Comparative methods for handling missing data in large databases

摘要: Objective Analysis of complex survey databases is an important tool for health services researchers. Missing data elements are challenging because the reasons "missingness" multifactorial, especially categorical variables such as race. We simulated missing race and analyzed bias from five methods used in predicting major amputation patients with critical limb ischemia (CLI). Methods Patient discharges fully observed containing lower extremity revascularization or CLI were selected 2003 to 2007 Nationwide Inpatient Sample, a database (weighted n = 684,057). Considering several random schemes, we compared methods: complete case analysis, replacement frequencies, indicator variable, multiple imputation, reweighted estimating equations. created 100 sets, 5%, 15%, 30% subjects' drawn be full set. Bias was estimated by comparing regression coefficients averaged over sets (β miss ) each method vs estimates set ), relative calculated – β /β ) × 100%. Results Our results demonstrate that equations produce least biased variable produces most coefficients. Complete imputation resulted moderate bias. Sensitivity analysis demonstrated optimal choice depends on quantity type encountered. Conclusions analytic topic research large databases. The commonly introduces severe should caution. present empiric evidence guide selection handling data.