Literature search and study selection

For this systematic review, we included all RCTs that included adults with any type of depression. These RCTs must have compared a group that received a decision aid that aimed to help patients decide about their treatment for any kind of depression treatment (as a stand-alone intervention, or as the main element within a complex intervention) with a group that did not, and directly assessed any beneficial or adverse effects in adults with depression. We excluded RCTs that had as population only pregnant women because they have different risks that should be considered when deciding whether to use antidepressants.^{Reference Wichman and Stern24} Also, we excluded conference papers. There were no restrictions on language or publication date.

Decision aids were defined as tools or technologies used to help patients make informed decisions by offering information about treatment options, and help them to construct, clarify and communicate their values and preferences.^{Reference Volk, Llewellyn-Thomas, Stacey and Elwyn25} However, sometimes it is difficult to differentiate from other information-based interventions.^{Reference Lenz, Buhse, Kasper, Kupfer, Richter and Muhlhauser26} To define if the proposed intervention was a decision aid, we used the six-item qualifying criteria for decision aids developed by the International Patient Decision Aid Standards Collaboration, as it provides the definition standards for decision aids: (a) describes the health condition or problem for which the index decision is required, (b) states the decision that needs to be considered, (c) describes the options available for the index decision, (d) describes the positive features of each option, (e) describes the negative features of each option and (f) describes what it is like to experience the consequences of the options.^{Reference Joseph-Williams, Newcombe, Politi, Durand, Sivell and Stacey27}

The decision aid assessed by the RCTs needed to meet all six criteria to be included in our systematic review.

A literature search was performed in two steps: a systematic review of five databases, and a review of all documents cited by any of the studies included in the first step. For the first step, we performed a literature search in five databases: Medline, EMBASE, Scopus, Web of Science and ClinicalTrials.gov. We used terms related to decision support, decision aid, decision-making, depression and clinical trials. The complete search strategies for each database are available in Supplementary File 1 available at https://doi.org/10.1192/bjb.2020.130. The last update of this database search was performed on 5 January 2019. Duplicated records were removed with EndNote version X8 for Windows (Clarivate Analytics, Thomson Reuters, New York; see https://endnote.com/). After that, titles and abstracts were independently screened by two pairs of independent reviewers (C.A.A.-R. with M.E.D.-B., and N.B.-C. with C.J.T.-H.) to identify potentially relevant articles for inclusion. This was performed with the online software Rayyan version 01 for Windows (Qatar Computing Research Institute, Qatar Foundation, Qatar; see https://rayyan.qcri.org).^{Reference Ouzzani, Hammady, Fedorowicz and Elmagarmid28} Disagreements were resolved through a discussion with a third reviewer (J.H.Z.-T.). Then, the full text of potentially relevant articles were assessed to evaluate their eligibility. This process was also independently performed by two researchers. The complete list of excluded articles at this full-text stage is available in Supplementary File 2.

For the second step, two independent reviewers (M.E.D.-B. and N.B.-C.) assessed all documents listed in the references section of the studies selected in the first step, and collected all articles that fulfilled the inclusion and exclusion criteria.

Data extraction

Two independent researchers (C.A.A.-R. and M.E.D.-B.) extracted the following information from each of the included studies into a sheet of Microsoft Excel version 2018 for Windows: author, year of publication, title, population (inclusion and exclusion criteria), setting, intervention (name, type, the methodology of application and length of use), comparator (name, type, the methodology of application and length of use), time of follow-up and effects of decision aid in all included outcomes.

Intervention information was collected with the Template for Intervention Description and Replication (TIDieR) checklist.^{Reference Hoffmann, Glasziou, Boutron, Milne, Perera and Moher29} The checklist originally was designed for pharmacological interventions; thus, we included only the following items, adapted for more complex interventions: name of intervention, rationale, location of delivery, materials, procedures, who provided, modes of delivery (grouped or individual), frequency (sessions) and possible options to choose within the decision aid. In case of disagreement, the full-text article was reviewed again by the researchers, to reach a consensus.

Study quality and certainty of the evidence

Two independent researchers (C.A.A.-R. and N.B.-C.) used the Cochrane Risk of Bias Tool for RCTs to assess systematic errors (or bias) in the design, conduct, analysis and reporting of the RCT that could potentially underestimate or overestimate the results.^{Reference Higgins, Altman, Gotzsche, Juni, Moher and Oxman30} We followed the instructions stated in the Cochrane Handbook for Systematic Reviews of Interventions and evaluated selection bias, performance bias, detection bias, attrition bias and reporting bias to assess each of the six domains of the tool as low, high or unclear risk of bias, by each RCT included in the systematic review.^{Reference Deeks, Higgins, Altman and Green31}

To assess the certainty of the evidence, we used the Grading of Recommendations Assessment, Development and Evaluation (GRADE) methodology,^{Reference Balshem, Helfand, Schunemann, Oxman, Kunz and Brozek32} which classifies it in a very low, low, moderate or high certainty of the evidence each outcome in the systematic review. This classification is based on the following criteria: risk of bias (evaluated through the Cochrane Risk of Bias Tool), inconsistency (heterogeneity between the RCT results and in terms of population, intervention, comparator and outcome; additionally assessed by the I ² test), indirectness (how different are the included RCTs to the question of interest) and imprecision (wideness of the confidence interval). The certainty of the evidence was assessed for all meta-analysed outcomes and non-meta-analysed outcomes that were important for decision- making. Additionally, when two or more RCTs assessed the same outcome, but a meta-analysis was not performed, we summarised the individual data of each RCT narratively, and then assessed the certainty of the evidence following the recommendations proposed by Murad et al.^{Reference Murad, Mustafa, Schunemann, Sultan and Santesso33}

Statistical analyses

We performed meta-analyses to summarise the results of the RCTs that evaluated the same outcomes. When outcomes were measured with different scales across studies, we calculated standardised mean differences (SMD) to compare and meta-analyse these studies; otherwise, we calculated weighted mean differences (WMD). For outcomes that had been measured more than once, we only considered the final measurement to perform the meta-analyses, as suggested in the Cochrane Handbook.^{Reference Deeks, Higgins, Altman and Green31} We assessed heterogeneity with the I ² statistic, and we considered that heterogeneity might not be significant when I ² < 40%.^{Reference Deeks, Higgins, Altman and Green31} We considered it appropriate to use random-effects models in all the meta-analysis because of the overall heterogeneity in terms of population, intervention and comparators.^{Reference DerSimonian and Laird34} We executed a sensitivity analysis, taking into account contradictory results within studies. We did not considerer to exclude studies with high risk for bias for sensitivity analysis, because all the included RCTs had at least one domain of the Cochrane Risk of Bias Tool with a high risk of bias. Also, we did not execute a subgroup analysis because of the low number of studies by each meta-analysis. Publication bias was not statistically assessed because the number of studies pooled for each meta-analysis was less than ten.^{Reference Deeks, Higgins, Altman and Green35} The data were processed with Stata version 15.0 for Windows.