Sample and data sources

We conducted a retrospective cohort study that included all VA patients with a clinical diagnosis of PTSD (ICD-9: 309.81; ICD-10: F43.1x) plus at least one PCL score between the start of fiscal year 2000 (1 October 1999) and the end of calendar year 2018 (31 December 2018) in the VA Corporate Data Warehouse (CDW) (n = 754 197). Patients entered the cohort in the year of their first VA use and remained in the cohort until death or 31 December 2018, whichever came first (minimum of 1 year and maximum of 20 years). We assessed suicide mortality using the VA/Department of Defense Mortality Data Repository.¹⁹ We assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All procedures involving human participants/patients were approved by the Veteran's Institutional Review Board of Northern New England (USA). A waiver of consent and authorisation was granted for the study.

PTSD symptom data

We integrated two different versions of the PCL, captured from up to two data sources within the CDW. The two data sources included scores obtained from structured data produced by psychometric assessment software in the VA medical record and scores documented by clinicians in 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.^{Reference Holder, Shiner and Li20} Scores abstracted from structured data and from NLP of clinical notes were integrated into a single data-set, which has been described in detail elsewhere.^{Reference Shiner, Levis, Dufort, Patterson, Watts and DuVall21} The two versions of the PCL were aligned to the DSM-IV and DSM-5, and we will call them the PCL-IV and the PCL-5.^{Reference Bovin, Marx, Weathers, Gallagher, Rodriguez and Schnurr17,Reference Weathers, Litz, Herman, Huska and Keane22} Validation work shows a correlation of 0.87 between PCL versions in a large sample of veterans.^{Reference Bovin, Marx, Weathers, Gallagher, Rodriguez and Schnurr17} We used a validated crosswalk (ICC = 0.96) to convert all values to PCL-5 scoring.^{Reference Moshier, Lee and Bovin23}

The PCL-5 has a range of 0–80 and scores of ≥31–33 are considered diagnostic for PTSD.^{Reference Bovin, Marx, Weathers, Gallagher, Rodriguez and Schnurr17} Although a threshold for remission has not been established, the largest prospective treatment study using this version of the PCL allowed early termination owing to symptomatic remission for scores of 18 or lower.^{Reference Schnurr, Chard, Ruzek, Chow, Shih and Resick24} The mean baseline score for VA patients starting PTSD treatment in this PCL data-set is approximately 50.^{Reference Shiner, Gui, Rozema, Cornelius, Dufort and Schnurr25} Based on these thresholds, we created exposure groups using four PCL score ranges: minimal (0–18), low (19–30), moderate (31–49) and high (≥50).

A clinically meaningful change in PCL score is approximately 15 points.^{Reference Marx, Lee, Norman, Bovin, Sloan and Weathers26} Because this change criterion was calculated using the Jacobson and Truax Reliable Change Index (1.96 times the standard error of the difference in change, which is the ‘distribution of change scores that would be expected if no actual change had occurred’^{Reference Jacobson and Truax27}), we used a corollary that changes of 7 points or less could be due to measurement error. Based on these thresholds, we created exposure groups using three PCL score ranges, each delineating a change in PTSD symptom severity: worse (increase of 15 points or more), no change (change within 7 points) and improved (decrease of 15 points or more). For patients with more than two PCL scores, we defined the change category based on the last two consecutive PCL scores that were documented between 8 weeks and 1 year apart, prioritising the shorter period. Additionally, we performed subgroup analysis based on whether patients who improved achieved symptomatic remission, which we defined as a follow-up PCL score ≤18.

Covariates

We measured patient characteristics at the time of their last PCL score, including age, gender, ethnicity, marital status, service-connected disability, and burden of mental and physical illness. In addition to PTSD diagnosis, we summarised physical and mental diagnoses in the preceding two calendar years (supplementary Appendix available at https://doi.org/10.1192/bjp.2022.110) using counts of diagnostic categories developed in prior research using VA electronic medical record (EMR) data.^{Reference Shiner, Peltzman, Cornelius, Gui, Forehand and Watts28} Specifically, we used ICD codes in patients’ medical records to create a modified Elixhauser Comorbidity Index for physical health,^{Reference Quan, Sundararajan, Halfon, Fong, Burnand and Luthi29} which was orthogonal to a mental health index based on the DSM-5 diagnostic groups.¹ Patients were categorised as having a low, medium or high burden of physical or mental illness if they had diagnoses from 0, 1–2 or 3+ relevant groups. We also specifically assessed whether patients had comorbid depressive or substance use disorder diagnoses.

Study outcomes

We measured suicide mortality from 1 day after the last PCL administration date through to 31 December 2018 using ICD-10 codes U03, X60–X84 and Y87.0.

Statistical analysis

For descriptive analyses, we calculated frequencies, both overall and by suicide mortality status, for all variables among patients in the cohort with at least one PCL score. We calculated suicide rates per 100 000 person-years at the patient level by multiplying the number of observed suicides by 100 000, then dividing that value by the number of person-years at risk from the day after each patient's last PCL until death or 31 December 2018, whichever occurred first.

For proportional hazards regression analyses examining the effect of PTSD symptom category or change in PTSD symptom category on suicide mortality rate, we used each patient's last PCL score or change in PCL score as the unit of analysis. The patient's time at risk began the day after the last PCL was administered and ended at death or on 31 December 2018, whichever occurred first. We used partially conditional proportional hazards regression models; by treating each PCL as the unit of analysis, the models were conditioned on the baseline covariates (age, gender, ethnicity, marital status, service-connected disability, and mental and physical health diagnoses) but not on the time-varying covariates (PCL responses). We conducted a set of stratified analyses for patients with and without comorbid depressive or substance use disorders for the PTSD symptom severity analysis but not the change in severity analyses, because of limitations in sample size.

Covariates were included in the models to adjust for potential confounding variables in the association between PCL responses and suicide mortality. We adjusted for covariates using a stepped process to better understand the associations between the covariates, the exposure groups and the outcome. Data collection and analyses were conducted from 1 May 2021 to 9 September 2021. Data analyses were performed using SAS software (Windows), version 9.4 and the SAS Enterprise Guide, version 7.1 (SAS Institute Inc.).