Imputation for Missing Physiological and Health Measurement Data: Tests and Applications

We evaluated alternative approaches to imputation for univariate estimates and multivariate regression analyses of physiological health measures collected in the 2003-2004 National Health and Nutrition Examination Survey (NHANES). From the NHANES public use data files we selected 5041 respondents age 20+ who provided questionnaire or medical exam data. Measures collected at interview (e.g., demographics, self-reported health status) and measures collected at physical examination (e.g., height, weight, blood pressure, cholesterol, hemoglobin, Hematocrit, and iron) were evaluated for rates of item missing data (i.e., item nonresponse). The properties of several imputation methods (including single and multiple imputation) were evaluated with respect to univariate estimates and a regression model using age, sex, race, height, weight, cholesterol, and marital status to predict blood pressure. Only small differences were found between imputation methods, and no major systematic differences between methods were observed. The findings suggest that for the missing data problems considered in our investigation, the specific imputation method makes little difference on univariate and multivariate estimates and standard errors.