Types of studies

We will only include randomised controlled trials of postoperative patients from all surgical specialties, including emergency and elective surgery, receiving a CGA intervention compared with a control group receiving standard care. To reduce the likelihood of publication bias, we will not limit included articles to the English language, and we will screen articles from trial databases and the grey literature for eligibility.

Types of participants

The focus of this review will be people age 65 years or older in hospital under the care of an inpatient surgical ward. Although there is not a standard numerical criteria to define old age, 65 years old is widely accepted as the chronological age to be considered an older person.

People admitted to hospital for elective or emergency surgery, or for an acute medical condition or injury requiring close observation and expectant management by a surgical team, will be eligible for inclusion in the analysis.

Studies containing a subset of surgical patients above the age of 65 will be eligible for inclusion; however, only study data pertaining to our population of interest will be included in the meta‐analysis.

Types of interventions

We will include studies in which a geriatrician, internist, hospitalist, or geriatric nurse has performed a multicomponent geriatric assessment in hospital, and in which patients receiving the intervention were compared with patients receiving standard postoperative care. The CGA intervention can be performed as part of a mobile, multidisciplinary team consulted to provide patient management recommendations, or as part of a specialised ward dedicated to providing multidisciplinary care to geriatric surgical patients. The CGA intervention may be carried out preoperatively, postoperatively, or throughout the patient's stay in hospital.

We will exclude studies in which CGA was used only as a tool to predict adverse postoperative events. We will also exclude studies examining only one aspect of the CGA instead of employing a multidimensional assessment, and we will exclude cross‐over studies. We will exclude enhanced recovery after surgery programmes because CGA is not a routine component of these programmes.

Types of outcome measures

Electronic searches

We will use a sensitive search strategy designed to retrieve studies from electronic databases. We will search the following databases, with publication dates ranging from inception to search date.

• Cochrane Central Register of Controlled Trials (CENTRAL), including the Cochrane Effective Practice and Organisation of Care (EPOC) Group Specialised Register, part of the Cochrane Library (www.cochranelibrary.com)

• MEDLINE In‐Process & Other Non‐Indexed Citations, OvidSP (1946 ‐ )

• Embase, OvidSP (1974 ‐ )

• PsycINFO, OvidSP (1987‐ )

• CINAHL (Cumulative Index to Nursing and Allied Health Literature), EBSCO (1980 ‐ )

The search terms will combine Medical Subject Headings (MeSH) and free text words as shown in the MEDLINE strategy in Appendix 1. The MEDLINE strategy will be translated using appropriate syntax and controlled vocabulary headings for other databases. We will not place any restrictions on language, publication type, or publication year.

Searching other resources

We will conduct a grey literature search to identify non‐indexed studies not appearing in the databases listed above. Sources will include:

• World Health Organization International Clinical Trials Registry Platform (WHO ICTRP) (www.who.int/ictrp/en/); and

• US National Institutes of Health Ongoing Trials Register ClinicalTrials.gov (clinicaltrials.gov).

We will use Science Citation Index to search the cited and citing articles of included studies.

Selection of studies

Two review authors will screen titles and abstracts to identify potentially eligible articles for full‐text review. We will assess potential eligibility based on design, participants, intervention, and outcomes as described. We will exclude studies that do not meet the inclusion criteria at this stage. Two review authors will independently carry out full‐text review. We will resolve conflicts between review authors at all stages of article screening and data extraction by discussion and consensus. We will report the number of studies excluded in the title and abstract review and the full‐text review. We will report articles that appeared relevant to the review question, but were excluded from analysis, with the reasons for exclusion as per Section 7.2.5 of the Cochrane Handbook (Higgins 2011).

Data extraction and management

Two review authors will independently extract data onto web‐based electronic data collection forms (Covidence.org). Disagreements between review authors will be resolved by discussion and consensus. We will translate included articles not published in English using available software prior to data extraction. We will then export data to Review Manager 5 for analysis (RevMan 5.3).

During data extraction, review authors will take note of the study source, eligibility, methods, participants, interventions, outcomes of interest, results, and other information as defined in Table 7.3.a of the Cochrane Handbook, in Higgins 2011, and the EPOC good‐practice data extraction form (EPOC 2015a). We will report all costs in euros.

Assessment of risk of bias in included studies

Two review authors will independently assess risk of bias using Cochrane's 'Risk of bias' tool (Higgins 2011), with modifications based on the EPOC guidance for risk of bias criteria (EPOC 2015b). We will evaluate each study using a bias assessment tool, rating each of the following criteria for each study as low risk, high risk, or uncertain risk.

1. Random sequence generation ‐ was the allocation sequence adequately generated.

2. Allocation concealment ‐ was allocation concealment adequate.

3. Baseline demographics between groups ‐ were baseline outcomes measured before the intervention and were they similar between groups.

4. Incomplete data ‐ were loss to follow‐up or dropouts low enough to limit risk of bias.

5. Blinding of participants and personnel ‐ were participants and personnel blind to the intervention.

6. Blinding of outcome assessment ‐ were outcome assessors blind to the intervention.

7. Protection from cross‐contamination ‐ were there safeguards to cross‐contamination of the control group.

8. Selective reporting ‐ were all outcomes in the methods reported in the results.

9. Other risks of bias ‐ were any additional risks noted during bias assessment.

Measures of treatment effect

We will report dichotomous outcome data such as the effect of CGA on patient mortality and discharge destination as risk ratios with 95% confidence intervals. We will report continuous outcome data such as the effect of CGA on length of stay using the mean difference between the CGA intervention and standard care with a 95% confidence interval. For all continuous‐variable outcomes, we will report the mean and standard deviations or standard error of the outcome measurements in each intervention group, as well as the number of participants on which the outcome was measured.

Unit of analysis issues

We will perform analyses at the participant level to avoid unit of analysis errors. If we identify cluster randomised controlled trials, we will use a ratio estimator approach to reduce the size of each cluster trial to its effective sample size (Rao 1992), which is its original sample size divided by design effect. The design effect is 1 + (M – 1) ICC, where M is the average cluster size and ICC is the intracluster correlation coefficient. For dichotomous data, the number of participants and the number of events will be divided by the design effect. For continuous data, the sample size will be divided by the design effect. Missing ICCs may be selected from other cluster randomised controlled trials included in the review or obtained from similar external studies. We will conduct sensitivity analyses to investigate whether removing clustered trials affects the conclusions.

If the results of a study cannot be adjusted for the unit of analysis error, we will exclude it from the pooled analysis. We will pool data based on time since admission to discharge and end of follow‐up as predefined outcome measurement points.

Dealing with missing data

Where feasible, we will obtain missing data from authors. We will investigate attrition rates (e.g. dropouts, losses to follow‐up, and withdrawals), and critically appraise issues of missing data and imputation methods (e.g. last observation carried forward). Where standard deviations for outcomes are not reported, we will impute these values by assuming the standard deviation of the missing outcome to be the average of the standard deviations from those studies where this information was reported. We will investigate the impact of imputation on meta‐analyses by means of sensitivity analysis.

Assessment of heterogeneity

Where studies are considered similar enough based on population, study design, and setting to allow pooling of data using meta‐analysis, we will assess the degree of heterogeneity by visual inspection of forest plots and by examining the Chi² test for heterogeneity. We will quantify heterogeneity between studies using the I² test. An I² of less than 40% will be considered unimportant; 40% to 60% may indicate moderate heterogeneity; 60% to 75% may indicate substantial heterogeneity; and 75% to 100% will indicate considerable heterogeneity. Where we detect substantial clinical, methodological, or statistical heterogeneity across included studies, we will not retain the pooled results from meta‐analysis but will instead use a narrative approach to data synthesis.

Assessment of reporting biases

Where information from 10 or more studies is available, we will generate a funnel plot to look for asymmetry. Where asymmetry exists, we will assess the reason for asymmetry based on Section 10.4 of the Cochrane Handbook (Higgins 2011).

We will assess publication bias by searching trial registries and the grey literature. For studies published after 1 July 2005, we will note lack of registration of the trial protocol with the WHO ICTRP in the 'Risk of bias' table. We will also note selective reporting of predefined outcomes.

Data synthesis

We will compare random‐effects and fixed‐effect models to assess if smaller studies affect the results. Given the complex and multidimensional nature of CGA, variation is expected in measured outcomes due to sampling error and differing patterns of implementation of CGA. If there is a difference between fixed‐effect and random‐effects models, we will assess the impact of small studies on the estimate of effect before deciding which model to use.

Subgroup analysis and investigation of heterogeneity

We will conduct subgroup analysis for the a priori defined variables listed below.

1. Orthopaedic versus other surgical specialties.

2. CGA timing ‐ is the CGA conducted preoperatively, postoperatively, or throughout an admission.

3. Emergency versus elective surgery.

We will analyse these subgroups at discharge and end of follow‐up. We will determine if the subgroups differ significantly by inspecting the overlap of confidence intervals and testing for subgroup differences using Review Manager 5 (RevMan 5.3).

Timing of the CGA in relation to surgery could affect patient outcomes because the potential benefits of CGA intervention could arise from optimising patient medical and social issues before surgery; by providing a better level of care following surgery; or both pre‐ and postoperative intervention may be necessary to see benefits. Most studies of CGA in surgical patients have been performed in orthopaedic trauma (hip fracture); the effect of CGA may play an important role in recuperation from hip surgery but not in other surgical interventions or populations. Finally, elective versus emergency surgery can give rise to different risk profiles. Determining if there is a benefit in one population versus another is important.

Sensitivity analysis

We will perform sensitivity analysis to explore changes in effect size after removing studies with a high risk of bias and comparing the use of a fixed‐effect and random‐effects model.