[R] Residual plots and residual deviance in the SURVEY package when only 'working deviance' available

[Marko Stojovic]

Hello -


I am analysing some survey data using the svyglm() command in the survey package. Since I am doing binomial regression, the family I'm choosing is 'quasibinomial', since this suppresses the warning that comes about from the inclusion of non-integer outcomes due to weights.


I am looking at doing some proper diagnostics of my model. However, the only goodness-of-fit method I can see from the mailing list is regtermtest(), which essentially tests the significance of additional terms, or the hypothesis that the model is better than the intercept.


My questions are:

- is it possible to obtain residual 'working' deviance after the Rao-Scott method, which would indicate deviance with respect to the saturated model?

- by way of corrollary (or substitute), can deviance residuals be obtained?

- given that svyglm appears to give the raw residuals through the residuals([svyglm object]) command, does anyone know of a good reference on how to effectively use pearson residuals for model diagnostics (on the basis that deviance ones are unavailable).


Many thanks.

Marko Stojovic

MSc Applied Statistics student, Birkbeck College, London

	[[alternative HTML version deleted]]