3. Unused methods.

Method	Details (for future updates)
Time‐to‐event outcomes	We will use HRs as our measure of treatment effect for any time‐to‐event outcomes and will present these with 95% CIs.
Individually randomised trials with clustering	Clustering may arise in individually randomised trials where each therapist treats multiple patients. For these trials, we use the same approach as described for cluster randomised trials to inflate the variance of the intervention estimates (using a design effect) when clustering has not been accounted for in the trial analysis. However, we would only apply this correction in trials where we can establish the mean number of participants per therapist, and where this number is large enough to affect the variance importantly (Higgins 2022a).
Cross‐over trials	If cross‐over designs had been used to evaluate any of our eligible interventions, other than pharmacotherapy, we would only use the data from the first period (if available). A cross‐over design for the interventions eligible for this review (aside from pharmacotherapy) is inappropriate because these interventions can lead to permanent change. In cross‐over trials evaluating pharmacological interventions, where an appropriate paired analysis is not available, we will attempt to approximate a paired analysis by imputing missing statistics (e.g. missing standard deviation, correlation). The values of these statistics will be informed by other trials included in the review, or trials outside the meta‐analysis (Elbourne 2002; Higgins 2022a). We would only include the first period data (if possible) in cross‐over trials in which there is less than two weeks' washout, because in this circumstance there is a serious risk of carry‐over effects arising from the effects of the first‐period antidepressant or antipsychotic persisting into subsequent period(s) (Hosenbocus 2011; Hulshof 2020).
Assessment of reporting biases	The risk of missing studies (termed 'unknown‐unknowns'): we will consider qualitative signals of non‐publication of studies (e.g. research area is in the early stages) and statistical signals of missing results. To examine the latter, we plan to investigate the potential for small‐study effects using contour‐enhanced funnel plots. Contour‐enhanced funnel plots aid in determining whether funnel plot asymmetry is due to publication bias or other factors (Peters 2008).
Data imputation	We will impute missing summary data (e.g. ICCs, standard deviations), where we are unable to obtain these data from the trial authors, and document the methods used and any assumptions made
Meta‐analyses	Given that random‐effects models can yield CIs that are too small, particularly in meta‐analyses with few trials, we will undertake sensitivity analyses using the restricted maximum likelihood (REML) estimator of between trial heterogeneity variance and the Hartung‐Knapp‐Sidik‐Jonkman CI methods (Hartung 2001; Sidik 2002).

CI: confidence interval; HR: hazard ratio; ICC: intraclass correlation coefficients; REML: restricted maximum likelihood