Assessing the impact of different imputation methods on serial measures of renal function: the Strong Heart Study

Missing data are a common problem in epidemiologic studies. This study had two aims: (a) to determine which method for imputing missing renal function data provides estimates closest to those made with complete data and (b) to determine which measure of renal function better estimates cardiovascular disease (CVD) risk. For these analyses, a subset of Strong Heart Study participants with complete data for renal function was identified. Data were randomly dropped from this complete set at three rates: 30, 45, and 60%. Five common techniques for handling missing data were compared: imputation using the mean, adjacent value (AV), single imputation, multiple imputation, and listwise deletion. Differences between the imputed sets and the complete set were determined for each method. Imputation methods were used to fill in missing values for serum creatinine (Scr) in one model and estimated glomerular filtration rate (eGFR) in another. For both Scr and eGFR, the AV method provided the most favorable results in predicting CVD risk, regardless of the rate of missing data.