A practical guide to multiple imputation of missing data in nephrology

Abstract

Health data are often plagued with missing values which can greatly reduce the sample size if only complete cases are considered for analysis. Furthermore, analyses which ignore the missing data have the potential to introduce bias in the parameter estimates. Multiple imputation techniques have been developed to recover the information that would otherwise be lost when excluding observations with missing data and to help minimise bias. However, the validity of analyses using imputed data relies on the imputation model having been correctly specified. The aim of this guide is to aid the reader in the decision-making process when conducting an analysis with multiply imputed data, in the context of nephrology research. We discuss: 1) missing mechanism assumption; 2) imputation method; 3) imputation model; 4) derived variables; 5) number of imputed datasets; 6) diagnostic checks; 7) analysis and pooling of results; and 8) reporting the results. This process is demonstrated using data from the National Health and Nutrition Examination Survey (NHANES) to explore the association between hypertension and kidney disease in adults from the general population. Example code is provided for SAS and the mice package in R.