Abstract

Outlier and influence statistics play an important role in assessing individual or grouped observations that may have undue impact on the parameter estimates of a statistical model. The methods are well-developed for linear and linear mixed effects models, and are easily implemented in most statistical packages. Though similar statistics exist for univariate survival models, not much has been done for models of multivariate survival data. The objective of this PhD work was to derive outlier and influence statistics for multivariate survival data models, by extending limited research work on such statistics for linear mixed-effects and univariate survival models. The derived statistics were evaluated using simulation studies and illustrated with an analysis of child survival data in Malawi, which had 56 sub-districts (clusters), from both rural and urban areas. The proposed statistics had a high performance of well over 90% in identifying correctly the outlying or influential clusters, and the performance improved with increasing cluster size. In the application to clustered survival data, mostly off-city clusters were identified as having a different child survival pattern and impactful on regression coefficients and variance estimates. This study recommends incorporating outlier and influence assessments when analysing clustered survival data, otherwise the estimates of both regression slopes and variance components could be biased.

More details

School	: School of Natural and Applied Sciences
Issued Date	: 2020

Download full document