Multiple Imputation of Missing Data: A Simulation Study on a Binary Response

Abstract

Currently, a growing number of programs become available in statistical software for multiple imputation of missing values. Among others, two algorithms are mainly implemented: Expectation Maximization (EM) and Multiple Imputation by Chained Equations (MICE). They have been shown to work well in large samples or when only small proportions of missing data are to be imputed. However, some researchers have begun to impute large proportions of missing data or to apply the method to small samples. A simulation was performed using MICE on datasets with 50, 100 or 200 cases and four or eleven variables. A varying proportion of data (3% -63%) was set as missing completely at random and subsequently substituted using multiple imputation by chained equations. In a logistic regression model, four coefficients, i.e. non-zero and zero main effects as well as non-zero and zero interaction effects were examined. Estimations of all main and interaction effects were unbiased. There was a considerable variance in the estimates, increasing with the proportion of missing data and decreasing with sample size. The imputation of missing data by chained equations is a useful tool for imputing small to moderate proportions of missing data. The method has its limits, however. In small samples, there are considerable random errors for all effects.

Missing Data in Alcohol Clinical Trials with Binary Outcomes

“…The present findings are consistent with other research that has simulated MCAR and MAR binary data in randomized trials and found minimal bias with MI (Hardt et al 2013) and little difference between MI and CCA (Caille et al, in press, Ma et al 2011). However, these studies did not test MNAR conditions, where MI consistently outperformed CCA in the present study, nor did they test LOCF or WCS approaches.…”

Nearest neighbor imputation algorithms: a critical evaluation

“…Even the final test we conducted in the SPECTF heart dataset is not fully exhaustive of what a researcher may encounter in the real-life, as we considered only a MCAR mechanism to create the missing data. The conclusions we draw applies to cases with moderate sizes of missingness, no lower than 15 % and no higher than 30 %; we intentionally limited our evaluations to this range as for small amounts of missing data, under the MAR or MCAR mechanisms, imputation may be useless and for larger amounts caution should always be applied because estimates may become very imprecise [21]. Thus, despite the efficiency of NN imputation under these conditions, it should remembered that imputation should be carefully applied and cannot solve all the problems of incomplete data [22] and that NN imputation can have serious drawbacks as we showed for instance considering the risk of distorting data distribution or the lack of precision in imputing variables with no dependencies in a dataset or, conversely, the possibility to introduce spurious associations considering dependencies where they do not exist.…”

Health and Well-Being of Adolescents in Different Family Structures in Germany and the Importance of Family Climate

“…We assumed the mechanism leading to missing values to be missing at random (MAR), and therefore integrated multiple imputations into the analyses to minimize bias stemming from missing data. This method was very well suited to this task, regarding the sample size, the number of variables included in the imputation model, and the analyses to be conducted [ 61 , 62 , 63 , 64 ]. The imputation model contained raw data for all variables used in the analysis, before items were combined into scales or dichotomized.…”