Skip to main content
Modern Multiple Imputation Approaches to Missing Data Problems
Organizer and Chair: Russell Steele (McGill University)
[PDF]
BEN GOODRICH, Columbia University
Evaluating the Missing-At-Random Assumption after Multiply Imputing Missing Data  [PDF]

Most algorithms for multiply imputing missing values make a ``Missing-At-Random'' (MAR) assumption, which loosely states that the probability that a data point is missing is conditionally independent from the values of the missing data. Thus, many researchers believe that the resulting completed data cannot falsify the MAR assumption. However, the presence of observed data among the completed data can provide some leverage. This study investigates ways to falsify the MAR assumption by applying selection models and covariance structure analysis to the completed data. Simulations are conducted to determine how effective these techniques are.

NATHANIEL SCHENKER, National Center for Health Statistics and Centers for Disease Control and Prevention
Multiple Uses for Multiple Imputation at the U.S. National Center for Health Statistics  [PDF]

This talk describes three recent or potential applications of multiple imputation at the National Center for Health Statistics, to illustrate various types of problems for which the technique can be used. One application involves missing data on body scans in the National Health and Nutrition Examination Survey (NHANES). The second seeks to improve on analyses of self-reported data from the National Health Interview Survey by imputing clinical values using models fitted to the smaller NHANES. The third uses observed birthweights and mixture models to identify questionable gestational ages in U.S. natality data, and seeks to impute more plausible gestational ages.

YAJUAN SI, Duke University
Multiple Imputation in Panel Studies with Attrition and Refreshment Samples  [PDF]

Panel data provides advanced understanding of mass behavior along time with multiple follow-up waves. However, panel studies suffer from attrition, which reduces the effective sample size and can lead to biased estimates if the tendency to drop out is systematically related to the substantive outcome of interest. A refreshment sample as an external data source includes new, randomly sampled respondents who are provided with the same questionnaire at the same time as a subsequent wave of the panel. We utilize refreshment samples to assess the effects of panel attrition and to correct for biases via multiple imputation.