| Abstract: | Techniques for influence analysis in generalized linear models have largely been limited to approximating the effect of deleting a single case, and most often, the effect upon estimation of the regression coefficients. Case-deletion methods seldom yield much insight into the role of those cases identified as influential. Several diagnostics are presented which are based on the local influence method using small perturbations of various elements of the data, including explanatory variables, the response, case weights, and a single case. These diagnostics can reveal how certain elements of data and interact with the model to gain influence. The diagnostics are illustrated with a logistic regression and a proportional hazards model. |
|---|