EndpointsWe will use response rates, defined as a 50% or more score reduction from baseline to last observation, for three scales: the 17-item Hamilton Depression Rating Scale (HAMD-17), the 31-item Hamilton Depression Rating Scale (HAM-31), and the Montgomery-Asberg Depression Rating Scale (MADRS). To handle missing data we will use last-observation-carried-forward imputation.Step 1: Defining subgroupsMelancholic depression. We see from study documents that this diagnosis was recorded as part of the baseline screens for studies SCAA100223, SCA30924, and SCA40910, which we assume were taken from responses to the Structure Clinical Interview of DSM-IV. If this variable is available for all the datasets then we will use it to define melancholic depression at baseline. If it is not, then we will use the definition diagnostic method from our previous study (1), based on HAMD and MADRS scores according to the DSM criteria:(a) One of: (1) Anhedonia - MADRS Question (Q) 8 (Answers [A] 4-6) (2) Lack of reactive mood - MADRS Q 1 (A 6) or Q 2 (A 6)(b) And at least three of: (1) Mood worse in morning - HAMD RS Q 18 (A 1-2 if specify in am) (2) Terminal insomnia - HAMD Q 6 (A 1-2) (3) Psychomotor changes - HAMD Q 9 (A 2-4) or Q 8 (A 2-4) (5) Decreased appetite or weight - HAMD Q 12 (A 2) or Q 16 (A 2) (6) Excessive guilt - HAMD Q 2 (A 2-4)In the three studies for which we were able to review the accompanying documents, the prevalence of melancholic depression was 44-53%. If our definition yields prevalence less than 40% or greater than 60% we will make it more or less inclusive until it is between 40-60%. The prevalence in our previous study was 60%, so we have already tightened the score requirement for MADRS Question 8 (we used data from SCA100223, which had the highest prevalence of the three trials).Depression severity. We will split baseline HAMD-17, HAMD-31, and MADRS scores at their mean, similar to Geddes et al. (2).Psychomotor retardation. We will use the baseline HAMD-17 Question 8.Step 2. Sample descriptivesThe sample will have age, sex, and baseline depression scores described according to melancholic and placebo/drug status with appropriate tests for difference.Step 3. Examination of lamotrigine and placebo response rates by subgroupChi-squared proportion tests will be used to compare lamotrigine and placebo response rates by melancholic status. These are not tests of treatment effects because they do no compare lamotrigine to placebo, but rather, for differences in lamotrigine and/or placebo response between subgroup categories. We will also conduct a test of difference-of-difference in proportions to examine if the treatment effect of lamotrigine varies by depression subtype by calculating Wilson-Newcombe confidence intervals.We will also create a series of 2x2 tables (treatment group by response) stratified by melancholic status and baseline depression severity, with associated number-needed-to-treat statistics, relative risks increases (i.e., risk of response) and their associated 95% confidence intervals. Step 4. Logistic regression analysesWe will conduct a series of logistic regression analyses predicting each outcome variable. Models for each outcome variable will be stratified by melancholic status, and then repeated un-stratified with a melancholic by treatment interaction. Baseline depression scores, measured with the same scale that defined each outcome, will then be added as a covariate to these models.Step 4. Exploratory analysesWe will repeat steps 2-4 but with psychomotor retardation as a subgroup variable.We will also create Receiver operating characteristic (ROC) curves for melancholic depression and psychomotor retardation, with placebo response and treatment response as the outcomes (in other words, stratified by treatment group). To answer Question 3, we will calculate the areas under the ROC curve, first with depression subtype as predictor and repeat the process with the single question for psychomotor retardation. Whichever predictor has greater area under the ROC curve is the better index for lamotrigine response.Essentially we are trying to see if it is possible to split the sample based on psychomotor retardation, and for this variable to perform similarly as the melancholic depression as a predictor of treatment response and/or placebo response. If so, this would be useful to know, as psychomotor retardation can be more easily and quickly assessed than melancholic depression.We would also like to examine the treatment effect (change score for lamotrigine vs. placebo) on different HAMD and MADRS items, and to use logistic regression to determine if certain baseline items are associated with likelihood of response to lamotrigine/placebo. It is very likely that the effect size for lamotrigine on core depressive symptoms (e.g., low mood, anhedonia) was masked by floor effects/lack of change on other nonspecific items (e.g., sleep, appetite, etc.). This has been found for most common antidepressants, particularly those that are not sedating or do not have marked effects on appetite like lamotrigine (e.g., SSRIs/SNRIs; see Lisinski et al., 2019
https://www.nature.com/articles/s41386-019-0523-4). This is also important because we are defining melancholia with several baseline MADRS/HAMD items, and we are assuming that it is the combination of these items that are related to treatment response, but in reality it may only be one or two items with the rest being superfluous. This can be seen as constituting a sensitivity analysis evaluating our melancholia diagnosis: if it is the entire melancholic syndrome that is the main predictor of response, removing items from the classification scheme should reduce predictive power, and one or two core items (e.g., anhedonia, nonreactive mood) should not outperform the entire syndrome. For any significant models using the pooled sample, it would also be useful to do sensitivity analyses by systematically removing each trial from the pooled sample to determine if the effect disappears.