1081 — Comparison of Statistical Methods to Account for Missing Race in VA Administrative Data
Mor MK (VA Pittsburgh Healthcare System ), Cheng C
(VA Pittsburgh HCS), Bean-Mayberry B
(VA Greater Los Angeles HSR&D Center of Excellence), Tang G
(VA Pittsburgh HCS), Stone RA
(VA Pittsburgh HCS)
Objectives:
VA administrative sources commonly are used to assess quality and health equity within the VA Healthcare System. However, missing racial data continue to be problematic, even after the implementation of improved data collection methods in FY2003. The aim of this study was to assess the sensitivity of estimated racial differences to the statistical method used to account for missing racial data.
Methods:
Using VA External Peer Review Program data for a national sample of veterans (FY2001-2003), we compared alternative missing data procedures to estimate racial disparities in the receipt of flu vaccination and foot examination in two cohorts of older patients for which these measures were recommended. We assigned patient race using the most recent self-reported racial data available in the VA administrative data. Race and demographic information obtained from VA inpatient and outpatient administrative data were supplemented with race data from the Centers for Medicare and Medicaid Services. Self-reported race was available for 63% of the sample, race was available from prior VA administrative data (FY1997-2002) or from Medicare for an additional 27.6%, with the remaining 2.8% were missing race from all data sources. For each outcome, analyses were conducted using complete case analysis, weighted analysis using inverse probability weights, multiple imputation, hot deck imputation, and the EM Algorithm. We compared point estimates of racial differences, standard errors, and tests of significance across all methods.
Results:
Although qualitative changes in racial estimates were rare, the choice of statistical method affected the tests of statistical significance and the magnitude of estimated racial coefficients. Differences in estimated coefficients of approximately one standard error or more were observed for all racial groups for at least one outcome. The estimated standard errors were similar across methods in these data.
Implications:
The choice of statistical methods used to account for missing race data can affect the magnitude of estimates and tests of significance for racial differences.
Impacts:
Estimates of racial differences based on VA administrative data are affected by the methodology used to account for missing racial data. Further work is needed to identify when each method is most appropriate.
