I am a bit confused about how regression modelling works.

I have a response $y$, 3 continuous predictors, and 3 factors. I don't have anything else available.

I fit the model

y ~ cont1 + cont2 + cont3 + factor1 + factor2 + factor3

to check for fixed effects.

Here's what I don't understand: What now? Should I change anything about this model before moving on to studying its output?

I am getting conflicting answers from reading threads on this forum: many, many say that I should remove insiginficant covariates using p-values. Others say that I should perform backwards stepwise AIC to reduce the model. And yet others say that I shouldn't do anything at all about any of the insignificant covariates.

Those are 3 different and common answers to one simple question!!

  • $\begingroup$ Please register &/or merge your accounts (you can find information on how to do this in the My Account section of our help center), then you will be able to edit & comment on your own question. $\endgroup$ – gung Jun 2 '17 at 23:00
  • $\begingroup$ What's your research question? What kinds of inferences do you want to draw from your model? Without knowing what you're trying to do, it's impossible to help you figure out how to get there. $\endgroup$ – Dan Hicks Jun 3 '17 at 0:00

I am attempting to provide an answer to a question I think you are asking, but you should know that similar questions to yours were already asked, and some excellent answers have already been provided: look here, here, and here.

There are conflicting views about this, but generally, the selection of predictors included in your model should be theory driven, and as such you should not remove predictors just because they are not significant or the removal of them improves the model fit; removing any variables from your model, without a good reason, can be considered as data snooping. Moreover, the effect of a predictor can be high, even if it is not significant -- so don't look at p values alone.

Coefficients of your predictors in your model are conditional upon the remaining variables in your model. As such, the removal of any predictors will likely affect the remaining covariates; it can inflate your coefficients and bias AIC, not to mention it will likely alter the associated p values -- see this about some of the problems associated with stepwise regression.

I would recommend thinking carefully about which variables are of theoretical importance in your research. If you are performing a confirmatory study, with variables of theoretical interest, then I would advise against the removal of any covariates from your model. If you are conducting an exploratory study with many variables, and if some of them are of no theoretical importance, were not specified in your literature review and hypothesis, and have marginal effect on the outcome and the model, then some would say that it would be acceptable to remove them.


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.