Diagnostic criteria for diabetes mellitus

In order to be consistent with changes in the classification of and diagnostic criteria for diabetes mellitus over the years, the diagnosis should be established using the standard criteria valid at the time of the trial commencing (for example ADA 2003; ADA 2008; WHO 1998). Ideally, the diagnostic criteria should have been described. We will use the trial authors' definition of diabetes mellitus if necessary. We plan to subject diagnostic criteria to a sensitivity analysis.

Intervention

• Metformin plus sulphonylurea (second‐ or third‐generation) combination therapy.

Comparator

• Metformin plus another glucose‐lowering intervention as a combination therapy (e.g. metformin plus dipeptidylpeptidase‐4 inhibitor, metformin plus insulin).

• Metformin plus placebo.

• Metformin monotherapy.

Concomitant interventions will have to be the same in both the intervention and comparator groups to establish fair comparisons.

If a trial includes multiple arms, we will include any arm that meets the review's inclusion criteria.

Minimum duration of intervention

We will include trials with a minimum duration of intervention of 52 weeks. Because we are primarily interested in patient‐important outcomes, we focus on longer‐term trials due to the fact that macrovascular and microvascular complications develop over time.

Minimum duration of follow‐up

The minimum duration of follow‐up will be 52 weeks. Extended follow‐up periods (also called open‐label extension studies) defined as a follow‐up of participants once the original trial was terminated, as specified in the power calculation for this trial, are frequently of an observational nature and will only be evaluated for adverse events (Buch 2011; Megan 2012).

Summary of specific exclusion criteria

• We will exclude combinations of more than two glucose‐lowering agents.

• We will exclude people diagnosed with gestational diabetes.

Primary outcomes

• All‐cause mortality

• Health‐related quality of life

• Serious adverse events

Secondary outcomes

• Cardiovascular mortality

• Non‐fatal myocardial infarction

• Heart failure

• Non‐fatal stroke

• Amputation of lower extremity

• Blindness or severe vision loss

• End‐stage renal disease

• Non‐serious adverse events

• Hypoglycaemia

• Socio‐economic effects

Dealing with duplicate and companion publications

In the event of duplicate publications, companion documents or multiple reports of a primary trial, we will maximise the information yield by collating all available data and will use the most complete dataset aggregated across all known publications. We will list duplicate publications, companion documents, multiple reports of a primary trial and trial documents of included trials (such as trial registry information) as secondary references under the study identifier (ID) of the included trial. Furthermore, we will also list duplicate publications, companion documents, multiple reports of a trial and trial documents of excluded trials (such as trial registry information) as secondary references under the study ID of the excluded trial.

Data from clinical trial registers

In case data of included trials are available as study results in clinical trial registers such as ClinicalTrials.gov, we will make full use of this information and extract data. If there is also a full publication of the trial, we will collate and critically appraise all available data. If an included trial is marked as a completed study in a clinical trial register but no additional information is available, we will add this trial to the table 'Characteristics of studies awaiting classification'.

Random sequence generation (selection bias due to inadequate generation of a randomised sequence) ‐ assessment at trial level

We will describe for each included trial the method used to generate the allocation sequence in sufficient detail to allow an assessment of whether it should produce comparable groups.

• Low risk of bias: sequence generation was achieved using computer random number generation or a random number table. Drawing of lots, tossing a coin, shuffling cards or envelopes, and throwing dice are adequate if performed by an independent person not otherwise involved in the trial. Use of the minimisation technique will be considered as equivalent to being random.

• Unclear risk of bias: insufficient information about the sequence generation process.

• High risk of bias: the sequence generation method was non‐random (e.g. sequence generated by odd or even date of birth; sequence generated by some rule based on date (or day) of admission; sequence generated by some rule based on hospital or clinic record number; allocation by judgement of the clinician; allocation by preference of the participant; allocation based on the results of a laboratory test or a series of tests; allocation by availability of the intervention).

Allocation concealment (selection bias due to inadequate concealment of allocation prior to assignment) ‐ assessment at trial level

We will describe for each included trial the method used to conceal allocation to interventions prior to assignment and will assess whether intervention allocation could have been foreseen in advance of, or during recruitment, or changed after assignment.

• Low risk of bias: central allocation (including telephone, interactive voice‐recorder, web‐based and pharmacy‐controlled randomisation); sequentially‐numbered drug containers of identical appearance; sequentially‐numbered, opaque, sealed envelopes.

• Unclear risk of bias: insufficient information about the allocation concealment.

• High risk of bias: using an open random allocation schedule (e.g. a list of random numbers); assignment envelopes were used without appropriate safeguards; alternation or rotation; date of birth; case record number; any other explicitly unconcealed procedure.

We will also evaluate trial baseline data to incorporate assessment of baseline imbalance into the 'Risk of bias' judgement for selection bias (Corbett 2014). Chance imbalances might also affect judgements on the risk of attrition bias. In case of unadjusted analyses we will distinguish between studies rated as at low risk of bias on the basis of both randomisation methods and baseline similarity, and studies rated as at low risk of bias on the basis of baseline similarity alone (Corbett 2014). We will reclassify judgements of unclear, low or high risk of selection bias as specified in Appendix 2.

Blinding of participants and study personnel (performance bias due to knowledge of the allocated interventions by participants and personnel during the trial) ‐ assessment at outcome level

We will evaluate the risk of detection bias separately for each outcome (Hróbjartsson 2013). We will note whether outcomes were self‐reported, investigator‐assessed or adjudicated outcome measures (see below).

• Low risk of bias: blinding of participants and key study personnel ensured, and unlikely that the blinding could have been broken; no blinding or incomplete blinding, but the review authors judge that the outcome is not likely to be influenced by lack of blinding.

• Unclear risk of bias: insufficient information about the blinding of participants and study personnel; the trial did not address this outcome.

• High risk of bias: no blinding or incomplete blinding, and the outcome is likely to be influenced by lack of blinding; blinding of trial participants and key personnel attempted, but likely that the blinding could have been broken, and the outcome is likely to be influenced by lack of blinding.

Blinding of outcome assessment (detection bias due to knowledge of the allocated interventions by outcome assessment) ‐ assessment at outcome level

We will evaluate the risk of detection bias separately for each outcome (Hróbjartsson 2013). We will note whether outcomes were self‐reported, investigator‐assessed or adjudicated outcome measures (see below).

• Low risk of bias: blinding of outcome assessment ensured, and unlikely that the blinding could have been broken; no blinding of outcome assessment, but the review authors judge that the outcome measurement is not likely to be influenced by lack of blinding.

• Unclear risk of bias: insufficient information about the blinding of outcome assessors; the trial did not address this outcome.

• High risk of bias: no blinding of outcome assessment, and the outcome measurement is likely to be influenced by lack of blinding; blinding of outcome assessment, but likely that the blinding could have been broken, and the outcome measurement is likely to be influenced by lack of blinding.

Incomplete outcome data (attrition bias due to amount, nature or handling of incomplete outcome data) ‐ assessment at outcome level

We will describe for each included trial, and for each outcome, the completeness of data including attrition and exclusions from the analysis. We will state whether attrition and exclusions were reported and the number included in the analysis at each stage (compared with the number of randomised participants per intervention/comparator groups), if reasons for attrition or exclusion were reported, and whether missing data were balanced across groups or were related to outcomes. We will consider the implications of missing outcome data per outcome such as high drop‐out rates (e.g. above 15%) or disparate attrition rates (e.g. difference of 10% or more between trial arms).

• Low risk of bias: no missing outcome data; reasons for missing outcome data unlikely to be related to true outcome (for survival data, censoring unlikely to be introducing bias); missing outcome data balanced in numbers across intervention groups, with similar reasons for missing data across groups; for dichotomous outcome data, the proportion of missing outcomes compared with observed event risk not enough to have a clinically‐relevant impact on the intervention effect estimate; for continuous outcome data, plausible effect size (difference in means or standardised difference in means) among missing outcomes not enough to have a clinically‐relevant impact on observed effect size; appropriate methods, such as multiple imputation, were used to handle missing data.

• Unclear risk of bias: insufficient information to assess whether missing data in combination with the method used to handle missing data were likely to induce bias; the trial did not address this outcome.

• High risk of bias: reason for missing outcome data likely to be related to true outcome, with either imbalance in numbers or reasons for missing data across intervention groups; for dichotomous outcome data, the proportion of missing outcomes compared with observed event risk enough to induce clinically‐relevant bias in intervention effect estimate; for continuous outcome data, plausible effect size (difference in means or standardised difference in means) among missing outcomes enough to induce clinically‐relevant bias in observed effect size; ‘as‐treated’ or similar analysis done with substantial departure of the intervention received from that assigned at randomisation; potentially inappropriate application of simple imputation.

Selective reporting (reporting bias due to selective outcome reporting) ‐ assessment at trial level

We will assess outcome reporting bias by integrating the results of the appendix 'Matrix of trial outcomes (publications and trial documents)' (Boutron 2014; Mathieu 2009), with those of the appendix 'High risk of outcome reporting bias according to ORBIT classification' (Kirkham 2010). This analysis will form the basis for the judgement of selective reporting.

• Low risk of bias: the trial protocol is available and all of the trial’s prespecified (primary and secondary) outcomes that are of interest in the review have been reported in the prespecified way; the study protocol is not available but it is clear that the published reports include all expected outcomes (ORBIT classification).

• Unclear risk of bias: insufficient information about selective reporting.

• High risk of bias: not all of the trial’s prespecified primary outcomes have been reported; one or more primary outcomes is reported using measurements, analysis methods or subsets of the data (e.g. subscales) that were not prespecified; one or more reported primary outcomes were not prespecified (unless clear justification for their reporting is provided, such as an unexpected adverse effect); one or more outcomes of interest in the review are reported incompletely so that they cannot be entered in a meta‐analysis; the trial report fails to include results for a key outcome that would be expected to have been reported for such a trial (ORBIT classification).

Other bias (bias due to problems not covered elsewhere) ‐ assessment at trial level

• Low risk of bias: the trial appeared to be free of other sources of bias.

• Unclear risk of bias: insufficient information to assess whether an important risk of bias existed; insufficient rationale or evidence that an identified problem introduced bias.

• High risk of bias: had a potential source of bias related to the specific trial design used; has been claimed to have been fraudulent; had some other serious problem.

We will present a 'Risk of bias' graph and a 'Risk of bias' summary figure.

We will distinguish between self‐reported, investigator‐assessed and adjudicated outcome measures.

We define the following outcomes as self‐reported.

• Health‐related quality of life.

• Non‐serious adverse events.

• Hypoglycaemia, if reported by participants.

• Weight, if reported by participants.

We define the following outcomes as investigator‐assessed.

• All‐cause mortality.

• Serious adverse events.

• Cardiovascular mortality.

• Non‐fatal myocardial infarction.

• Heart failure.

• Non‐fatal stroke.

• Amputation of lower extremity.

• Blindness or severe vision loss.

• End‐stage renal disease.

• Hypoglycaemia, if measured by trial personnel.

• Socio‐economic effects.

• Weight, if measured by trial personnel.

• HbA1c.

Summary assessment of risk of bias

Risk of bias for a trial across outcomes: some risk of bias domains, such as selection bias (sequence generation and allocation sequence concealment), affect the risk of bias across all outcome measures in a trial. In case of high risk of selection bias, all endpoints investigated in the associated trial will be marked as high risk. Otherwise, we will not perform a summary assessment of the risk of bias across all outcomes for a trial.

Risk of bias for an outcome within a trial and across domains: we will assess the risk of bias for an outcome measure by including all entries relevant to that outcome (i.e. both trial‐level entries and outcome‐specific entries). We consider low risk of bias to denote a low risk of bias for all key domains, unclear risk to denote an unclear risk of bias for one or more key domains and high risk to denote a high risk of bias for one or more key domains.

Risk of bias for an outcome across trials and across domains: these are our main summary assessments that we will incorporate into our judgements about the quality of evidence in the 'Summary of findings' tables. We define outcomes as at low risk of bias when most information comes from trials at low risk of bias, unclear risk when most information comes from trials at low or unclear risk of bias, and high risk when a sufficient proportion of information comes from trials at high risk of bias.

Trial sequential analyses

In a single trial, sparse data and interim analyses increase the risk of type I and type II errors. To avoid type I errors, group sequential monitoring boundaries are applied to decide whether a trial could be terminated early because of a sufficiently small P value, that is the cumulative Z‐curve crosses the monitoring boundaries (Lan 1983). Likewise, before reaching the planned sample size of a trial, the trial may be stopped due to futility if the cumulative Z‐score crosses the futility monitoring boundaries (Higgins 2011a). Sequential monitoring boundaries for benefit, harm, or futility can be applied to meta‐analyses as well, called trial sequential monitoring boundaries (Higgins 2011; Wetterslev 2008). In a trial sequential analysis (TSA), the addition of each trial in a cumulative meta‐analysis is regarded as an interim meta‐analysis and helps to clarify if significance is reached or futility is reached or whether additional trials are needed (Wetterslev 2008).

TSA combines a calculation of the diversity‐adjusted required information size (cumulated meta‐analysis sample size to detect or reject a specific relative intervention effect) for meta‐analysis with the threshold of data associated with statistics. We will perform TSA on all outcomes included in the 'Summary of findings' table (see above) (Pogue 1997; Wetterslev 2008).

The idea in TSA is that if the cumulative Z‐curve crosses the boundary for benefit or harm before a diversity‐adjusted required information size is reached, a sufficient level of evidence for the anticipated intervention effect has been reached with the assumed type I error and no further trials may be needed. If the cumulative Z‐curve crosses the boundary for futility before a diversity‐adjusted required information size is reached, the assumed intervention effect can be rejected with the assumed type II error and no further trials may be needed. If the Z‐curve does not cross any boundary, then there is insufficient evidence to reach a conclusion. To construct the trial sequential monitoring boundaries, the required information size is needed and is calculated as the least number of participants needed in a well‐powered single trial and subsequently adjusted for diversity among the included trials in the meta‐analysis (Wetterslev 2008). We will apply TSA as it decreases the risk of type I and II errors due to sparse data and multiple updating in a cumulative meta‐analysis, and it provides us with important information in order to estimate the risks of imprecision when the required information size is not reached. Additionally, TSA provides important information regarding the need for additional trials and the required information size of such trials (Wetterslev 2008).

We will apply trial sequential monitoring boundaries according to an estimated clinical important effect. The required information size based will be based on an a priori effect corresponding to a 10% relative risk reduction for beneficial effects of the intervention and a 30% relative risk increase for harmful effects of the interventions.

TSA for continuous outcomes will be performed with mean differences, by using the trials applying the same scale to calculate the required sample size. For the continuous outcomes we will test the evidence for the achieved differences in the cumulative meta‐analyses.

For the heterogeneity adjustment of the required information size we will use the diversity (D²) estimated in the meta‐analyses of included trials. In case diversity is zero in a meta‐analysis we will perform a sensitivity analysis using an assumed diversity of 20% when future trials are included, possibly changing future heterogeneity among trials.

Quality of evidence

We will present the overall quality of the evidence for each outcome specified under 'Types of outcome measures: Summary of findings' according to the GRADE approach, which takes into account issues related not only to internal validity (risk of bias, inconsistency, imprecision, publication bias) but also to external validity, such as directness of results. Two review authors (KSM and PK or LK) will independently rate the quality of evidence for each outcome. We will present a summary of the evidence in a 'Summary of findings' table. This will provide key information about the best estimate of the magnitude of the effect, in relative terms and as absolute differences, for each relevant comparison of alternative management strategies, numbers of participants and trials addressing each important outcome and rating of overall confidence in effect estimates for each outcome. We will create the 'Summary of findings' table based on the methods described in the Cochrane Handbook for Systematic Reviews of Interventions by means of Review Manager's table editor (RevMan 2014). We will include an appendix titled 'Checklist to aid consistency and reproducibility of GRADE assessments' (Meader 2014) to help with standardisation of the 'Summary of findings' tables (Higgins 2011a). Alternatively, we will use the GRADEpro Guideline Development Tool (GDT) software (GRADEproGDT 2015) and present evidence profile tables as an appendix. We will present results for the outcomes as described in the Types of outcome measures section. If meta‐analysis is not possible, we will present the results in a narrative format in the 'Summary of findings' table. We will justify all decisions to downgrade the quality of studies using footnotes, and we will make comments to aid the reader's understanding of the Cochrane review where necessary.