Propensity Score Weighting Accounts Well for Placebo Response in Two Trials of Medical Therapy for MDD
Key findings
- Propensity score weighting (PSW), a novel statistical technique for analyzing randomized controlled trial data, aims to achieve balance between exposed and unexposed trial arms in the distribution of patients who are likely to have a placebo response
- This study evaluated the performance of an AI approach to PSW when it was used to re-analyze the results of two randomized controlled trial (RCT) of therapies for major depressive disorder (MDD) that presented negative or borderline results
- The PSW methodology confirmed the inefficacy of the investigational drug in a negative trial, whereas it identified a strong efficacy signal in a trial that had been considered failed, in which the placebo response had been extraordinarily high
- Although this was a post hoc study, the PSW methodology could be prospectively applied to any conventional RCT in MDD that meets certain design criteria
In 2012 a meta-analysis of 204 pharmaceutical trials in major depressive disorder (MDD), published in The Journal of Clinical Psychiatry, established that the higher the placebo effect, the lower the estimated treatment effect will be. Yet statistical analyses implicitly assume the placebo effect is the same for all study participants.
Subscribe to the latest updates from Psychiatry Advances in Motion
Propensity score weighting (PSW) is a novel statistical technique for analyzing randomized controlled trials (RCT) that aims to balance the impact of the placebo effect between exposed and unexposed trial arms.
Massachusetts General Hospital researchers previously reported in Translational Psychiatry an artificial intelligence (AI) approach to PSW they successfully applied to analyze a trial of paroxetine in MDD.
Now, Maurizio Fava, MD, psychiatrist-in-chief and executive director of the Clinical Trials Network and Institute at Massachusetts General Hospital, and colleagues have used PSW to re-analyze two RCTs in MDD that were complicated by high placebo response rates. In a new paper in Translational Psychiatry they show that PSW targets the "sweet spot" between false-negative and false-positive RCT results.
Background on the Trials
The researchers examined data from two double-blind multicenter studies:
Study SEP360-029—514 patients with MDD were randomly assigned to 0.5 mg or 2.0 mg dasotraline, 150 mg venlafaxine (active comparator), or placebo for eight weeks. No clinically meaningful or statistically significant treatment effect was noted for the primary endpoint, the total score on the Hamilton Rating Scale for Depression (HAMD-17) at week 8, so the study was considered negative.
Study SEP380-201—289 outpatients with a current MDD episode associated with type 1 bipolar disorder were randomly assigned to use non-racemic amisulpride 200 or 400 mg/day or placebo for six weeks.
Both amisulpride groups showed numerical improvement on the primary endpoint, the change from baseline on the Montgomery–Åsberg Depression Rating Scale (MADRS) at week 6, but neither change was statistically significant. In the placebo group, the MADRS score improved by almost 15 points, an extraordinarily high placebo response, and the study was considered to have failed.
Methods of the Re-Analyses
The PSW methodology for re-analyzing the trials had four main steps:
- For each individual assigned to the placebo groups, an artificial neural network (ANN) calculated the change from screening to baseline on each item in the HAMD-17 and each of the 10 items in the MADRS, then predicted the subject's probability of being a placebo responder (defined here as ≥50% change from baseline in HAMD-17 total score at week 8 or MADRS total score at week 6).
- The ANN was validated by comparing its predicted probabilities of placebo response to the observed values.
- For each study, the ANN was used to predict the probability of every participant (all trial arms) being a placebo responder.
- The inverse of the estimated individual probability was included as a weighting factor in the model used to analyze the treatment effect (the baseline-corrected change from placebo at study end).
Results
The re-analyses confirmed the absence of drug effect in the negative trial but indicated drug efficacy in the failed trial:
Study SEP360-029—The re-analysis identified a high proportion of placebo responders, comparable to the number in study SEP380-201, but there was no separation between placebo and either dose of dasotraline, and the efficacy of the active comparator (venlafaxine) was confirmed.
Study SEP380-201—The re-analysis showed substantial separation of amisulpride from placebo, and treatment effect sizes in the two amisulpride arms were about twice as large as the values estimated in the original analysis.
Future Directions
The PSW methodology could be prospectively applied to any conventional RCT as long as the trial is designed to collect screening and baseline data, the criteria for assessing the clinical placebo response are prespecified, and the criteria for implementing and qualifying the predictive performance of the ANN model are prespecified.
Pre-randomization parameters other than a change in individual HAMD-17 or MADRS scores could be used to predict placebo response. Examples include demographic data, quality-of-life ratings, and disease-related information.
view original journal article Subscription may be required