Topics in Longitudinal Studies with Coarsened Data

longitudinal studies;measurement error;missing data;multiple imputation;repeated measures;semiparametric approach

Jiang, Liqiu ; Anastasios A. Tsiatis, Committee Chair,John F. Monahan, Committee Member,Wenbin Lu, Committee Member,Marie Davidian, Committee Member,Jiang, Liqiu ; Anastasios A. Tsiatis ; Committee Chair ; John F. Monahan ; Committee Member ; Wenbin Lu ; Committee Member ; Marie Davidian ; Committee Member

University:North Carolina State University

关键词: longitudinal studies;    measurement error;    missing data;    multiple imputation;    repeated measures;    semiparametric approach;   

Others  :  https://repository.lib.ncsu.edu/bitstream/handle/1840.16/5405/etd.pdf?sequence=1&isAllowed=y

美国|英语

来源: null

【 摘 要 】

In the first part of the dissertation, we derive two methods for responders analysis in longitudinal data with random missing data. Often a binary variable is generated by dichotomizing an underlying continuous variable measured at a specific point in time according to a prespecified threshold value. Ordinarily, a logistic regression model is used to estimate the effects of covariates on the binary response. In the event that the underlying continuous measurements are from a longitudinal study, the repeated measurements are often analyzed using a repeated measures model because of mathematical and computational convenience of available off-the-shelf software. This practical advantage motivates us to propose two methods: one is to use repeated measures model as an imputation approach in the presence of missing data on the responder status as a result of patient drop-out before completion of the study. We then apply the logistic regression model on the observed or otherwise imputed responder status; the other is to construct estimating equations based on the relationship of repeated measures model and logistic regression model. Large sample properties of the resulting estimators are derived and simulation studies carried out to assess the performance of the estimators in situations where either the model for the continuous repeated measurements is misspecified as following a multinormal distribution, when, in truth, it follows a logistic distribution that is compatible with the logistic regression model for the probability of response or when the model that the probability of response following a logistic regression model is misspecified because, in truth, the longitudinal data follow a multinormal distribution. We show that the resulting estimators are robust to misspecification and apply them to data from a clinical trial on a toenail disease.We adopt a semiparametric estimator to a longitudinal data with measurement error in the second part of the dissertation. In longitudinal studies, we are often interested in the relationship between a primary response and the profile of repeated measurements collected over time for a subject, which can be dictated by individual random effects in the framework of a generalized linear model. For example, if the longitudinal profile is linear, the relationship of individual intercept and slope and primary response would be of interest. The naive method by fitting a regression model to obtain estimates for individual random effects can lead to biased results. Li, Zhang, and Davidian (Biometrics 2004) developed conditional score approaches for generalized linear models which require no assumption on the distribution of the random effects and yield consistent inference regardless of the true distribution. However, the estimator can only be used for generalized linear models in canonical form with normally distributed measurement error. To overcome this limitation, we adopt locally efficient semiparametric estimators proposed by Tsiatis and Ma (Biometrika 2004) for functional measurement error models to use for such longitudinal studies. The distribution of random effects is allowed to be misspecified and the method will still yield consistent inference. Simulation studies are carried out to assess the performance of the estimator. We show that the estimator can give much better inference than the naive method in terms of bias and empirical probability coverage. The approach is applied to data from a study on woman's bone disease.