Improving Disparity Estimates for Rare Racial/Ethnic Groups with Trend Estimation and Kalman Filtering: An Application to the National Health Interview Survey

Abstract

Objective. Single-year estimates of health disparities in small racial/ethnic groups are often insufficiently precise to guide policy, whereas estimates that are pooled over multiple years may not accurately describe current conditions. While collecting additional data is costly, innovative analytic approaches may improve the accuracy and utility of existing data. We developed an application of the Kalman filter in order to make more efficient use of extant data. Data Source. We used 1997We used -2004 National Health Interview Survey data on the prevalence of health outcomes for two racial/ethnic subgroups: American Indians/ Alaska Natives and Chinese Americans. Study Design. We modified the Kalman filter to generate more accurate current-year prevalence estimates for small racial/ethnic groups by efficiently aggregating past years of cross-sectional survey data within racial/ethnic groups. We compared these new estimates and their accuracy to simple current-year prevalence estimates. Principal Findings. For 18 of 19 outcomes, the modified Kalman filter approach reduced the error of current-year estimates for each of the two groups by 20-35 percent--equivalent to increasing current-year sample sizes for these groups by 56-135 percent.Conclusions. This approach could increase the accuracy of health measures for small groups using extant data, with virtually no additional cost other than those related to analytical processes.Key Words. Health disparities, Chinese, American Indians, borrowing strength, cross-sectional data Eliminating racial/ethnic health disparities is a major federal policy goal; however, most national health surveys have limited ability to assess the health of racial/ethnic population subgroups (e.g., national origin subgroups of r Health Research and Educational Trust

Methodological Issues in the Collection, Analysis, and Reporting of Granular Data in Asian American Populations: Historical Challenges and Potential Solutions

“…Combining estimation procedures from various surveys helps to address non-coverage and non-response issues and to estimate prevalence rates of other factors. 54 …”

Smoothing across time in repeated cross‐sectional data

“…For stroke, the relative standard error (standard error divided by the mean) exceeds 0.20 for 8 of the 11 groups and exceeds 0.30 for six of the groups. As discussed by Elliott et al [1], standard errors of these magnitudes are too large to meet National Center for Health Statistics recommended standards for estimating health disparities.…”

“…However, even large surveys like the NHIS typically provide only small annual samples of rarer subgroups, making their annual sample means imprecise. For example, Table I provides 2004 NHIS estimates of the prevalence of stroke and the average body mass index (BMI) for 11 racial/ethnic groups defined by Elliott et al [1]. The sampling errors are large for all but the most populous groups.…”

Accounting for Misclassification Bias in Binary Outcome Measures of Illness: The Case of Post-Traumatic Stress Disorder in Male Veterans

“…Empirical analyses of the determinants of adult health frequently rely on large‐scale general population surveys, where illness is often measured by a dichotomous indicator of the respondent's reported presence or absence of a diagnosis or condition or by a categorical variable based on a self‐rating scale from excellent to good to fair to poor. Some recent examples include studies by Cutler and Lleras‐Muney (2008) of adult mortality and morbidity; by Krieger, Barbeau, and Soobader (2005) of occupational class and health; by Elliot et al (2009) of racial disparities in health; and by Dobkin and Shabani (2009) of veterans’ health, all of which are based largely on the self‐reported health conditions in the U.S. National Health Interview Survey; Angrist, Chen, and Frandsen's (2010) analysis of the health consequences of military service using self‐reported health in U.S. Census files; and analyses of the relationship between socioeconomic status and health conditions in older adults using the U.S. Health and Retirement Survey (Bowen 2009; Link et al 2008).…”