Stata statistical software version 12

We estimated the risk of premature death after having been diagnosed as having TBI with matched or sibling controls^{24 (link),35 (link)} using the clogit command in Stata statistical software, version 12 (StataCorp). The clogit command fits conditional (fixed-effects) logistic regression models to matched case-control groups. Traumatic brain injury patients and controls were followed up from the same time point (ie, from 6 months after TBI). We included 3 confounders (low income, single marital status, and immigrant status) on theoretical grounds based on related work on other neuropsychiatric disorders^{35 (link),36 (link)} and tested whether they were each associated (at P < .05) with caseness and outcome measures, respectively.^{37 (link)} Separately, we adjusted for emigration rates after TBI. In sensitivity analyses, we stratified by sex, psychiatric comorbidity (lifetime and separately for preexisting and new diagnoses), and co-occurring injuries at the time of TBI (Author Table 1). The STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines were followed (Author Table 2).

We conducted logistic regression models overall and by patient subgroups while controlling for age, exam year, race/ethnicity, timing of last screen and indicator variables for screening facility. For specificity, we also used generalized estimating equations to account for clustering of exams by patient. Model-based standardization (predictive margins) was then used to estimate the overall and stratum-specific performance characteristics [5 (link),6 (link)]. Ninety-five percent confidence intervals were estimated using the delta method as implemented in the margins command within Stata. All analyses were conducted using Stata statistical software, version 12 (Stata Corp, College Station, TX). All p-values are two-sided.

All analyses were conducted using STATA statistical software, version 12 (StataCorp) with its survey routines. To generate prevalence estimates and inferential statistics that reflect CCJTDC’s population, each participant was assigned a sampling weight augmented with a nonresponse adjustment to account for missing data. Because minorities are disproportionately incarcerated, weighted estimates for males and females overall are similar to those for African American males and females.
We present prevalence estimates for participants who were still living at follow-up. (Five and 12 years after detention, 50 and 97 participants had died, respectively.) Because incarceration prevents people from achieving many positive outcomes, we also present prevalence only for participants living in the community during the recall period (see eMethods). We used logistic regression to examine sex and racial/ethnic differences in outcomes, adjusting for age at detention and legal status. We used the Latent Class Analysis Stata plugin²⁴ to empirically identify “classes” of participants who exhibited similar patterns of positive outcomes 12 years after detention. Three participants who self-identified as “other” race/ethnicity were excluded from all analyses. We conducted separate analyses for males and females because combining them could obfuscate important differences.