Missing data in multivariate longitudinal studies: Comparing results from different missing data techniques using an empirical data set.

摘要

Problems of missing data arise in almost all developmental research. One common practice in solving the problem of missing data has been to delete those cases with missing data points. However, statisticians have shown that this type of missing data method is not adequate, since it can significantly decrease the sample size, thus reducing power, and provide results that are not representative for the population, hence introducing bias. Depending on the amount and causes of missingness, as well as the pattern of missing data, different approaches can be used. At this writing, the recommended missing data procedures are the maximum likelihood (ML) method and the multiple imputation (MI) method, particularly when dealing with multivariate and/or longitudinal datasets. However, most of the recommendations are based on simulation studies.; The goal of this dissertation was to provide empirical evidence of the performance of recommended missing data techniques as well as the traditional method of deleting cases when they are used on a data set from an empirical study that tries to find answers to real research questions. Three missing data techniques (i.e., listwise deletion, the ML method, and the MI method) were used with an empirical data set in order to evaluate the extent of differences in the results and, therefore, in the interpretations of results. In particular, the focus was on two sets of analyses: linear growth modeling and predicting a specific outcome from a set of prediction variables at previous points in time.; Data from the first three waves of the 4-H Study of Positive Youth Development were used for the analyses. The results showed that the three missing data techniques did not yield comparable results for research questions assessing different aspects of development (i.e., change over time or prediction effects). However, the results suggested that the listwise deletion method produces quite different results from the recommended missing data techniques, which was evident in large discrepancies in parameter estimates between the listwise deletion method and the recommended techniques, as well as substantially larger standard errors in the listwise deletion method. In addition, smaller sample size and loss of statistical power negatively affected results of the listwise deletion method. Nevertheless, the results also showed that in some cases the MI method produces different results from the ML method. One possible explanation of this difference might be the use of auxiliary variables with the MI method only.; Generalization of these conclusions is limited to those studies that have a large sample size with similar amounts of missing data for each variable, as was the case in the longitudinal sample of the 4-H Study. In addition, further studies should be explored in which the MI method is conducted with appropriate software designed specifically for imputing longitudinal data. Such analyses could allow researchers to determine the extent that results differ depending on the imputation model that is used.; The results of the present empirical analyses suggest that researchers should not use the listwise deletion method, but use both recommended missing data techniques (i.e., the ML and the MI method) in order to see if and how the results of their analyses change. In addition, a longitudinal researcher should not just depend on the missing data techniques to solve the problems of missing data; he or she should also take actions before data collection, such as creating a plan to develop an appropriate questionnaire and make an effort to increase the sample retention rate, in order to decrease attrition and missing data.