We included all studies (n = 10) submitted to the Dutch Medicines Evaluation Board during an 11-year period (1991–2004) as part of market authorization applications in Europe for the indication acute manic episode of BD. All studies were short-term double-blind randomized, placebo-controlled trials involving patients diagnosed with an acute manic episode of BD (DSM IV criteria). Pharmaceutical companies provided data to enable an IPD meta-analysis. All studies were approved by the ethic committees of their respective research-centers, and all participants gave informed consent. Availability of data on ethnicity was a prerequisite for inclusion. In the database, the following predefined ethnic groups were available: Caucasian, Black, Asian, Oriental, Hispanic, Native American, and Other. The original studies were identified for the purpose of collecting manuscripts and corresponding authors were contacted in case of unclarities or missing information (e.g., on the definition of ethnic subgroups). In addition, the ethnic subgroups were examined and redefined according to the current JAMA Network guidelines as: American Indian or Alaska Native, Asian, ‘Black’, Native Hawaiian or Other Pacific Islander, ‘White’ and ‘Some Other Race’ (Flanagin et al. 2021). In the original studies, the term ‘race’ was used, or may have been used interchangeably with ‘ethnicity’.

We further restricted the analyses to subjects that were given a potential effective dose of the medication, as indicated in the Summary of Product Characteristics (SmPC). Patients receiving dosages under the minimal effective dose or over the maximum doses as stated in the SmPC) were excluded. Risk of bias was assessed using the Risk of Bias tool (Stewart et al. 2015). The guideline for the preferred reporting items for systematic reviews and meta-analyses (PRISMA) was followed except for items pertaining specifically to systematic reviews (Flanagin et al. 2021). We pre-specified our methods and analysis plan in the PROSPERO database for systematic reviews (ID: CRD42024543671). There were no deviations from the analysis-plan state in the PROSPERO protocol.

The severity of the acute manic episode of BD at baseline and at study endpoint was assessed with two instruments. The Young Mania Rating Scale (YMRS) comprises 11 items: seven items are scored on a 0–4 scale and four items are scored on a 0–8 scale. The total score thus ranges from 0 (no symptoms) to 60 (severe symptoms) (Young et al. 1978). The Mania Rating Scale from the Schedule for Affective Disorders and Schizophrenia – Change Version (MRS from SADS-C) also comprises 11 items: one item is scored on a 0–2 scale and ten items are scored on a 0–5 scale (higher score indicates higher severity). The total score thus ranges from 0 (no symptoms) to 52 (severe symptoms) (Endicott and Spitzer 1978).

We used two efficacy outcomes: the standardised difference in mean change score on the (Y)MRS from baseline to follow-up and the difference in percentage of responders. Patients were considered a responder if their score on the YMRS or MRS decreased by 50% or more from baseline to follow-up (Endicott and Spitzer 1978). Since two different rating scales were used in the studies, we decided to use mean percent improvement as primary outcome measure. The endpoint was defined as the three-week post-baseline assessment, since this is the time point recommended for establishing short-term efficacy in the EMA Committee for Proprietary Medicinal Products (CPMP) guideline on the clinical investigation of medicinal products for the treatment of mania (CPMP 2001). For any missing individual (Y)MRS item, we imputed the average of the other (Y)MRS items for that patient for that visit. For patients who dropped out before week three, the last observation was carried forward (LOCF) to week three. The difference in mean improvement and the difference in percent responders between active treatment and placebo at endpoint (LOCF) were considered as the main outcome measures.

To examine possible confounding effects, baseline severity, age, and gender were used as covariates in the various statistical models. To provide an impression of the possible influence of the publication date of the individual studies on the results, the studies were ranked in chronological order.

We used a one-stage, random effects IPD meta-analysis. Traditional methods for meta-analysis synthesis use aggregate study level data that are generally obtained from publications. Meta-analysis synthesis of individual patient data uses the crude data from individual patients from each study. To explore participant-level variations and to control for potential confounders and between-study heterogeneity, IPD meta-analysis was chosen over pooled linear regression analysis or study aggregate meta-analysis Because of existing heterogeneity between studies (e.g., different patient populations, different types of medications, and different companies), random effect rather than fixed effect models were used. The one-stage approach was chosen over the two-stage for primary analysis, as it can analyse predictors at the subject-level, has more power to detect treatment-covariate interactions and leads to less bias when few studies or studies with small sample sizes are included.

In case of missing data for some subjects in one or more studies, the last observation will be carried forward (LOCF). This approach was chosen to maintain consistency across different studies that provided incomplete data, ensuring comparability in the pooled estimates. Also, this approach was chosen in all original efficacy trials, making our resulst better interpretable in context. However, we acknowledge that LOCF can introduce bias—particularly if dropout is related to lack of effect or adverse events—and may therefore overestimate or underestimate the true treatment effect. In addition, imputing missing (Y)MRS items using LOCF may reduce variance inappropriately or produce misleading effect sizes if attrition is non-random. Despite these limitations, LOCF was deemed appropriate based on our short-term data set constraints and the need to harmonize data handling across multiple trials. Forest plots for each primary outcome will be visualized as it provides an overview in which combined estimates, inconsistency across studies and the precision of individual studies can be examined.

Between-study heterogeneity was in part be mitigated by the study-specific fixed intercept and random treatment effect, and was further be corrected by adjusting for confounders such as baseline severity, age, and gender in order to determine the effect of a treatment*ethnicity interactionterm in addition to an ethnicity variable.

As a sensitivity analysis for the one-stage IPD, a two stage approach was used. The two-stage IPD analysed studies that included at least 5 patients from each ethnicity group in each study arm in order to ensure adequately powered analysis. In the two stage analysis ethnicity was defined as a dichotomous variable, alternating two ethnicity subgroups so every effect between them can be compared.

For the first step, we calculated the total scores on the respective questionnaires at baseline and week three. We also calculated response rate, defined as at least 50% reduction of (Y)MRS score between baseline and week three. Subsequently, for each study, multivariate linear regression analyses were performed with mean percentual (Y)MRS change from baseline as the dependent variable and treatment condition, ethnicity, and treatment condition*ethnicity as the independent variables.

Similarly, for each study, a multivariate logistic regression analysis was performed with response as dependent variable. Thus, in both analyses (symptom change from and response as the dependent variable), for each study, the interaction of ethnicity*treatment condition (active medication vs placebo) was added to the main effects (ethnicity and treatment condition) as an independent variable for a modifier effect of ethnicity on treatment effect. Subsequently, to examine the effect of baseline severity, age, and gender these variables were cumulatively added as independent variables to the main effects and the interaction of ethnicity by treatment. For these analyses, the Statistical Package for the Social Sciences (SPSS) version 26 was used.

As the second step, a random effect meta-analysis was performed with the regression-coefficients and odds ratios (ORs) for the treatment condition*ethnicity interactions in the different studies. In these analyses, the 95% confidence interval (CI) indicates the scope of uncertainty in the effect estimate of the treatment condition*ethnicity interactions considering heterogeneity between studies. For these analyses, the Comprehensive Meta-Analysis (CMA) version 2 software was used.

Finally, the treatment effect in the different ethnic subgroups (White, Black and Asian) was examined separately. A conventional IPD meta-analysis for each ethnic group was performed, yielding ethnicity-specific pooled mean differences in outcomes (symptom change and response) between participants receiving active medication and participants receiving placebo.