Comparisons of Methods for Analysis of Repeated Binary Responses with Missing Data

Abstract

It is important yet challenging to choose an appropriate analysis method for the analysis of repeated binary responses with missing data. The conventional method using the last observation carried forward (LOCF) approach can be biased in both parameter estimates and hypothesis tests. The generalized estimating equations (GEE) method is valid only when missing data are missing completely at random, which may not be satisfied in many clinical trials. Several random-effects models based on likelihood or pseudo-likelihood methods and multiple-imputation-based methods have been proposed in the literature. In this paper, we evaluate the random-effects models with full- or pseudo-likelihood methods, GEE, and several multiple-imputation approaches. Simulations are used to compare the results and performance among these methods under different simulation settings.

Key Words:

• GEE
• Generalized linear mixed model
• Missing data
• Multiple imputation
• Random-effects model

ACKNOWLEDGMENTS

We thank an associate editor and two referees for their constructive comments and suggestions and for providing additional references that led to significant improvement of this paper. We also thank Xiaoming Li for the SAS program for the propensity score multiple imputation approach.

Notes

Note. Results based on log odds ratios are presented as main entries of the table, and results based on difference in proportions are in parentheses. Boldface font indicates that the type I error rate is beyond 2 standard errors of the simulations. MAR1 and MAR3 correspond to the missing at random cases 1 and 3, respectively. RD, random-effects model with a normal random intercept; RD_WL, random-effects model with a random intercept following Wang and Louis's bridge distribution. MI_logit, MI_PS, and MI_RD correspond to multiple imputation with the logistic regression model, propensity score method, and predicted distribution from a random-effects model, respectively.

Note. Results based on log odds ratios are presented as main entries of the table, and results based on difference in proportions are in parentheses. Boldface font indicates that the type I error rate is beyond 2 standard errors of the simulations. MAR1, MAR2, and MAR3 correspond to the missing at random cases 1, 2, and 3, respectively.

Note. Results based on log odds ratios are presented as main entries of the table, and results based on difference in proportions are in parentheses. MAR1, MAR2, and MAR3 correspond to the missing at random cases 1, 2, and 3, respectively.

Note. MAR1, MAR2, and MAR3 correspond to the missing at random cases 1, 2, and 3, respectively.

Note. MAR1, MAR2, and MAR3 correspond to the missing at random cases 1, 2, and 3, respectively.

Note. MAR1, MAR2, and MAR3 correspond to the missing at random cases 1, 2, and 3, respectively.

Note. MAR1, MAR2, and MAR3 correspond to the missing at random cases 1, 2, and 3, respectively.