The research, conducted by Nicholas Goldrosen and colleagues at the Pennsylvania Commission on Sentencing and Pennsylvania State University’s Criminal Justice Research Center, tackled the problem with a four-step data linkage procedure. First, the team assembled every sentence reported to the Pennsylvania Commission on Sentencing between 2016 and 2021—418,446 sentences handed down by the state’s county Courts of Common Pleas. Next, staff obtained full criminal history records for each defendant from the Pennsylvania State Police, records that contain each individual’s Social Security Number along with biographic details. Those identifiers were then submitted to the Veterans Re-entry Search Service, or VRSS, a web-based application operated by the VA that allows criminal justice agencies to verify whether an individual ever served in the U.S. military. Finally, the veteran status flag generated for each person was attached to every case in which that person appeared as a defendant.

The technical details of this pipeline matter, because they define who counts as a veteran. VRSS identifies anyone with any amount of service in any branch of the military, including reserve components and the National Guard—a deliberately broad definition that contrasts sharply with self-identification, which research shows produces substantial undercounts. Some veterans with brief service, with other-than-honorable discharges, or without eligibility for benefits do not consider themselves veterans at all. Others, charged or convicted of a crime, may feel their behavior reflects poorly on their veteran identity and decline to disclose their service. Prison-based sampling introduces a different distortion: in Pennsylvania in 2023, only about 12 percent of criminal sentences resulted in state imprisonment, meaning any prison-only study excludes the overwhelming majority of convictions, most of them for lower-level offenses resolved in the community.

The linked dataset revealed that veterans made up 6.0 percent of sentenced defendants, almost exactly matching their roughly 6.4 percent share of the general adult population—a finding consistent with prior research showing that veterans and non-veterans are arrested at roughly equal rates. But beneath that headline equivalence, the demographic and offense profiles of the two groups diverged in ways that carry real consequences for policy. Veteran defendants were older, more likely to be male, more likely to be white, and more likely to be convicted in rural counties. Their criminal histories were markedly lighter than those of non-veterans: lower average prior record scores, fewer prior felony convictions, and—most strikingly—nearly six times less likely to include a juvenile adjudication, a pattern the authors attribute to the fact that a significant juvenile record is typically a bar to military enlistment in the first place.

Perhaps the most scientifically intriguing finding concerns age. Among non-veterans, age at sentencing follows the classic unimodal, right-skewed age-crime curve familiar from a century of criminology, peaking around age 26. Veterans, by contrast, show a distinctly bimodal distribution, with peaks at approximately 33 and 53 years of age, the younger peak slightly larger. The gap of roughly seven years between the non-veteran peak and the younger veteran peak is suggestive: seven years is also the average length of enlistment in the Iraq and Afghanistan era, raising the possibility that military service mechanically delays entry into offending simply by removing individuals from the civilian risk set during their early twenties. The second peak, however, resists easy explanation. Offending that first appears around age 53 does not fit established life-course taxonomies, such as Moffitt’s adolescence-limited and life-course-persistent offender types or the trajectory categories identified by Nagin and Land. The authors suggest that the differential composition of the veteran population, combined with effects of military service itself, may produce forms of later-life offending that existing criminological frameworks fail to capture.

Offense types told an equally distinctive story. Veterans were overrepresented among those convicted of driving under the influence, crimes against minors, and offenses classified as dangers to persons. They were underrepresented among those convicted of drug, firearms, and property offenses. Some of these patterns echo earlier prison-based research, particularly the overrepresentation in serious offenses against minors, but the DUI finding illustrates precisely what prison samples miss: DUI convictions frequently result in non-custodial sentences, so a study restricted to incarcerated populations would largely fail to detect this veteran-specific offense signature. Because the new dataset covers the full conviction pipeline—only about 10 percent of cases led to state prison and about 31 percent to county jail, with the remainder served entirely in the community—it captures offenses that older methods systematically overlooked.

For the growing national infrastructure of veterans’ treatment courts, which now operate in hundreds of jurisdictions, and for states like Kansas and Minnesota that have written veteran-specific sentencing alternatives into law, these findings carry practical weight. Problem-solving courts have traditionally been built around drug and property offenses, the offenses that dominate their dockets. If veteran defendants are disproportionately convicted of alcohol-involved offenses such as DUI, veterans’ treatment courts may need to recalibrate their treatment capacity accordingly, with greater emphasis on alcohol-related intervention. Similarly, the veteran-specific housing units that Pennsylvania and several other states operate within their prisons must be prepared to deliver programming for a population that includes a larger share of individuals convicted of serious, often sexual, offenses than the generic treatment court model anticipates.

The study also quantifies a troubling gap in program reach. In Pennsylvania, only about 4.5 percent of veteran defendants participate in veterans’ treatment courts. Evaluating whether such courts outperform ordinary case processing requires knowing who the veterans are across the entire system, not just at the point where someone voluntarily identifies as one. Without systematic identification, the authors argue, policymakers cannot know whether these interventions work, for whom, or where in the process veterans are falling through the cracks. The 2022 initiative by the Council on Criminal Justice, which produced a model policy framework for veterans in the justice system, has emphasized exactly this kind of evidence gap.

Methodologically, the authors are candid about limitations. Because the data capture a single time window, they cannot disentangle cohort effects—for example, whether Vietnam-era veterans differ in age patterns from those who served more recently. The dataset also lacks information on the timing and length of individual military service, which limits how precisely the bimodal age distribution can be interpreted. And because Hispanic ethnicity is unreliably reported in sentencing data, the team used surname-based probabilistic classification with the “predictrace” package in R, drawing on Census and Social Security Administration data, an approach that introduces its own measurement uncertainty.

Even so, the study functions as a proof of concept with implications far beyond Pennsylvania. The authors call for criminal justice agencies to use VA administrative tools—VRSS and the newer SQUARES system—to query every person they contact, ideally from the point of arrest onward, using the broadest possible definition of veteran status, and to make resulting veteran status identifiers regularly available to researchers. The alternative, they note, is a continued reliance on data structures that misrepresent the very population policymakers are trying to serve. As interest in veteran-specific justice responses continues to grow, the Pennsylvania experiment suggests that the path to better policy runs through something unglamorous but powerful: connecting the administrative records that already exist, and letting the linked data reveal patterns that no single database could show on its own.