0
$\begingroup$

Whereby no interaction terms or main effects are significant in a model, should model simplification be carried out and the interaction terms removed sequentially until the remaining variables are either significant/insignificant?

Info on the model I am running: I am currently carrying out a GLMM with negative binomial distribution looking at the effects of different factors on the number of farmland birds along the edges of fields.

The main fixed effects included in the model are: Crop type (3 level factor, which was experimentally manipulated) Hedgerow structure (continuous) Percentage gaps (continuous) Month (2 level factor) with all 2 way interactions and 3 way interactions with month included. Field is included as a random factor to account for non-independence and offset is field length.

This is the first time i have deal with complex models and as i am partly hypothesis testing (whether crop type has an effect) I am unsure what the correct protocol is. The model output is included below. Thanks Model output

$\endgroup$
1
$\begingroup$

No, significance testing of coefficients is not a good method of model selection. Better methods of model selection include AIC, Bayes factors, and cross-validated prediction error. What you should use depends on exactly what you want to do. Trying to find out whether crop type has a nonzero effect is not a sensible goal because $10^{-100}$ is nonzero, but also effectively 0 for all human purposes, so knowing that the effect is nonzero would be uninformative. If you want to estimate the effect of crop type, you could try searching for a predictively accurate model with cross-validation and then looking at the appropriate coefficients.

$\endgroup$
1
  • $\begingroup$ Thanks for your answer. You make a valid point so i have taken your advice. $\endgroup$
    – Arran
    Aug 1 '16 at 15:35

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.