Epidemiology
· 2026 May · PMID 42148595
·
Publisher ↗
BACKGROUND: Data from validation substudies help quantify bias due to misclassification of an analytic variable by estimating classification parameters. Studies with long enrollment or follow-up may be susceptible to cha...BACKGROUND: Data from validation substudies help quantify bias due to misclassification of an analytic variable by estimating classification parameters. Studies with long enrollment or follow-up may be susceptible to changes in classification parameters over time, which would have important implications for validation substudy design and quantitative bias analysis. Herein, we provide guidance for sampling validation data to account for time trends in classification parameters. METHODS: Using a simulated cohort of 10,000 observations, we induced exposure misclassification under three scenarios: absence of a time trend, but expected change in exposure prevalence; linear time trends in exposure misclassification; and logarithmic time trends in exposure misclassification. Validation sampling was conducted at the beginning, middle, and end of follow-up. These validation data were then used to impute positive and negative predictive values for the entire cohort over the study period as a function of time. We compared these imputed values with the true values. RESULTS: We demonstrated that, in the presence of a time trend, purposeful sampling allows for estimation of the changing positive and negative predictive values over the course of the study. Estimation of predictive values was accurate under both the linear and logarithmic scenarios, and was closest to the truth when a large proportion of exposure/outcome strata was sampled for the validation substudy. CONCLUSIONS: When a time trend in classification parameters exists, designs that allow estimation of time-varying predictive values should be used instead of conventional validation study designs that estimate a single summary classification parameter over the study period.
Bórquez I, Allen B, Basaraba C
… +4 more, Renson A, Moore B, Marshall BDL, Cerdá M
Epidemiology
· 2026 May · PMID 42138361
·
Publisher ↗
BACKGROUND: Overdose prevention centers (OPCs) may reduce public drug use and, with it, policing of people who use drugs in the communities surrounding these sites. We applied an augmented synthetic control method to ass...BACKGROUND: Overdose prevention centers (OPCs) may reduce public drug use and, with it, policing of people who use drugs in the communities surrounding these sites. We applied an augmented synthetic control method to assess changes in pedestrian stops before and after the November 2021 opening of two OPCs (Washington Heights and East Harlem) in New York City (NYC). METHODS: We retrieved pedestrian stop information from the New York Police Department's (NYPD) Stop, Question, and Frisk, program from January 2017 to December 2024, and created bimonthly averages using five- and ten-minute walking buffers surrounding the OPCs and 57 donor sites (syringe service and opioid treatment programs) as outcomes. Covariates were derived from American Community Survey, NYPD Calls for Services, and SafeGraph pedestrian mobility estimates. RESULTS: The opening of the Washington Heights OPC was associated with a reduction of 2.8 bimonthly average pedestrian stops in the post-intervention period when using five-minute walking buffers, although results were compatible with increases and reductions (95%CI=-9.4, 4.0). For ten-minute walking buffers, results were compatible with a wide range of reductions (ATT=-9.2 [95%CI=-18.3, -1.3]). East Harlem OPC showed larger point estimates when examining both distances (ATT=-8.4 [95%CI=-12.2, -4.5] and ATT=-13.7 [95%CI=-22.1, -4.2] with five- and ten-minute walking buffers, respectively). For both sites, permutation tests suggested that these reductions fell within the range of possible donor-unit placebo effects. CONCLUSIONS: Our study shows limited evidence of an effect of NYC's first two OPCs on pedestrian stops in their immediate vicinity, with a potential decrease concentrated in the first two years at the East Harlem location.
Cole SR, Zivich PN, Shook-Sa BE
… +3 more, Richardson DB, Hudgens MG, Edwards JK
Epidemiology
· 2026 May · PMID 42138357
·
Full text
Epidemiology stands to benefit greatly from combining data sources with complementary strengths. We propose a study design and estimators to combine information from multiple observational studies to simultaneously addre...Epidemiology stands to benefit greatly from combining data sources with complementary strengths. We propose a study design and estimators to combine information from multiple observational studies to simultaneously address confounding and outcome measurement error. Using inverse probability weighted, g computation, and augmented inverse probability weighted estimators, we show how to combine information from two studies wherein the first study is subject to outcome misclassification (but has adequate confounder control) and the second study has gold-standard outcomes (but inadequate confounder control). Monte Carlo experiments demonstrate that the proposed estimators remove both confounding and measurement biases and provide appropriate 95% confidence interval coverage, while standard analyses fall short. Fusion designs offer a principled approach to combine data from multiple sources to address multiple biases in epidemiologic research.
Young JC, Dahabreh IJ, James S
… +6 more, Erlinge D, Fröbert O, Berglund A, Rylance R, Hernán MA, Matthews AA
Epidemiology
· 2026 May · PMID 42081803
·
Publisher ↗
BACKGROUND: When decision makers use evidence from a randomized trial to inform population-level decisions, the target population they envision rarely aligns with the population of individuals who enrolled in the trial....BACKGROUND: When decision makers use evidence from a randomized trial to inform population-level decisions, the target population they envision rarely aligns with the population of individuals who enrolled in the trial. Here, we extend inferences from the VALIDATE-SWEDEHEART randomized trial (hereafter, the index trial), which compared the effects of bivalirudin and heparin during percutaneous coronary intervention on the risk of death, reinfarction, and bleeding, to two clinically relevant target populations: first, the trial-eligible population of individuals eligible for the index trial regardless of enrollment, and second, the treatment-candidate population of individuals who are considered candidates for bivalirudin and heparin under routine care, regardless of eligibility for the index trial. METHODS: Using data from the index trial, we fit logistic regression models for the outcome at 180 days in each group based on assigned treatment. We then standardized risk estimates to the baseline covariate distribution of the trial-eligible and treatment-candidate target populations, which were characterized using data from Swedish healthcare registries. RESULTS: The estimated risk difference comparing bivalirudin vs. heparin was -1.1% (-3.1%, 0.9%) in the trial-eligible population and -1.0% (-3.0%, 1.0%) in the treatment-candidate population. The corresponding risk ratios were 0.92 (0.80, 1.07) and 0.93 (0.80, 1.07), respectively, aligning closely with estimates from the index trial. Absolute risks in each treatment group were, however, between 0.8 and 1.2 percentage points higher in comparison with the index trial. CONCLUSIONS: Estimated risk ratios for the broader trial-eligible and treatment-candidate populations generally align with the findings from the index trial. While trials provide essential evidence for healthcare, questions often arise about wider, clinically relevant populations beyond the population of trial participants. By leveraging data from trials and observational data sources, we can attempt to address questions in these wider target populations.
Epidemiology
· 2026 May · PMID 42068138
·
Publisher ↗
BACKGROUND: Depression causes a substantial burden on a person's health and well-being. However, evidence is limited regarding whether depression of one person in a marital relationship may affect the other partner's hea...BACKGROUND: Depression causes a substantial burden on a person's health and well-being. However, evidence is limited regarding whether depression of one person in a marital relationship may affect the other partner's health. We assessed whether depression in one partner within a married couple might contribute to the other partner's mortality hazard, also considering whether only one or both partners have depression. METHODS: We examined a nationally representative sample comprising 8,442 older US adults within 4,225 couples from the Health and Retirement Study. Using the Center for Epidemiologic Studies Depression Scale, we characterized depression status as follows: i) no depression in either partner, ii) depression in respondent only, iii) depression in spouse only, and iv) depression in respondent and their spouse. Associations between couples' depression status on individuals' mortality hazards over 11 years were estimated using Cox proportional hazards models adjusted for 27 characteristics of the individual, their spouse, and their household. RESULTS: We observed a higher mortality hazard when only the respondent had depression (hazard ratio (HR): 1.44 [95% confidence interval (CI): 1.26, 1.66]), as well as a modestly elevated hazard when only their spouse exhibited depression (HR: 1.18 [95% CI: 1.01, 1.39]). When both partners had depression, we observed a jointly elevated mortality hazard (HR: 1.53 [95% CI: 1.26, 1.86]). CONCLUSIONS: Results suggest the harmful effects of depression could extend beyond the individual to spouses' physical health. Future studies on health effects of depression should incorporate familial contexts.
Song S, Hitchings MDT, Yang Y
… +2 more, Longini IM, N3C consortium
Epidemiology
· 2026 Jul · PMID 42017631
·
Full text
The test-negative design (TND) has become a widely used observational study design for evaluating vaccine effectiveness, especially during the COVID-19 pandemic. Traditionally, TND has been viewed as a variant of the cas...The test-negative design (TND) has become a widely used observational study design for evaluating vaccine effectiveness, especially during the COVID-19 pandemic. Traditionally, TND has been viewed as a variant of the case-control study and largely limited to use with logistic regression models. In this paper, we first establish that TND can be framed as a special case of a cohort study, thereby opening the door to a wider range of analytical approaches. We then introduce the Prentice, Williams, and Peterson gap-time (PWP-GT) frailty model as a novel method for analyzing TND data, accounting for recurrent infections and time-dependent vaccination status. Through extensive simulation studies, we demonstrate that the proposed model outperforms conventional models commonly applied in TND-based vaccine effectiveness studies. Finally, we apply our method to data from the National COVID Cohort Collaborative, estimating the effectiveness of full and booster doses of Pfizer's COVID-19 vaccines against both initial infection and reinfection during the Omicron variant circulation period in a real-world setting.
Pamplin Ii JR, Wheeler-Martin K, Perry A
… +8 more, Mannes Z, Krawczyk N, Crystal S, Hasin DS, Martins SS, Shroff R, Cerdá M, Neill DB
Epidemiology
· 2026 Apr · PMID 41979535
·
Full text
BACKGROUND: Overdose rates in the U.S. rose dramatically during the COVID-19 pandemic. Well-documented racial and sociodemographic inequities in the impact of the pandemic suggest the potential for similar inequities for...BACKGROUND: Overdose rates in the U.S. rose dramatically during the COVID-19 pandemic. Well-documented racial and sociodemographic inequities in the impact of the pandemic suggest the potential for similar inequities for overdose. Our objective was to identify subgroups of New York State Medicaid enrollees who experienced the greatest increases in non-fatal opioid overdose risk following onset of the COVID-19 pandemic. METHODS: Data are from a retrospective cohort of 1,021,889 people enrolled in New York State Medicaid from 2019-2020. To identify subgroups with the greatest increased risk of non-fatal overdose following onset of the COVID-19 pandemic, we used Heterogeneous Treatment Effect (HTE)-Scan, a novel machine learning method developed for accurate and computationally efficient discovery of heterogeneous treatment effects in complex data. RESULTS: In the total sample, risk of non-fatal opioid overdose increased 22% after onset of the pandemic. We also identified two subgroups with elevated risk relative to the total sample: subgroup 1 (Black and Hispanic males aged 45-64 years old with no baseline documentation of opioid use disorder (OUD); N = 53,065) and subgroup 2 (people aged 45-64 years old with documented aged/blind/disabled status and no baseline documentation of OUD; N = 73,694). These subgroups experienced a 54% and 57% increase in non-fatal overdose risk, respectively. CONCLUSIONS: We estimated heterogeneous effects of onset of the COVID-19 pandemic on non-fatal overdose, with elevated risks estimated for older working-aged, structurally disadvantaged adults without documented OUD. These findings illustrate the importance of structural factors in driving heterogeneous risk of overdose following complex social events.
BACKGROUND: To minimize the health impacts of power outages, which have steadily increased with climate change and aging infrastructure, it is crucial to recognize their range of effects and identify the most vulnerable...BACKGROUND: To minimize the health impacts of power outages, which have steadily increased with climate change and aging infrastructure, it is crucial to recognize their range of effects and identify the most vulnerable individuals and communities. Here, we investigate the health consequences of outages among those with asthma residing in New York City. METHODS: We combined administrative emergency department (ED) visit data from 2018 to 2022 with two granular power outage data sources: the New York State Department of Public Service (locality-level) and the New York City Housing Authority (NYCHA, building-level). We applied an augmented synthetic control approach to four large-scale outages in the New York State Department of Public Service data to assess their association with asthma-related ED visits. We used a case-crossover design to assess the impact of same-day, building-level outage exposure in the NYCHA data on asthma-related ED visits. RESULTS: A large, localized outage on 21 July 2019 was associated with 0.20 (95% confidence interval: 0.00, 0.34) per 1000 sub-daily increases in asthma-related ED visits the same evening of the outage start. For NYCHA residents, in the summer months, outage days were associated with elevated odds of asthma-related ED visits (2.23, 95% confidence interval: 0.92, 5.37), and these effects were particularly relevant for children and for lag days 0-1. CONCLUSIONS: Climate change impacts communities through severe weather events and their subsequent disruptions to critical infrastructure, such as electricity service. We add to the limited evidence that these disruptions carry health risks, in this case acute care asthma visits.
Wang SV, Hahn G, Sushama Kattinakere S
… +13 more, Mahesri M, Pillai HS, Aldis R, Lii J, Dutcher SK, Eniafe R, Jones JT, Kim K, He J, Lee H, Toh S, Desai RJ, Yang J
BACKGROUND: One of the ways to enhance analyses conducted with large claims databases is by validating the measurement characteristics of the code-based algorithms used to identify health outcomes or other key study para...BACKGROUND: One of the ways to enhance analyses conducted with large claims databases is by validating the measurement characteristics of the code-based algorithms used to identify health outcomes or other key study parameters of interest. These metrics can be used in quantitative bias analyses to assess the robustness of results for an inferential study, given potential bias from outcome misclassification. However, performing this validation through manual chart review of free-text notes from linked electronic health records requires extensive time and resource allocation. METHODS: We describe an expedited process for validating code-based algorithms that introduces efficiency using two distinct mechanisms: (1) use of natural language processing to reduce the time spent by human reviewers to review each chart, and (2) a multiwave adaptive sampling approach with predefined criteria to stop the validation study once performance characteristics are identified with sufficient precision. We illustrate this process in a case study that validates the performance of a claims-based outcome algorithm for intentional self-harm in patients with obesity. RESULTS: We empirically demonstrate that the natural language processing-assisted annotation process reduced the time spent on review per chart by 40%, and the use of the predefined stopping rule with multiwave samples would have prevented review of 77% of patient charts with limited compromise to the precision of performance characteristics. CONCLUSION: This approach could facilitate more routine validation of code-based algorithms used to define key study parameters, ultimately enhancing understanding of the reliability of findings derived from database studies.
The possibility of unmeasured confounding is one of the main limitations for causal inference from observational studies. There are different methods for (partially) empirically assessing the plausibility of unconfounded...The possibility of unmeasured confounding is one of the main limitations for causal inference from observational studies. There are different methods for (partially) empirically assessing the plausibility of unconfoundedness. However, most currently available methods require (at least partial) assumptions about the confounding structure, which may be difficult to know in practice. In this paper, we describe a simple strategy for empirically assessing the plausibility of conditional unconfoundedness (i.e., whether the candidate adjustment set of covariates suffices for confounding adjustment), which does not require any explicit assumptions about the confounding structure, relying instead on assumptions related to temporal ordering between covariates, exposure, and outcome (which can be guaranteed by design) and selection into the study. The proposed method essentially relies on testing the association between a subset of the covariates included in the adjustment set (those associated with the exposure, given all other covariates) and the outcome conditional on the remaining covariates and the exposure. We describe the assumptions underlying the method, provide proofs, use simulations to corroborate the theory, and illustrate the method with an applied example assessing the causal effect of delivery mode and intelligence quotient measured in adulthood using data from the 1982 Pelotas (Brazil) birth cohort. We also discuss the implications of measurement error and some important limitations of the suggested approach.
BACKGROUND: Observational studies often rely on nonprobability samples that may not represent the target population, limiting the generalizability of findings. Health examination data from the Korean National Health Insu...BACKGROUND: Observational studies often rely on nonprobability samples that may not represent the target population, limiting the generalizability of findings. Health examination data from the Korean National Health Insurance Service (NHIS) faces this challenge, as voluntary participation introduces sampling bias. Poststratification offers a potential solution by reweighting samples to match population distributions. METHODS: Using the NHIS-National Sample Cohort, we demonstrated three poststratification approaches-simple poststratification, raking, and multilevel regression with poststratification-to estimate obesity prevalence among adults 20-39 years of age in 2019. The Korea National Health and Nutrition Examination Survey served as the reference standard for population-level estimates. We compared these methods with inverse probability sampling weights. Additionally, we applied poststratification to evaluate the accuracy of self-reported disease history for hypertension, diabetes, dyslipidemia, and stroke. RESULTS: Crude obesity prevalence from NHIS-National Sample Cohort was 36.3% (95% confidence interval: 36.0, 36.6), substantially higher than the Korea National Health and Nutrition Examination Survey reference of 31.4% (28.8, 33.9). Simple poststratification using age and sex reduced this estimate to 33.9% (33.6, 34.2), ranking with additional smoking and alcohol variables yielded 34.7% (33.9, 35.5), and multilevel regression with poststratification incorporating region produced 34.2% (33.3, 35.0). Inverse probability sampling weights yielded comparable results (33.9%-34.1%). For self-reported disease history, poststratification consistently produced modest decreases in sensitivity estimates, suggesting that health examination participants may report disease history more accurately than the general population. CONCLUSION: Poststratification provides a principled framework for improving population-level inferences from nonprobability samples. These methods warrant broader application in epidemiological research using administrative and electronic health record data.
Meche A, Boonyasai RT, Hsu YJ
… +6 more, Greer RC, Mehta HB, Alexander GC, Segal JB, Cooper LA, Jackson JW
Epidemiology
· 2026 May · PMID 41921520
·
Full text
BACKGROUND: This tutorial paper demonstrates the application of a conceptual model to measure racial and ethnic disparities in treatment intensification among adults with hypertension consistent with the Institute of Med...BACKGROUND: This tutorial paper demonstrates the application of a conceptual model to measure racial and ethnic disparities in treatment intensification among adults with hypertension consistent with the Institute of Medicine (IOM) definition. The IOM defines disparity as differences in health care quality that are not due to access-related factors, clinical needs, preferences, and appropriateness of intervention (referred to as allowable covariates). METHODS: We used a conceptual model called the Target Study to estimate annual disparities in the probability of treatment intensification between Black and White patients seen at primary care clinics within a large healthcare system in the Mid-Atlantic region from 2018 to 2022, using electronic medical record data. We emulated the specified target study through an appropriate study design and the use of inverse probability weighting to balance allowable covariates while measuring disparities. RESULTS: Unadjusted analyses showed a higher percentage of treatment intensification for Black patients compared with White patients, with annual differences ranging from 2 to 4 percentage points. For example, in 2020, the unadjusted difference was three percentage points (95% confidence interval: 1%, 5%). After adjusting for allowable covariates via the emulated target study, Black patients had consistently lower percentages of treatment intensification by 3-4 percentage points each year. For example, in 2020, the adjusted disparity was -3 percentage points (95% confidence interval: -4%, -1%). CONCLUSIONS: The Target Study can implement the IOM definition for measuring disparities in health care. Future research should explore its application across different populations to better understand and address health disparities.
Epidemiology
· 2026 May · PMID 41921519
·
Full text
For a series of indicators used to assess psychosocial constructs, we propose reporting new types of correlation matrices to gain greater insight into the relation of the indicators with one another. What we define as th...For a series of indicators used to assess psychosocial constructs, we propose reporting new types of correlation matrices to gain greater insight into the relation of the indicators with one another. What we define as the observed residual correlation (ORC) matrix can give insight as to whether, when a given indicator is above the indicator-average scores across all indicators for that individual, what other indicators might be anticipated to be above that individual's average score as well. What we define as the relative excess correlation (REC) matrix can give insight, for each pair of indicators, whether the strength of that particular correlation is above or below what might have been anticipated based on the correlation of each of those two indicators with all of the others. The ORC and REC matrices will, generally, have numerous negative entries even if all of the raw correlations between each pair of indicators are positive. We discuss the properties of, and the relations between, these correlation matrices, and their analogues for covariances. The positive deviations of the REC matrix entries from zero also can help identify clusters of indicators that are more strongly related to one another, providing insights somewhat analogous to factor analysis, but without the need for decisions concerning rotations or the number of factors. However, the ORC and REC matrices can also be used purely descriptively to provide insights into understanding the relation of indicators with one another.
Epidemiology
· 2026 May · PMID 41921518
·
Publisher ↗
Although data may capture continuous event times or event times with high resolution (e.g., day), some statistical analyses require the discretization of time into intervals and assigning each event (i.e., outcome or los...Although data may capture continuous event times or event times with high resolution (e.g., day), some statistical analyses require the discretization of time into intervals and assigning each event (i.e., outcome or loss to follow-up [LTFU]) to the start or end of an interval. First, using a simulated example, we showed that outcomes should be assigned to the end of the interval. Next, we considered four approaches for assigning LTFU events in a simulated example and in 20 real datasets. Comparing the resulting cumulative risk curves with the curve using continuous time, one approach always had the least error: assigning LTFU to the start or end of the interval, depending on which was closest to the continuous event time. This approach was superior to always censoring at the beginning or end of the interval.
Hernán MA, Dickerman BA, Swanson SA
… +1 more, Dahabreh IJ
Epidemiology
· 2026 May · PMID 41921517
·
Publisher ↗
Descriptions of the target trial framework often appear to assume that investigators knew the protocol of the target trial from the start of the investigation. In practical applications, however, the protocol of the targ...Descriptions of the target trial framework often appear to assume that investigators knew the protocol of the target trial from the start of the investigation. In practical applications, however, the protocol of the target trial is constrained by the observational data that investigators have access to. When using preexisting observational databases, the target trial protocol typically needs to be iteratively developed as investigators learn about the data. Here we examine the process by which investigators adapt their original causal question, operationalized into the protocol of a hypothetical trial (the index trial), to the available data. This adaptive process results in a final causal question, operationalized into the protocol of another hypothetical trial (the target trial), that depends both on the original causal question and on the available data. As a result, prespecification of an emulatable target trial protocol is not generally possible because adaptations are expected after inspecting the data. The adaptive nature of the specification of the target trial protocol raises important questions about the types of data examinations that are permissible to guide the adaptations, the procedures for transparent reporting of each adaptation and its rationale, and the possibility of prespecifying the rules that will govern the investigators' decisions to adapt the protocol to the data.
You X, Zhu H, Jiang M
… +21 more, Gu Y, Jiang Y, Jiang T, Qin R, Lv H, Liu C, Xu X, Liu Y, Sun T, Xu B, Chen J, Jiang Y, Liu X, Zhou K, Jin G, Ma H, Hu Z, Lin Y, Zong Y, Liu H, Du J
BACKGROUNDS: The impact of maternal gestational diabetes mellitus (GDM) and its glycemic subtypes on the early visual development of infants remains unclear. METHODS: Based on the Jiangsu Birth Cohort, we included 2139 i...BACKGROUNDS: The impact of maternal gestational diabetes mellitus (GDM) and its glycemic subtypes on the early visual development of infants remains unclear. METHODS: Based on the Jiangsu Birth Cohort, we included 2139 infants born to 2041 mothers recruited from 2014 to 2018. Maternal GDM was categorized into three categories based on oral glucose tolerance test results: impaired fasting glucose, impaired glucose tolerance, and comorbid impaired fasting glucose and impaired glucose tolerance. Infants' grating visual acuity was measured at 1 year of age using Teller Acuity Cards II. We used generalized estimating equation models to examine the association between maternal GDM and infant visual acuity, accounting for the dependence of twin observations. We then conducted a metabolomics analysis to explore the metabolic profiles associated with GDM and infant visual acuity. RESULTS: Prenatal GDM exposure was associated with a 70% increase in risk of abnormal visual acuity in infants (relative risks [RR] = 1.7; 95% confidence interval [CI] = 1.2, 2.3). The increased risks were more pronounced among infants in the comorbid impaired fasting glucose and impaired glucose tolerance group (RR = 3.2; 95% CI = 1.4, 7.7), followed by the impaired glucose tolerance group (RR = 1.8; 95% CI = 1.1, 3.0) and the impaired fasting glucose group (RR = 1.6; 95% CI = 0.7, 3.6), compared with unexposed infants. The metabolomics analysis suggested glycine, serine, and threonine metabolism as an enriched pathway and identified creatine as a metabolite of interest. CONCLUSIONS: Maternal GDM is associated with an increased risk of abnormal grating visual acuity in infants at 1 year of age. Glycine, serine, and threonine metabolism pathways may contribute to this association.
BACKGROUND: Although death certificates are widely used to study mortality in the general population, their use to study mortality among individuals with military service has been limited in part by a lack of information...BACKGROUND: Although death certificates are widely used to study mortality in the general population, their use to study mortality among individuals with military service has been limited in part by a lack of information on the validity of military service reported on death certificates. Our objective was to estimate bias parameters for the misclassification of military service to facilitate quantitative bias analysis in studies of mortality in military populations. METHODS: We included 2014-2021 death certificates from decedents aged 18-64 years who resided and died in Alabama, Michigan, Minnesota, Montana, or Oregon. Death certificates were linked to military service records by the U.S. Defense Manpower Data Center using Social Security numbers. Military service was defined as any service in the Regular, Reserve, or National Guard components of the U.S. military in Defense Manpower Data Center records. We estimated sensitivity, specificity, and predictive values for military service reported on death certificates stratified by demographics, military characteristics, and manner of death. RESULTS: Among the 467,075 death certificates included, 9.3% indicated military service; however, military records showed that 10.9% had served (sensitivity 81.5%, 95% confidence interval [CI] = 81.2%, 81.8%; specificity 99.5%, 95% CI = 99.5%, 99.5%). Sensitivity was lower among female (72.3%) compared with male (82.2%) servicemembers and was particularly low for those who had never served on active duty (63.5%). CONCLUSIONS: Military service was underreported on death certificates, especially for female servicemembers and those without active-duty service. The bias parameters we estimated can be used to account for this misclassification when analyzing death certificates.