Interpersonal sensitivity and response to selective serotonin reuptake inhibitors

Step 1For each study, we will merge the relevant study data/variables into three datasets, one for each trial. Step 2We will calculate SCL-IS sum scores and HAMD-17 sum scores at baseline and each visit/treatment session. Missing scores will be carried forward using last-observation-carried-forward (LOCF) imputation, as per the original trials. SCL-IS and HAMD-17 change scores (baseline - final score) will then be calculated.The study variables and age/sex will be cross-referenced with the original trial reports to ensure data merging and calculation of LOCF outcomes occurred without error.Step 3We will pool/append the three trial datasets from into one large dataset, with the ITT populations (fluoxetine, pooled n = 640; paroxetine, pooled n = 761; placebo, pooled n = 381).Step 4To answer the first study objective, we will calculate the pooled effect size estimate for the difference between SCL-IS change scores between medication (paroxetine or fluoxetine) and placebo groups. The paroxetine and fluoxetine groups are being combined because we do not have any reason to suspect differences between these two medications (in other words, we hypothesize both to be significantly more effective at decreasing SCL-IS scores vs. placebo). However, as a sensitivity analysis to test this assumption, we will also calculate the effect sizes for fluoxetine vs. placebo (Trials 1 and 2 only) and paroxetine vs. placebo (all trials) separately to ensure both medications significantly differed from placebo, and the pooled effect was not accounted for entirely by one of the study medications.In order to further account for inter-study variability in the pooled effect size estimate, the two-stage individual patient data meta-analysis method by Fisher (9) will be used. The first stage involves estimating treatment effects in each study. In the second stage, these individual study effects will be pooled by using inverse-variance weights. For example, if the estimated effect of a medication (vs placebo) in a particular study is Xi, it will be assigned a weight equal to 1/(Xi)^2, with the denominator representing the study variance. This way, a pooled estimate Xp, can be calculated that takes into account inter-study differences. Fisher has written a Stata command ipdmetan which we will use. Other advantages of this command are that it can calculate the Q statistic (an index of heterogeneity), it accepts individual-level data, and is capable of an unbiased estimation of treatment x patient characteristic interaction, if the sample size allows. It also allows for forest plots to be created as an easy way of presenting the results, if desired. This part of the analysis is, essentially, a fixed-effects meta-analysis being conducted from IPLD data as opposed to aggregate trial-level data. Although we do not anticipate significant between-study heterogeneity given the similar trial methods and patient selection/sampling procedures, this will allow us to observe and test for between-study heterogeneity if it is present.Step 5The pooled effect size estimates from Step 4 will be calculated again, but HAMD-17 change scores will be included in the model as a covariate. The purpose of adding HAMD-17 change scores as a covariate in this analysis is to assess whether the effect of the study medications on interpersonal sensitivity occurred independently of their effect on depression symptoms.   Step 6To address the second study objective, we will graph SCL-IS and HAMD-17 baseline and change scores to examine their distributions. We will also graph the following pairs of scatter plots: baseline SCL-IS vs. baseline HAMD-17, SCL-IS change vs. HAMD-17 change, and baseline SCL-IS vs. HAMD-17 change. These will be done separately for each treatment group. The purpose of this step is to better understand the nature of the data and the SCL-IS variables regarding their distributions in the sample, associations with HAMD-17 scores, and statistical assumptions. This will be helpful both for interpreting the results of this study but also for future research in terms of developing pre-specified analysis plans, power analyses, covariate selection, etc.Step 7If the graphs from Step 6 suggest linear relationships with certain statistical assumptions (normality, homoscedasticity), we will use linear regression models to calculate the pooled standardized regression coefficient (Beta) for the following models:(a) Baseline SCL-IS scores predicting baseline HAMD-17 scores (all groups) - this model allows us to determine if baseline interpersonal sensitivity is correlated with baseline depression severity, and if so, we can control for baseline depression severity in subsequent models (see b, c, and d). This will need to be done to account for floor effects associated with baseline severity that can bias change score estimates (10). (b) HAMD-17 change scores predicting SCL-IS change scores (in the medication group) - this model will allow us to estimate the extent that depression improvement accounted for/explained interpersonal sensitivity improvement in the medication groups, and will complement the analysis conducted earlier in Step 5.(c) Baseline SCL-IS scores predicting HAMD-17 change scores (all groups) - this is done to explore whether baseline interpersonal sensitivity was associated with more or less depression improvement. (d) Repeat model ‘c' with treatment condition (medication vs. placebo) as a covariate and an interaction between baseline SCL-IS scores and treatment condition - this is done to test whether the correlation observed in model (c) differed between treatment groups (which would be evidenced by a significant interaction).Again, these models (‘a', ‘b', ‘c', and ‘d') will be estimated using the ipdmetan command to get the pooled estimate of interest (in this case, Beta) while also observing the estimate for each trial and testing for significant between-trial effect size heterogeneity, if it exists. The command follows the same analytical procedure as discussed in Step 4: the effect of interest is separately calculated for each study, these are pooled using inverse variant weights, and between-trial effect heterogeneity is observed both on the forest plot and formally tested with a Q statistic. References9. Fisher, D. J. (2015). Two-stage individual participant data meta-analysis and generalized forest plots. The Stata Journal, 15(2), 369-396.10. Vöhringer, P. A., & Ghaemi, S. N. (2011). Solving the antidepressant efficacy question: effect sizes in major depressive disorder. Clinical therapeutics, 33(12), B49-B61.