In the protocol for this review, suicide‐related behaviour (both ideation and attempt) was specified as secondary outcome. However, due to the concern that taking antidepressant medications may potentially result in suicidal behaviour, we made a decision to include such behaviours as a primary outcome.

For dichotomous outcomes, such as 'number relapsed', results from each trial are expressed as an odds ratio (OR) with 95% confidence intervals and combined in meta‐analysis. Although the protocol for the review stipulated that we would express relapse rates as a risk ratio (RR), ORs have more favourable mathematical properties.

Originally we intended to perform subgroup analyses on trials that included children and adolescents versus those that included participants of any age who had experienced a first episode of depression. However, as the search did not yield any trials of the latter type, we could not perform this analysis. We also intended to perform subgroup analyses on trials that contained children versus those that contained adolescents, but the nature of the trials included in the review did not contain enough data to allow for this subgroup analysis.

During the review process it became apparent that within the two types of trial design that we had anticipated, there was considerable diversity. In trials where participants who had responded or remitted from an episode of MDD or DD during an acute phase of treatment were re‐randomised into a continuation or maintenance phase, re‐randomisation commonly occurred either early (after an acute phase) or late (after either a continuation and/or maintenance phase). Due to the variability in the length of treatment before re‐randomisation, we felt that it was important to perform subgroup analyses based on time of re‐randomisation (early or late).

Originally, we intended to undertake sensitivity analyses to assess the effect of risk of bias that may be introduced due to the decisions made in the process of undertaking the review. In psychiatry trials it is important to investigate the impact of assumptions made in various imputation methods used to account for missing data, such as analysis using LOCF and OC. However, as there were limited data contained in trials, we were unable to perform these analyses.