Journal of Data Science

Missing data is a common occurrence in various fields, spanning social science, education, economics, and biomedical research. Disregarding missing data in statistical analyses can introduce bias to study outcomes. To mitigate this issue, imputation methods have proven effective in reducing nonresponse bias and generating complete datasets for subsequent analysis of secondary data. The efficacy of imputation methods hinges on the assumptions of the underlying imputation model. While machine learning techniques such as regression trees, random forest, XGBoost, and deep learning have demonstrated robustness against model misspecification, their optimal performance may necessitate fine-tuning under specific conditions. Moreover, imputed values generated by these methods can sometimes deviate unnaturally, falling outside the normal range. To address these challenges, we propose a novel Predictive Mean Matching imputation (PMM) procedure that leverages popular machine learning-based methods. PMM strikes a balance between robustness and the generation of appropriate imputed values. In this paper, we present our innovative PMM approach and conduct a comparative performance analysis through Monte Carlo simulation studies, assessing its effectiveness against other established methods.

Abstract: Missing data are a common problem for researchers working with surveys and other types of questionnaires. Often, respondents do not respond to one or more items, making the conduct of statistical analyses, as well as the calculation of scores difficult. A number of methods have been developed for dealing with missing data, though most of these have focused on continuous variables. It is not clear that these techniques for imputation are appropriate for the categorical items that make up surveys. However, methods of imputation specifically designed for categorical data are either limited in terms of the number of variables they can accommodate, or have not been fully compared with the continuous data approaches used with categorical variables. The goal of the current study was to compare the performance of these explicitly categorical imputation approaches with the more well established continuous method used with categorical item responses. Results of the simulation study based on real data demonstrate that the continuous based imputation approach and a categorical method based on stochastic regression appear to perform well in terms of creating data that match the complete datasets in terms of logistic regression results.

Summary: Longitudinal binary data often arise in clinical trials when repeated measurements, positive or negative to certain tests, are made on the same subject over time. To account for the serial corre lation within subjects, we propose a marginal logistic model which is implemented using the Generalized Estimating Equation (GEE) ap proach with working correlation matrices adopting some widely used forms. The aim of this paper is to seek some robust working correla tion matrices that give consistently good fit to the data. Model-fit is assessed using the modified expected utility of Walker & Guti´errez Pe˜na (1999). To evaluate the effect of the length of time series and the strength of serial correlation on the robustness of various working correlation matrices, the models are demonstrated using three data sets containing respectively all short time series, all long time series and time series of varying length. We identify factors that affect the choice of robust working correlation matrices and give suggestions under different situations.

Abstract: Mixed effects models are often used for estimating fixed effects and variance components in continuous longitudinal outcomes. An EM based estimation approach for mixed effects models when the outcomes are truncated was proposed by Hughes (1999). We consider the situation when the longitudinal outcomes are also subject to non-ignorable missing in addition to truncation. A shared random effect parameter model is presented where the missing data mechanism depends on the random effects used to model the longitudinal outcomes. Data from the Indianapolis-Ibadan dementia project is used to illustrate the proposed approach

Abstract: Here we develop methods for applications where random change points are known to be present a priori and the interest lies in their estimation and investigating risk factors that influence them. A simple least square method estimating each individual’s change point based on one’s own observations is first proposed. An easy-to-compute empirical Bayes type shrinkage is then proposed to pool information from separately estimated change points. A method to improve the empirical Bayes estimates is developed. Simulations are conducted to compare least-square estimates and Bayes shrinkage estimates. The proposed methods are applied to the Berkeley Growth Study data to estimate the transition age of the puberty height growth.