1
$\begingroup$

I am currently running quasipoisson models with a continuous response variable and 13 covariates. I am using glm() and summary(). Most of my covariates are continuous but I also have the YEAR and SEASON as categorical variables.

When I run the model it appears that the year 2008 is not significant while all the other years are. How can I remove only this year from my model? Do I have to remove the entire covariate?

$\endgroup$
  • 4
    $\begingroup$ Highly likely that you don't want to do this. summary() is showing the results of $t$ tests that the difference between the state level and the reference level. That one is not significantly different doesn't mean that that category is not different from some of the others. If you want to get rid of the category, you need to throw out those data and then you are cherry picking and possibly will also render some other categories "non-significant" due to the reduced degrees of freedom. $\endgroup$ – Gavin Simpson Jul 25 '13 at 3:22
  • $\begingroup$ Thank you for your answer. So is there a way to remove only the category that is not significant? $\endgroup$ – Elena Spark Jul 25 '13 at 4:31
  • 2
    $\begingroup$ That would not constitute good statistical practice. $\endgroup$ – Frank Harrell Jul 25 '13 at 11:04
  • 1
    $\begingroup$ @ElenaSpark You don't want to do that! First off you can't do a test or tests, decide what is not significant, get rid of insignificant things and then go on and redo the original tests. The p-value will not have its usual meaning because it doesn't reflect the data dredging you have done. Just leave that category in. You could do some post hoc tests to look and see which groups are different. But leave all categories in there! $\endgroup$ – Gavin Simpson Jul 25 '13 at 14:49
  • $\begingroup$ Thank you all very much. I will conserve my covariate for now and run a post hoc test. $\endgroup$ – Elena Spark Jul 26 '13 at 19:53