Practical Suggestions on Rounding in Multiple Imputation

• Activity Number: 28
• Type: Contributed
• Date/Time: Sunday, August 8, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Recai M. Yucel*+
• Companies: Institute for Health Policy
• Keywords: rounding ; categorical data ; missing data ; multiple imputation

Abstract

In the last decade, substantial progress has been made in the area of missing data. Many statistical methods and their implementations in software products (e.g., Splus 6 "missing'' library, SAS PROC MI, SOLAS) have become available for practitioners in a numerous research areas. The key idea underlying most of these methods is to "replace" missing values by random draws from the conditional distribution of the missing data given the observed data. For convenience some methods (e.g., norm module of missing library in Splus 6, SAS PROC MI) impose a multivariate normal distribution on the variables that are incompletely observed. When these variables are not of a "normal'' nature but rather categorical, practitioners are often advised to round the imputed value to the nearest integer (or category) that is within the defined region. We provide some practical suggestions for rounding using commonly available software while avoiding potential biases and more efficient results in terms of marginal distribution consistency. We also consider relevant complications in the structure of the variables of interest such as ordinal or nominal variables.