Background:

Posttraumatic stress disorder (PTSD) is a trauma- and stress-related disorder that causes substantial harm to patients and their families. Though serotonin reuptake inhibitors (SRIs) are effective in reducing PTSD symptoms, most will have an inadequate response to them. This often leads to the addition of other classes of medication (augmentation), such as other types of antidepressants, atypical antipsychotics, or prazosin (an α1 adrenoreceptor antagonist typically used to lower blood pressure). However, these augmenting medications have not been as well studied for use with PTSD, and several have potentially harmful metabolic and cardiovascular side effects. As no trials have directly compared the benefits and risks of these medications in PTSD, patients and providers are left in a challenging position when trying to make treatment decisions.

Objectives:

To use Veterans Health Administration electronic records data to compare mental health and metabolic/cardiovascular outcomes in patients with PTSD taking an SRI who are prescribed 4 augmenting medication categories: (1) atypical antipsychotics, (2) mirtazapine, (3) prazosin, and (4) tricyclic antidepressants. Specifically, our aims were as follows:

• Aim 1: To compare the impact of augmenting medications on the following mental health outcomes:

  –

  PTSD symptoms (primary outcome; measured with the PTSD Checklist [PCL])

  –

  Psychiatric hospitalizations and psychiatric emergency department (ED) visits

  –

  Suicidal ideation (measured through mandatory Veterans Affairs [VA] screens)

• Aim 2: To compare the impact of augmenting medications on the following metabolic and cardiovascular outcomes:

  –

  Weight (primary outcome)

  –

  Low-density lipoprotein (LDL) cholesterol, high-density lipoprotein (HDL) cholesterol, and triglyceride levels

  –

  Blood glucose and hemoglobin A_1c(HbA_1c)

  –

  Blood pressure

  –

  Incident diagnoses of obesity, dyslipidemia, diabetes, and hypertension

  –

  Use of medications to treat metabolic risk factors

  –

  Cardiovascular and cerebrovascular disease events

  –

  All-cause mortality

• Aim 3: To examine variations in the risks and benefits of augmenting medications in specific demographic subgroups:

  –

  Female veterans

  –

  Iraq and Afghanistan veterans

  –

  Veterans aged ≥65 years

Methods:

We constructed a cohort of 872 324 VA patients who received a diagnosis of PTSD from January 1, 2007-December 31, 2015, and who did not have a comorbid diagnosis of bipolar disorder or a psychotic disorder. We further restricted the sample to those taking an SRI who filled a new prescription for 1 of the 4 groups of augmenting medications for at least 60 days in a period of 120 days. We considered the date of starting this new augmenting medication as time “index date.” We then compared changes in several mental health outcomes from the year before the year following the index date. Our primary mental health outcome was change in PTSD symptoms as measured by the PCL. We also evaluated changes in ED visits and hospitalizations for mental health problems as well as changes in rates of suicidality. For metabolic and cardiovascular outcomes, we evaluated changes in weight, total cholesterol, HbA_1c, glucose, LDL cholesterol, HDL cholesterol, triglycerides, and blood pressure. We also compared rates of incident diagnoses of obesity, diabetes, dyslipidemia, and hypertension, as well as the use of medications to treat these conditions in the postindex year. Finally, we evaluated incident cardiovascular disease events and all-cause mortality in the postindex year. We also examined heterogeneity in these effects by sex, age, and era of military service. To control for potential confounding in this observational design, we adjusted for potentially confounding covariates as well as for propensity score weighting. We conducted numerous sensitivity analyses to test our conclusions.

Results:

We found that patients had statistically significant but clinically small improvements in their PTSD symptoms after receiving augmenting medications, and the effect was largely similar across classes (antipsychotics: −0.79 points on the PCL [score range, 17-85]; 95% CI, −1.04 to −0.54 points; mirtazapine: −0.77 points; 95% CI, −1.06 to −0.49 points; prazosin: −1.09 points; 95% CI, −1.28 to −0.91 points; tricyclics: −0.93 points; 95% CI, −1.41 to −0.44). We observed more dramatic reductions in rates of ED visits and hospitalizations for mental health conditions (antipsychotics: −3.15 visits/100 person-years; 95% CI, −3.80 to −2.51 visits/100 person-years; mirtazapine: −2.80 visits/100 person-years; 95% CI, −3.46 to −2.14; prazosin: −2.78 visits/100 person-years; 95% CI, −3.28 to −2.29; tricyclics −1.58 visits/100 person-years; 95% CI, −2.42 to −0.75) and in rates of suicidal ideation (antipsychotics: −2.63%; 95% CI, −3.41% to −1.85%; mirtazapine: −1.87%; 95% CI, −2.72% to −1.02%; prazosin: −2.64%; 95% CI, −3.25% to −2.03%; tricyclics: −1.07%; 95% CI, −2.39% to 0.26%). When we further evaluated the time course of these changes, it was apparent that medications were started after increases in symptoms, hospitalizations, and suicidal thoughts were seen on a population level. Following initiation of the medication, levels then returned to baseline but generally did not continue to improve beyond the baseline level. Changes in mental health outcomes tended to be similar across the 4 augmenting medication classes. We observed adverse metabolic effects that were most notable for mirtazapine, followed by antipsychotics; these effects were minimal with prazosin. Metabolic changes tended to be most dramatic for outcomes that were not as responsive to medications, such as weight and triglycerides. Rates of use of new medications or intensification of existing regimens to treat metabolic outcomes were also very high, which may have obscured the impact of the augmenting medications. We also found, relative to prazosin, that the other medications were associated with increased risk of incident cardiovascular disease risk factors and event diagnoses, as well as all-cause mortality. The effects on mental health and metabolic outcomes were generally similar across subgroups. However, we found that adverse metabolic changes were more dramatic in female and younger (aged <65 years) veterans.

Conclusions:

Our findings suggest that the use of medications to augment SRI treatment in patients with PTSD may lead to mental health benefits, particularly in those who have a flare in symptoms or in patients with more severe illness or who are otherwise unstable. This may come at a cost of worsening metabolic parameters, particularly for those prescribed mirtazapine or antipsychotics. Therefore, patients should be closely monitored for both symptom and functional improvement as well as metabolic impact. It will be important to confirm these findings with randomized clinical trials given the observational nature of our study.

Limitations:

These analyses are based on observational data, and therefore, we cannot conclude that the associations between the medications and outcomes are causal. In addition, though we remained powered to detect small effects, PTSD symptom outcomes were available for a minority (18%) of the sample.

Inadequate Response to First-Line Therapies

Though multiple behavioral and pharmacologic therapies are effective in treating PTSD symptoms, many patients will not respond to these treatments. In terms of psychotherapy, competing demands from work, home life, and other medical conditions may make completion challenging. Indeed, previous studies found that only 6% to 33% of veterans with PTSD receive a minimally adequate course of evidence-based psychotherapy.^21,22 The use of medications for PTSD may help overcome some of the challenges of finding time for or having difficulty accessing PTSD specialty care.²³ Medications also provide additive benefit for those who are engaging in PTSD psychotherapy.²⁴ Serotonin reuptake inhibitors (SRIs), including selective inhibitors such as sertraline reuptake inhibitors and serotonin/norepinephrine reuptake inhibitors (collectively referred to as SRIs in this report) such as venlafaxine, are proven to reduce PTSD symptoms in randomized controlled trials (RCTs) and are considered to be first- line medications for PTSD.^25-27 Unfortunately, many patients with PTSD will not respond to first-line therapy with SRIs, and SRIs may be less effective in treating chronic rather than acute PTSD symptoms.²⁸ In clinical trials, only 60% of patients taking SRIs will have significant improvements in PTSD symptoms, and only 20% to 30% will achieve complete remission.²⁹

Identifying Gaps in Our Understanding of How to Treat Patients Who Do Not Respond to First-Line Medications

We developed our research questions in collaboration with our patient and stakeholder partners, identifying questions and concerns important to patients through several group discussions and gaps in the evidence using a systematic review completed for the Veterans Administration/Department of Defense (VA/DoD) Clinical Practice Guideline for the Management of Posttraumatic Stress.²⁷ This guideline is widely disseminated and is the main prescribing resource for providers working with veterans with PTSD. The multidisciplinary working group that developed the guideline conducted a systematic review of all PTSD treatment trials. For patients who do not have an adequate response to SRI therapy, augmentation with an additional medication is recommended. Though the guideline makes recommendations about several medications that could be used in this setting, no RCTs have directly compared the risks and benefits of these strategies. Therefore, patients and providers are left in a difficult position without recommendations based on strong evidence.

To address this critical knowledge gap, we proposed to compare the risks and benefits of several medications that have proven efficacy (prazosin, mirtazapine, and tricyclic antidepressants [class B in the VA guideline]) or have conflicting evidence but are commonly used as augmenting strategies (atypical antipsychotics).²⁷ In any population of patients, it is critical to weigh potential benefits against the adverse effects of medications. Most of these augmenting strategies have serious potential harms, which include weight gain and increases in lipid levels, blood sugar, and blood pressure.^30-32 These metabolic and cardiovascular consequences are particularly concerning in veterans with PTSD, as research from our group and others has demonstrated that patients with PTSD have a substantially increased risk of developing and dying from CVD.^{5,10,13,14,33,34} Therefore, data on the comparative benefits and harms of augmenting medications for PTSD are urgently needed. Our team had extensive experience working with national VA electronic health records (EHRs) and proposed to build a cohort of veterans diagnosed with PTSD who had been prescribed 1 of the 4 augmenting medication classes and to follow them over time to compare mental health benefits and metabolic/cardiovascular harms.

Improving Care for Patients and Addressing Limitations of Previous Work

Though observational designs have limitations for making conclusions about causality, a clinical trial would take substantially more time and resources. As patients and providers had no information on the comparison of these medications, our team felt that the outcomes of this study would provide much-needed guidance to assist with shared decision-making while also highlighting the most promising studies that could be further evaluated with criterion-standard clinical trials. The benefits of using comprehensive VA data could also address some of the challenges in clinical trials. Even the largest clinical trials had been limited to a few hundred participants generally followed for 6 months, though most trials were smaller and shorter. The use of VA data would allow us to examine the long-term effects of these medications on important but rarer clinical outcomes, including CVD events and death. We could also examine the efficacy and risks of these medications in a real-world patient population as opposed to a highly selected, volunteer clinical trial population. This is particularly important in PTSD, a disorder that is highly comorbid with other psychiatric and medical conditions. For example, previous trials have typically excluded patients with active suicidal ideation and substance dependence, common comorbidities among veterans with PTSD that may affect the safety and efficacy of medications. Last, we could focus on important and understudied subgroups that may have distinct risk and benefit profiles. For example, women are the fastest-growing group of veterans, yet they have been grossly underrepresented in trials of PTSD, and most studies in veteran populations have not included any women.²⁸ Returning Iraq and Afghanistan veterans were another priority population, and their younger age may affect the balance of risks and benefits of treatment; again, no studies had specifically focused on these veterans. Finally, older veterans may be at particularly high risk for the adverse metabolic and cardiovascular consequences of augmenting medications.³⁵ However, because of their higher rate of underlying medical comorbidities, they are often excluded from clinical trials.³⁶ With these research gaps in mind, we designed a VA EHR-based observational study to address the following specific aims:

• Aim 1: To compare the impact of augmenting medications on the following mental health outcomes:

  –

  PTSD symptoms (primary outcome; measured with the PTSD Checklist [PCL])

  –

  Psychiatric hospitalizations and psychiatric emergency department (ED) visits

  –

  Suicidal ideation (measured through mandatory VA screens)

• Aim 2: To compare the impact of augmenting medications on the following metabolic and cardiovascular outcomes:

  –

  Weight (primary outcome)

  –

  Low-density lipoprotein (LDL) cholesterol, high-density lipoprotein (HDL) cholesterol, and triglyceride levels

  –

  Blood glucose and hemoglobin A_1c(HbA_1c)

  –

  Blood pressure

  –

  Incident diagnoses of obesity, dyslipidemia, diabetes, and hypertension

  –

  Use of medications to treat metabolic risk factors

  –

  Cardiovascular and cerebrovascular disease events

  –

  All-cause mortality

• Aim 3: To examine variations in the risks and benefits of augmenting medications in specific demographic subgroups:

  –

  Female veterans

  –

  Iraq and Afghanistan veterans

  –

  Veterans aged ≥65 years

Development of Study Protocol With Stakeholder Partners

Our stakeholder partners were instrumental in developing our research question and study protocol. Initially, we planned to quantify changes in metabolic parameters and CVD events in patients with PTSD who were prescribed antipsychotics vs other psychiatric medications. However, through consultation with our stakeholder partners, we identified a knowledge gap around the comparative safety and effectiveness of antipsychotics and several other medications that are commonly used when patients do not have an adequate response to first-line therapies for PTSD. In previous national surveys of providers, our partner Nancy Bernardy, PhD, learned of providers’ frustration with the lack of evidence and guidelines surrounding decision-making in this setting. Dr Bernardy was also involved in creating the VA/DoD PTSD treatment guideline and had firsthand knowledge of the challenges in advising providers. She expressed that the guideline committee needs studies comparing augmenting medications so that they can offer algorithms for augmentation, as well as better descriptions of the risks and benefits of specific strategies. Additionally, our partner Ilse Wiechers, MD, MHS, MPP, had been working to reduce potential harms of antipsychotic use in patients with PTSD but admitted that it was difficult to counsel providers to avoid these widely used medications without further evidence that other augmenting strategies were safer. Last, our patient partners described how patients often feel that when first-line therapies do not work, other medications are “thrown at them,” and they have little understanding of how choices are made or knowledge of risks/benefits to be able to participate in decision-making. This leads to frustration and, in many cases, discontinuation of mental health care. Based on these perspectives and feedback, we refined our goals to focus on comparing the risks and benefits of several augmenting medications.

Stakeholder meetings were held 3 times yearly throughout the course of the project. Veteran patient partners provided important input on health outcomes that are relevant to patients. Our provider partners were instrumental in providing insight and feedback on mental health conditions for cohort selection criteria, appropriate laboratory tests and diagnoses to examine as outcomes, and medication use definitions, ensuring that those decisions reflect and account for clinical realities. Stakeholder partners were actively involved in decisions regarding data cleaning and analysis, with Drs Bernardy and Wiechers providing particular expertise on VA pharmacy data and multiple issues related to the data and selection criteria. At each stage of the study, our stakeholder partners provided insight into our findings from primary analyses and sensitivity analyses, as well as feedback on areas that may require more detailed analysis or redefined criteria to more accurately capture outcomes.

Stakeholder Involvement in Dissemination

During stakeholder meetings, patient and stakeholder partners were directly involved in discussions regarding manuscript preparation, including which findings to present and how to frame the discussions. Partners have been actively communicating through email with the study team and are fully engaged in all aspects of manuscript preparation. All stakeholders have provided substantive feedback on the 2 manuscripts currently in preparation, as well as on past conference presentations.

Study Overview

We constructed a retrospective cohort using patients seen in VA care from January 1, 2007-December 31, 2015. We limited our study to veterans who received a diagnosis of PTSD in their medical record, were prescribed a first-line SRI medication for PTSD, and had 1 of the following medications added: atypical antipsychotics (quetiapine, risperidone, or olanzapine), prazosin (an α1 adrenoreceptor antagonist typically used to lower blood pressure), mirtazapine, and tricyclic antidepressants (amitriptyline or imipramine). We selected a limited number of atypical antipsychotics and tricyclic antidepressants in consultation with PCORI to better examine their comparative effectiveness. Within these classes, we selected the medications that had the best evidence of potential benefit for PTSD and were more widely used within the VA during the study period. We compared mental health and metabolic/cardiovascular outcomes for each of the augmentation strategies. Specifically, we considered the initial date the augmenting medication was filled as the “index date” and evaluated the change in mental health and metabolic/cardiovascular parameters from 1 year before and 1 year following the index date (Figure 1). In addition, we compared rates of CVD events and mortality in the year following the index date. We conducted multiple sensitivity analyses to assess the robustness of our conclusions.

Figure 1

Overview of Study Design.

Study Setting

Not applicable.

Participants

To form this retrospective cohort, we selected patients who had received a diagnosis of PTSD (ICD-9-CM code 309.81 or ICD-10 F43.10, F43.11, or F43.12) in 2 or more outpatient encounters or 1 inpatient encounter between January 1, 2007, and December 31, 2015. Patients were selected from the Veterans Health Administration (VHA) Corporate Data Warehouse (CDW), which is a comprehensive national repository of clinical and administrative data from within the VA and other external sources. We allowed at least 365 days (1 year) of follow-up time from the index date when the augmenting medication was added for outcome assessment. We restricted our population to those who had filled a prescription for first-line SRI antidepressant therapy for PTSD, as defined according to the latest VA/DoD clinical guideline (selective serotonin and serotonin/norepinephrine reuptake inhibitors).

Patients were included if they had filled at least 30 days of an SRI in the past 365 days (the “pre-index” year) and started an augmenting medication. Patients were required to be free of the augmenting medication for 180 days leading up to the index date so that we could examine the change in outcomes associated with initiation of the augmenting medications. The new augmenting medication had to be filled for at least 60 days within a 120-day period, and the patient needed to receive a prescription of an SRI for at least 60 days in the “postindex” year of 365 days. The date of the first prescription for the augmenting medication was used as the index date for the study. Veterans with a diagnosis of bipolar affective disorder or psychotic disorders (using 2 or more outpatient encounters or 1 inpatient encounter) were excluded, as these conditions are widely accepted, FDA-approved indications for the use of antipsychotic medications. To be representative of the real-world population of veterans who would benefit from our study, we kept our exclusion criteria minimal and did not exclude patients with additional psychiatric or medical comorbidities.

Interventions and Comparators or Controls

First-line SRI medications included sertraline, paroxetine, venlafaxine, fluoxetine, fluvoxamine, citalopram, escitalopram, desvenlafaxine, and duloxetine. Though only sertraline and paroxetine are FDA approved for the treatment of PTSD, other SRIs were frequently prescribed and therefore included. We selected 4 groups of medications that were commonly used and/or had evidence of benefit from previous clinical trials: (1) atypical antipsychotics (quetiapine, risperidone, or olanzapine), (2) prazosin, (3) mirtazapine, and (4) tricyclic antidepressants (amitriptyline or imipramine).

Though trials of atypical antipsychotic medications have yielded conflicting results, work from our group and others has found that they are frequently prescribed for PTSD.^37,38 The largest antipsychotic trial found that adjunctive risperidone had no benefit vs placebo on overall PTSD symptoms or QOL.²⁸ However, there were small but statistically significant reductions in reexperiencing and hyperarousal PTSD symptoms, and 2 meta-analyses of previous clinical trials found reductions in overall PTSD symptom scores with atypical antipsychotics.^39,40 Work from our group and others has demonstrated that veterans with PTSD commonly receive antipsychotics without another indication for their use.^41,42 To select the other augmenting medications for comparison, we used the expertise of our investigative team, stakeholder partners, and existing literature and clinical guidelines. We selected mirtazapine, prazosin, and tricyclic antidepressants, all of which have some trial evidence of benefit in PTSD and were recommended (class B) in the VA/DoD Clinical Practice Guideline at the time of the study design.²⁷ Nefazodone has shown some evidence of benefit but is not widely used due to the potential for life-threatening liver injury and other adverse effects. Phenelzine is also of benefit in PTSD but cannot be used as an augmentation strategy because it is contraindicated for use with SRIs. We did not include mood stabilizers or benzodiazepines, as these are considered ineffective or harmful (class D) in the guideline. Several other medications, such as propranolol and clonidine, were not included, as they are classified as having insufficient evidence and are not widely used for augmentation.

We used VA pharmacy prescription records that included medication type, dose, date of fill, amount dispensed, and days’ supply to create a record of medications that patients possessed on a given calendar day. To avoid gaps in medication use due to delays in prescription refill processing, patients may also request refills of prescriptions early. To account for this, we allowed “banking” of medications if a patient received a refill before they had run out of medication from prior fills. We also allowed use of these banked medications to prospectively cover gaps between prescriptions.

We defined a new augmenting medication as one of the medications listed above that had not been prescribed for the previous 180 days in a patient who had been prescribed an SRI for at least 30 days. We considered the index date for the study to be the fill date of the adjunctive medication. We compared patient outcomes the year before the index date with those in the year after the index date. We required that the adjunctive medication be used for at least 60 days within a 120-day period and that the SRI be used for at least 60 days in the postindex year. Polypharmacy is common in patients with PTSD, so in our primary analyses, we allowed patients to contribute data to more than 1 medication class and also to be taking other classes of augmenting medications as long as only 1 was “new” based on our definition. We conducted sensitivity analyses allowing patients to contribute data at only 1 time point and restricting to patients using only a single class of augmenting medication during the 2-year study observation period.

Study Outcomes

We worked with our patient and stakeholder partners to select outcome variables that were most important to patients and could be obtained with high integrity from our medical record data. Below are additional details on each of the outcomes.

Sample Size Calculations and Power

As this was an observational study using EHRs, we used all available data that met our inclusion and exclusion criteria rather than selecting a sample size based on power calculations. However, the large sample size of available VA patients did provide substantial power to perform the proposed analyses. For the majority of our outcomes, we had 90% power to detect a difference of 0.04 SDs in the mean outcome between treatment groups. For weight, our primary metabolic outcome, the SD of change from pre-index year to postindex year was 10.5 lb, yielding a minimum detectable effect at 90% power of 0.42 lb, which was below our MCID of 2.2 lb (1 kg). For binary outcomes, we were able to detect differences in outcome prevalence of 1%, 0.8%, and 0.4% for overall prevalences of 10%, 5%, and 1%, respectively, with 90% power. As described in the Changes to the Original Study Protocol section, we did have a greater number of missing PCL scores than expected. We went to great lengths to extract additional PCL scores, including using natural language processing (NLP). NLP is a text-processing algorithm that can search the text of clinical notes to identify terms, such as PCL score or PTSD Checklist score, and then extract the numeric score associated with these terms. Still, despite the rates of missing data, we had sufficient power to detect meaningful changes in this outcome. Our revised power calculations demonstrated that we had 90% power to detect differences of 0.08 SDs. This yielded a minimum detectable effect of a change in PCL score of 0.9 points, which was well below our MCID of 5 to 6 points.

Time Frame for the Study

The time frame for the primary analyses was 1 year (365 days) before the index date and 1 year follow-up after the index date, for a total of 2 years (730 days) for each eligible window. We selected a 1-year time frame in consultation with our stakeholder partners, because based on data from clinical trials, it was expected to be sufficient to capture improvements in PTSD symptoms and metabolic impacts of the medications under study. Although clinical trials typically have a shorter outcome assessment period, we felt that a full year of follow-up would also allow us to examine variations in effects based on the length of use of medication. The study included patients who received a diagnosis of PTSD on 2 or more outpatient or inpatient encounters at any VA health care facility between January 1, 2007, and December 31, 2015. Select covariates, including psychiatric and medical comorbidities, were searched through patients’ medical records for 2 years before the index date. Therefore, the overall study period included data from January 1, 2005, through December 31, 2016.

Data Collection and Sources

We used several data from the VHA CDW, Pharmacy Benefits Management database, and the VHA Vital Status Files. The VHA CDW is a comprehensive national repository of clinical and administrative data from within the VA and other external sources. The CDW contains information on VA inpatient and outpatient visits and associated clinical diagnoses, information on non-VA visits reimbursed by the VA, vital signs, VA pharmacy records, laboratory data, and other patient-level variables. The Pharmacy Benefits Management database contained additional detailed information about medication fills that we used to determine eligibility as described previously. The Vital Status Files combine vital status information from multiple sources, including the National Death Index data from the Suicide Data Repository, and were supplemented by mortality data from the CDW. Data from CDW and Vital Status Files were accessed through the VA Informatics and Computing Infrastructure (VINCI). All data sets were stored on secure servers behind VA firewalls in the VINCI workspace, a password-protected interface where approved team members could access the data. The raw and cleaned data set, statistical code, and results were saved and annotated to allow replication and verification of findings.

Covariates

Through our extensive prior research on VA patients with PTSD and review of the scientific literature, we identified multiple covariates that were important to consider in analyses of medication use and mental and physical health outcomes. Covariates included sociodemographic information, service use factors, mental health conditions, and medical comorbidities. The sociodemographic information collected for this study was age, sex, race/ethnicity, marital status, and rural vs urban status. Mental health and medical comorbidities collected for the study included depression, personality disorders, anxiety disorders, insomnia, substance abuse/dependence, alcohol abuse/dependence, traumatic brain injury, obesity, dyslipidemia, diabetes, hypertension, ischemic heart disease, congestive heart failure, and cerebrovascular disease determined by ICD-9/ICD-10 code diagnoses (see Appendix B for coding algorithms). The Charlson Comorbidity Index score was also calculated as a general measure of health status for patients. We used ICD-9/ICD-10 codes from the VA Northeast Program Evaluation Center to define mental health and substance abuse diagnoses.⁵⁰ Traumatic brain injury was defined using the DoD algorithm.⁵⁵ The service use factors we collected included distance to the nearest VA medical center, type of center (community-based outpatient clinic vs medical center), primary care use (number of primary care visits), level of VA service connection (a rating from 0% to 100% that reflects medical conditions related to military service and impacts patient costs for VA care), and mental health use (number of mental health visits).

Analytical and Statistical Approaches

Aim 1: To Compare the Impact of Augmenting Medications on the Following Mental Health Outcomes

Aim 2: To Compare the Impact of Augmenting Medications on Metabolic and Cardiovascular Outcomes

Aim 3: To Examine Variations in the Risks and Benefits of Augmenting Medications in Specific Demographic Subgroups

Changes to the Original Study Protocol

We reduced the number of days required to be free of the augmenting medication class before the index date from 365 days to 180 days. Our stakeholder partners were concerned that requiring patients to be free of an augmenting medication of a full year to identify a “new start” was unnecessarily conservative, as the psychological and metabolic effects would wash out more quickly. Also, given that our goal is to create a generalizable “real-world” cohort comparable with that of the veterans we treat rather than an idealized clinical trial cohort, our stakeholder partners felt we would have eliminated too many people, as this treatment-resistant population likely had tried multiple medications.

We also increased the minimum number of days augmenting medication use from 30 to 60 days. The minimum requirement was increased because some stakeholder partners were concerned that the effects of the augmenting medications might not be apparent before 60 days. They suggested that the 60-day minimum would better reflect the standard used in clinical trials to judge the safety and efficacy of medications. We restricted to a minimum of 60 days of use over a 120-day period to allow scattered use during a full year. We expanded the study time frame from the original January 1, 2010-September 30, 2014, to January 1, 2007-December 31, 2015, to compensate for the loss in the sample size that resulted from the increase in required medication use and missing data.

Based on discussions with team partners, we allowed and adjusted for the use of medications from 1 or more of the 4 classes of augmenting medications during the study. Allowing for this more-inclusive real-world sample reflected what our patients were actually using, and we felt it would allow us to better apply our results to the veterans we treat. To account for this change, we added a sensitivity analysis comparing results from a restricted population (patients using only 1 augmenting medication class during the 2-year observation period, which would more closely mimic a clinical trial) with the inclusive real-world sample (allowing and controlling for the use of other augmenting medication classes). All changes described above were approved by PCORI, and the contract was modified. The updated study protocol is included in Appendix A.

Once we applied our temporal restrictions for outcomes, we found that the amount of missing data for the PTSD symptom screens was greater than anticipated. For our primary outcome variable, the PCL, the VA has a specialized software program that providers are instructed to use to enter scores, which automatically populates a note in the medical record and sends data to the CDW database that we used to extract scores. However, our team was aware that many providers would enter the scores in the text of their notes instead of using the VA software, which would mean the scores would be missing in the CDW database. Therefore, we worked with colleagues at the Salt Lake City VA with expertise in NLP who were developing a method to extract PCL scores. We conducted validation studies comparing extracted scores with those that were available within CDW. We also examined text snippets for scores that were out of the PCL scoring range. In all cases, the algorithm extracted the score correctly, and the error was made by the provider entering the score. After eliminating these scores, use of the NLP method increased our total number of available scores for analysis by 10%. We repeated our analyses examining average change in PCL score from the pre-index year to postindex year for each of the augmenting medication groups. Our findings were similar to those from our previous analyses. These were included in our June 2019 Interim Progress Report, but our collaborators have requested that we not include these findings in any reports for publication because they feel additional validation is needed for the NLP algorithm. Therefore, for this report, we include PCL data from our original CDW extraction.

Finally, given the amount of polypharmacy and the complexity of cleaning and coding individual medication data, we decided not to conduct an originally planned sensitivity analysis that would have explored the impact of medication exposure over a 5-year period before the augmentation intervention. In discussion with our partners, it was agreed that our additional sensitivity analyses adjusting for use of other medication classes during the pre-index year was sufficient and that attempting to isolate the effects of medications used several years beforehand would be exceedingly complex and difficult to interpret.

Study Cohort Construction and Patient Characteristics

Figure 2 shows the sample size at each stage in the development of our cohort as we applied the inclusion and exclusion criteria as specified in the Participants section. In our primary analyses, we allowed patients to contribute data to more than 1 medication group as long as they met all inclusion and exclusion criteria at the time. Therefore, the number of 2-year study “windows” shown in the final step (green boxes) is greater than the number of patients (blue boxes). Therefore, it is possible that if 2 augmenting medications were initiated in similar time frames, the 2-year study windows for these observations could overlap. To address this, we also conducted sensitivity analyses using our “restricted sample” as described in the Changes to the Original Study Protocol section. In this sample, we only allowed patients to contribute a single window of data, selecting the first eligible window by calendar year.

Figure 2

Cohort Construction.

The characteristics of the 4 augmenting medication groups at the index date before and after the application of the propensity score weights are shown in Tables 2a and 2b. Before applying the weights, patients in the 4 augmenting medication groups were largely similar in terms of demographics, comorbidities, prescribing facility characteristics, and service use. Due to the large size of the data set, even small variations are statistically significant, but our investigative team and stakeholder partners felt that some clinically significant differences were the larger proportion of women in the group prescribed tricyclics (15.2% vs 7.7%-8.2% for the other medications) and the increased likelihood of substance (alcohol and other substances)/alcohol use disorders in those prescribed antipsychotics (36.6%/29.8%) or mirtazapine (34.9%/28.6%) compared with those prescribed prazosin (32.4%/27.4%) and tricyclics (26.2%/20.2%). Despite the similarities of the groups, we still pursued propensity score–weighted analyses as planned to further balance these and other covariates.

Table 2a

Patient Characteristics by Augmenting Medication Class: Unweighted.

Table 2b

Patient Characteristics by Augmenting Medication Class: Propensity Score Weighted.

Aim 1 Results

Given the large number of tables, complete results from all aim 1 primary and sensitivity analyses are included in Appendix C (primary analyses and sensitivity analyses described in the Analytical and Statistical Approaches section) and Appendix D (sensitivity analyses using the restricted sample taking medications from only 1 augmenting class during the study period as described in the Changes to the Original Study Protocol section). In the tables below, we present the average pre- and average postindex year values and change from pre-index year to postindex year for all outcomes. As the unweighted and propensity score–weighted pre- and postindex year averages were highly similar, we present only the propensity-weighted results below (unweighted and weighted results are shown side by side in the “Tables - Change in Outcomes” section of Appendices C and D). For each outcome, we then show the results from models comparing the change from pre- to postindex year in each of the augmenting medication groups with the antipsychotic groups. We show the unadjusted models, models adjusted for all covariates in Table 2a, and propensity-weighted models. As noted in the Analytical and Statistical Approaches section, given the number of hypotheses tested, we considered results of P < .001 to be statistically significant.

PTSD Symptom Score

For our primary outcome of PCL score, the mean scores in the pre-index year were in the low 60s, indicating moderate PTSD (Table 3a). Across all four augmenting medication groups, PTSD symptom scores changed minimally from the pre- to postindex year, decreasing by approximately 1 point, which was well below our MCID. This may reflect the challenges in treating this population that already likely had an inadequate response to first-line SRIs, or it may represent an inherent lack of efficacy of these augmenting medications for PTSD symptoms. Changes were similar for each augmenting medication group and did not differ after adjustment for the factors shown in Table 2a or by using a propensity-weighted approach (Table 3b). The 95% CIs exclude any MCID in change in PCL score between the augmenting medication groups, meaning no group had a differential change in PCL score that exceeded our MCID. In addition, analyses of all possible pairwise comparisons did not find any differences in PCL change by medication type that met our significance threshold (Appendix I). We found similar results when using the PC-PTSD to assess PTSD symptoms (Tables 3c and 3d).

Table 3a

Propensity Score–Weighted Changes in PCL Score by Augmenting Medication Group (n = 44 729 Pre-index Windows, n = 44 295 Postindex Windows).

Table 3b

Unadjusted and Adjusted Models for Change in PCL Score by Augmenting Medication Group.

Table 3c

Propensity Score–Weighted Changes in PC-PTSD Score by Augmenting Medication Group (n = 81 106 Pre-index Windows, n = 41 566 Postindex Windows).

Table 3d

Unadjusted and Adjusted Models for Change in PC-PTSD Score by Augmenting Medication Group.

Given the very small change in PCL score, we decided to explore the association using all available time points of PCL over the 2-year period rather than comparing the average scores in the pre-index and postindex years. Figure 3 displays smoothed lines representing the population mean PCL score in each of the 4 groups over the 2-year study period. The addition of the augmenting medication (the index date) is marked with a vertical line. For each medication group, PTSD symptoms increase before the addition of the augmenting medication and then decrease to roughly the baseline level after 3 to 4 months. Beyond this point, symptoms do not continue to improve but seem to remain close to the baseline level. This suggests that providers may be adding medications in response to increases in symptoms rather than simply to patients who did not respond to SRIs. It is not clear whether the return to baseline symptom level is a result of the augmenting medication or of regression to a baseline mean over time.

Figure 3

PCL Score Over the Pre-index and Postindex Year by Augmenting Medication Group.

Mental Health ED Visits and Hospitalizations

We found that mental health ED visits were highest among those taking antipsychotics and lowest among those augmented with tricyclics (Table 4a). All groups had declines in rates of mental health ED visits (relative decreases of 16%-20%) in the year following the addition of augmenting medication. After accounting for covariates with traditional and propensity-weighted approaches, we found no significant differences by medication class in the decrease in ED visits (Table 4b).

Table 4a

Propensity Score–Weighted Changes in Mental Health ED Visits (Per 100 Person-Years) by Augmenting Medication Group (n = 247 825 Pre- and Postindex Windows).

Table 4b

Unadjusted and Adjusted Models for Change in Mental Health ED Visits (Per 100 Person-Years) by Augmenting Medication Group.

Mental health hospitalizations followed a similar pattern, with the rates dropping in all groups in the year following the addition of an augmenting medication (Tables 4c and 4d). These more dramatic reductions in mental health ED visits/hospitalizations contrast with the minimal change in PTSD symptom score. As with symptoms, we conducted additional analyses to more closely examine the temporal pattern of ED visits and hospitalizations. Again, we found that the medications seemed to be started after a spike in use in all groups except tricyclics (Figure 4). After augmentation, the rate returned to baseline levels. It may be that mental health ED and hospitalization use reflects response only in the subset of patients with more severe PTSD, whereas PTSD symptoms reflect change in the full population. Tricyclics also seemed to be used in a more stable population, which may reflect the fact that they are being used for comorbid conditions, such as chronic pain, rather than being targeted and titrated to effect for PTSD symptoms.

Table 4c

Propensity Score–Weighted Changes in Mental Health Hospitalizations (Per 100 Person-Years) by Augmenting Medication Group (n = 247 825 Pre- and Postindex Windows).

Table 4d

Unadjusted and Adjusted Models for Change in Mental Health Hospitalizations (Per 100 Person-Years) by Augmenting Medication Group.

Figure 4

Mental Health ED Visits Over the Pre- and Postindex Year by Augmenting Medication Group.

Suicidal Thoughts and Plan

At baseline, nearly 1 in 5 patients who were screened endorsed having suicidal thoughts. In each group, significantly fewer patients endorsed suicidality in the year following the prescription of an augmenting medication. In fully adjusted and propensity-weighted models (Table 5a), those prescribed tricyclics had less reduction in suicidality than did those prescribed antipsychotics but the reduction did not meet our <.001 cutoff (Table 5b). There were no significant changes from pre-index year to postindex year in the proportion of patients endorsing having a suicidal plan (Tables 5c and 5d). As with PCL score and mental health ED visits and hospitalizations, all groups except for those taking tricyclics had augmenting medications started after an increase in suicidal thoughts (Figure 5). The proportion endorsing suicidal thoughts then decreased to slightly below the baseline level for all groups.

Table 5a

Propensity Score–Weighted Changes in Suicidal Thoughts (% Endorsing) by Augmenting Medication Group (n = 70 454 Pre-index Windows, n = 56 940 Postindex Windows).

Table 5b

Unadjusted and Adjusted Models for Change in Suicidal Thoughts by Augmenting Medication Group.

Table 5c

Propensity Score–Weighted Changes in Suicidal Plan (% Endorsing) by Augmenting Medication Group (n = 22 635 Pre-index Windows, n = 17 668 Postindex Windows).

Table 5d

Unadjusted and Adjusted Models for Change in Suicidal Plan by Augmenting Medication Group.

Figure 5

Proportion Endorsing Suicidal Thoughts Over the Pre-index Year and Postindex Year by Augmenting Medication Group.

Aim 2 Results

Given the large number of tables, the complete results from all aim 2 primary and sensitivity analyses are included in Appendix E (primary analyses and sensitivity analyses described in the Analytical and Statistical Approaches section) and Appendix F (sensitivity analyses using the restricted sample taking medications from only 1 augmenting class during the study period as described in the Changes to the Original Study Protocol section). In the tables below, we present the pre- and postindex year changes in all outcomes. As the unweighted and propensity score–weighted results were highly similar, we present only the propensity-weighted results below (unweighted and weighted results are shown side by side in the “Tables - Change in Outcomes” section of Appendices E and F). For each outcome, we then show the results from models comparing the change from pre- to postindex year in each of the augmenting medication groups with the prazosin group. Prazosin was selected as the reference for aim 2, as we hypothesized it would have the smallest effect on metabolic and cardiovascular outcomes. We show the unadjusted models, models adjusted for all covariates in Table 2a, and propensity-weighted models. As noted in the Changes to the Original Study Protocol section, given the number of hypotheses tested, we considered results of P < .001 to be statistically significant.

Weight

We found that patients who were prescribed antipsychotics or mirtazapine had increases in weight >2.2 lb (1 kg) (our MCID) from the pre- to postindex year, with weight gain being greatest in the group augmented with mirtazapine, followed by those augmented by antipsychotics (Tables 6a and 6b). Of note, these 2 groups had the lowest pre-index weights, which may reflect providers avoiding prescribing these medications in overweight or obese patients. In fully adjusted and propensity score–weighted analyses, all groups had significantly greater weight gain than with prazosin. In models with all pairwise comparisons, weight gain with mirtazapine was significantly greater than in all other groups (Appendix I). Examining weight change over time (Figure 6), the most dramatic weight gain in all groups occurred within the first 4 months.

Table 6a

Propensity Score–Weighted Changes in Weight (lb) by Augmenting Medication Group (n = 236 641 Pre-index Windows, n = 233 605 Postindex Windows).

Table 6b

Unadjusted and Adjusted Models for Change in Weight by Augmenting Medication Group.

Figure 6

Change in Weight Over the Pre- and Postindex Year by Augmenting Medication Group.

Lipids

Triglycerides increased in all groups except prazosin, indicating increased cardiometabolic risk (Table 7a), but the change in triglyceride levels exceeded our MCID of 10 mg/dL only for mirtazapine. In adjusted and propensity score–weighted models, all other groups had significantly greater increases in triglycerides than did those taking prazosin (Table 7b). Protective HDL cholesterol also significantly decreased in those prescribed antipsychotics and mirtazapine (Tables 7c and 7d), but this did not exceed our MCID of 5 mg/dL. Despite the adverse effects seen on these lipids and the increases in body weight, total and LDL cholesterol decreased by a small but statistically significant amount (less than our MCID of 10 mg/dL) in all augmenting medication groups, which would indicate less metabolic dysfunction (Tables 7e-7h). Total cholesterol also decreased in all groups except mirtazapine (Tables 7g and 7h), but again, this was less than our MCID of 10 mg/dL. Evaluating these results together, we hypothesized that patients were being started on statins or other cholesterol-lowering medications, which tend to impact total and LDL cholesterol more than HDL cholesterol and triglycerides. We present results evaluating the proportion of patients who were newly started on these medications or had an existing regimen intensified (described in the Use of Medications to Treat Metabolic Risk Factors section below). Similar to our analyses of mental health outcomes for aim 1, we also examined these changes over time. We have presented a representative figure for triglycerides, given that these had the largest changes (see Figure 7). The majority of increases occurred within the first 4 to 6 months after starting an augmenting medication.

Table 7a

Propensity Score–Weighted Changes in Triglycerides by Augmenting Medication Group (n = 198 246 Pre-index Windows, n = 192 683 Postindex Windows).

Table 7b

Unadjusted and Adjusted Models for Change in Triglycerides by Augmenting Medication Group.

Table 7c

Propensity Score–Weighted Changes in HDL Cholesterol by Augmenting Medication Group (n = 196 817 Pre-index Windows, n = 191 699 Postindex Windows).

Table 7d

Unadjusted and Adjusted Models for Change in HDL Cholesterol by Augmenting Medication Group.

Table 7e

Propensity Score–Weighted Changes in LDL Cholesterol by Augmenting Medication Group (n = 195 657 Pre-index Windows, n = 190 728 Postindex Windows).

Table 7f

Unadjusted and Adjusted Models for Change in LDL Cholesterol by Augmenting Medication Group.

Table 7g

Propensity Score–Weighted Changes in Total Cholesterol by Augmenting Medication Group (n = 194 734 Pre-index Windows, n = 189 760 Postindex Windows).

Table 7h

Unadjusted and Adjusted Models for Change in Total Cholesterol by Augmenting Medication Group.

Figure 7

Change in Triglycerides Over the Pre-index Year and Postindex Year by Augmenting Medication Group.

Glucose and HbA_1c

Glucose and HbA_1c increased on average from the pre- to postaugmentation year in all medication groups, but these did not exceed our MCIDs of 0.5% for HbA_1c and 5 mg/dL for glucose (Tables 8a and 8c). These findings were maintained when we used fully adjusted/propensity score–weighted models, and parallel what we observed with the changes we saw in weight (Tables 8b and 8d).

Table 8a

Propensity Score–Weighted Changes in HbA_1c by Augmenting Medication Group (n = 128 453 Pre-index Windows, n = 135 092 Postindex Windows).

Table 8b

Unadjusted and Adjusted Models for Change in HbA_1c by Augmenting Medication Group.

Table 8c

Propensity Score–Weighted Changes in Glucose by Augmenting Medication Group (n = 214 104 Pre-index Windows, n = 209 158 Postindex Windows).

Table 8d

Unadjusted and Adjusted Models for Change in Glucose by Augmenting Medication Group.

Blood Pressure

Given the increase observed in weight, we would have expected a small increase in blood pressure. We did see a significant increase in the group that was prescribed mirtazapine, but it was well under our MCID. However, we observed small but significant declines in systolic blood pressure in those prescribed antipsychotics and prazosin (Tables 9a and 9b), with prazosin having significantly greater declines in adjusted and propensity-weighted models. Diastolic blood pressure declined slightly in those prescribed prazosin and antipsychotics and increased slightly in those prescribed mirtazapine or tricyclics (Tables 9c and 9d). Similar to our findings for lipids, we believed these results needed to be examined in conjunction with changes in antihypertensive medication, which could have masked the adverse impact of augmenting medications.

Table 9a

Propensity Score–Weighted Changes in Systolic Blood Pressure by Augmenting Medication Group (n = 240 298 Pre-index Windows, n = 237 794 Postindex Windows).

Table 9b

Unadjusted and Adjusted Models for Change in Systolic Blood Pressure by Augmenting Medication Group.

Table 9c

Propensity Score–Weighted Changes in Diastolic Blood Pressure by Augmenting Medication Group (n = 240 298 Pre-index Windows, n = 237 794 Postindex Windows).

Table 9d

Unadjusted and Adjusted Models for Change in Diastolic Blood Pressure by Augmenting Medication Group.

Incident Diagnoses of CVD Risk Factors

We evaluated incident diagnoses for cardiometabolic risk factors in the year following augmentation (Tables 10a-10e). As we hypothesized, prazosin had the lowest rate of incident cardiometabolic risk factor diagnoses. The exception was obesity, where mirtazapine had the lowest rates, though in adjusted models, these were not significantly different from prazosin. Given that weight gain is a known adverse effect of mirtazapine, providers are likely avoiding prescribing this medication in patients who are already overweight.

Table 10a

Propensity Score–Weighted New Diagnoses of Cardiovascular Risk Factors (Per 100 Person-Years) by Augmenting Medication Group (n =247 825 Pre- and Postindex Windows).

Table 10b

Difference in Incident Diagnoses of Diabetes Per 100 Person-Years.

Table 10c

Difference in Incident Diagnoses of Dyslipidemia Per 100 Person-Years.

Table 10d

Difference in Incident Diagnoses of Hypertension Per 100 Person-Years.

Table 10e

Difference in Incident Diagnoses of Obesity Per 100 Person-Years.

Use of Medications to Treat Metabolic Risk Factors

New Medications/Dose Increases to Treat Diabetes

Only tricyclics were associated with a significantly higher rate of medication intensification for diabetes compared with prazosin (Tables 11a and 11b). A substantially higher proportion of patients prescribed tricyclics had diabetes at baseline, and differences were reduced though still significant in propensity-weighted models that accounted for this (Table 11b).

Table 11a

New Prescription or Dose Increase of a Diabetes Medication by Augmenting Medication Group (n = 247 825 Postindex Windows).

Table 11b

Comparison of New Prescription or Dose Increase of a Diabetes Medication.

New Medications/Dose Increases to Treat Dyslipidemia

We evaluated the number of patients in each augmenting medication class that were prescribed new medications to treat cholesterol in the year after starting the augmenting medication. We also examined the number that increased the dose of an existing medication or switched to a more potent medication (eg, changed from simvastatin 10 mg to rosuvastatin 10 mg). We found high rates of new/increased use of medications to treat dyslipidemia among all groups (Table 11c). In fully adjusted, propensity-weighted models (Table 11d) and when restricting cholesterol medications to statins (Tables 11e and 11f), those prescribed antipsychotic or tricyclic medications had significantly higher rates of new dyslipidemia medication than did those prescribed prazosin.

Table 11c

New Prescription or Dose Increase of a Cholesterol Medication by Augmenting Medication Group (n = 247 825 Pre- and Postindex Windows).

Table 11d

Comparison of New Prescription or Dose Increase of a Cholesterol Medication.

Table 11e

New Prescription or Dose/Potency Increase of a Statin by Augmenting Medication Group (n = 247 825 Pre- and Postindex Windows).

Table 11f

Comparison of New Prescription or Dose/Potency Increase of a Statin.

New Medications/Dose Increases to Treat Hypertension

Approximately 1 in 3 patients had a new antihypertensive medication started or an existing regimen intensified in the year following the addition of an augmenting medication (Table 11g). Rates were highest for those prescribed tricyclics, for which increases in blood pressure are a known adverse effect. Antipsychotics and mirtazapine had similar rates of hypertension treatment intensification. As expected, those taking prazosin, which may have a small lowering effect on blood pressure, had the lowest rates of hypertension treatment intensification. The higher rates of hypertension treatment intensification in comparison with prazosin were confirmed in adjusted and propensity score–weighted models (Table 11h).

Table 11g

New Prescription or Dose Increase of an Antihypertensive Medication by Augmenting Medication Group (n = 247 825 Pre- and Postindex Windows).

Table 11h

Comparison of New Prescription or Dose Increase of an Antihypertensive Medication.

Incident CVD Events

In adjusted and propensity-weighted models, rates of CVD events (based on diagnostic codes) were significantly higher in those prescribed mirtazapine, antipsychotics, or tricyclics than in those prescribed prazosin (Tables 12a-12g). This fits with our findings mentioned previously that demonstrated that mirtazapine and antipsychotics had the greatest impact on risk factors, such as weight, blood pressure, lipids, and blood sugar, while patients augmented with prazosin had the smallest changes.

Table 12a

Myocardial Infarction Hazard Ratio (n = 247 825 Postindex Windows).

Table 12b

CABG Hazard Ratio (n = 247 825 Postindex Windows).

Table 12c

PCI Hazard Ratio (n = 247 825 Postindex Windows).

Table 12d

Heart Failure Hazard Ratio (n = 247 825 Postindex Windows).

Table 12e

Unstable Angina Hazard Ratio (n = 247 825 Postindex Windows).

Table 12f

Stroke Hazard Ratio (n = 247 825 Postindex Windows).

Table 12g

Any CVD or Cerebrovascular Disease Event Hazard Ratio (n = 247 825 Postindex Windows).

All-Cause Mortality

In unadjusted and fully adjusted/weighted analyses, we found that all-cause mortality was significantly higher in those taking antipsychotics, mirtazapine, and tricyclics than in the reference group of prazosin (Table 13). Higher rates of all-cause mortality have been observed with antipsychotics for several years. An updated meta-analysis in 2018 found that antipsychotic medication use was associated with higher rates of all-cause mortality in patients with and without dementia. Atypical antipsychotics are also associated with QT prolongation, ventricular arrhythmias, and sudden cardiac death.^58-62

Table 13

All-Cause Mortality Hazard Ratio (n = 247 825 Postindex Windows).

Aim 3 Results

The third aim of our project was to examine subgroup differences, including sex, age, and era of military service, in mental health and metabolic/cardiovascular outcomes associated with various augmenting medications. This involved repeating all of our aim 1 and 2 analyses for each of the 3 subgroup categories. For each outcome tested, we examined 2-way interactions (if the outcome differed by subgroup; eg, In patients taking an antipsychotic, does weight change from the pre- to postindex year differ by sex?). We also examined 3-way interactions, which would tell us if the subgroup differences also varied by the augmenting medication group (eg, Does sex difference in weight gain differ in patients taking mirtazapine vs those taking an antipsychotic?). Given the extensive number of analyses, which generated several hundred results tables, we here highlight clinically and statistically significant subgroup findings. Full analyses are presented in Appendix G with a separate section for each mental health and metabolic/cardiovascular outcome.

Mental Health Outcomes

We did not find subgroup differences in our primary mental health outcome of PCL score (see Appendix G). For our secondary outcome of PC-PTSD score, we observed small differences by sex and age, but these did not meet our criteria for statistical or clinical significance. Looking at mental health hospitalizations as another indication of response to PTSD treatment augmentation, we did not find results that differed by sex. Veterans aged <65 years and those who served in Iraq/Afghanistan tended to have greater absolute decreases in mental health hospitalizations. However, because of their higher baseline rates of hospitalizations, this represented a lower percentage decline than with the comparison groups (see Tables 14a and 14b and 15a and 15b). Typically, these differences did not meet our clinical and statistical significance thresholds. The findings were in a similar direction but also not significant for rates of mental health ED visits (see Appendix G).

Table 14a

Unweighted Change in Mental Health Hospitalization (Incidence Per 100 Person-Years) in Those Aged ≥65 Years by Augmenting Medication Group.

Table 14b

Unweighted Change in Mental Health Hospitalization (Incidence Per 100 Person-Years) in Those Aged <65 Years by Augmenting Medication Group.

Table 15a

Unweighted Change in Mental Health Hospitalization (Incidence Per 100 Person-Years) In Non-Iraq/Afghanistan Veterans by Augmenting Medication Group.

Table 15b

Unweighted Change in Mental Health Hospitalization (Incidence Per 100 Person-Years) in Iraq/Afghanistan Veterans by Augmenting Medication Group.

Metabolic/Cardiovascular Outcomes

We also examined subgroup differences in the metabolic and cardiovascular adverse effects we analyzed in aim 2. Again, given that these represent numerous models and comparisons, we will highlight the relevant findings here. Full analyses are presented in Appendix G. The greatest subgroup differences were observed in our primary outcome of weight. Women had significantly greater weight gain than did men in all 4 augmenting medication classes (see Tables 16a and 16b), but the 3-way interactions did not meet our clinical and statistical significance thresholds, indicating that the sex discrepancy in weight gain was similar across medication classes. We also observed differences in weight gain by age and military service era. Younger veterans and Iraq/Afghanistan veterans had greater weight gain than did those aged ≥65 years or from other service eras (P < .0001 for all 2-way interactions; see Tables 17a and 17b and 18a and 18b). The 3-way interactions for age met our significance threshold and were driven by greater weight gain in those aged <65 years taking antipsychotics or mirtazapine.

Table 16a

Unweighted Change in Weight (lb) in Men by Augmenting Medication Group.

Table 16b

Unweighted Change in Weight (lb) in Women by Augmenting Medication Group.

Table 17a

Unweighted Change in Weight (lb) in Those Aged ≥65 Years by Augmenting Medication Group.

Table 17b

Unweighted Change in Weight (lb) in Those Aged <65 Years by Augmenting Medication Group.

Table 18a

Unweighted Change in Weight (lb) in Non-Iraq/Afghanistan Veterans by Augmenting Medication Group.

Table 18b

Unweighted Change in Weight (lb) in Iraq/Afghanistan Veterans by Augmenting Medication Group.

Subpopulation Considerations

We found that mental health outcomes were largely similar across men and women, though we did observe a smaller effect on the PC-PTSD symptom score in women. We also found that adverse cardiovascular effects were more dramatic in female, younger, and Iraq/Afghanistan veterans. These results add important information given that existing trials have tended to enroll homogeneous populations, such as all male veterans. Our findings that younger and Iraq/Afghanistan veterans had less-robust responses need to be confirmed in other settings but may still be important. It is plausible that this is a group with more proximal trauma being seen earlier in their course of PTSD. This may make their symptoms more volatile and perhaps more difficult to treat. It is also possible that the other stressors experienced by younger, returning veterans, such as readjustment to family and work, may complicate treatment.

In terms of metabolic heterogeneity, although the subgroup effects of these medications have not been examined in PTSD, previous evidence does exist for the use of antipsychotics in schizophrenia. Consistent with our findings, female and younger patients have greater weight gain and metabolic risk with antipsychotics.⁶⁴ It is possible that this could lead to reduced effectiveness in these subgroups.⁶⁵ Metabolic consequences, such as the mean 5 pounds of weight gain experienced by women on mirtazapine, may lead patients to discontinue medications early or prevent them from titrating to an adequate dose.

Future Directions

Our observational findings in a VA population would of course benefit from additional study in other samples and in clinical trial settings. In addition, they suggest several other avenues of work that could improve clinical care for patients with PTSD. First, though we present mean responses, within each group, there will be patients who experience more benefits and fewer adverse reactions. An important next step of this work could be the development of models to predict which patients will have better responses to specific medications. This could allow providers to more efficiently select medications and avoid adverse effects or lack of improvement in PTSD symptoms that can frustrate patients and prevent their further engagement with PTSD treatment. Large data sets, such as the VA health records, allow for advanced modeling techniques to conduct such studies.

Second, our project highlights several opportunities for improvements to clinical infrastructure. The VA has instituted guidelines and clinical reminders to improve rates of monitoring for metabolic consequences in antipsychotics. Our findings of high rates of metabolic medication use suggest that providers are noticing and responding to these adverse outcomes. However, less attention has been paid to monitoring the primary PTSD symptom response. Our results indicate that some patients will have minimal benefit with these medications and that the majority of patients do not have repeated assessments with validated instruments after initiation of medication. Therefore, it would be helpful to have systems in place for regular reassessments of PTSD symptom scores and check-ins with patients on their perception of improvements after augmentation. If there is no substantial improvement after a reasonable time frame, monitored titration off the medication could be considered to prevent potential metabolic harms. We are fortunate to have a multidisciplinary team that includes veteran patient partners and stakeholder partners involved in national VA drug safety monitoring and educational initiatives. As we disseminate our findings, we will work to inform additional stakeholders of these recommendations for future work.

Limitations

Our study has several limitations; for one, we used observational data, and although we used traditional methods and propensity score weighting to help reduce confounding, patients were not randomly assigned to the augmenting medications, and we did not have a control group continuing their SRI without the addition of an augmenting medication. This limits our ability to draw conclusions about the causality of the associations we observed, as they could be due to temporal changes or factors other than the augmenting medication or could primarily represent regression to the mean. This cohort includes veterans who were mostly male, and the findings therefore might not be generalizable to other populations. Although ascertainment of medication use was thorough, we relied on prescription fill data, which may not represent actual use, and we did not have access to non-VA prescription data. The indications for the prescriptions are not included, so it is possible that some of these medications were prescribed to target conditions other than PTSD. We also required that patients take the augmenting medication for at least 60 days, as our team felt this would allow adequate time for impact on our mental health and metabolic outcomes. However, this left us unable to examine the frequency and impact of shorter trials and particularly discontinuations that may have been due to adverse effects of medications. Finally, while data were complete for covariates and many of our outcomes (see tables in Appendices C-H that show the number of outcomes for each variable and augmenting medication group), a substantial number of assessment windows (203 096 [82%]) did not have PCL scores in the required time frame before and following augmenting medication initiation. Findings from a sensitivity analysis using additional scores obtained by NLP were similar to our initial analyses, with small changes in PTSD symptom score. In addition, the patterns of PTSD symptoms pre- and postinitiation of augmentation were similar to those for our other mental health outcomes, and we retained substantial power to detect changes well below our MCID threshold. Still, this remains a limitation and highlights the need for improved monitoring of PTSD symptoms. Many of our laboratory variables, such as cholesterol, are routinely monitored and available for a majority of the sample. However, others, such as HbA_1c, though useful for metabolic monitoring with antipsychotics, may be collected less often or only in patients with existing conditions, such as diabetes. For example, in our study, HbA_1c was present for 50% to 55% of the windows across drug classes. Though missing data reduce our ability to fully evaluate the impact of these medications, we do not believe they will bias our comparisons of the effects of these medications, as the rates of missing data were similar in the 4 augmenting medication groups. The use of ICD codes in administrative data may have also led to misclassification or underascertainment of outcomes. However, it is unlikely that the outcomes were classified differentially based on augmenting medication group.

Appendix A.

Study Protocol (PDF, 212K)

Appendix B.

Coding Algorithms (PDF, 136K)

Appendix C.

Aim 1 Full Sample Analyses (PDF, 780K)

Appendix D.

Aim 1 Single Augmenting Class Analyses (PDF, 778K)

Appendix E.

Aim 2 Full Sample Analyses (PDF, 990K)

Appendix F.

Aim 2 Single Augmenting Class Analyses (PDF, 864K)

Appendix G.

Aim 3 Full Sample Analyses (PDF, 1.8M)

Appendix H.

Aim 3 Single Augmenting Class Analyses (PDF, 1.3M)

Appendix I.

Pairwise Comparisons (PDF, 766K)

Appendix J.

Propensity Score Distributions (PDF, 696K)