Types of studies

We included individual and clustered randomised trials that compared intervention to usual care.

We included full‐text studies, conference abstracts and unpublished data. The minimum follow‐up period for included studies was six months, as this is the minimum amount of time until impact on outcomes such as nursing home admission should be assessed.

We excluded the following types of studies.

• Studies that focused solely on a single particular disease or syndrome (e.g. heart failure, falls, stroke).

• Studies of interventions after discharge from hospital.

• Studies designed to test hospital avoidance in exacerbations of chronic conditions.

• Studies involving participants who were not community‐dwelling.

Types of participants

We included participants aged 65 years or older (or 55 years or older if the mean age of study participants was over 70 years) who satisfied each of the following criteria.

• Community dwelling.

• Not acutely unwell (i.e. not currently an inpatient in an acute hospital and not presenting to an emergency department or general practitioner (GP) for unscheduled care).

• Identified as at risk of nursing home admission or defined as frail.

Community dwelling was defined as living outside an environment where 24‐hour nursing care was provided onsite. This encompassed participants living in their own home (with or without assistance), in a relative's home, or in a retirement village or sheltered accommodation but excluded participants living in nursing or care homes or full‐time residential care.

Frailty is defined as a vulnerability to sudden health state changes triggered by relatively minor stressor events (Clegg 2011), and was measured by an objective scale/screening tool in included studies.

Types of interventions

We included trials that compared CGA meeting the following criteria with usual care.

• Delivered by a healthcare professional with gerontological expertise. This included a geriatrician, specialist nurse or therapist with gerontological expertise and must have been explicitly stated in the study methodology.

• Used to inform a holistic care plan.

• A single assessment or multiple visits.

• Delivered in a community setting (i.e. participant's home, general practice, community‐based clinic, etc.).

Traditionally CGA requires a minimum number of disciplines to be involved within the team (Ellis 2017). Given the community‐based nature of the intervention involved in this Cochrane Review, one healthcare professional with an expertise in geriatric medicine, and, therefore, multidisciplinary work and links, was considered sufficient.

CGA was either provided in the participant's own home (i.e. dCGA) or in a community setting other than the participant's home (i.e. cCGA).

Usual care was defined as the current standard care received by frail, older people where the study is carried out. This generally involved usual care by the GP or family doctor in the community, with CGA provided only when the older person was admitted to the acute hospital or when they were referred by their GP to a geriatric medicine clinic with issues such as functional decline, cognitive decline, falls, etc.

Types of outcome measures

Electronic searches

The Information Specialist of the Cochrane EPOC Group developed the search strategies in consultation with the review authors. We searched the Cochrane Database of Systematic Reviews (CDSR) and the Database of Abstracts of Reviews of Effects (DARE) for related systematic reviews.

We searched the following databases (from inception) for primary studies on 28 April 2020.

• Cochrane Central Register of Controlled Trials (CENTRAL; 2020, Issue 4) in the Cochrane Library.

• MEDLINE, Ovid (including Epub ahead of print, in‐process and other non‐indexed citations, 1946 onwards).

• Embase, Ovid (1974 onwards).

• CINAHL, EBSCO (Cumulative Index to Nursing and Allied Health Literature; 1980 onwards).

Search strategies comprised natural language and controlled vocabulary terms. We applied no language limits. We used a study design filter to identify randomised trials. All strategies used are provided in Appendix 1.

Searching other resources

Selection of studies

We downloaded all titles and abstracts retrieved by electronic searching to a reference management database and removed duplicates. Two review authors (RB and AM) independently screened titles and abstracts for inclusion. We retrieved the full‐text study reports/publications. Two review authors (RB and AM) independently screened the full texts and identified studies for inclusion, and identified and recorded reasons for exclusion of the ineligible studies. We resolved any disagreements through discussion or, if required, consulted a third review author (DR). We listed any studies that initially appeared to meet the inclusion criteria but that we later excluded, and their reasons for exclusion, in the Characteristics of excluded studies table. We collated multiple reports of the same study, so that each study rather than each report, was the unit of interest in the review. We also provided any information we could obtain about ongoing studies. We recorded the selection process in sufficient detail to complete a PRISMA flow diagram (Liberati 2009), and Characteristics of included studies table.

Data extraction and management

We used a standard data collection form adapted from the EPOC data collection form, described in the EPOC‐specific resources for review authors (EPOC 2013), for study characteristics and outcome data. Two review authors (RB and AM) independently extracted the following study characteristics from the included studies.

• Methods: year, location, study personnel.

• Participants: number, mean age, sex, inclusion and exclusion criteria.

• Interventions: description of interventions.

• Outcomes: main and other outcomes specified and collected, time points reported.

• Notes: funding for trial, notable conflicts of interest of trial authors, ethical approval.

Two review authors (RB and AM) independently extracted outcome data from included studies. If an included study reported outcome data in an unusable way, we noted this in the Characteristics of included studies table. We resolved disagreements in extracted data by consensus or by involving a third review author (DR) according to the Cochrane Handbook for Systematic Reviews of Interventions (Section 7.6.5; Higgins 2011a).

RB and AM repeated data extraction and reconciliation for all studies during peer review.

Assessment of risk of bias in included studies

Two review authors (RB and AM) independently assessed the risk of bias for each included study using the criteria outlined in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011b). We resolved any disagreements by discussion or by involving a third review author (DR). We assessed the risk of bias according to the following domains.

• Random sequence generation.

• Allocation concealment.

• Blinding of participants and personnel.

• Blinding of outcome assessment.

• Incomplete outcome data.

• Selective outcome reporting.

• Baseline characteristics.

• Other bias.

We judged 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 the risk of bias judgements across different studies for each of the domains listed. We considered blinding separately for different key outcomes where necessary. We did not exclude studies on the basis of their risk of bias, but we clearly reported the risk of bias when presenting the results of the studies.

Studies were defined as high risk of bias based for unbalanced baseline characteristics when baseline differences were considered likely to impact on outcomes of interest, allowing for adjustments made in the study analysis.

When we considered treatment effects, we took into account the risk of bias for the studies that contributed to that outcome.

For cluster‐randomised trials, we considered the following additional factors when considering risk of bias.

• Bias arising from the randomisation process, including bias arising from the recruitment of participants into clusters.

• Bias due to deviation from intended intervention.

• Bias due to missing outcome data, including loss of clusters.

• Bias in measurement of outcome, including whether outcome assessors were aware participants were in a trial.

Measures of treatment effect

We estimated the effect of the intervention using the risk ratio (RR) for dichotomous data, together with the appropriate associated 95% confidence interval (CI) and mean difference (MD) or standardised mean difference (SMD) for continuous data, together with the 95% appropriate associated CI. We ensured that an increase in scores for continuous outcomes was interpreted in the same way for each outcome, explaining the direction to the reader and reporting where we reversed the directions if this was necessary. When both change scores and postintervention outcomes were available, we used postintervention outcomes as long as there was no significant difference in the continuous variable of interest at baseline.

Unit of analysis issues

We included cluster randomised trials in the review. We attempted to obtain a direct estimate of the required effect measure from the cluster randomised trials (e.g. an RR with CI) if the analysis properly accounted for the cluster design. If this was not available, we extracted the following data from cluster trials:

• mean size of each cluster;

• outcome data ignoring cluster design.

This was combined with an estimated intracluster correlation coefficient (ICC) from similar studies to ascertain a design effect and an effective sample size, which were then incorporated into supplementary analyses.

The ICC estimates were 0.001 for death (Kul 2014), 0.007 for nursing home admission (Buys 2013), 0.004 for unplanned hospital admission (Kul 2014), 0.003 for emergency department visits (Howard 2008), 0.007 for serious adverse events (falls) (Cumming 2008), 0.05 for change in function (Metzelthin 2013), and 0.002 for QoL (Hoogendijk 2016).

We based the methods we used to include cluster randomised trials on guidance from the Cochrane Handbook for Systematic Reviews of Interventions (Deeks 2011).

For studies with a stepped wedge cluster design, we included participants who remained either in the control or intervention arm for 12 months. This is due to the nature of CGA, which aims to have benefits relatively far beyond the time it is applied (e.g. long‐term prevention of nursing home admission or hospital admission), as well as the outcomes of interest. Therefore, we only included participants who did not 'cross‐over' up to 12 months.

Dealing with missing data

If data were missing, we contacted the study authors to request them. If it was not possible to retrieve complete data, we reported this in the risk of bias assessment and addressed missing outcomes and summary data as a source of bias in the data analyses. We contacted study investigators to verify key study characteristics and obtain missing outcome data where possible.

If standard deviation (SD) values were missing, we used any other available information (CIs, standard errors) to derive SD values. If this information was not available, we contacted the study authors to request the missing information. If we were unable to obtain any information, we imputed the SD values using the available information.

As per the guidance in Chapter 16 of the Cochrane Handbook for Systematic Reviews of Interventions, we assumed that missing data were missing at random (Lefebvre 2011).

Assessment of heterogeneity

We used the I² statistic to measure heterogeneity among the trials in each analysis. If we identified significant heterogeneity (i.e. I² values greater than 75%; Higgins 2003), we examined the subgroup analyses (high versus low risk of bias) to determine if this explained the heterogeneity (Deeks 2011).

Assessment of reporting biases

We assessed whether study authors published a study protocol detailing study methodology and outcomes in advance of the index study.

Data synthesis

If we found a sufficient number of studies (at least five), we conducted a meta‐analysis. We undertook meta‐analyses only where it was meaningful (i.e. if the treatments, participants and the underlying clinical question were similar enough for pooling to make sense). Where a trial reported multiple arms, we included only the relevant trial arms.

In the absence of between‐study heterogeneity (I² = 0), we used fixed‐effect estimates; however, when I² > 0, we used random‐effects estimates.

Subgroup analysis and investigation of heterogeneity

At the protocol stage, we had planned to perform subgroup analyses of CGA alone versus CGA with additional structured intervention, but this was not practical as all but one study (Clarkson 2006) involved structured interventions afterwards. Further, we had planned a subgroup analysis analysing single assessment versus multiple visits but again, all but one study (Clarkson 2006) involved at least one or more episodes of contact with the participant so this subgroup analysis was not conducted.

Sensitivity analysis

We applied the Cochrane RoB 1 tool (Deeks 2011), and selected the domains of allocation concealment (selection bias), incomplete outcome data (attrition bias) and blinding of outcome assessment (detection bias) as priorities for defining studies as high or low risk of bias as bias in these domains was most likely to exaggerate the effect seen in the studies. We defined 'low risk' as being at low risk of bias for at least two of the three domains, as long as the other domain was not considered at 'high risk'. We defined 'high risk' as having one or more of these domains considered at 'high risk' or having one or two of the three domains considered 'low risk'.

We performed sensitivity analysis defined a priori to assess the robustness of our conclusions and explore its impact on effect sizes. This involved restriction of the analysis to studies at low risk of bias.

We had also planned at the protocol stage to perform sensitivity analysis restricting the analysis to published studies only but this was not possible as all included studies were published. Additionally, we had planned a sensitivity analysis for imputation of missing data but no data was imputed.

Summary of findings and assessment of the certainty of the evidence

We summarised the findings of the main intervention comparison for the study outcomes (death, nursing home admission, unplanned hospital admission, emergency department visits and serious adverse events) in a summary of findings table to draw conclusions about the certainty of the evidence within the text of the review.

Two review authors independently assessed the certainty of the evidence (high, moderate, low or very low) using the five GRADE considerations (study limitations, consistency of effect, imprecision, indirectness and publication bias). We used methods and recommendations described in Section 8.5 and Chapter 12 of the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011b; Schünemann 2011), and the Cochrane EPOC worksheets (EPOC 2016).

Selective reporting of outcomes is assessed at the study level as part of the assessment of risk of bias, so for the studies contributing to the outcome in the summary of findings table, this was addressed by study limitations. We used the GRADE approach for assessing the certainty of evidence (Higgins 2011c), and resolved disagreements on the assessments of the certainty of the evidence by discussion. We provided justification for our decisions to downgrade or upgrade the certainty of the evidence using footnotes in the summary of findings tables. We used plain language statements to report these findings in the review.

We considered whether there was any additional outcome information that we were unable to incorporate into the meta‐analyses. We noted this in the comments and stated if it supported or contradicted the information from the meta‐analyses. If it was not possible to meta‐analyse the data, we summarised the results in the text.