Types of studies

We included all randomised controlled trials (RCTs) in people with a clinical diagnosis of stroke (Hatano 1976), where an SSRI had been given within the first year of stroke onset i.e. 1) studies had to state that participants were recruited within 12 months of stroke onset, or 2) studies where the mean (or median) time since stroke was less than 12 months. Studies had to have reported at least one of our outcomes of interest in order to be included in the meta‐analysis. For those studies which did not report data in a form that we could use, we attempted to get the raw data from the authors, and if this was not possible, we retained the studies in the list of included studies, even though they could not be included in the meta‐analysis .

We included studies:

• with more than two arms (e.g. SSRI versus another active treatment versus placebo). We included data from the SSRI arm and the placebo arm (or usual‐care arm if a placebo was not used), and discarded data from the other active treatment arms;

• in all languages.

We excluded studies:

• using a quasi‐experimental design (i.e. where investigators describe a non‐random component in the sequence generation process, such as date of admission);

• using a cross‐over design.

There was no restriction on the eligibility of RCTs on the basis of sample size or duration of follow‐up.

Types of participants

We included studies that had recruited survivors of a stroke, defined as a sudden‐onset focal neurological disturbance, assumed to be vascular in origin, and lasting more than 24 hours (Hatano 1976). Studies had to recruit participants within 12 months of stroke onset, or the mean/median time since stroke had to be less than 12 months. We intended to include studies in subarachnoid haemorrhage and perform a subgroup analysis but we did not find any such studies. We intended to exclude trials of mixed populations (e.g. stroke and head injury) unless separate results for those with stroke were available, but we found no such studies.

Types of interventions

We included any drug classified as a SSRI (e.g. fluvoxamine, fluoxetine, sertraline, citalopram and paroxetine). We included any dose or mode of delivery, given for any duration and for any reason (e.g. to aid neurological recovery, to treat depression or anxiety or emotionalism, or to prevent depression or anxiety or other mood disorders). We did not include drugs that have mixed effects that include SSRI actions.

The comparator arm could include usual care or a placebo.

We excluded studies in which an SSRI was compared with another active intervention (e.g. another type of drug or herb or acupuncture). We also excluded studies that combined an SSRI with another active treatment and compared with the active treatment alone, because of the potential for interaction between the SSRI and other active treatment.

If studies had two SSRI arms, we combined these and compared with control.

Types of outcome measures

We included several outcomes.

Electronic searches

We developed the MEDLINE search strategy with the help of the Cochrane Stroke Group Information Specialist and adapted it for the other databases.

We searched the following electronic bibliographic databases.

• Cochrane Stroke Group Trials Register (7 January 2021)

• Cochrane Central Register of Controlled Trials (CENTRAL; 2021, Issue 1) in the Cochrane Library (searched 7 January 2021) (Appendix 1)

• MEDLINE (from 1948 to 7 January 2021) (Appendix 2)

• Embase Ovid (from 1980 to 7 January 2021) (Appendix 3)

• CINAHL EBSCO (Cumulative Index to Nursing and Allied Health Literature ; 1982 to 7 January 2021) (Appendix 4)

• AMED Ovid (Allied and Complementary Medicine) (from 1985 to 7 January 2021) (Appendix 5)

• PsycINFO Ovid (from 1967 to 7 January 2021) (Appendix 6)

• PsycBITE Psychological Database for Brain Impairment Treatment Efficacy (www.psycbite.com/) (16 July 2018)

In addition, we searched the following ongoing trials registers (Appendix 7).

• Stroke Trials Registry (www.strokecenter.org/trials) (26 June 2018)

• US National Institutes of Health ongoing Trials Register ClinicalTrials.gov (www.ClinicalTrials.gov) (7 January 2021)

• World Health Organization International Clinical Trials Registry Platform (WHO ICTRP) (apps.who.int/trialsearch/) (7 September 2020)

Evidence for this update included search results from the previous version of this review), combined with results from the above searches. In Mead 2012 there had been no date limits and searches had been applied from inception of databases.

Following editorial review, one review author (GM) searched the WHO ICTRP using the term 'stroke' on 8 August 2021 to identify any new trials registered after 7 July 2020 i.e. > 1 year since last search.

Searching other resources

In an effort to identify further published, unpublished and ongoing studies, we:

• searched reference lists of included trials and relevant reviews when full texts were retrieved for detailed scrutiny;

• contacted researchers in the field.

Selection of studies

Joshua Cheyne (Cochrane Stroke Group Information Specialist), ran the searches of CENTRAL, MEDLINE, Embase, CINAHL, AMED, and PsycINFO, and downloaded the resulting references into Reference Manager. These were imported into Covidence, which automatically removed some, but not all, of the duplicates.

Any two review authors (from GM, EL, LAL, MK, SW, RT, A‐SR, LL,C‐FH, MH or XH) independently scrutinised the resulting titles and abstracts and excluded obviously irrelevant reports and duplicates. We obtained full texts of potentially eligible studies. Any two review authors (from GM, AB, EL, LAL, MK, SW, RT, A‐SR, LL C‐FH or XH) independently applied inclusion and exclusion criteria; if there was lack of consensus, a third review author (usually GM unless she had already scrutinised the paper) also applied inclusion and exclusion criteria.

We include a flow diagram that includes the number of unique references identified by the searches, the number of records excluded after preliminary screening of titles and abstracts, the number of records retrieved in full text, and the number of studies fulfilling our inclusion criteria (Figure 1).

Figure 1

---

---

PRISMA flow diagram for this update

Data extraction and management

For the new eligible studies that we had identified, any two review authors (from GM, EL, LAL, MK, MH, SW, RT, LL, A‐SR, LL, C‐FH or XH) independently extracted data from each new study.

We extracted the following data:

• the report: author, year and source of publication;

• the study: sample characteristics, social demography;

• the participants: stroke sequence (first‐ever versus recurrent), social situation, time since stroke onset, prior history of psychiatric illness, current neurological status, stroke severity, whether people with aphasia were recruited, the proportion with depression at baseline (if recorded by trialists), we did not extract information on location or size of lesion as this was unlikely to have been recorded by the trialists, and brain imaging often does not show a visible infarct in people with minor strokes;

• the research design and features: adherence, non‐response and length of follow‐up;

• the intervention: type, duration, dose, timing and mode of delivery;

• the effect size: sample size, nature of outcome, estimate and standard deviation (SD) (or standard error (SE));

• source of funding.

Methods in previous versions were broadly similar and are fully reported in previous versions (Legg 2019; Mead 2012)

Assessment of risk of bias in included studies

We assessed risks of bias using the Cochrane 'Risk of bias' tool (Higgins 2017). We assessed the methods used in each study to control for the following potential sources of bias: sequence generation (selection bias); allocation concealment (selection bias); blinding of participants, personnel and outcome assessors (performance and detection bias); incomplete outcome data (attrition bias); selective outcome reporting (reporting bias); and other potential threats to validity.

Where there was a disagreement between the two review authors on risk of bias, a third review author (LAL, GM) reviewed the data and reached a consensus in discussion with the first two review authors.

For incomplete outcome data, we categorised as 'low risk' if missing data were imputed using appropriate methods or if missing outcome data overall were less than 5%. For this update, we decided to categorise studies as low risk of bias if more than 5% of data were missing providing the number of participants with missing data was balanced between groups and the reasons for missing data were unrelated to treatment allocation.

We extracted data on source of funding, and listed this under 'Other sources of bias'. If the source of funding was not given, or if there were links with the pharmaceutical industry and no explicit statement that the funder had no input into the design or analysis of the study, we classified this as 'unclear risk'. We also recorded any other potential threats to validity.

We also extracted data on how adverse effects were reported, and listed these in the descriptions of the studies.

If a study author was also one of the review authors, a review author who was not involved in the study extracted data and assessed quality.

For this 2021 update, we contacted study authors to obtain more information to enable us to make judgements about risk of bias if this was unclear from the paper, AND if this was going to influence whether a study was categorised as low risk of bias across all seven domains i.e. changing a judgement of 'uncertain' to 'low risk' of bias.

Measures of treatment effect

For dichotomous data, we reported risk ratios (RRs). For ordinal scales, where there was a well‐recognised cut‐point in the scale (e.g. mRS) we analysed the data as a dichotomous outcome (dependent or independent).

For ordinal scales with no recognised cut‐point, we analysed the data as continuous data. The data required for meta‐analyses of continuous data in Review Manager 5 are means and standard deviations (SDs) (Review Manager 2014). When extracting continuous data from the study reports, we checked whether trials reported the SD or the standard error (SE). We had planned to use the SE or 95% confidence interval (CI) to compute the SD when SDs were missing, but this was not needed as all the trials reported SDs.

For ordinal scales and continuous data, we calculated standardised mean differences (SMDs), because different scales were used for the same outcomes (e.g. Barthel Index (BI) and Funtional Independence Measure (FIM) for disability score, the Beck Depression Inventory (BDI) or the Hamilton Rating Scale for Depression (HAMD) for depression). The SMD does not correct for differences in the direction of the scale. Some scales increase with disease severity and others decrease, so we multiplied the mean value from one set of trials by −1. For example, in the National Institute of Health Stroke Scale (NIHSS), a low score indicates a less severe stroke, whilst a low score in the Scandinavian Stroke Scale (SSS) indicates a more severe stroke.

If there was more than one outcome measure in the same domain (e.g. two different depression scales), we made a post‐hoc decision to select the one with the most complete data.

Unit of analysis issues

The number of observations in the analysis should match the number of 'units' that were randomised. We considered outcomes measured at the end of treatment and at the end of follow‐up in separate analyses. For side effects, we considered the number of participants developing a specified side effect rather than the total number of side effects in each group.

Dealing with missing data

For this update, we contacted authors of new trials to obtain any data that we needed for our meta‐analysis that had not been included in a published full‐text article or an abstract.

Note that for some analyses, the sample size is slightly different for different outcomes, this is because of the missing responses for some outcomes.

Assessment of heterogeneity

We assessed whether there was evidence of inconsistency in our results by considering possible clinical, methodological, and statistical heterogeneity. We assessed clinical and methodological heterogeneity by comparing similarities in our included studies between study designs, participants, interventions, and outcomes.

We quantified the effect of heterogeneity using the I² statistic. We assessed statistical heterogeneity by visually examining forest plots. Thresholds for the interpretation of the I² statistic can be misleading, since the importance of inconsistency depends on several factors. A rough guide to interpretation in the context of meta‐analyses of randomised trials is as follows: (Section 9.5.2; Deeks 2021):

• 0% to 40% is not considered important;

• 30% to 60% suggests moderate heterogeneity;

• 50% to 90% suggests substantial heterogeneity;

• 75% to 100% is considerable heterogeneity.

Assessment of reporting biases

We searched clinical trials registers to identify published protocols for each of our included studies. We checked for selective reporting of results by comparing the published protocol with the published full‐text article and by scrutinising the aims and methods of the trials and comparing these with outcomes reported. We found several papers by the same authors, and contacted the authors to check whether the publications were duplicates or to check if the included study populations were unique. If it was not possible to determine whether different publications reported overlapping groups of participants, we included just one of the papers and listed the others as awaiting assessment.

If we had identified a sufficient number of included studies at low risk of bias (i.e. more than 10 studies (Sterne 2017), we would have generated a funnel plot of low risk of bias studies, to assess risk of publication bias, an asymmetrical funnel plot might have suggested publication of only positive results (Egger 1997). We made a post‐hoc decision to perform a funnel plot for all studies, irrespective of risk of bias, for the co‐primary outcome of disability.

We deployed a comprehensive search strategy in an effort to avoid reporting biases in our review methodology. See Search methods for identification of studies.

We tried to avoid language bias by including all trials, irrespective of language: we sought translation where needed.

Data synthesis

We completed meta‐analysis of outcomes for which we had comparable effect measures from more than one study, and when measures of heterogeneity indicated that pooling of results was appropriate. We used the statistical calculator provided in Review Manager 5 to perform meta‐analysis (Review Manager 2014).

We used a fixed‐effect model (Mantel 1959), rather than a random‐effects model because of the dominance of the three largest trials (AFFINITY 2020, EFFECTS 2020; FOCUS 2019); random effects would have given too much weight to the smaller trials. The dominance of the three large trials makes a fixed‐effect model a more reliable indicator of the effect than the average across the smaller trials. We assessed the robustness of the results to choice of model using a sensitivity analysis for our primary outcomes.

In the initial 2012 review, we had performed multiple meta‐analyses of all the primary and secondary outcomes, included all trials irrespective of risk of bias, and then explored the influence of each aspect of bias on estimates of effects in a series of sensitivity analyses.

In the 2019 review we limited our analyses of all outcome measures to studies at low risk of bias (Higgins 2017), as we wanted reliable data, not confounded by bias, to find out whether SSRIs were more effective than placebo or usual care at improving a range of important outcomes. Also, had we included all studies, irrespective or risk of bias, for all available outcomes, the number of analyses would have become unmanageable within the resources we had. However, in the 2019 review, we performed a sensitivity analysis by using data on our two primary outcomes (disability and independence) from all trials, irrespective of risk of bias. For this current update, we use the same approach.

We reached decisions on overall risk of bias by study by consideration of six risk of bias domains: sequence generation, allocation concealment, blinding of participants and trial personnel, blinding of outcome assessment, incomplete outcome data, selective outcome reporting and other risk of bias. We required a study to have a judgement of low risk of bias in all domains in order to categorise it as having an overall low risk of bias.

Subgroup analysis and investigation of heterogeneity

If there had been at least two studies at low risk of bias across all six domains of risk of bias, we would have explored variability in the participants, interventions, and outcomes among studies using the following subgroups.

• Type of SSRI.

• Studies with depression at baseline as an inclusion criterion and those where depression was not an inclusion criterion.

• Time since stroke at recruitment. We categorised these as less than three months (0 ‐ 90 days), three to six months (91 to 108 days), six to nine months (181 to 271 days) or nine to 12 months (272 to 365 days).

Finding high statistical heterogeneity (I² > 50%) would not have prevented us from performing a subgroup analysis, rather we would have considered the reason for this heterogeneity.

We did not perform subgroup analyses by type of SSRI or time since stroke as all studies included in the main analyses were similar with regard to these characteristics.

Sensitivity analysis

We explored the potential effects of decisions made as part of the review process as follows.

• We included all studies regardless of risk of bias judgement for our primary outcomes of disability score and independence.

• We conducted meta‐analysis using the alternate meta‐analytical effects model (fixed‐effect or random‐effects) for just those studies at low risk of bias across all six domains.

• We intended to conduct a meta‐analysis using the alternate 'last available follow‐up' time point if studies had reported more than one follow‐up after the end of treatment, but none of the included studies did this.

• We compared effect estimates from the above results with effect estimates from the main analysis. We reported differences that altered the interpretation of effects.

Summary of findings and assessment of the quality of the evidence

We created a summary of findings table using the following outcomes: disability; dependent according to the mRS; neurological deficit score; depression (continuous data); death; seizures; and gastrointestinal side effects (summary of findings Table 1). We chose these outcomes as they are of high clinical relevance. We used the five GRADE considerations (study limitations, consistency of effect, imprecision, indirectness and publication bias) to assess the quality of the body of evidence as it relates to the studies that contribute data to the meta‐analyses for the prespecified outcomes (Atkins 2004). We used methods and recommendations described in Chapter 11 of the Cochrane Handbook for Systematic Reviews of Interventions (Schünemann 2017), using GRADEproGDT software (GRADEproGDT 2015). We justified all decisions to downgrade the quality of studies using footnotes, and made comments to aid the reader's understanding of the review where necessary.

Summary of findings and assessment of the certainty of the evidence

We created a 'Summary of findings' table using the following outcomes: disability; dependent according to the mRS; neurological deficit score; depression (continuous data); death; seizures; and gastrointestinal side effects (summary of findings Table 1). We chose these outcomes as they are of high clinical relevance. We used the five GRADE considerations (study limitations, consistency of effect, imprecision, indirectness and publication bias) to assess the quality of the body of evidence as it relates to the studies that contribute data to the meta‐analyses for the prespecified outcomes (Atkins 2004). We used methods and recommendations described in Chapter 11 of the Cochrane Handbook for Systematic Reviews of Interventions (Schünemann 2017), using GRADEproGDT software (GRADEproGDT 2015). We justified all decisions to downgrade the quality of studies using footnotes, and made comments to aid the reader's understanding of the review where necessary.