4

Methods

The committee was tasked with conducting an independent scientific assessment of potential relationships between exposures experienced during military service and specific mental, behavioral, and neurologic health outcomes and chronic multisymptom illness (CMI). Congress directed the Department of Veterans Affairs (VA) to provide information from health care and other administrative records for use in original analyses to be conducted by the committee. Because the results of any single data analysis may be limited by data quality and availability, the committee supplemented its original data analyses with a structured literature review to facilitate a more holistic assessment of potential relationships of military-related exposures and selected health outcomes. While it was not tasked with conducting a comprehensive systematic literature review, its review was used to identify existing evidence as a source of external, supporting evidence to contextualize the statistical findings from its data analyses. This chapter first describes the methods used for the original epidemiologic analyses and structured literature review. It then presents the committee’s approach to assessing and interpreting both sources of evidence. The chapter closes with the committee’s approach for integrating multiple sources of evidence to develop its conclusions about possible relationships between the selected military exposures and specified health outcomes.

ORIGINAL DATA ANALYSES

This section first describes the sources of data and types of variables that the committee requested and received to perform its data analysis.

Based on its conceptual model and the available data, the design and details of the analysis are then described, including decisions for specific outcome, exposure, and covariate variables. Limitations of this original data analysis are grouped by exposures, outcomes, other covariates, and the data sources in general.

Data and Data Sources

The committee submitted its data request to the Veterans Health Administration (VHA) in June 2024; after required processes for institutional review board and access approvals, Westat, the subcontractor, received access through the VA Informatics and Computing Infrastructure (VINCI) system in November 2024.¹ The dataset comprised veterans who had at least one inpatient or one outpatient encounter at any VA health care facility between January 1, 2016, and November 15, 2024. This date range was selected to limit miscoding errors in the 15 months following the transition of International Classification of Diseases, Ninth Revision (ICD-9) to Tenth Revision (ICD-10) diagnosis codes that occurred in October 2015. Data from 2016 and 2024 were requested for concordance purposes only. For example, the additional years before and after the study period (2017–2023) were used to increase linkages between the exposure and outcomes data, exclude cases with prior diagnoses, and examine health care engagement. Although some studies have used two outpatient visits with the same ICD-10 code diagnoses to identify cases of a specific outcome, the committee chose to use one outpatient encounter to increase sensitivity of identifying cases. Specifically, because the committee’s charge was to broadly assess possible relationships, the study period is short and recent, and some of the outcomes of interest are rare, this approach errs on being more inclusive to capture all possible cases. Moreover, since many of the rare outcomes require relatively severe and persistent presentation to receive a diagnosis, these diagnoses were more likely to be retained in a patient’s record after just one diagnosis. Exposure data from the Individual Longitudinal Exposure Record (ILER), a joint VA and Department of Defense (DoD) web-based application that compiles data from multiple sources to provide an exposure history for each individual, were obtained through a request to DoD’s Defense Health Agency (DHA) in June 2024. ILER data were approved for use in October 2024 and received in February

___________________

¹ VINCI contains electronic health records (EHRs) and other VA datasets that are cleaned and linked for research purposes. Westat received approval from VA to use a virtual private network for first remote desktop computers and then VINCI itself. All analyses were completed within VINCI by Westat, and the results were reviewed to protect personally identifiable information before they were extracted and shared with the committee. No raw data or any tables with any cell sizes of less than 10 veterans were shared.

2025. VA linked the VHA and ILER data using Social Security numbers, dates of birth, and veteran benefits identifiers; generated random identifiers; removed the personal identifiers; and made the final limited dataset available to Westat in the VINCI space. ILER records that did not have a VA record match were dropped from the sample population for analysis. Table

TABLE 4-1 Summary of Data and Data Sources Requested and Included in the Limited Dataset VA Provided to the Committee

NOTE: ADUSH = Assistant Deputy Under Secretary for Health; CDW = Corporate Data Warehouse; DHA = Defense Health Agency; DMDC = Defense Manpower Data Center; DOEHRS = Defense Occupational and Environmental Health Readiness System; EHR = electronic health record; ILER = Individual Longitudinal Exposure Record; OMHSP = Office of Mental Health and Suicide Prevention; SBOR = Suicide Behavior and Overdose Report; USVETS = United States Veterans Eligibility Trends and Statistics; VA = Department of Veterans Affairs; VBA = Veterans Benefits Administration; VHA = Veterans Health Administration.

4-1 summarizes the data sources requested and received, the data owners, and a brief description of each source. Two datasets curated and maintained by the VA Office of Mental Health and Suicide Prevention (OMHSP) were requested but not provided, and the effect of that lack is described in the next section on outcomes data.

Outcomes Data

The committee investigated the outcomes shown in Box 4-1. Appendix D lists the ICD-10 codes that correspond to each outcome. The Statement of Task specified 10 outcomes. The committee combined two outcomes and added seven: anxiety disorders, adjustment disorders, sleep disorders, substance use disorders (SUD), amyotrophic lateral sclerosis (ALS), Parkinson’s disease, and multiple sclerosis (MS). The committee chose to add anxiety

disorders, adjustment disorders, sleep disorders, and SUD because of known or perceived interest in their etiology among post-9/11 veteran communities. It interpreted “neurocognitive disorders” in the Statement of Task to include ALS, Parkinson’s disease, and MS in addition to dementia, as they are neurologic diseases known to have presentations that can include changes in cognition. Although traumatic brain injury (TBI) was specified as an outcome in the Statement of Task, during VA’s charge to the committee, VA stated that because TBI is defined as “a traumatically induced structural injury and/or physiological disruption of brain function as a result of an external force” (VA and DoD, 2021) and by definition is not caused by toxic exposure, it could be included as a secondary outcome or other type of variable (Hastings, 2024). The committee chose to use TBI as a covariate.

Data on each of the outcomes came from ICD-10 diagnostic codes recorded in inpatient and outpatient records available in VHA’s Corporate Data Warehouse (CDW). Pharmacy data were not available to the committee and not used as an additional verification for ICD-10 code-based diagnosis. In presenting the committee’s charge, VA specified that it should focus only on diagnoses and not on symptom severity (Hastings, 2024). Because several ICD-10 diagnostic codes may be associated with a class of outcomes, the codes that were used for adjustment disorders, ALS, attention disorders, anxiety disorders, bipolar disorder, dementia, depression, MS, Parkinson’s disease, psychosis, posttraumatic stress disorder (PTSD), schizophrenia, sleep disorders, and SUD² were drawn from the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, a recent National Academies study and other literature that also used ICD-10 codes for outcomes, and committee expertise (APA, 2013; Cheng et al., 2020; Goldman et al., 2023; Larson et al., 2018; Leung et al., 2022; Mehta et al., 2023; NASEM, 2025; Rawat et al., 2023). Cases of TBI were also identified by ICD-10 codes (DoD, 2019; Rawat et al., 2023) and are a measure of diagnosed TBI; these codes do not include all people who have been screened for TBI and do not distinguish TBI severity. Only outcomes diagnosed between January 1, 2017, and December 31, 2023, were examined.

CMI, also known as Gulf War illness among those who served in the 1990–1991 Gulf War, is typically assessed using self-reported symptom-based surveys and does not have a singular diagnostic code or case definition (IOM, 2014). Lacking an existing case definition or measure for CMI based on diagnostic data, the committee developed a measure to identify cases. A National Academies committee (IOM, 2014) tasked with developing a case definition

___________________

² SUD includes alcohol-related disorders; opioid-related disorders; cannabis-related disorders; sedative-, hypnotic-, or anxiolytic-related disorders; cocaine-related disorders; other stimulant-related disorders; hallucinogen-related disorders; nicotine dependence; inhalant-related disorders; and other psychoactive substance–related disorders.

for CMI recommended either the Kansas (Steele, 2000) or Centers for Disease Control and Prevention (CDC) definition (Fukuda et al., 1998) depending on specific needs or circumstances. The Kansas definition requires chronic symptoms in three of six domains (fatigue/sleep, pain, gastrointestinal, respiratory, neurologic/cognitive/mood, and skin symptoms), with onset since 1990; the CDC definition requires one or more symptoms present for at least 6 months from at least two categories: fatigue, mood and cognition, and musculoskeletal. Given the limitations of using ICD codes rather than symptom-based information as is typically seen in research examining CMI as an outcome, the committee applied the stricter Kansas definition to minimize potential misclassification. It developed a list of codes for each of the six domains (see Appendix D) and identified a CMI case as the first diagnosis in the third of the six domains. In the absence of standardized CMI definitions or measures, the committee conducted a series of sensitivity analyses to assess the validity of its measure. First, it restricted and loosened the number of domains used to identify a case to one and four, respectively. In addition, because CMI symptoms overlap with long COVID symptoms, the committee also conducted sensitivity analyses comparing CMI cases with and without COVID and their respective controls. The committee determined that there were no meaningful differences across these different measures.

The committee requested but did not receive data from VA’s OMHSP on suicide fatalities, which capture completed suicides reported to VA. Although incomplete compared to the National Death Index (NDI) data, the committee opted to pursue these data because they are owned by VA and require fewer data sharing and other permissions, and the committee believed it would receive these data sooner than the NDI data, which was important given the committee’s rapid timeline. When these data were not provided, the committee identified ICD-10 codes for nonfatal suicide attempts and intentional self-harm to capture suicide behaviors (Hedegaard et al., 2018).

Exposure Data

Data on military-related environmental and occupational exposures and military service history came from ILER. The committee worked with DHA to develop a custom data extract comprising exposure information linked to personnel and deployment records. The committee requested all relevant deployment information from the geographic areas of interest³ and two sources of exposure data from Defense Occupational and

___________________

³ The Southwest Asia Theater of Operations and Afghanistan. The Southwest Asia Theater comprises Iraq, Kuwait, Saudi Arabia, the neutral zone between Iraq and Saudi Arabia, Bahrain, Qatar, the United Arab Emirates, Oman, the Gulf of Aden, the Gulf of Oman, the Persian Gulf, the Arabian Sea, the Red Sea, and the airspace above these locations.

Environmental Health Readiness System: environmental monitoring data (the exposure pathway dataset), which were collected by environmental science and engineering officers, and industrial hygiene data on workplace hazards. The data-use request included a list of predefined exposures, which was chosen to protect personally identifiable information and to improve analysis time by reducing data processing time, file size, and effort required to link to VA records. The exposure pathway data are linked to personnel based on their shared documented dates and locations (e.g., stationed at a given military base when an exposure was recorded). The group-level industrial hygiene data are linked to personnel through membership in a similar exposure group—those with a similar exposure profile to a given agent because of the similarity in their occupational materials, tasks, and how they perform their tasks. These two ILER sources were linked to personnel and deployment records from the Defense Manpower Data Center (DMDC) by ILER staff. The committee requested all deployment records in ILER; DHA only provided records for individuals with a documented exposure of interest due to privacy concerns.

For both datasets, the committee developed keywords to identify and classify exposures. In the exposure pathway data, exposures were documented in a narrative across multiple unstructured variables; for the query, regular expression pattern matching was used to identify keywords associated with the exposure and source (e.g., “dust” and “road” for road dust). Some exposures of initial interest to the committee such as pesticides were subsequently omitted due to low frequency and lack of consistent data. The committee used its expertise to best capture and categorize exposures based on the data available, despite some challenges in categorization. For example, because “dust” exposures were numerous, the committee further classified it as road (documentation of both “dust” and dirt roads, airfields, motor vehicles, or similar) or desert (only information on “dust” or “soil” without accompanying documentation of the source) dust. Additionally, the committee wanted to classify burn barrels as separate from burn pits; however, burn barrels were only captured intermittently. To include this exposure with sufficient sample size for comparison analyses, it was aggregated with other open waste burning, including burn pits and burning landfills. However, the committee recognizes that the materials burned in barrels, including paper products, wood, and plastic objects, may not be comparable to items burned in larger, open locations, such as machinery. In contrast to the exposure pathway data, the industrial hygiene data were structured and encompassed a finite list of the assessed chemical or compound. Therefore, the committee’s industrial hygiene query used keywords specific to the chemicals typically associated with each exposure.

The committee’s initial analysis plan was to combine both ILER exposure data sources into 16 categories for analysis. After preliminary

analyses, the committee determined that differences in the two datasets and sample sizes required a different approach. For example, in the exposure pathway data, exhaust exposures were identified using keywords such as “exhaust” and “combustion.” In contrast, keywords in the industrial hygiene data identified specific chemicals typically associated with exhaust, such as “nitrogen dioxide,” “sulfur dioxide,” and “polycyclic aromatic compounds.” Without clear source attribution, many of the industrial hygiene keywords were associated with multiple exposures of interest and so not directly comparable to those in the other dataset (e.g., specific chemicals versus broader sources in the exposure pathway dataset), so analysis results would have been substantially more imprecise. Additionally, after examining sample sizes, several categories in the industrial hygiene data had low counts that would require most analyses with them to be suppressed. Moreover, more than 95% of veterans had already been identified through the exposure pathway data. The committee determined that it would only use exposure pathway data and refined the exposure categories to nine.⁴ Box 4-2 shows the list of exposures. Solid bullets denote committee-determined groupings of exposure categories, and indented, empty bullets denote the specific types of exposures as named in the exposure pathway data that are included in the broader category designation.

Covariate Data

Covariate data came from both ILER and VA. Social and demographic information (e.g., age, sex, race, and ethnicity) is available from multiple sources, including the DMDC personnel files in ILER, CDW EHR, and the United States Veterans Eligibility Trends and Statistics (USVETS), a VA data source that consists of administrative data from the Veterans Benefits Administration (VBA), VHA, and DMDC combined with purchased commercial data (NCHS, 2025; VA, n.d.). Additional measures of health care use (counts of nonpharmacy encounters) came from CDW inpatient and outpatient records. Military information, such as service component and branch, came from both ILER personnel files and USVETS. Additional covariates used from USVETS include the VA enrollment date, VA disability rating (a measure of service-connected disability), and VA priority group (a score that combines military service history, disability rating, income level, and other VA benefits). When demographic data differed between sources

___________________

⁴ The committee also requested an exposure to foodborne illness to serve as a negative control because it originally designed sensitivity analyses to assess the potential magnitude of bias. The committee chose not to conduct these analyses because they would not produce meaningful estimates given the limitations of the data.

for the same individual, VHA demographics were used. The demographics recorded closest to an individual’s first VHA encounter within the study period were used.

Study Design Description

Ideally, the committee’s analytic approach would follow the conceptual framework (shown in Figure 2-1 and described in detail in Chapter 2) and account for factors that occur before or concurrently with military exposures during deployment and may contribute to the development of mental, behavioral, or neurologic outcomes or CMI. However, the data available did not have several of the elements in the framework, particularly those that clearly occur and are measured before entering military service and before deployment and affect risk of exposure and outcome, such as mental health diagnoses or adverse childhood experiences (ACEs). Some information was available on factors such as TBI and combat-related or other traumas that may have been concurrent with deployment-related environmental and occupational exposures and could influence the associations of interest; however, the data do not specify the timing of these exposures. Therefore,

it is not clear if they could be confounders or mediators of the associations. For example, although VHA health records include a measure indicating whether someone had reported to VHA ever experiencing military sexual trauma while serving, they only include the incidents and not information on when trauma occurred. For these reasons, such factors were considered but not included in the model. Some factors that may occur after exposure or deployment also can be strong predictors of the outcomes of interest and may be in the causal pathway. For example, VA priority and disability ratings assigned at separation influence access to VA health and other benefits and the likelihood of receiving health care at VA. The committee designed a streamlined analysis accounting for variables in the model that were measured in the ILER or VA datasets, including matching on and adjusting for demographic factors (age at encounter, sex, enrollment year in VBA, race, ethnicity, and last known service branch) and developing a series of sensitivity analyses to help interpret the findings (see the next sections for more information).

Case-Control Sampling

The committee determined that a case-control design was the most appropriate given the outcomes and data sources. Specifically, a case-control design is practical and appropriate given the short outcome window, rare outcomes, and sample size. In addition, compared to a cohort design, a case-control study is more computationally efficient given VHA server limitations and the committee’s timeline. Statistical results from a case-control study (i.e., odds ratios [ORs]) are also easily interpretable. The sample was post-9/11 veterans who were enrolled in VA health care; received care at a VA facility between January 1, 2017, and December 31, 2023; had a record in ILER after September 11, 2001; and were deployed to the relevant locations (see footnote 3). Cases were defined as VA health care users who had an ICD-10 recorded diagnosis of one of the outcomes that was received in either inpatient or outpatient VA settings during the specified period. Some veterans had received diagnoses before January 1, 2017, first documented in an outside health system; therefore, cases should be interpreted as the first documentation, not as incident events. Controls were individuals with an ILER exposure record who used VA health care in the same period but did not have a diagnosis of the outcome of interest.

For its analysis, the committee chose an approach used for phenome-wide association studies in EHRs: veterans who had any of the outcomes of interest or related outcomes at any point during the follow-up period were not eligible to match as controls (Denny et al., 2010; Wu et al., 2019a). It determined related outcomes of interest by mapping outcome ICD-10 codes to phecodes (groupings of ICD billing codes that capture clinically

___________________

³ The Southwest Asia Theater of Operations and Afghanistan. The Southwest Asia Theater comprises Iraq, Kuwait, Saudi Arabia, the neutral zone between Iraq and Saudi Arabia, Bahrain, Qatar, the United Arab Emirates, Oman, the Gulf of Aden, the Gulf of Oman, the Persian Gulf, the Arabian Sea, the Red Sea, and the airspace above these locations.

meaningful concepts for clinical research using EHR data). Records with any ICD-10 codes associated with that phecode were then excluded from the pool of eligible controls. For example, PTSD was defined as having an ICD-10 diagnosis of F43.1 Post-traumatic stress disorder. The phecode mapping of PTSD resulted in excluding both those with an F43.1 diagnosis and those with the diagnosis of Z86.51 Personal history of combat and operational stress reaction, an outcome related to PTSD, from the pool of potential controls for analyses with PTSD as the outcome. The list of outcomes and associated phecodes is in Appendix E. This strategy is meant to partly address a major limitation of EHR data—namely, that people may have health conditions and related symptoms present for a time before diagnosis within the health system.

Each case was matched to up to four controls on age at encounter of first diagnosis, sex, and enrollment year in VBA. Matching was limited to these three variables because data were not available on several potential confounders that might have been considered, such as rank and date of enlistment. An individual’s status as a case or control was determined based on diagnoses during 2017–2023. For each outcome, veterans with the other outcomes examined in the study were included in the pool of potential controls. For example, a veteran with a depression diagnosis could match as a control for one with attention-deficit/hyperactivity disorder (ADHD). A control could also be matched to multiple cases. Cases and controls were compared on key demographic variables (e.g., military component and rank, race and ethnicity, health care utilization, health status) to determine similarity of the case and control groups, and no notable differences were identified.

Analytic Methods

Before running preliminary descriptive statistics, the committee determined that several models would be run for each exposure–outcome pair, as each model would add complexity or account for different factors. For each pair, unadjusted and adjusted conditional logistic regression models were used to estimate the odds of exposure for matched case-control pairs. The unadjusted model accounted for the matching characteristics: age at encounter, sex, and enrollment year in the VBA. The adjusted model also accounted for race, ethnicity, and last known service branch. As the committee only had access to certain military characteristics at discharge, it assumed that the branch at separation was the same as at beginning of service. Factors reserved for adjustment included those with high rates of missingness and that were potentially related to the exposure and outcome. Due to general missingness, use of variables like race and ethnicity in the matching would exclude potential study participants for whom data

on the matching characteristics are needed to create a match. Moreover, including these variables instead as covariates for adjustment allowed the committee to observe whether there were meaningful differences on these characteristics (i.e., between races, ethnicities, and service branches). In comparison, matching variables do not produce a statistic. To be fully responsive to its Statement of Task, the committee repeated this analysis stratifying each pair by TBI status. Veterans were included in the TBI stratification if there was a TBI-related encounter and diagnosis code (see Appendix D) during the study period.

Given the limitations and availability of exposure data (see Limitations section), the committee treated exposure as a binary measure (ever or never), except for cumulative exposure, which is based on additive counts of individual binary exposures. To address the potential for multiple exposures, the committee designed analyses to understand the risk of disease when a veteran had more than one exposure. ORs for each outcome were compared for no exposures to at least one, two, three, four, five, six, and seven-plus exposure groups.

The committee undertook sensitivity analyses to assess the robustness of findings from the primary analysis models. First, all models were repeated selecting controls from VHA with combat service post-9/11. Service flags from USVETS and combat service recorded in CDW were used to identify controls. Using these flags, the controls for this analysis served during the same period and saw combat (likely deployed); however, their location may not have been in Southwest Asia or Afghanistan or their deployments may have been shorter or in other ways different from individuals who appear in ILER. Second, estimates for the neurocognitive outcomes were stratified by age to determine whether the patterns of disease differed in older age groups.

Quality Assurance and Quality Control

Westat followed committee-defined protocol to define case and control populations and conduct analyses; analysts performed variable validation tests to verify exposure, outcome, and covariate data. Iterations of the study protocol were implemented and made with committee consent. Coding was performed in SAS, with codes reviewed by at least one additional analyst. Upon completing analyses, Westat drafted results memos for the committee that included methodologic and programming notes. The project director and analysts wrote the memos, with one serving as the primary writer and the other as quality control. The results were also reviewed by other project analysts and/or the primary investigator for quality control and adherence to VHA privacy policy.

Limitations

The committee recognizes the many limitations of its data analysis. Importantly, the sample only generalizes to the population who have a recorded exposure in ILER, separated from the military, and received VHA health care between 2017 and 2023. It is not representative of all post-9/11 service members or veterans. The study population does not capture health effects of military exposures from those who are still in the military. Separation may be affected by any health effects of military exposures (e.g., those with serious health conditions caused or exacerbated by exposures may be more likely to separate sooner). It also may not be representative of Marines, because their exposure data were collected and recorded differently in ILER and their information is limited. VHA users differ systematically from veterans who do not use VHA health care services in age, insurance coverage, and general health status (Liu et al., 2005; Meffert et al., 2019; Wagner et al., 2024; Wang et al., 2021). As described in Chapter 1, the task was to examine military exposures that happened on overseas deployment to Southwest Asia and Afghanistan and not other possible exposures during deployment elsewhere or domestically (i.e., on military bases in the United States or on ships at sea). The remainder of this section describes key study limitations arising from the quality and availability of the underlying exposure and outcome data.

The committee also identified additional mediators that it neither matched on nor adjusted for, because they are strongly associated with the outcome and in the causal pathway connecting exposures to the outcome. For example, if a particular exposure caused serious mental illness, the associated symptoms would likely affect marital status, employment, income, and homelessness. Adjusting for these variables would attenuate the estimated association between exposure and outcome, so the committee did not include them in the statistical model.

Limitations Related to Exposure Measurement

The committee only had access to data on military exposures recorded in ILER, and it recognizes that some were not documented. As noted previously, information on some exposures of potential interest like pesticides, ultrafine particles, and per- and polyfluoroalkyl substances was not available or insufficient. Similarly, although the committee captured exposure to general solvents, it did not consider specifically trichloroethylene in the final analysis, as this level of detail was not always available. Thus, the general exposure categories reflect sources of exposure to chemical mixtures with unknown and potentially heterogeneous composition. Moreover, a substantial proportion of the exposures in ILER lacked sufficient location

information. Approximately 10% of industrial hygiene and 5% of exposure pathway locations were missing or entered as buildings, rooms, or other areas with no matching geographic information to determine country.

Ideally, the committee would have person-level data that documented whether any specific veteran was exposed to a given hazard and the amount of exposure. Person-level self-reports of potential harmful exposures are available from a variety of sources, such as VA exposure registries and the Toxic Exposure Screen, but the committee did not request these data as they are not reliable for population health research (Mattras, 2024; NASEM, 2017; VA, 2024). The available exposure data were collected at the area level (such as a military base) and linked to individuals based on their recorded dates and locations or membership in a similarly exposed group. Because the committee had no specific information on which part of a base someone may have been assigned (or outside of it), specific tasks designated to personnel, or how long or frequently they performed these tasks, ecologic exposure data inhibited its ability to assess individual associations. Living conditions varied across unit types and missions, and many of these structures had asbestos and other hazardous structural exposures. The risk of misclassification from assigning measures captured at the area and group levels to individuals is a major concern. Relying on group-average exposure measurements rather than person-level measurements reduces statistical power to detect associations between toxic exposures and health effects, and mismeasurement of exposure will bias these estimates of association (Armstrong, 1998; Yland et al., 2022). Additionally, the area-based exposure data in ILER available for the committee’s analysis are based on periodic and inconsistent surveillance, and many exposures were not documented systematically and often as unstructured text with limited detail. This reflects operational considerations, such as the accessibility of deployment locations and physical space limitations that shaped decisions of environmental science officers when choosing what monitoring equipment to bring and thus what exposures to measure (Kolivosky, 2024). Consequently, the committee developed rules for classification based on its experience and expertise but had to make assumptions about the documented exposures, and this may have introduced additional misclassification (see the section on Interpreting the Results of the Committee’s Original Analyses in this chapter), which could attenuate or magnify observed associations between a given exposure and outcome. However, the lack of underlying detail in the documentation did not permit more precise or accurate exposure measurement.

Use of area-level exposure measures also meant that there was no reliable method to account for frequency or duration of exposures. The committee attempted to account for this by using the duration of exposure potential measured in days at a location with a known exposure (APHC,

2013). However, approximately 40% of exposure pathway and 50% of industrial hygiene exposure end dates were missing or invalid. Likewise, there is no information on dose of exposures. In some cases, ILER included documentation of reported hazard severity and probability. However, differing levels of exposure were not reported in 10% of cases. Similarly, many of these exposures were co-occurring (see Chapter 5). The committee was unable to tease apart the contribution of individual exposures in the presence of other highly correlated exposures or interaction between them, but it recognizes that advanced statistical techniques could be used in future analyses to assess this and that its inability to provide evidence of a dose–response relationship may not accurately convey the risk presented by these exposures.

Furthermore, limited information on dose, duration, and frequency inhibits the committee’s ability to evaluate the impact of the healthy warrior effect. Like the “healthy worker” effect observed in civilians, this is a form of selection bias in which military personnel may experience more exposures due to their physical fitness and access to medical care during deployment (Arrighi and Hertz-Picciotto, 1994; Buckley et al., 2015; Chowdhury et al., 2017; Nuyts et al., 1993). This type of selection bias will often cause service members with occupational exposures at lower levels to exhibit less risk for disease until they have a much higher dose, duration, or accumulation of exposures. Additionally, younger and healthier service members are more often assigned to more physically taxing responsibilities with the potential for more or higher levels of exposures, and if not medically discharged, will continue to be deployed, increasing frequency of exposures. These higher and more frequent exposures can compound over time, leading to higher risk of more prominent adverse health outcomes, including mental health disorders, later in life. Although this effect is routinely accounted for in occupational analyses, the committee’s inability to account for dose impeded its ability to assess this effect. The challenge of the healthy warrior effect is compounded by “length bias”—service members with longer service and more time deployed are more likely to have at least one exposure observed and recorded. In the context of this analysis, mental, behavioral, and neurologic health outcomes with symptoms that emerge in early adulthood (e.g., bipolar disorder or MS) would be expected to lead to fewer deployments or earlier military separation, thus reducing opportunities for exposures to be observed and recorded.

Although the committee analyzed associations with multiple binary (ever/never) exposure types, it was unable to evaluate the impact of frequency or distribution of the same exposures. Specifically, all exposures were assigned as “ever” versus “never” recorded, which did not account for repeated exposures to the same toxicant. Considering that certain

exposures, such as burn pits, may have occurred daily for some service members or reoccurred during subsequent deployments, longer and more frequent exposures could produce bioaccumulation of toxicants, increased cellular damage, and elevated risk of specific health outcomes. Relatedly, some exposures may have occurred only once, and a one-time exposure to moderate concentrations of bioaccumulative toxicants is not likely to provide measurable risk compared to repeated exposures over time. These scenarios cannot be accounted for in the data analysis, but future studies could apply longitudinal data structures to better assess risk from chronic exposures in this population.

Finally, the committee considered analyses to evaluate potential differences in exposure windows in relation to diagnostic dates. Study participants may have experienced symptoms during periods of active duty and would not receive a diagnosis until after deployment or discharge. Additionally, mental and behavioral health outcomes are ascertained via VHA medical records, and initial diagnoses may have occurred using other medical resources, such as DoD TRICARE or private insurance. Since the committee could not accurately assess symptom onset or disease incidence, it was unable to determine appropriate lag times for known exposure windows or conduct any analysis of time to diagnosis.

Limitations Related to Outcome Measurement

The committee only had access to data on health outcomes from VHA administrative databases, allowing only examination of associations between military exposures and mental, behavioral, and neurologic health outcomes documented in the VHA system. Notably, the committee did not receive data on suicide fatalities and was limited to assessing nonfatal events. Not all veterans receive care (and diagnoses) at VHA, and they may do so for a subset of their needs (e.g., certain procedures or services) (Eibner et al., 2015; Hynes et al., 2021; Radomski et al., 2022; Rasmussen and Farmer, 2023; Tsai et al., 2015; VA, 2018a,b; Wagner et al., 2024). In particular, veterans with private insurance are less likely to use VHA care (Shen et al., 2008). Use of VHA may be related to the effects of military exposure and mental and behavioral health outcomes, as deployment experiences (including combat trauma and military exposures) and their sequelae influence VA priority and disability ratings assigned at deployment. Additionally, disabling health conditions (including mental, behavioral, and neurologic health outcomes) may reduce the ability to obtain private insurance, limiting the options for care outside of VHA. The committee’s analysis was further limited to clinical conditions documented in VHA records beginning January 1, 2017. Thus, diagnoses before 2017 are not reflected in the analysis, including incident mental, behavioral, and

neurologic health diagnoses in VA clinical records data before that date. Similarly, individuals who died from their exposures prior to 2017 also would not be captured in the data.

A proportion of outcomes may have been misdiagnosed or inconsistently diagnosed due to underdiagnosis (a condition is missed), overdiagnosis (a condition is diagnosed before it is verified), and differences in how diagnoses are coded (a diagnosis was miscoded or used an alternate code), as is common in EHRs, especially for psychiatric disorders. For example, an existing condition may be undiagnosed if the veteran were being seen for a different complaint, leading to underdiagnosis. Alternately, diagnosis may be made using a code that was not included in the list for this analysis. It is also possible for diagnoses to be prematurely entered into the EHR; clinicians may assign a diagnosis to an encounter before it is verified, resulting in overdiagnosis. Correct recording of diagnoses in the EHR is also not random but, instead, may be related to the study exposures or outcome severity. For example, patients with more frequent VHA engagement may be more likely to have mental and behavioral health conditions correctly documented, and, as noted, frequency of VHA use may be associated with these exposures and outcomes.

Limitations Related to Lack of Data on Confounders

As described in the committee’s conceptual framework in Chapter 2, a broad range of factors, including military-related environmental and occupational exposures, contribute to developing the mental, behavioral, and neurologic outcomes of interest. An ideal study would account for the influence of these other factors to determine the unique effects of the environmental and occupational exposures. Key factors that may contribute to such adverse outcomes but on which the committee had no data include preexisting psychiatric diagnoses. The committee did not have health records from before entering the military and limited health and administrative information from active duty. Therefore, it had no information on whether individuals received premilitary mental or behavioral health diagnoses. However, because many mental, behavioral, and neurologic health outcomes are disqualifying for military service, it is likely that these were incident at some point within service or after separation.

STRUCTURED LITERATURE REVIEW

In recognition of the limitations of drawing conclusions from any singular analysis, particularly when the quality of the underlying data is limited, the committee conducted a review of the epidemiologic literature as a source of supporting evidence to contextualize the results of its statistical analyses. This section describes the committee’s structured literature review,

including the processes for identifying and screening studies for inclusion. The next section describes how the committee evaluated the literature and integrated its analyses and the supporting literature to develop conclusions.

Identification of Epidemiologic Evidence

A structured search was conducted to assess published epidemiologic literature on possible associations between the exposures and outcomes of interest. Given the committee’s timeline, the committee began before receiving the exposure data and completed screening and evaluation in parallel to the data request and analyses for efficiency. Owing to the paucity of literature on military-specific and -related exposures and the health outcomes, the committee also included analogous literature on environmental and occupational exposures in general populations. The committee acknowledges that reliance on general epidemiologic literature rather than veteran-specific literature may in some cases have limited relevance to military exposures. For example, because the committee only had information on exposure sources and not specific toxicants, the chemical composition of military exposures may differ from civilian exposures (e.g., particulate matter [PM] from burn pits may not be chemically similar to PM from ambient air). At the same time, because exposures during deployment are typically more intense, frequent, and sometimes longer duration, evidence from civilian exposures may represent a lower range of what one might expect to observe from military exposures (e.g., PM from burn pits may be more intense than PM from ambient air). The committee further recognizes that extrapolating literature from the general population to the military may not be able to do justice to the complexity of the military experience, especially the added stressful aspects of deployment to a combat zone, which may enhance vulnerability.

The exposures for the literature search were burn pits, PM, fuels, metals, mold, radiation, and solvents. Within each category, the committee specified more precise search terms (e.g., the search for metals included terms for cadmium, chromium, lead, manganese, etc.). Of note, the exposure categories used for the literature review do not align perfectly with the exposure categories developed for the analyses. Whereas the committee assessed dust and PM, exhaust, and incinerator exposures in its statistical analyses, it relied on literature on PM (which captured evidence on unspecified PM, PM_2.5, PM₁₀, specific pollutants [e.g., ozone, nitrogen oxides], diesel exhaust, and dust) to complement its analytic results. It decided not to conduct additional searches on dust, exhaust, and incinerator exposure after accessing the ILER data for three reasons: (1) PM being a major component of these exposures; (2) the relative robustness of the literature on PM and the outcomes of interest compared with what the

committee anticipated would be available on dust, exhaust, and incinerator exposures; and (3) time constraints after receiving the data.

A National Academies research librarian searched four databases—Embase, Medline, ProQuest, and Scopus—for published articles based on title, abstract, and keywords for each exposure–outcome pair. Because TBI was not included as a primary outcome in the data analysis, it was excluded. The committee provided a list of search terms for each exposure and outcome, and the librarian identified similar and related search terms. Although veterans are the focus of the data analysis, the population parameter included adults aged 18+ to capture studies of occupational cohorts, environmental exposures, and case-control studies from any country. For attention-deficit disorder/ADHD, the age restriction was dropped to include studies of adolescents because ADHD is often diagnosed in adolescence and defined as a neurodevelopmental disorder with onset before age 12 (APA, 2013; NIMH, 2021). Box 4-3 summarizes the general search parameters. For each exposure and health outcome, several additional related terms were searched. For example, using PM, the following terms were included: “particulate matter”, “particulate matter 10”, “particulate matter 2.5”, “PM 10”, “PM 2.5”, “air pollutants”, “air pollution”, “airborne hazards”, “ambient air”, “ambient particulate matter”, “atmospheric pollution”, “organic pollutants”, “air pollutants occupational”, “air pollution indoor”, “air pollution radioactive”, “inhalation exposure”, “persistent organic pollutants”, “waste burning, smoke”, and “soot”. Using ALS, search terms included “ALS” and “Lou Gehrig’s disease”.

Results were limited to studies published in peer-reviewed English-language journals between January 1, 2013, and August 1, 2024. This approximately 10-year search strategy was considered appropriate because the committee was not tasked with conducting a systematic review and weight-of-the-evidence assessment of the published literature. Rather, the published epidemiologic literature was intended to be used as supportive evidence of the findings from the committee’s original data analysis to inform whether exposure–outcome relationships were possible. The meta-analyses and systematic reviews may have included individual studies published before 2013.

Literature Screening

Studies were screened for inclusion based on the following criteria: original peer-reviewed research reports (primary studies, systematic reviews, meta-analyses, or umbrella reviews); published between January 1, 2013, and August 1, 2024; English language; conducted in adult human populations (except for literature on ADHD, which included adolescents, as described above); and evaluated the relationship between an exposure of

interest and an outcome of interest. Studies could have measured incidence, prevalence, worsening of symptoms, or mortality related to an outcome. Excluded literature included narrative reviews, commentaries, case reports, and conference abstracts. Systematic reviews were excluded if they did not include the criteria for selecting studies or if they did not assess the risk of bias in the included studies. Primary analyses in a review were not further included in the committee’s assessment of evidence separate from the review.

EVALUATION OF EVIDENCE

The committee developed a tiered framework for evidence synthesis before seeing the results of any analysis and used this to assess the results of its analyses and associations reported in the literature, integrate and synthesize findings across these sources of evidence, and develop conclusions about possible relationships between the exposures and outcomes. First, the committee assessed the results of its analyses using statistical measures and considered the potential influence of systematic and random error to develop findings. Second, it assessed supporting evidence for associations found in its data analysis from the literature and related systematic reviews. Finally, to develop its conclusions, it synthesized evidence from its analyses and the literature based on the strength and consistency of evidence across evidence types.

Interpreting the Results of the Committee’s Original Analyses

The committee’s primary focus was a thorough and robust statistical analysis of data provided by DoD and VA. The first step of the framework

was to assess the strength of the evidence from its original analysis of these data. It examined multiple statistical metrics that describe both association strength and data quality: the point estimate and magnitude of effect, 95% Wald-type confidence interval (CI), and p-value (with no threshold for statistical significance). For each exposure–outcome pair, an error in interpreting the results is possible due to imperfections in the observational data. Therefore, the committee also considered the multiple sources of biases from systematic and random error that are omnipresent in such analyses.

The committee considered three forms of systematic error—selection bias, confounding, and misclassification. Selection bias occurs when those who participate in a study differ from the target population and may affect the committee’s results due to documentation of exposures in ILER, use of VA health care, and/or data on persons that are otherwise captured in DoD and VA administrative records. Consequently, although the study’s results may accurately describe relationships between exposures documented in ILER and outcomes among veterans who receive health care at VA, they likely are not generalizable to all service members and veterans. Selection bias is challenging to overcome once data collection is complete, particularly in situations with limited information about the base population from which the sample was drawn (all veterans who served in Southwest Asia Theater of Operations or Afghanistan after 9/11). Accordingly, the results of the committee’s primary analyses must be interpreted with caution related to the possible influence of selection bias, particularly when making inferences beyond the study population.

Confounding occurs when a third variable is associated with both the exposure and outcome of interest and occurs before both can obscure (magnify or attenuate) their true relationship if unaccounted for. The committee lacked information about prior traumas that could influence both the likelihood of deployment exposure and occurrence of mental, behavioral, and other neurologic health outcomes. Such traumas include those before entering service (such as ACEs) and those during service (such as traumas during military training or combat). Several possible methodologic approaches exist to mitigate the effect of confounding, including sample restriction, matching, and statistical adjustment, all of which the committee used in its analysis.

Misclassification, or information bias, refers to inaccurately categorizing a study participant’s status on a variable. For example, the committee may have misclassified exposure status due to limited and inconsistent surveillance data and outcome status if someone received diagnoses outside of VHA. Exposure measurement error can attenuate or magnify observed associations between an exposure and outcome (Hart et al., 2015; Kioumourtzoglou et al., 2014; Wu et al., 2019b). Although the committee expects substantial error in the assessed exposures for its analyses, any

error is likely not related to the outcomes of interest and would therefore bias the estimated effects toward the null. If the degree of misclassification differs between the exposed and unexposed groups, an association may be observed when a true association does not exist.

The committee also considered the potential for random error, or error due to chance, which could cause it to incorrectly interpret its results. Because the committee conducted many comparisons, it was aware that it might incorrectly identify a statistical association between a given exposure and outcome. In other words, the results could show an observed association between an exposure–outcome pair when that was instead due to chance. This is known as the “multiple-comparisons problem,” and a common way to address this is a statistical procedure to adjust for multiple comparisons. However, the committee ultimately decided not to do so for three key reasons. First, it did not rely on p-values to determine the importance of findings but considered multiple statistical metrics. P-values used for null hypothesis statistical testing are limited because they rely, in part, on sample size and therefore have less meaning in a sample as large as this one (Gómez-de-Mariscal et al., 2021; Greenland, 2019; Greenland et al., 2016; Yaddanapudi, 2016). Thus, if p-values are not the focus of the committee’s interpretations, there is no need to adjust them out of concern for incorrect interpretations due to chance alone. Second, the committee’s charge was to assess “… possible [emphasis added] relationships between toxic exposures experienced during military service and mental health conditions and other health outcomes.” The committee interpreted “possible” to mean that VA was interested in holistic evidence regarding whether associations were possible and not formal determinations about evidence based on specific statistics, such as p-values. Third, the utility of p-values in general, and adjustments for multiple comparisons specifically, is debated in epidemiologic literature (Kang et al., 2017; Rothman, 1990). In a foundational paper, Rothman (1990) argues that such adjustments confer important limitations, including increasing false negatives for associations that are not null.

Based on the committee’s consideration of multiple statistical measures and its assessment of the likelihood of the influence of random and systematic error on its confidence in its results, it determined whether the results of its original analyses supported the existence of a statistical association between each exposure and each outcome. Given the restricted literature on exposure–outcome pairs and limitations of the exploratory analysis, the committee used a point estimate to facilitate prioritizing results that would consider known biases of the dataset and common exposures to identify a likely true relationship. Therefore, in the adjusted analyses, the committee adopted a point estimate threshold of 1.10 or greater in which the 95% CI excluded 1.0 as a criterion for the prioritization of results described in

Chapters 6, 7, and 8. An OR of 1.10 offers a relatively conservative threshold for identifying possible associations given the small anticipated effect sizes in environmental epidemiology. At the same time, it is sufficient to capture weak evidence of true possible risk-conferring relationships while also mitigating against the possibility of making an error when interpreting the data due to the limitations of the underlying data. Thus, these criteria allow the committee to capture possibly meaningful associations for further investigation. Only positive associations that meet this threshold are noted; statistically significant point estimates less than 1.0 are not described because the committee chose to focus on identifying signals of potentially meaningful risk-conferring relationships. For the stratified analyses, the committee prioritized results where (1) adjusted OR between strata were different; (2) at least one stratum’s OR was above 1.0 and the lower bound of the 95% CI was also above 1.0; and (3) the 95% CI did not overlap between strata. The 95% CI for the stratum without elevated risk may include 1.0.

Assessing Supporting Evidence: Epidemiologic Literature

The second step in the committee’s tiered framework was to draw from the epidemiologic evidence. To inform its approach to assessing this literature, the committee reviewed the process for evidence synthesis used by National Academies (IOM, 2016; NASEM, 2018b, 2020) and other expert committees (IARC, 2012) evaluating the health effects of environmental and occupational exposures. Because the committee was not tasked with a weight-of-the-evidence review, each individual study captured in its search was not assessed; rather, their number, general quality, and consistency were used to support findings for each exposure–outcome pair in the original data analysis. The committee used these factors to determine whether an association between a given exposure and outcome provided generally suggestive, limited, mixed, inadequate, or insufficient evidence of a risk-conferring relationship. The committee also made conclusions of no identified literature.

Consideration of Mechanisms and Biological Plausibility

Consistent with prior National Academies committees, this committee determined that conclusions about potential relationships can be drawn without or irrespective of mechanistic information or animal studies when evidence is consistent from epidemiologic studies (NASEM, 2018a; NRC, 2014). Therefore, the committee did not conduct a structured literature review of the mechanistic or animal literature. Instead, it reviewed literature

based on mechanistic and animal studies on general mechanisms (neurotoxicity, neuroinflammation, and oxidative stress) by which environmental and occupational toxicants may cause physiologic changes to the brain, which may lead to the mental, behavioral, and other neurologic outcomes of interest (see Chapter 3). The committee relied on these general mechanisms to support the biologic rationale for a potential relationship between exposures and outcomes only when the epidemiologic evidence was sparse or absent.

Integrating Multiple Sources of Evidence

The committee was tasked to assess possible relationships, so its conclusions are based on the results of its analyses and the supporting evidence in the scientific literature regarding associations between military exposures and mental, behavioral, and neurologic outcomes. In accordance with this approach and informed by expert committees examining both causal and noncausal associations between environmental and occupational exposures and health outcomes, the committee determined a “possible relationship” to mean that a link may exist between a given exposure and developing a particular outcome among post-9/11 veterans who are VA health care users.

To draw overall conclusions about possible relationships between each exposure and outcome of interest, the committee compared its analytic results with the results of the literature review. For each exposure–outcome association examined, the evidence from the literature may or may not align with the results of the committee’s analyses. The committee considered four possible hypothetical scenarios of consistency of findings between its original analyses and the epidemiologic literature, in terms of strength and magnitude of the observed association. For each one, the committee classified a possible relationship into one of two conclusions:

• The committee concluded that there was a possible risk-conferring relationship between a given exposure and a given outcome for the following three scenarios:
  • The evidence from both the data analyses and available epidemiologic literature showed consistent risk-conferring effects;
  • The evidence from the data analysis showed a possible risk-conferring relationship and there was mixed, limited, or no identified epidemiologic literature; or
  • The data analysis did not show a possible risk-conferring relationship, but the epidemiologic literature showed suggestive evidence of a risk-conferring relationship.

• The committee concluded there was inadequate or insufficient evidence of a possible risk-conferring relationship between a given exposure and a given outcome for the following scenario:
  • The data analysis did not show a possible risk-conferring relationship and there was mixed, limited, or no identified epidemiologic literature.

Literature refers to PM and results of the data analyses refer to dust and PM. Literature on PM, which captures evidence on PM_2.5, PM₁₀, specific pollutants (e.g., ozone, nitrogen oxides), diesel exhaust, and dust, was used to support and contextualize analytic results on dust and PM, exhaust, and incinerator emissions. Conclusions relying on the data analyses alone or together with the literature refer to the exposure categories used in the analysis—dust and PM, exhaust, and incinerator emissions. Conclusions relying solely on suggestive evidence in the literature refer to PM alone. The committee did not make conclusions of inadequate or insufficient evidence based on the literature alone, so there are no conclusions of inadequate or insufficient evidence for PM alone.

SUMMARY

The committee’s task was to assess possible relationships between mental, behavioral, and neurologic health outcomes and CMI and environmental and occupational exposures experienced among post-9/11 veterans deployed to Southwest Asia or Afghanistan. Mental and behavioral health outcomes include adjustment disorders, attention disorders, anxiety disorders, depression, PTSD, serious mental illness including bipolar disorder and psychosis and schizophrenia, sleep disorders, SUD, and nonfatal suicide attempts and intentional self-harm. Neurologic outcomes include ALS, dementia, MS, and Parkinson’s disease. The committee identified burn pits, dust and PM, exhaust, fuels, incinerator emissions, metals, mold, radiation, and solvents as the military-related environmental and occupational exposures based on its expertise and the available data.

For its analyses, 135 case-control studies using health care and other administrative data from DoD and VA were conducted on the possible relationship between each exposure and outcome. The committee also examined whether experiencing more than one exposure, measured based on additive counts of individual binary exposures, was associated with increased risk of the outcomes. Data comprised records of post-9/11 veterans who received a diagnosis at an inpatient or outpatient encounter at VHA between 2017 and 2023 and had a record in ILER. Conditional logistic regression models estimated ORs among matched cases and controls. The committee conducted sensitivity analyses to assess the robustness of

its findings and stratified all exposure–outcome pair results by age and TBI status to assess differences between groups.

The committee recognizes many limitations of its analysis, which primarily arise from shortcomings in the quality and availability of the data. The committee’s sample only generalizes to the population of veterans who have a recorded exposure in ILER, separated from the military, and received VHA health care between 2017 and 2023—not to all post-9/11 service members or veterans, including those who are still in the military. Other limitations related to exposure measurement, the outcome measurement, and lack of data on potential confounders could have influenced the committee’s ability to observe true relationships. For example, limitations related to exposure measurement include only having data recorded in ILER, which is incomplete, periodic, inconsistent, not documented systematically, and lacking information on dose, duration, and frequency. In addition, although ILER was developed to provide person-level exposure data, the data are recorded at the area rather than person level and do not permit assessing an individual’s risk of developing a condition.

In recognition of the limitations of the data quality and availability and limitations of relying on any single study and thus to facilitate a more holistic assessment of potential relationships of military exposures and selected health outcomes, the committee completed a structured literature review and also considered biologic plausibility in its assessments. To reach conclusions, the committee synthesized the results of its analyses and of the literature search. It prioritized reporting analytic results with an adjusted OR of at least 1.10 with the exclusion of 1.0 in the associated 95% CI. Statistically significant point estimates less than 1.0 are not described because the committee chose to focus on identifying signals of potentially risk-conferring relationships for further investigation. The committee concluded that there were possible risk-conferring relationships under three scenarios: (1) both the analyses and literature results showed a possible risk-conferring relationship; (2) the analyses showed such a relationship but there was mixed, limited, or no identified literature; and (3) the analyses did not show a possible risk-conferring relationship, but the literature showed suggestive evidence of one. Results and the committee’s conclusions about possible relationships are in Chapters 6–8 and Appendix G.

REFERENCES

Armstrong, B. G. 1998. Effect of measurement error on epidemiological studies of environmental and occupational exposures. Occupational and Environmental Medicine 55(10):651–656.

Arrighi, H. M., and I. Hertz-Picciotto. 1994. The evolving concept of the healthy worker survivor effect. Epidemiology 5(2):189–196.

Buckley, J. P., A. P. Keil, L. J. McGrath, and J. K. Edwards. 2015. Evolving methods for inference in the presence of healthy worker survivor bias. Epidemiology 26(2):204–212.

Cheng, Y., A. Ahmed, E. Zamrini, D. W. Tsuang, H. M. Sheriff, and Q. Zeng-Treitler. 2020. Alzheimer’s disease and Alzheimer’s disease-related dementias in older African American and white veterans. Journal of Alzheimer’s Disease 75(1):311–320.

Chowdhury, R., D. Shah, and A. R. Payal. 2017. Healthy worker effect phenomenon: Revisited with emphasis on statistical methods: A review. Indian Journal of Occupational & Environmental Medicine 21(1):2–8.

Denny, J. C., M. D. Ritchie, M. A. Basford, J. M. Pulley, L. Bastarache, K. Brown-Gentry, D. Wang, D. R. Masys, D. M. Roden, and D. C. Crawford. 2010. PheWAS: Demonstrating the feasibility of a phenome-wide scan to discover gene-disease associations. Bioinformatics 26(9):1205–1210.

DoD (Department of Defense). 2019. ICD-10-CM Coding Guidance for Traumatic Brain Injury. https://health.mil/Reference-Center/Publications/2020/07/31/ICD10-Coding-Guidance-for-TBI (accessed June 6, 2025).

Eibner, C., H. Krull, K. M. Brown, M. Cefalu, A. W. Mulcahy, M. S. Pollard, K. Shetty, D. M. Adamson, E. F. L. Amaral, P. Armour, T. Beleche, O. Bogdan, J. L. Hastings, K. A. Kapinos, A. M. Kress, J. Mendelsohn, R. Ross, C. M. Rutter, R. M. Weinick, D. Woods, S. D. Hosek, and C. M. Farmer. 2015. Current and projected characteristics and unique health care needs of the patient population served by the Department of Veterans Affairs. Santa Monica, CA: RAND Corporation.

Fukuda, K., R. Nisenbaum, G. Stewart, W. W. Thompson, L. Robin, R. M. Washko, D. L. Noah, D. H. Barrett, B. Randall, B. L. Herwaldt, A. C. Mawle, and W. C. Reeves. 1998. Chronic multisymptom illness affecting Air Force veterans of the Gulf War. JAMA 280(11):981–988.

Goldman, S. M., F. M. Weaver, K. T. Stroupe, L. Cao, B. Gonzalez, K. Colletta, E. G. Brown, and C. M. Tanner. 2023. Risk of Parkinson disease among service members at Marine Corps base Camp Lejeune. JAMA Neurology 80(7):673–681.

Gómez-de-Mariscal, E., V. Guerrero, A. Sneider, H. Jayatilaka, J. M. Phillip, D. Wirtz, and A. Muñoz-Barrutia. 2021. Use of the p-values as a size-dependent function to address practical differences when analyzing large datasets. Scientific Reports 11(1):20942.

Greenland, S. 2019. Valid p-values behave exactly as they should: Some misleading criticisms of p-values and their resolution with s-values. The American Statistician 73(Supp 1):106–114.

Greenland, S., S. J. Senn, K. J. Rothman, J. B. Carlin, C. Poole, S. N. Goodman, and D. G. Altman. 2016. Statistical tests, p values, confidence intervals, and power: A guide to misinterpretations. European Journal of Epidemiology 31(4):337–350.

Hart, J. E., X. Liao, B. Hong, R. C. Puett, J. D. Yanosky, H. Suh, M. A. Kioumourtzoglou, D. Spiegelman, and F. Laden. 2015. The association of long-term exposure to PM2.5 on all-cause mortality in the Nurses’ Health Study and the impact of measurement-error correction. Environmental Health 14:38.

Hastings, P. 2024. Charge to Committee on the Mental Health Effects of Toxic Exposures Among Veterans. Presentation to the NASEM Committee on the Mental Health Effects of Toxic Exposures Among Veterans, Meeting 1 (February 22, 2024). https://www.nationalacademies.org/event/41865_02-2024_mental-health-effects-of-toxic-exposures-among-veterans-meeting-1 (accessed June 24, 2025).

Hedegaard, H., M. Schoenbaum, C. Claassen, A. Crosby, K. Holland, and S. Proescholdbell. 2018. Issues in developing a surveillance case definition for nonfatal suicide attempt and intentional self-harm using International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) coded data. National Health Statistics Report (108):1–19.

Hynes, D. M., S. Edwards, A. Hickok, M. Niederhausen, F. M. Weaver, E. Tarlov, H. Gordon, R. L. Jacob, B. Bartle, A. O’Neill, R. Young, and A. Laliberte. 2021. Veterans’ use of Veterans Health Administration primary care in an era of expanding choice. Medical Care 59:S292–S300.

IARC (International Agency for Research on Cancer). 2012. Chemical agents and related occupations. Vol. 100F. Lyon, France: International Agency for Research on Cancer.

IOM (Institute of Medicine). 2014. Chronic multisymptom illness in Gulf War veterans: Case definitions reexamined. Washington, DC: The National Academies Press.

IOM. 2016. Gulf War and health: Volume 10: Update of health effects of serving in the Gulf War, 2016. Washington, DC: The National Academies Press.

Kang, J., J. Hong, P. Esie, K. T. Bernstein, and S. Aral. 2017. An illustration of errors in using the p value to indicate clinical significance or epidemiological importance of a study finding. Sexually Transmitted Diseases 44(8):495–497.

Kioumourtzoglou, M.-A., D. Spiegelman, A. A. Szpiro, L. Sheppard, J. D. Kaufman, J. D. Yanosky, R. Williams, F. Laden, B. Hong, and H. Suh. 2014. Exposure measurement error in PM2.5 health effects studies: A pooled analysis of eight personal exposure validation studies. Environmental Health 13(1):2.

Kolivosky, J. 2024. Environmental Sampling and the Periodic Occupational and Environmental Monitoring Summary. Presentation to the NASEM Committee on the Mental Health Effects of Toxic Exposures Among Veterans, Meeting 2 (April 3, 2024). https://www.nationalacademies.org/documents/embed/link/LF2255DA3DD1C41C0A42D3BEF0989ACAECE3053A6A9B/file/D1294DE64D7A878F7E48C4152A39B1D3148C694D08CF?noSaveAs=1 (accessed June 9, 2025).

Larson, T. C., W. Kaye, P. Mehta, and D. K. Horton. 2018. Amyotrophic lateral sclerosis mortality in the United States, 2011–2014. Neuroepidemiology 51(1–2):96–103.

Leung, L. B., K. Chu, D. Rose, S. Stockdale, E. P. Post, K. B. Wells, and L. V. Rubenstein. 2022. Electronic population-based depression detection and management through universal screening in the Veterans Health Administration. JAMA Network Open 5(3):e221875.

Liu, C. F., M. L. Maciejewski, and A. E. Sales. 2005. Changes in characteristics of veterans using the VHA health care system between 1996 and 1999. Health Research Policy and Systems 3(1):5.

Mattras, T. F. 2024. Honoring Our Promise to Address Comprehensive Toxics Act of 2022—Toxic Exposure Screening. Presentation to the NASEM Committee on the Mental Health Effects of Toxic Exposures Among Veterans, Meeting 2 (April 3, 2024). https://www.nationalacademies.org/documents/embed/link/LF2255DA3DD1C41C0A42D3BEF0989ACAECE3053A6A9B/file/D7C6907DDA6B9FC9BFFAF087C086AB2085F9B2A1784E?noSaveAs=1 (accessed May 28, 2025).

Meffert, B. N., D. M. Morabito, D. A. Sawicki, C. Hausman, S. M. Southwick, R. H. Pietrzak, and A. J. Heinz. 2019. US veterans who do and do not utilize Veterans Affairs health care services: Demographic, military, medical, and psychosocial characteristics. Primary Care Companion for CNS Disorders 21(1):18m02350.

Mehta, P., J. Raymond, M. Han, R. Punjani, T. Larson, J. D. Berry, B. Rix Brooks, B. Oskarsson, S. A. Goutman, and K. Horton. 2023. A revision to the United States National ALS Registry’s algorithm to improve case-ascertainment. Amyotrophic Lateral Sclerosis and Frontotemporal Degeneration 24(3–4):230–236.

NASEM (National Academies of Sciences, Engineering, and Medicine). 2017. Assessment of the Department of Veterans Affairs Airborne Hazards and Open Burn Pit Registry. Washington, DC: The National Academies Press.

NASEM. 2018a. Progress toward transforming the Integrated Risk Information System (IRIS) program: A 2018 evaluation. Washington, DC: The National Academies Press.

NASEM. 2018b. Veterans and Agent Orange: Update 11 (2018). Washington, DC: The National Academies Press.

NASEM. 2020. Respiratory health effects of airborne hazards exposures in the Southwest Asia theater of military operations. Washington, DC: The National Academies Press.

NASEM. 2025. Veterans, prescription opioids and benzodiazepines, and mortality, 2007–2019: Three target trial emulations. Washington, DC: The National Academies Press.

NCHS (National Center for Health Statistics). 2025. The linkage of National Center for Health Statistics survey data to United States Department of Veterans Affairs administrative data: Linkage methodology and analytic considerations. Hyattsville, MD: National Center for Health Statistics.

NIMH (National Insitute of Mental Health). 2021. Attention-Deficit/Hyperactivity Disorder in Adults: What You Need to Know. https://www.nimh.nih.gov/health/publications/adhd-what-you-need-to-know (accessed July 16, 2024).

NRC (National Research Council). 2014. Review of EPA’s Integrated Risk Information System (IRIS) process. Washington, DC: The National Academies Press.

Nuyts, G. D., M. M. Elseviers, and M. E. De Broe. 1993. Healthy worker effect in a cross-sectional study of lead workers. Journal of Occupational Medicine 35(4):387–391.

Radomski, T. R., X. Zhao, E. Z. Lovelace, F. E. Sileanu, L. Rose, A. L. Schwartz, L. J. Schleiden, A. H. Oakes, A. N. Pickering, D. Yang, J. A. Hale, W. F. Gellad, M. J. Fine, and C. T. Thorpe. 2022. Use and cost of low-value health services delivered or paid for by the Veterans Health Administration. JAMA Internal Medicine 182(8):832–839.

Rasmussen, P., and C. M. Farmer. 2023. The promise and challenges of VA community care: Veterans’ issues in focus. RAND Health Quarterly 10(3):9.

Rawat, B. P. S., J. Reisman, T. K. Pogoda, W. Liu, S. Rongali, R. H. Aseltine, Jr., K. Chen, J. Tsai, D. Berlowitz, H. Yu, and K. F. Carlson. 2023. Intentional self-harm among U.S. veterans with traumatic brain injury or posttraumatic stress disorder: Retrospective cohort study from 2008 to 2017. JMIR Public Health Surveillance 9:e42803.

Rothman, K. J. 1990. No adjustments are needed for multiple comparisons. Epidemiology 1(1):43–46.

Shen, Y., A. Hendricks, F. Wang, J. Gardner, and L. E. Kazis. 2008. The impact of private insurance coverage on veterans’ use of VA care: Insurance and selection effects. Health Services Research 43(1 Pt 1):267–286.

Steele, L. 2000. Prevalence and patterns of Gulf War illness in Kansas veterans: Association of symptoms with characteristics of person, place, and time of military service. American Journal of Epidemiology 152(10):992–1002.

Tsai, J., N. P. Mota, and R. H. Pietrzak. 2015. U.S. female veterans who do and do not rely on VA health care: Needs and barriers to mental health treatment. Psychiatric Services 66(11):1200–1206.

VA (Department of Veterans Affairs). n.d. Data Governance and Analytics. https://department.va.gov/administrations-and-offices/enterprise-integration/data-governance-and-analytics/ (accessed May 27, 2025).

VA. 2018a. Profile of post-9/11 veterans: 2016. Washington, DC: National Center for Veterans Analysis and Statistics, Department of Veterans Affairs.

VA. 2018b. Profile of veterans: 2016 data from the American Community Survey. Washington, DC: National Center for Veterans Analysis and Statistics, Department of Veterans Affairs.

VA. 2024. Environmental Health Registry Evaluation for Veterans. https://www.publichealth.va.gov/exposures/benefits/registry-evaluation.asp (accessed June 18, 2025).

VA and DoD. 2021. VA/DoD clinical practice guideline for the management and rehabilitation of post-acute mild traumatic brain injury. Version 3.0—June 2021. Washington, DC: Department of Veterans Affairs and Department of Defense.

Wagner, T. H., A. Schmidt, F. Belli, M. Aouad, E. Gehlert, M. Desai, L. Graham, and L. Rose. 2024. Health insurance enrollment among US veterans, 2010-2021. JAMA Network Open 7(8):e2430205.

Wang, Z. J., P. Dhanireddy, C. Prince, M. Larsen, M. Schimpf, and G. Pearman. 2021. 2021 survey of veteran enrollees’ health and use of health care. Monterey, CA: Advanced Survey Design.

Wu, P., A. Gifford, X. Meng, X. Li, H. Campbell, T. Varley, J. Zhao, R. Carroll, L. Bastarache, J. C. Denny, E. Theodoratou, and W. Q. Wei. 2019a. Mapping ICD-10 and ICD-10-CM codes to phecodes: Workflow development and initial evaluation. JMIR Medical Informatics 7(4):e14325.

Wu, X., D. Braun, M. A. Kioumourtzoglou, C. Choirat, Q. Di, and F. Dominici. 2019b. Causal inference in the context of an error prone exposure: Air pollution and mortality. Annals of Applied Statistics 13(1):520–547.

Yaddanapudi, L. N. 2016. The American Statistical Association statement on p-values explained. Journal of Anaesthesiology and Clinical Pharmacology 32(4):421–423.

Yland, J. J., A. K. Wesselink, T. L. Lash, and M. P. Fox. 2022. Misconceptions about the direction of bias from nondifferential misclassification. American Journal of Epidemiology 191(8):1485–1495.