Abstract Background. American Indians experience higher stroke morbidity and mortality compared to US general population, but are underrepresented in public health research. Data on incident stroke in American Indians derive mainly from the Strong Heart Study (SHS), a population-based cohort study of cardiovascular disease in 4549 American Indians who were 45-74 years old when baseline exams were conducted from 1988-1990. The SHS had higher stroke rates than reported for Whites and African Americans in external comparisons to other cohorts. These findings suggested similar disparities in covert vascular brain injury (VBI), an often asymptomatic form of cerebrovascular disease that precedes clinical events. Accordingly, from 2010-2013 the Strong Heart Stroke Study (SHSS) used structural cranial magnetic resonance imaging to assess covert VBI in 1033 surviving members of the SHS. Goals. In this dissertation we addressed three limitations to using SHS and SHSS data for analysis of stroke and covert VBI in American Indians: Manuscript 1) lack of research that directly compares stroke incidence and mortality in American Indians vs. other racial groups, and which limits current knowledge to external comparisons that do not account for differences in stroke risk factors; Manuscript 2) potential selection bias in SHSS data when survival and participation of cohort members depends on both the exposures and outcomes of interest; and Manuscript 3) an inherent limitation in effect measures estimates that condition on categories defined by progressively older age or longer time since exposure, and which leads to observed point estimates that are potentially biased estimates of the true effects. Manuscript 1. Methods: We pooled data from the SHS and the Atherosclerosis Risk in Communities Study (ARIC) to compare stroke risk and post-stroke mortality in American Indians vs. Blacks and Whites. We used Cox regression to estimate hazard ratios (HR) with attained age as the time scale to account for differences in baseline age at enrollment, and adjusted estimates for baseline factors that included prevalent hypertension and diabetes. Due to effect modification, analyses were stratified by birth year tertile (1914-1930, 1931-1937, and 1938-1947). We used logistic regression to compare 30-day and 1-year post-stroke mortality among participants from both studies who experienced stroke during follow-up. Results: Stroke risk among American Indians in the SHS was lower than among Blacks for all birth year tertiles (1914-1930: HR = 0.9 (95% CI = 0.7, 1.1); 1931-1937: HR = 0.9 (95% CI = 0.7, 1.2); 1938-1947: HR = 0.9 (95% CI = 0.7, 1.2)), but higher than among Whites (1914-1930: HR = 1.6 (95% CI = 1.3, 2.0); 1931-1937: HR = 2.2 (95% CI = 1.7, 2.8); 1938-1947: HR = 2.7 (95% CI = 2.0, 3.6)) in ARIC. Adjusting for risk factors including prevalent diabetes at baseline resulted in strengthening of associations compared to Blacks (oldest to youngest tertile HR = 0.8 (95% CI = 0.6-1.0); 0.7 (95% CI = 0.5-1.0); and 0.6 (95% CI = 0.4-0.8)), and attenuation of associations compared to Whites (oldest to youngest tertile HR = 1.1 (95% CI = 0.9-1.5); 1.2 (95% CI = 0.9-1.6); and 1.1 (95% CI = 0.8-1.5)). American Indians had higher risk of 30-day and 1-year mortality compared to Blacks (relative risk = 2.2 (95% CI = 1.4-3.0) and 1.4 (95% CI = 1.1-1.8), respectively) and Whites (relative risk = 1.8 (95% CI = 1.2-2.3) and 1.5 (95% CI = 1.1-1.8), respectively). These comparisons persisted after adjusting for risk factors. Manuscript 2. Methods: We used marginal structural models with inverse probability weighting to adjust for selection bias in the SHSS, applied to the analysis of prevalent hypertension and covert VBI as measured by white matter hyperintensities. Predicted probabilities of survival from 1988-2010 and participation of survivors were estimated and inverted to create weights, and stabilized using conventional methods to reflect the distribution of hypertension in cohort participants. In addition, we computed novel stabilized weights that account for each person’s probability of meeting the inclusion criterion of remaining stroke-free up to their SHSS exam. These weights allowed us to avoid over-correcting for attrition of individuals who would have subsequently gone on to experience clinical stroke. We applied these weights to estimate the prevalence difference (PD) for the association of hypertension with a binary indicator of abnormal VBI, as well as the mean difference (MD) for a continuous variable reflecting the ratio of white matter/total intracranial volume; the ratio estimates were multiplied by 1000 to simplify presentation of results. Hypertension was evaluated as both a cross-sectional risk factor and accounting for longitudinal trends in prevalence since baseline. Results: In the cross-sectional analysis, hypertension was associated with higher prevalence of abnormal VBI in unweighted models (PD = 7.9% (95% CI = -2, 17)). The point estimate increased 13% after selection weighting (PD = 8.9% (95% CI = 0, 18)). Prevalent hypertension was likewise associated with a higher proportion of white matter volume compared to the total intracranial volume in unweighted models (MD = 0.8 (95% CI = -0.4, 2.0)) and after selection weighting (MD = 0.9 (95% CI = -0.3, 2.1)). Adjusting weights to account for the stroke-free inclusion criterion did not change results compared to the conventional stabilized estimates. In the analysis treating hypertension as a longitudinal exposure, prevalent hypertension at all three study exams was associated with higher prevalence of abnormal VBI (PD = 8.0% (95% CI = -6, 22)) and higher ratio of white matter/total intracranial volume (MD = 1.7 (95% CI = 0.0, 3.4)) compared to not having hypertension at any exam. Selection weighting had no appreciable impact on point estimates in the longitudinal analysis. Manuscript 3. Methods: We used Mathematica software with constrained optimization to identify bounds for the risk difference (RD) when conditioning on event-free survival to some minimum age or time since exposure. Bounds were identified assuming only causative exposure effects in the target population, and allowing for exposure to prevent disease in some individuals so long as the causative effects were proportionally greater in the overall population. We applied these bounds to the analysis of post-stroke survival from Manuscript 1, with follow-up time divided into 0-30 days, 31-180 days, and 181-365 days after the stroke event. Results: The RD attenuated across follow-up periods for American Indians vs. Blacks (0-30 days: RD = 14% (95% CI = 6, 23); 31-180 days: RD = -1% (95% CI = -7, 4); 181-365 days: RD = -3% (95% CI = -7, 2)) and Whites (0-30 days: RD = 12% (95% CI = 3, 21); 31-180 days: RD = 1% (95% CI = -5, 6); 181-365 days: RD = -2% (95% CI = -6, 3)). With assumptions of only causative exposure effects, bounds on the the conditional risk difference for American Indians vs. Blacks were 0-16% for 0-30 days post-stroke event, and 1-13% for 181-365 days post-stroke. For American Indians vs. Whites the bounds were 0-14% for 0-30 days post stroke, and 0-13% for 31-180 days post-stroke. Allowing for preventive effects that were equal to or less than causative effects yielded bounds that were too wide for meaningful interpretation (all lower bounds = 0; all upper bounds ≥ 30). Conclusions. We found that American Indians in the SHS had lower stroke risk than Blacks, but not than Whites, in ARIC after adjusting for risk factors that included prevalent diabetes. These findings suggest that diabetes may be a factor behind stroke disparities in some American Indian communities. American Indians had higher post-stroke mortality than Blacks and Whites especially in the first 30 days after stroke onset, but cumulative risk comparisons and analyses using bounds for conditional effects were consistent with elevated risk persisting for at least 1 year. Among long-term survivors of the SHS who participated in the SHSS assessment of covert VBI, selection bias may be of concern for some analyses. Although adjusting selection weights for the stroke-free inclusion criterion did not change results in this example, other studies with inclusion criteria that result in excluding larger proportions of the study population may wish to include sensitivity analyses with similar adjustments.
University of Minnesota Ph.D. dissertation. December 2015. Major: Epidemiology. Advisor: Richard MacLehose. 1 computer file (PDF); xi, 91 pages.
Stroke disparities and selection bias in an American Indian cohort: the Strong Heart and Strong Heart Stroke Studies.
Retrieved from the University of Minnesota Digital Conservancy,
Content distributed via the University of Minnesota's Digital Conservancy may be subject to additional license and use restrictions applied by the depositor.