This work may not be copied, distributed, displayed, published, reproduced, transmitted, modified, posted, sold, licensed, or used for commercial purposes. By downloading this file, you are agreeing to the publisher’s Terms & Conditions.

Original Research

Patient and Clinical Factors Associated With Response to Medications for Posttraumatic Stress Disorder


Objective: Fluoxetine, paroxetine, sertraline, topiramate, and venlafaxine have previously shown efficacy for posttraumatic stress disorder (PTSD) in randomized clinical trials. Two prior studies using Department of Veterans Affairs (VA) medical records data show these medications are also effective in routine practice. Using an expanded retrospective cohort, we assessed the possibility of differential patterns of response based on patient and clinical factors.

Methods: We identified 6,839 VA outpatients with clinical diagnoses of PTSD between October 1999 and September 2019 who initiated one of the medications and met pre-specified criteria for treatment duration and dose, combined with baseline and endpoint PTSD checklist (PCL) measurements. We compared 12-week changes in PCL score within clinical subgroups defined by sex, race and ethnicity, and military exposures, as well as comorbidities. Comorbidities were identified using International Classification of Diseases diagnostic codes and grouped according to major diagnostic classifications in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (eg, Psychotic Disorders, Depressive Disorders). We used a propensity score weighting approach to balance covariates among medication arms within each clinical subgroup. In our exploratory analyses using unweighted data for the overall cohort, we built penalized logistic regression models to identify covariates that predicted meaningful improvement.

Results: There were no significant differences between medications in our weighted subgroup analyses. In unweighted exploratory analyses, higher baseline PCL scores and concurrent receipt of evidence-based psychotherapy predicted meaningful improvement, while high levels of disability predicted not realizing meaningful improvement.

Conclusions: In the largest real-world study of medications for PTSD to date, we did not observe a pattern of differential response among clinical subgroups. All patients taking medications for PTSD, especially those with the highest levels of disability, should consider combined treatment with evidence-based psychotherapy.

J Clin Psychiatry 2021;82(6):21m13913

To cite: Shiner BR, Gui J, Rozema L, et al. Patient and clinical factors associated with response to medications for posttraumatic stress disorder. J Clin Psychiatry. 2021;82(6):21m13913.
To share:

© Copyright 2021 Physicians Postgraduate Press, Inc.

aNational Center for PTSD, White River Junction, Vermont
bVeterans Affairs Medical Center, White River Junction, Vermont
cDepartments of Psychiatry and The Dartmouth Institute for Health Policy & Clinical Practice, Geisel School of Medicine at Dartmouth, Hanover, New Hampshire
dDepartments of Biomedical Data Science, Community & Family Medicine, and The Dartmouth Institute for Health Policy & Clinical Practice, Geisel School of Medicine at Dartmouth, Hanover, New Hampshire
eNational Center for PTSD, White River Junction, Vermont
fDepartment of Psychiatry, Geisel School of Medicine at Dartmouth, Hanover, New Hampshire
gOffice of Systems Redesign and Improvement, Washington, DC
*Corresponding author: Brian R. Shiner, MD, MPH, VA Medical Center; 215 North Main St, 116D, White River Junction, VT 05009 (



Randomized controlled trials (RCTs) show that effective treatments for posttraumatic stress disorder (PTSD) include both pharmacologic and psychotherapeutic approaches.1,2 Several individual medications have shown efficacy as PTSD treatment in placebo-controlled RCTs,1,2 but there have not been head-to-head prospective comparisons of the most effective agents. Two prior retrospective studies using Department of Veterans Affairs (VA) data to examine 5 of these medications—fluoxetine, sertraline, paroxetine, topiramate, and venlafaxine—indicate that they are also effective in routine clinical practice.3,4

There is a great deal of variation among patients with PTSD in terms of demographic characteristics and comorbidities that appear to drive medication selection and may influence outcomes.5,6 For example, patients with PTSD and comorbid pain disorders, headache disorders, and alcohol use disorder (AUD) are increasingly likely to receive anticonvulsants such as topiramate.5 While expert opinion has focused on factors such as comorbidity in the selection of specific medications for PTSD,7 there is a lack of definitive data to support such an approach.8 In the absence of RCTs examining treatment effectiveness in specific clinical scenarios, retrospective studies using health care data can inform clinical decision making based on patient and clinical factors.9

Building on our prior work,3,4 we conducted a retrospective comparative effectiveness study of the same 5 medications for PTSD in clinically important subgroups defined by sex, race and ethnicity, and military service characteristics, as well as comorbidities. In addition to examining these pre-identified groups, we conducted exploratory analyses to identify other potential predictors of meaningful improvement in symptoms. As there has been no prior research comparing these agents within clinical subgroups, we did not have a hypothesis. Rather, our goal was to provide clinicians with preliminary information about how to best select a medication based on demographic characteristics and comorbidities. While we did not have formal hypotheses, we expected to find differences in the pattern of response related to differences between the medications. For example, this might include superior PTSD symptom reduction in patients with comorbid headache disorders when they receive topiramate, presumably representing a synergistic effect of addressing both problems.


Data Sources

This was a retrospective medical record review. We used the VA Corporate Data Warehouse (CDW) to identify all VA users with a clinical diagnosis of PTSD (309.81, F43.1x) from October 1, 1999–September 30, 2019. We obtained information on services use, clinical diagnoses, prescription fills, and patient-reported outcome measures (PROMs) from the CDW for these patients. This study was approved by the Veterans Institutional Review Board of Northern New England.

Cohort Selection

We identified patients who initiated a course of fluoxetine, sertraline, paroxetine, topiramate, or venlafaxine. The study sample was further restricted to those who met our criteria for adequate acute phase medication management. Patients receiving continuous treatment with sertraline, fluoxetine, paroxetine, venlafaxine, or topiramate daily for ≥ 12 weeks at an adequate dose were considered to have received an adequate medication trial (AMT). Adequate doses, which were required for the final 8 weeks only to allow for titration, were as follows: fluoxetine ≥ 20 mg, paroxetine ≥ 20 mg, sertraline ≥ 100 mg, topiramate ≥ 100 mg, and venlafaxine ≥ 150 mg. We further restricted to those who received baseline PTSD symptom measurement within 4 weeks prior to or 2 weeks after treatment initiation, received follow-up symptom measurement within 2 weeks prior to or 4 weeks after the 12-week point, and met our criterion for PTSD severity at baseline (defined below).

PTSD Symptoms

In order to maximize sample size within our clinical subgroups, we integrated 2 different versions of a PROM for PTSD, captured from up to 2 data sources within the CDW, to obtain our baseline and follow-up symptom measurements. This included scores obtained from structured data produced by psychometric assessment software in the VA medical record and scores documented by clinicians in their treatment notes. We used a previously published natural language processing (NLP) algorithm with 98% precision in identifying the correct score and version of the PCL to abstract scores from clinical notes.10,11 Scores abstracted from structured data and from NLP of clinical notes were integrated into a single dataset.

The two PROMs were the PTSD Checklist (PCL) versions aligned to the Diagnostic and Statistical Manual of Mental Disorders (DSM), Fourth and Fifth Editions,12,13 which we will henceforth call the PCL-IV and the PCL-5.14,15 Validation work shows a correlation of 0.87 between PCL versions in a large sample of Veterans.16 We used a validated crosswalk (intraclass correlation coefficient = 0.96) to convert all values to PCL-5 scoring17 and required a baseline severity score of ≥ 31 out of 80 to classify participants as having PTSD.16 We did not examine individual symptomatic criteria for PTSD both because individual item scores were not available when abstracting PCL data from note text using NLP and because the PCL version scoring crosswalk we used was based on total scores. We created a covariate for whether the original score was from the PCL-IV or PCL-5 due to prior findings that venlafaxine may have superior effects on PTSD as assessed using DSM-5, but not DSM-IV, which may be a result of additional items related to negative alterations in cognitions and mood.3 In addition to calculating continuous change from baseline to follow-up, we assessed a categorical outcome of clinically meaningful improvement, which was a decrease of 15 points or more from baseline to follow-up.18

Our a priori power calculations were based on a random sample of 200 patients from our first published study of the comparative effectiveness of evidence-based medications for PTSD in routine VA practice who had 5 repeated PCL measurements over their initial 8 weeks of an AMT (correlation, ρ = 0.7).4 We modeled between group differences in effect size per 2-week period for change in PCL with a power of 80% and a Bonferroni-corrected 2-sided type I error rate of 0.005 to account for 10 potential comparisons at each time point using a generalized estimating equation. We found minimum cell sizes of 288, 104, and 41 for small, medium, and large effects (d = 0.3, 0.5, and 0.8, respectively). Therefore, we eliminated subgroups with less than 41 AMTs in any of the 5 medication cells, as greater effect size differences were implausible based on a prior meta-analysis of RCT results.2

Independent Variables

We measured 6 groups of covariates that could plausibly affect the relationship between treatment and outcome. See Table 1 for details.


To conduct our primary analysis, we were guided by expert opinion in dividing the sample into putative clinical subgroups based on sex, race and ethnicity, military service era, military exposures, and comorbidities.7 We defined comorbidities by the presence of 2 or more outpatient diagnoses or 1 or more inpatient diagnosis in the year prior to medication initiation. We repeated 3 analytic steps described below separately for each clinical subgroup. As point of reference for subgroup results, we also conducted analyses in the overall group.

The first step in our primary analysis was to account for differences in covariate profile among trials of each of the 5 medications. We used the RAND Toolkit for Weighting and Analysis of Nonequivalent Groups (TWANG).20 The TWANG package supports causal modeling of observational data through the estimation and evaluation of propensity scores and associated weights. In our application, the propensity score represented the probability that a particular trial would be of each medication.21 We estimated propensity scores with multinomial logistic regression using generalized booster effects,22 in which the dependent variable is an indicator for each of the 5 medications and the independent variables are an antiparsimonious specification of variables that have a plausible correlation with the outcome (ie, our 6 groups of covariates).21,22 Using these propensity scores, we weighted participants in order to balance the covariate distributions across medications.

The second step in our primary analysis was to compare continuous and categorical outcomes among the 5 medications with weighted regression analyses, using medication received as the sole independent variable. In general, weighted means can have greater sampling variance than unweighted means. Therefore, we used survey commands, which account for the weights, to perform the outcomes analyses when comparing the weighted medication groups. These weighted medication groups were defined by the inverse of the propensity scores and adjusted covariates unbalanced at the P < .01 level after TWANG weighting. In balancing our extensive list of covariates, a Bonferroni correction would indicate a corrected α of P < .001. However, we conservatively maintained an α threshold of P < .01 for significant differences to avoid type II error. For our continuous outcome of change in total PCL score, we used weighted linear regression analysis, whereby the coefficient of the variable tests the hypothesis that each of the 5 psychotropic medications has the same mean change from baseline to follow-up. For our categorical outcome of clinically meaningful improvement, we used weighted logistic regression analysis, whereby the coefficient of the variable tests the hypothesis that each of the 5 psychotropic medications results in the same percentage of patients achieving clinically meaningful improvement. P values were calculated from Wald test in the propensity score weighted regression models.

The third step in our primary analysis was to the potential contribution of unmeasured confounding on significant baseline to follow-up differences by calculating E-values, which indicate the minimum strength of association on the risk ratio scale that an unmeasured confounder would need to have with both the exposure and the outcome, conditional on the measured covariates, to fully explain away a specific exposure-outcome association.23,24

For our exploratory analyses, we conducted penalized logistic regression models to identify the strongest predictors of meaningful improvement and remission using least absolute shrinkage and selection operators (LASSO).25 As we were interested in predictors within (rather than between) comparison groups, we used raw (rather than propensity-score weighted) covariates described in Table 1. We chose LASSO because it provides information about predictors that are most important when many covariates are available. We set the tuning parameter to select the most regularized model such that error is within 1 standard error of the cross-validated minimum. We evaluated the robustness of our feature selection using 100 bootstrapped samples. At the extreme ends of the distribution of bootstrapped replications, some features that are important in the full model are dropped by LASSO. We ran LASSO models in 6 groups: overall (including an indicator for medication received) and patients who received each of the 5 medications. We performed data management in SAS version 9.4 (SAS Institute)

Volume: 82

Quick Links: PTSD , Trauma