Propensity score-based sensitivity analysis method for uncontrolled confounding

Estimating generalized propensity scores with survey and attrition weighted data

Prior work in causal inference has shown that using survey sampling weights in the propensity score estimation stage and the outcome model stage for binary treatments can result in a more robust estimator of the effect of the binary treatment being analyzed. However, to date, extending this work to continuous treatments and exposures has not been explored nor has consideration been given for how to handle attrition weights in the propensity score model. Nonetheless, generalized propensity score (GPS) analyses are being used for estimating continuous treatment effects on outcomes when researchers have observational data, and those data sets often have survey or attrition weights that need to be accounted for in the analysis. Here, we extend prior work and show with analytic results that using survey sampling or attrition weights in the GPS estimation stage and the outcome model stage for continuous treatments can result in a more robust estimator than one that does not. Simulation study results show that, although using weights in both estimation stages is sufficient for robust estimation, it is not necessary and unbiased estimation is possible in some cases under various approaches to using weights in estimation. Analysts do not know if the conditions of our simulation studies hold, so use of weights in both estimation stages might provide insurance for reducing potential bias. We discuss the implications of our results in the context of an empirical example.

Unreliable Continuous Treatment Indicators in Propensity Score Analysis

Propensity score analyses (PSA) of continuous treatments often operationalize the treatment as a multi-indicator composite, and its composite reliability is unreported. Latent variables or factor scores accounting for this unreliability are seldom used as alternatives to composites. This study examines the effects of the unreliability of indicators of a latent treatment in PSA using the generalized propensity score (GPS). A Monte Carlo simulation study was conducted varying composite reliability, continuous treatment representation, variability of factor loadings, sample size, and number of treatment indicators to assess whether Average Treatment Effect (ATE) estimates differed in their relative bias, Root Mean Squared Error, and coverage rates. Results indicate that low composite reliability leads to underestimation of the ATE of latent continuous treatments, while the number of treatment indicators and variability of factor loadings show little effect on ATE estimates, after controlling for overall composite reliability. The results also show that, in correctly specified GPS models, the effects of low composite reliability can be somewhat ameliorated by using factor scores that were estimated including covariates. An illustrative example is provided using survey data to estimate the effect of teacher adoption of a workbook related to a virtual learning environment in the classroom.

Estimating the impact of birth interval on under-five mortality in east african countries: a propensity score matching analysis

BACKGROUND

Under-five mortality remains a global public health concern, particularly in East African countries. Short birth interval is highly associated with under-five mortality, and birth spacing has a significant effect on a child's likelihood of survival. The association between short birth intervals and under-five mortality was demonstrated by numerous observational studies. However, the effect of short birth intervals on under-five mortality has not been investigated yet. Therefore, this study aimed to investigate the impact of short birth intervals on under-five mortality in East Africa using Propensity Matched Analysis.

METHODS

A secondary data analysis was conducted based on the most recent Demographic and Health Survey (DHS) data of 12 East African countries. A total weighted sample of 105,662 live births was considered for this study. A PSM analysis was carried out to evaluate the effect of short birth intervals on under-five mortality. Under-five mortality was the outcome variable, while the short birth interval was considered a treatment variable. To determine the Average Treatment Effect on the population (ATE), Average Treatment Effect on the treated (ATT), and Average Treatment Effect on the untreated (ATU), we performed PSM analysis with a logit-based model using the psmatch2 ate STATA function. The quality of matching was assessed statistically and graphically. The common support assumption was checked and fulfilled. We have employed Mantel-Haenszel bounds to examine whether the result would be free from hidden bias or not.

RESULTS

The prevalence of short birth intervals in East Africa was 44%. The under-five mortality rate among mothers who had optimal birth intervals was 39.9 (95% CI: 38.3, 41.5) per 1000 live births while it was 60.6 (95% CI: 58.5, 62.8) per 1000 live births among mothers who had a short birth intervals. Propensity score matching split births from mothers into treatment and control groups based on the preceding birth interval. In the PSM analysis, the ATT values in the treated and control groups were 6.09% and 3.97%, respectively, showed under-five mortality among births to mothers with short birth intervals was 2.17% higher than births to mothers who had an optimal birth interval. The ATU values in the intervention and control groups were 3.90% and 6.06%, respectively, indicating that for births from women who had an optimal birth interval, the chance of dying within five years would increase by 2.17% if they were born to mother with short birth interval. The final ATE estimate was 2.14% among the population. After matching, there was no significant difference in baseline characteristics between the treated and control groups (p-value > 0.05), which indicates the quality of matching was good.

CONCLUSIONS

We conclude that enhancing mothers to have optimal birth spacing is likely to be an effective approach to reducing the incidence of under-five mortality. Our findings suggest that births to mothers with short birth intervals have an increased risk of death in the first five years of life than births to mothers who had an optimal birth interval. Therefore, public health programs should enhance interventions targeting improving birth spacing to reduce the incidence of under-five mortality in low-and middle-income countries like East African countries. Moreover, to achieve a significant reduction in the under-five mortality rate, interventions that encourage birth spacing should be considered. This will improve child survival and help in attaining Sustainable Development Goal targets in East African countries.

Assessing the effects of duration of birth interval on adverse pregnancy outcomes in sub-Saharan Africa: a propensity score-matched analysis

OBJECTIVES

Unlike high-income countries, sub-Saharan African countries have the highest burden of adverse pregnancy outcomes such as abortion, stillbirth, low birth weight and preterm births. The WHO set optimal birth spacing as a key strategy to improve pregnancy outcomes. Estimating the impact of short and long birth intervals on adverse pregnancy outcomes based on an observational study like the Demographic and Health Survey (DHS) is prone to selection bias. Therefore, we used the propensity score-matched (PSM) analysis to estimate the actual impact of short and long birth intervals on adverse pregnancy outcomes.

DESIGN

A community-based cross-sectional study was conducted based on the DHS data.

SETTING

We used the recent DHS data of 36 sub-Saharan African countries.

PARTICIPANTS

A total of 302 580 pregnant women for stillbirth and abortion, 153 431 for birth weight and 115 556 births for preterm births were considered.

PRIMARY OUTCOME MEASURES

To estimate the impact of duration of birth interval (short/long) on adverse pregnancy outcomes, we used PSM analysis with logit model using psmatch2 ate STATA command to find average treatment effect on the population (ATE), treated and untreated. The quality of matching was assessed statistically and graphically. Sensitivity analysis was conducted to test the robustness of the PSM estimates using the Mantel-Haenszel test statistic.

RESULTS

The prevalence of short and long birth intervals in sub-Saharan Africa was 46.85% and 13.61%, respectively. The prevalence rates of abortion, stillbirth, low birth weight, macrosomia, and preterm births were 6.11%, 0.84%, 9.63%, 9.04%, and 4.87%, respectively. In the PSM analysis, the differences in ATE of short birth intervals on abortion, stillbirth, low birth weight, and preterm births were 0.5%, 0.1%, 0.2%, and 0.4%, respectively, and -2.6% for macrosomia. The difference in ATE among the treated group was 1%, 1%, and 1.1% increased risk of abortion, low birth weight, and preterm births, respectively, while there was no difference in risk of stillbirth between the treated and control groups. The ATEs of long birth intervals on abortion, stillbirth, low birth weight, macrosomia and preterm births were 1.3%, 0.4%, 1.0%, 3.4%, and 0.2%, respectively. The ATE on the treated group had 0.9%, 0.4%, 2.4%, 2.8%, and 0.2% increased risk of abortion, stillbirth, low birth weight, macrosomia, and preterm births, respectively. The estimates were insensitive to hidden bias and had a good quality of matching.

CONCLUSION

Short and long birth intervals had a significant positive impact on stillbirth, abortion, low birth weight, macrosomia and preterm births after matching treated and control groups by observed variables. These findings highlighted maternal and newborn healthcare programmes and policies to empower reproductive-aged women to exercise optimal birth spacing to reduce the incidence of stillbirth, abortion, low birth weight, macrosomia and preterm births.

Intravesical BCG and Incidence of Alzheimer Disease in Patients With Bladder Cancer: Results From an Administrative Dataset

INTRODUCTION

Alzheimer disease (AD) is a common neurodegenerative disease, and immunomodulation offers treatment opportunities. Preclinical data suggest that intravesical Bacillus Calmette-Guerin (BCG) treatment could delay AD development. We investigated this relationship in a population-based cancer database.

SAMPLE AND METHODS

We queried the Surveillance, Epidemiology, and End Results-Medicare database for patients with high-risk nonmuscle-invasive bladder cancer (hrNMIBC). BCG dosage and subsequent Alzheimer diagnosis were collected through ICD-9/10 codes. Multivariable Cox regression was performed to assess the association between BCG therapy and subsequent Alzheimer diagnosis.

RESULTS

We identified 26,584 hrNMIBC patients; 51% received BCG and 8.3% were diagnosed with Alzheimer. BCG exposure was significantly associated with lower Alzheimer occurrence (hazard ratio: 0.73, P <0.05), which was dose-dependent. Increasing age, female sex, Black race, and increasing comorbidity index were significantly associated with a greater risk of subsequent Alzheimer diagnosis.

DISCUSSION

Treatment with intravesical BCG among patients with hrNMIBC was associated with a significantly lower risk for subsequent Alzheimer diagnosis, which seemed dose-dependent.

Comparative effectiveness of hysteroscopic and laparoscopic sterilization for women: a retrospective cohort study

OBJECTIVE

To compare real-world effectiveness of hysteroscopic to laparoscopic sterilization.

DESIGN

Retrospective cohort of Medicaid claims for hysteroscopic or laparoscopic sterilization procedures performed in California, 2008-2014. After excluding postpartum procedures, we applied log-linear (Poisson) event-history regression models for clustered person-period data, weighted for propensity to receive either sterilization procedures, and adjusted for sociodemographic and clinical variables to examine the poststerilization pregnancy rates.

SETTING

Clinics, hospitals.

PATIENT(S)

Women aged 18-50 years with Medicaid claims between January 1, 2008, and August 31, 2014.

INTERVENTION(S)

Hysteroscopic or laparoscopic sterilization procedure.

MAIN OUTCOME MEASURE(S)

Poststerilization pregnancy measured by pregnancy-related claims.

RESULT(S)

Among women with hysteroscopic (n = 5,906) or laparoscopic (n = 23,965) sterilization, poststerilization pregnancy claims were identified for 4.74% of women after hysteroscopic sterilization and 5.57% after laparoscopic sterilization. The pregnancy rates decreased over time after either procedure. Twelve months after the procedure, the crude incidence of pregnancy claims was higher for hysteroscopic sterilization than for laparoscopic sterilization (3.26 vs. 2.61 per 100 woman-years), but the propensity-weighted adjusted incidence rate ratio was 1.06 (95% confidence interval [CI], 0.85-1.26). Between 13 and 24 months after the procedure, there were fewer pregnancies for women after hysteroscopic sterilizations than for those after laparoscopic sterilizations (adjusted incidence rate ratio, 0.63 [95% CI, 0.45-0.88]), with no statistically significant differences in later years. The cumulative pregnancy rates 5 years after sterilization were lower with hysteroscopic sterilization than with laparoscopic sterilization (6.26 vs. 7.22 per 100 woman-years; propensity-weighted, adjusted risk ratio, 0.76 [95% CI, 0.62-0.90]). The poststerilization pregnancy rates varied by age and race/ethnicity.

CONCLUSION(S)

The pregnancy rates after female sterilization are higher than expected, whether performed hysteroscopically or laparoscopically. These findings are reassuring that the effectiveness of hysteroscopic sterilization was not inferior to laparoscopic sterilization.

CLINICAL TRIAL REGISTRATION NUMBER

NCT03438682.

A FLEXIBLE SENSITIVITY ANALYSIS APPROACH FOR UNMEASURED CONFOUNDING WITH MULTIPLE TREATMENTS AND A BINARY OUTCOME WITH APPLICATION TO SEER-MEDICARE LUNG CANCER DATA

In the absence of a randomized experiment, a key assumption for drawing causal inference about treatment effects is the ignorable treatment assignment. Violations of the ignorability assumption may lead to biased treatment effect estimates. Sensitivity analysis helps gauge how causal conclusions will be altered in response to the potential magnitude of departure from the ignorability assumption. However, sensitivity analysis approaches for unmeasured confounding in the context of multiple treatments and binary outcomes are scarce. We propose a flexible Monte Carlo sensitivity analysis approach for causal inference in such settings. We first derive the general form of the bias introduced by unmeasured confounding, with emphasis on theoretical properties uniquely relevant to multiple treatments. We then propose methods to encode the impact of unmeasured confounding on potential outcomes and adjust the estimates of causal effects in which the presumed unmeasured confounding is removed. Our proposed methods embed nested multiple imputation within the Bayesian framework, which allow for seamless integration of the uncertainty about the values of the sensitivity parameters and the sampling variability, as well as use of the Bayesian Additive Regression Trees for modeling flexibility. Expansive simulations validate our methods and gain insight into sensitivity analysis with multiple treatments. We use the SEER-Medicare data to demonstrate sensitivity analysis using three treatments for early stage non-small cell lung cancer. The methods developed in this work are readily available in the R package SAMTx.

Sensitivity analysis of treatment effect to unmeasured confounding in observational studies with survival and competing risks outcomes

No unmeasured confounding is often assumed in estimating treatment effects in observational data, whether using classical regression models or approaches such as propensity scores and inverse probability weighting. However, in many such studies collection of confounders cannot possibly be exhaustive in practice, and it is crucial to examine the extent to which the resulting estimate is sensitive to the unmeasured confounders. We consider this problem for survival and competing risks data. Due to the complexity of models for such data, we adapt the simulated potential confounder approach of Carnegie et al (2016), which provides a general tool for sensitivity analysis due to unmeasured confounding. More specifically, we specify one sensitivity parameter to quantify the association between an unmeasured confounder and the exposure or treatment received, and another set of parameters to quantify the association between the confounder and the time-to-event outcomes. By varying the magnitudes of the sensitivity parameters, we estimate the treatment effect of interest using the stochastic expectation-maximization (EM) and the EM algorithms. We demonstrate the performance of our methods on simulated data, and apply them to a comparative effectiveness study in inflammatory bowel disease. An R package "survSens" is available on CRAN that implements the proposed methodology.

Assessing the impact of unmeasured confounding for binary outcomes using confounding functions

A critical assumption of causal inference is that of no unmeasured confounding: for estimated exposure effects to have valid causal interpretations, a sufficient set of predictors of exposure and outcome must be adequately measured and correctly included in the respective inference model(s). In an observational study setting, this assumption will often be unsatisfied, and the potential impact of unmeasured confounding on effect estimates should be investigated. The confounding function approach allows the impact of unmeasured confounding on estimates to be assessed, where unmeasured confounding may be due to unmeasured confounders and/or biases such as collider bias or information bias. Although this approach is easy to implement and pertains to the sum of all bias, its use has not been widespread, and discussion has typically been limited to continuous outcomes. In this paper, we consider confounding functions for use with binary outcomes and illustrate the approach with an example. We note that confounding function choice encodes assumptions about effect modification: some choices encode the belief that the true causal effect differs across exposure groups, whereas others imply that any difference between the true causal parameter and the estimate is entirely due to imbalanced risks between exposure groups. The confounding function approach is a useful method for assessing the impact of unmeasured confounding, in particular when alternative approaches, e.g. external adjustment or instrumental variable approaches, cannot be applied. We provide Stata and R code for the implementation of this approach when the causal estimand of interest is an odds or risk ratio.

Do preoperative depressive symptoms predict quality of life after laparoscopic cholecystectomy: A longitudinal prospective study

OBJECTIVE

The impact of preoperative depressive symptoms on quality of life (QOL) after laparoscopic cholecystectomy (LC) remains unclear. The purpose of this study was to develop a benchmark for capturing the burden of depressive symptoms on QOL after LC and for supporting evidence-based clinical interventions for remediating these effects.

METHODS

Patients diagnosed with depressive symptoms (Beck Depression Inventory score > 13) after LC (n = 336) were classified into a depressive symptoms group. Propensity score matching was then used to match them with 336 patients in a non-depressive symptoms group for all potential confounding factors. All patients completed the 36-item Short Form Health Survey (SF-36) and the Gastrointestinal Quality of Life Index (GIQLI) at baseline and at 2 years postoperatively. The 95% confidence intervals (CIs) for differences in responsiveness estimates were derived by bootstrap estimation.

RESULTS

The GIQLI results revealed that the non-depressive symptoms group had relatively stronger responses for emotional impairment (4.10, 95% CI 2.81 to 5.39) and social impairment (4.06, 95% CI 2.65 to 5.46) in comparison with the depressive symptoms group. In the SF-36, the non-depressive symptoms group also had stronger responses for role emotional (12.63, 95% CI 10.73 to 14.54), social functioning (11.25, 95% CI 9.85 to 12.65), vitality (3.81, 95% CI 2.82 to 4.81), mental health (11.97, 95% CI 10.36 to 13.56) and general health (3.84, 95% CI 2.95 to 4.75).

CONCLUSIONS

Depressive symptoms complicate the management of LC patients and are associated with poorer outcomes. Because depressive symptoms are very common, further studies are needed to evaluate integrated and comprehensive approaches for assessing and treating these symptoms.

Propensity Scores in Pharmacoepidemiology: Beyond the Horizon

Purpose of review

Propensity score methods have become commonplace in pharmacoepidemiology over the past decade. Their adoption has confronted formidable obstacles that arise from pharmacoepidemiology's reliance on large healthcare databases of considerable heterogeneity and complexity. These include identifying clinically meaningful samples, defining treatment comparisons, and measuring covariates in ways that respect sound epidemiologic study design. Additional complexities involve correctly modeling treatment decisions in the face of variation in healthcare practice, and dealing with missing information and unmeasured confounding. In this review, we examine the application of propensity score methods in pharmacoepidemiology with particular attention to these and other issues, with an eye towards standards of practice, recent methodological advances, and opportunities for future progress.

Recent findings

Propensity score methods have matured in ways that can advance comparative effectiveness and safety research in pharmacoepidemiology. These include natural extensions for categorical treatments, matching algorithms that can optimize sample size given design constraints, weighting estimators that asymptotically target matched and overlap samples, and the incorporation of machine learning to aid in covariate selection and model building.

Summary

These recent and encouraging advances should be further evaluated through simulation and empirical studies, but nonetheless represent a bright path ahead for the observational study of treatment benefits and harms.

Use of Fibrates Monotherapy in People with Diabetes and High Cardiovascular Risk in Primary Care: A French Nationwide Cohort Study Based on National Administrative Databases

Background and Aim

According to guidelines, diabetic patients with high cardiovascular risk should receive a statin. Despite this consensus, fibrate monotherapy is commonly used in this population. We assessed the frequency and clinical consequences of the use of fibrates for primary prevention in patients with diabetes and high cardiovascular risk.

Design

Retrospective cohort study based on nationwide data from the medical and administrative databases of French national health insurance systems (07/01/08-12/31/09) with a follow-up of up to 30 months.

Methods

Lipid-lowering drug-naive diabetic patients initiating fibrate or statin monotherapy were identified. Patients at high cardiovascular risk were then selected: patients with a diagnosis of diabetes and hypertension, and >50 (men) or 60 (women), but with no history of cardiovascular events. The composite endpoint comprised myocardial infarction, stroke, amputation, or death.

Results

Of the 31,652 patients enrolled, 4,058 (12.8%) received a fibrate. Age- and gender-adjusted annual event rates were 2.42% (fibrates) and 2.21% (statins). The proportionality assumption required for the Cox model was not met for the fibrate/statin variable. A multivariate model including all predictors was therefore calculated by dividing data into two time periods, allowing Hazard Ratios to be calculated before (HR_<540) and after 540 days (HR_>540) of follow-up. Multivariate analyses showed that fibrates were associated with an increased risk for the endpoint after 540 days: HR_<540 = 0.95 (95% CI: 0.78–1.16) and HR_>540 = 1.73 (1.28–2.32).

Conclusion

Fibrate monotherapy is commonly prescribed in diabetic patients with high cardiovascular risk and is associated with poorer outcomes compared to statin therapy.

Can you party your way to better health? A propensity score analysis of block parties and health

While other indicators of social capital have been linked to health, the role of block parties on health in Black neighborhoods and on Black residents is understudied. Block parties exhibit several features of bonding social capital and are present in nearly 90% of Philadelphia's predominantly Black neighborhoods. This analysis investigated: (1) whether or not block parties are an indicator of bonding social capital in Black neighborhoods; (2) the degree to which block parties might be related to self-rated health in the ways that other bonding social indicators are related to health; and (3) whether or not block parties are associated with average self-rated health for Black residents particularly. Using census tract-level indicators of bonding social capital and records of block parties from 2003 to 2008 for 381 Philadelphia neighborhoods (defined by census tracts), an ecological-level propensity score was generated to assess the propensity for a block party, adjusting for population demographics, neighborhood characteristics, neighborhood resources and violent crime. Results indicate that in multivariable regression, block parties were associated with increased bonding social capital in Black neighborhoods; however, the calculation of the average effect of the treatment on the treated (ATT) within each propensity score strata showed no effect of block parties on average self-rated health for Black residents. Block parties may be an indicator of bonding social capital in Philadelphia's predominantly Black neighborhoods, but this analysis did not show a direct association between block parties and self-rated health for Black residents. Further research should consider what other health outcomes or behaviors block parties may be related to and how interventionists can leverage block parties for health promotion.

A comparison of confounding adjustment methods with an application to early life determinants of childhood obesity

We implemented six confounding adjustment methods: (1) covariate-adjusted regression, (2) propensity score (PS) regression, (3) PS stratification, (4) PS matching with two calipers, (5) inverse probability weighting and (6) doubly robust estimation to examine the associations between the body mass index (BMI) z-score at 3 years and two separate dichotomous exposure measures: exclusive breastfeeding v. formula only (n=437) and cesarean section v. vaginal delivery (n=1236). Data were drawn from a prospective pre-birth cohort study, Project Viva. The goal is to demonstrate the necessity and usefulness, and approaches for multiple confounding adjustment methods to analyze observational data. Unadjusted (univariate) and covariate-adjusted linear regression associations of breastfeeding with BMI z-score were -0.33 (95% CI -0.53, -0.13) and -0.24 (-0.46, -0.02), respectively. The other approaches resulted in smaller n (204-276) because of poor overlap of covariates, but CIs were of similar width except for inverse probability weighting (75% wider) and PS matching with a wider caliper (76% wider). Point estimates ranged widely, however, from -0.01 to -0.38. For cesarean section, because of better covariate overlap, the covariate-adjusted regression estimate (0.20) was remarkably robust to all adjustment methods, and the widths of the 95% CIs differed less than in the breastfeeding example. Choice of covariate adjustment method can matter. Lack of overlap in covariate structure between exposed and unexposed participants in observational studies can lead to erroneous covariate-adjusted estimates and confidence intervals. We recommend inspecting covariate overlap and using multiple confounding adjustment methods. Similar results bring reassurance. Contradictory results suggest issues with either the data or the analytic method.

The Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS)

IMPORTANCE/OBJECTIVE: Although the suicide rate in the U.S. Army has traditionally been below age-gender matched civilian rates, it has climbed steadily since the beginning of the Iraq and Afghanistan conflicts and since 2008 has exceeded the demographically matched civilian rate. The Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS) is a multicomponent epidemiological and neurobiological study designed to generate actionable evidence-based recommendations to reduce Army suicides and increase knowledge about risk and resilience factors for suicidality and its psychopathological correlates. This paper presents an overview of the Army STARRS component study designs and of recent findings. DESIGN/SETTING/PARTICIPANTS/INTERVENTION: Army STARRS includes six main component studies: (1) the Historical Administrative Data Study (HADS) of Army and Department of Defense (DoD) administrative data systems (including records of suicidal behaviors) for all soldiers on active duty 2004-2009 aimed at finding administrative record predictors of suicides; (2) retrospective case-control studies of fatal and nonfatal suicidal behaviors (each planned to have n = 150 cases and n = 300 controls); (3) a study of new soldiers (n = 50,765 completed surveys) assessed just before beginning basic combat training (BCT) with self-administered questionnaires (SAQ), neurocognitive tests, and blood samples; (4) a cross-sectional study of approximately 35,000 (completed SAQs) soldiers representative of all other (i.e., exclusive of BCT) active duty soldiers; (5) a pre-post deployment study (with blood samples) of soldiers in brigade combat teams about to deploy to Afghanistan (n = 9,421 completed baseline surveys), with sub-samples assessed again one, three, and nine months after returning from deployment; and (6) a pilot study to follow-up SAQ respondents transitioning to civilian life. Army/DoD administrative data are being linked prospectively to the large-scale survey samples to examine predictors of subsequent suicidality and related mental health outcomes.

MAIN OUTCOME MEASURES

Measures (self-report and administratively recorded) of suicidal behaviors and their psychopathological correlates.

RESULTS

Component study cooperation rates are comparatively high. Sample biases are relatively small. Inefficiencies introduced into parameter estimates by using nonresponse adjustment weights and time-space clustering are small. Initial findings show that the suicide death rate, which rose over 2004-2009, increased for those deployed, those never deployed, and those previously deployed. Analyses of administrative records show that those deployed or previously deployed were at greater suicide risk. Receiving a waiver to enter the Army was not associated with increased risk. However, being demoted in the past two years was associated with increased risk. Time in current deployment, length of time since return from most recent deployment, total number of deployments, and time interval between most recent deployments (known as dwell time) were not associated with suicide risk. Initial analyses of survey data show that 13.9% of currently active non-deployed regular Army soldiers considered suicide at some point in their lifetime, while 5.3% had made a suicide plan, and 2.4% had attempted suicide. Importantly, 47-60% of these outcomes first occurred prior to enlistment. Prior mental disorders, in particular major depression and intermittent explosive disorder, were the strongest predictors of these self-reported suicidal behaviors. Most onsets of plans-attempts among ideators (58.3-63.3%) occurred within the year of onset of ideation. About 25.1% of non-deployed U.S. Army personnel met 30-day criteria for a DSM-IV anxiety, mood, disruptive behavior, or substance disorder (15.0% an internalizing disorder; 18.4% an externalizing disorder) and 11.1% for multiple disorders. Importantly, three-fourths of these disorders had pre-enlistment onsets.

CONCLUSIONS

Integration across component studies creates strengths going well beyond those in conventional applications of the same individual study designs. These design features create a strong methodological foundation from which Army STARRS can pursue its substantive research goals. The early findings reported here illustrate the importance of the study and its approach as a model of studying rare events particularly of national security concern. Continuing analyses of the data will inform suicide prevention for the U.S. Army.

Design of the Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS)

The Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS) is a multi-component epidemiological and neurobiological study designed to generate actionable evidence-based recommendations to reduce US Army suicides and increase basic knowledge about the determinants of suicidality. This report presents an overview of the designs of the six components of the Army STARRS. These include: an integrated analysis of the Historical Administrative Data Study (HADS) designed to provide data on significant administrative predictors of suicides among the more than 1.6 million soldiers on active duty in 2004-2009; retrospective case-control studies of suicide attempts and fatalities; separate large-scale cross-sectional studies of new soldiers (i.e. those just beginning Basic Combat Training [BCT], who completed self-administered questionnaires [SAQs] and neurocognitive tests and provided blood samples) and soldiers exclusive of those in BCT (who completed SAQs); a pre-post deployment study of soldiers in three Brigade Combat Teams about to deploy to Afghanistan (who completed SAQs and provided blood samples) followed multiple times after returning from deployment; and a platform for following up Army STARRS participants who have returned to civilian life. Department of Defense/Army administrative data records are linked with SAQ data to examine prospective associations between self-reports and subsequent suicidality. The presentation closes with a discussion of the methodological advantages of cross-component coordination.

The use of propensity scores and observational data to estimate randomized controlled trial generalizability bias

Although randomized controlled trials are considered the 'gold standard' for clinical studies, the use of exclusion criteria may impact the external validity of the results. It is unknown whether estimators of effect size are biased by excluding a portion of the target population from enrollment. We propose to use observational data to estimate the bias due to enrollment restrictions, which we term generalizability bias. In this paper, we introduce a class of estimators for the generalizability bias and use simulation to study its properties in the presence of non-constant treatment effects. We find the surprising result that our estimators can be unbiased for the true generalizability bias even when all potentially confounding variables are not measured. In addition, our proposed doubly robust estimator performs well even for mis-specified models.