 (0.12 to 0.56)
LABA/LAMA vs LAMA	Moderate	0.48 (0.34 to 0.62)	0.48 (0.34 to 0.62)	0.54 (0.36 to 0.73)
LABA/LAMA vs LABA	Low	0.52 (0.31 to 0.74)	0.52 (0.31 to 0.74)	0.44 (0.20 to 0.67)
LABA/ICS vs LAMA	Very low	0.51 (−0.39 to