I am sorry if this has been asked elsewhere. I could find it when looking.

I am running a quasi-binomial generalised linear model (to correct for overdispersion) weighted by a count of sample size. I have simplified the model using F tests.

model2 <-update(model1,~.-A)  

I now have a model with 3 interaction terms and 6 main effects (4 categorical and 2 continuous). All remaining interaction are significant using F tests as are the 2 fixed effects not bound up in interactions. No further model simplification is possible.

This is an example of how I have run my model;

  glm(X ~ (A+ B + C + D + E + F + B:F + B:C + B:D),family=quasibinomial(link=logit), weights=Y)

I understand that the R output for quasi-binomial generalised linear model gives t values and Pr(>|t|) however it seems to be convention to report p values.

I know that it is possible to get P values of terms by comparing models with and without the terms using F tests in the case of a quasi-binomial distribution (this is what I have done to simplify my model). However as I understand it, it doesn’t make sense to remove fixed terms which are bound up in the significant interactions. For example to assess the significance of fixed term F when B:F is a significant interaction.

Is it true that it’s not possible to get the p value for the original fixed terms that are bound up in the interactions, only the interaction itself?

Is there another way to test that the fixed term is significant? I understand that the t values reported are for each level of the terms rather than a single test of the term. There could be a problem with multiple testing?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.