Types of studies

Types of participants

We included studies involving adults (at least 18 years of age) with a DSM‐IV (Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition) (APA 2000), or equivalent diagnosis of SAD (major depressive disorder (MDD) with a seasonal pattern). Studies that enrolled SAD participants with bipolar depression were excluded; because patients with bipolar disorders are primarily treated with mood stabilisers, not with antidepressants, while SGAs are the first line treatment for patients with unipolar depression. Studies with participants with secondary co‐morbidities were considered for inclusion.

Types of interventions

Types of outcome measures

Electronic searches

An Information Specialist with the Cochrane Common Mental Disorders Group ran an updated search on the following databases to identify reports of RCTs (29 January 2020) (Appendix 1):

• Cochrane Central Register of Controlled Trials (CENTRAL; 2020, Issue 1) in the Cochrane Library (all years);

• Ovid MEDLINE (2011 to 29 January 2020);

• Ovid Embase (2011 to 29 January 2020);

• Ovid PsycINFO (2011 to 29 January 2020);

• Cochrane Common Mental Disorders Controlled Trials Register (CCMDCTR) (archived database) (2011 to June 2016 only).

International trial protocols registered on the US National Institutes of Health Ongoing Trials Register ClinicalTrials.gov and the World Health Organization International Clinical Trials Registry Platform (WHO ICTRP) were captured via CENTRAL on the Cochrane Library.

The Information Specialist also ran a separate search for adverse events (Ovid databases) (24 February 2020) (Appendix 2).

We applied no restrictions to language or publication status.

Searches for the earlier version of this review, published in 2011, are shown in Appendix 3.

Searching other resources

In addition, we handsearched the references of all included studies and pertinent review articles. We ran a separate search of the US National Institutes of Health Ongoing Trials Register ClinicalTrials.gov (www.clinicaltrials.gov/) to detect ongoing or unpublished studies. Finally, we contacted experts in the field of SAD to ensure that we were aware of all published and unpublished trials.

Selection of studies

Two review authors (BNS, KT, AC, TP, AS, or GG) independently reviewed the titles and abstracts of records identified by the search. We retrieved full‐text copies of all studies that potentially met the inclusion criteria based on abstract review, and two review authors (BNS, KT, AC, TP, AS, or GG) independently reviewed these to determine their eligibility. Any disagreements between review authors were discussed with and resolved by a third party (BNS or GG). We used Covidence software to facilitate title/abstract and full‐text screening (Covidence).

No studies published only as abstracts were considered for inclusion in the review, hence there was no need to contact trial authors for information required to permit an appraisal of trial quality.

Data extraction and management

Two review authors (BNS, KT) independently abstracted data from the included trials. A third review author (AC) checked the data for accuracy and completeness. Any discrepancies were resolved by consensus or by involving a third review author.

We extracted the following study characteristics.

• Methods: study design, duration of study, duration of treatment period, number of study centres, Country.

• Participants: number of participants, age (mean or range), proportion of women, diagnostic criteria, inclusion and exclusion criteria.

• Interventions: intervention, comparison.

• Outcomes: primary and secondary outcomes specified.

• Notes: funding for studies and conflicts of interest of authors.

All possible comparisons based on our inclusion criteria were as follows.

1. SGAs versus placebo

2. SGAs versus light therapy

3. SGAs versus other SGAs

4. SGAs versus psychotherapy

In the case of missing data, we would have contacted the study authors, but this was not necessary.

Assessment of risk of bias in included studies

We assessed the risk of bias of randomised trials using the Cochrane 'Risk of bias' tool, as described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2009). The tool includes assessment of: random sequence generation; allocation concealment; blinding of participants, personnel, and outcome assessors; incomplete outcome data; selective outcome reporting; and other potential threats to validity. Specifically, we assessed attrition (loss of participants during a study) in the trials and the reasons for it, particularly where the attrition rates between two groups in a trial were substantially different. In addition, we assessed whether all relevant outcomes for the trial were reported in the published articles. We assessed each domain as low, high, or unclear risk of bias.

For non‐randomised studies, we used criteria involving selection of cases or cohorts and controls, adjustment for confounders, methods of outcomes assessment, length of follow‐up, and statistical analysis (Deeks 2003). We entered our 'Risk of bias' assessment into the standard 'Risk of bias' tables (used to generate figures in the report), as well as in an Additional table (Table 1; Table 2; Table 3; Table 4; Table 5). Two review authors (BNS, KT) independently assessed the risk of bias. Any discrepancies were resolved by consensus or by involving a third review author.

Measures of treatment effect

We used the extracted data from the original studies to construct 2 x 2 tables (e.g. clinical response versus no response for antidepressant versus comparator). Where multiple studies allowed for pooling, we calculated the risk ratio (RR) with 95% confidence intervals (CI) for each outcome. We planned to calculate the number needed to treat for an additional beneficial outcome (NNTB) and risk difference (RD). In addition, we planned to pool continuous data using the mean difference (MD) or standardised mean difference (SMD).

Unit of analysis issues

For any included trials with multiple treatment groups (e.g. differing doses of one SGA versus placebo or multiple SGAs versus placebo), we would include the data for the treatment arms and halve the data from the placebo arm, or collapse the data for different doses into one group where this was considered to be clinically appropriate (Hansen 2009). We planned to include data from cluster‐randomised studies and from the first period of cross‐over studies.

Dealing with missing data

We used an intention‐to‐treat analysis where data were missing from participants who dropped out of trials before completion, assuming in such cases that participants did not achieve a clinical response or remission. We planned that where data regarding an assessed outcome were not reported, we would contact authors of publications or pharmaceutical companies to obtain the missing results, but this was not necessary. We have reported the proportion of participants for whom no outcome data were obtained (because of attrition) in the 'Risk of bias' assessment of each trial, and considered the potential impact of the missing data in our interpretation of the results.

Assessment of heterogeneity

We used the Cochran Chi² test (Q‐test) to assess heterogeneity, considering a P value less than 0.10 as statistically significant. We used the I² statistic to estimate the degree of heterogeneity, which describes the percentage of total variation across studies that results from heterogeneity rather than chance. We interpreted the importance of any heterogeneity in terms of its magnitude and the direction of effects. We did not consider thresholds, instead adopting the overlapping bands suggested in the Cochrane Handbook. For example, we considered an I² of:

• 0% to 40% as probably not important;

• 30% to 60% as representing moderate heterogeneity;

• 50% to 90% as substantial heterogeneity;

• 75% to 100% as considerable heterogeneity (Higgins 2020).

Assessment of reporting biases

Had we found more than 10 studies, we would have undertaken an analysis of a funnel plot, which is a graph used to detect publication bias. We wanted to know whether the largest studies were near the average with small studies spread on both sides of the average. Variations from this assumption can indicate the existence of publication bias, but asymmetry may not necessarily be caused by publication bias.

Data synthesis

We analysed data using Review Manager 5 (Review Manager 2020). We pooled data for meta‐analysis where the participant groups were similar, and the studies assessed the same treatments with the same comparator and had similar definitions of outcome measures over a similar duration of treatment. We used a fixed‐effect model because we expected studies to be very similar. Studies reporting dichotomous data were weighted using the Mantel‐Haenszel method. We rated the strength of the evidence based on the system developed by the GRADE Working Group (Guyatt 2011).

We performed a narrative analysis of data on adverse events comparing crude rates. We planned to conduct quantitative analysis of the rate of adverse events only if we located a sufficient number of prospective observational studies or randomised trials that contained data on adverse events suitable for pooling.

Subgroup analysis and investigation of heterogeneity

We planned to perform subgroup analyses of SSRIs and SNRIs, but this was not possible due to the limited number of identified studies.

Sensitivity analysis

We performed sensitivity analysis using a random‐effects model  to see if this would change the pooled results. We planned to conduct additional sensitivity analyses excluding small studies, studies with a high risk of bias, and studies published in abstract form; however, due to the limited number of trials available it was not possible to perform these analyses.

Summary of findings and assessment of the certainty of the evidence

Two review authors (BNS, KT) assessed the certainty of evidence using the GRADE approach, which considers risk of bias, inconsistency, indirectness, imprecision, and publication bias for each outcome. Evidence from RCTs starts at a level of high certainty. Concerns about any of the aforementioned domains would result in downgrading of the evidence by one or two levels; explanations for our decisions would be provided in the 'Summary of findings' tables. For example, we would downgrade one level for imprecision if the 95% CI included both a beneficial or non‐beneficial effect. We assessed the certainty of the evidence as high, moderate, low, or very low according to the GRADE approach (Guyatt 2011). We created two 'Summary of findings' tables, one for the comparison SGAs versus placebo (summary of findings Table 1), and one for SGAs versus light therapy (summary of findings Table 2).