Statistical Diagnosis And Impact Of The Linear Model Of The Longitudinal Data Analysis

Recently, much concern has been raised about studies of the models for longitudinal data. The problems about statistical diagnostics and influence analysis of a class of specific models for longitudinal data are considered in this thesis. The models are involved in which the random error is divided into three parts i.e. measurement error, random effects and serial correlation, so the covariance matrix can be modeled with four parameters.First for the global influence, REML estimates of parameters in covariance matrix are produced by means of Newton-Raphson iteration, with that, the WLS estimates of mean parameters can be found and statistics can be developed to measure the influence of each subject, subsequently, the influence subjects can be exploited. The statistics is Cook's distance in the thesis. Moreover, outlier subject's testing is studied in the thesis.For the local influence analysis, statistics are developed respectively to measure the influence for mean and covariance parameters under subject and case perturbation of measurement error, random effects, whole random error and response variable. These statistics are employed to exploit the influence subjects or cases.Real datasets analysis is used to illustrate the studies for global and local influence analysis.