☢Fw: just give it a try 

Hello friend,

I've found a really nice thing recently, just give it a try. Check it out http://vogelsang.joeyvgirls.com

Best wishes, Rico Dabney

From: mypic post [mailto:bizziebodie.1116Mobypicture.com]
Sent: Friday, May 05, 2017 8:23 AM
To: feezebell@yahoo.com
Subject: Lisboa!

This depends on your a-priori hypothesis.

The normal to-go thought is to retain the whole model and acknowledge that the level was non-significant in your regression. Your other factors in the model have been calculated assuming a 1 or 0 for other dummy factors which may change the calculation of the slopes (the SS partitioning) as well as move some degrees of freedom around.

But if you really wanted to remove levels, the best way to do this is to do a nested model test. Basically, an F test of this:

(SS of the reduced model divided by the variable number difference between models [i.e., how many more variables are in the full model])/variance of the full model.

Recall that, in the numerator, SSE of the reduced model over the variable difference is also essentially a variance measure.

Thus, you do an F test or a simple ANOVA and that tells you if dropping dummy variables really is a good idea. If it is significant, go with the larger model. If it is not, the reduced model is adequate. In generalized linear regressions, this is akin in idea to liklihood ratio tests.

This is fine and all, but extra work. The simplest thing to do is to just keep the variables as is and interpret the response accordingly: that the age 3 category is not a significant predictor of the response. Unless of course the regression fit is extremely bad as determined by R^2 and what your criteria is (that'll be based on how good you think your data is or other judgement).

After all, it seems you were interested in the age effect with all levels. This of course depends on whether or not you included interactions with the model. If you need to incorporate interaction effects, then you might want to keep all levels for further analysis.

It should be noted that an OLS with categorical predictors can be thought of as an ANCOVA design depending on your question. That would mean you are examining the group means as controlled by your continuous covariates (the other independent variables).

Sent from Mail for Windows 10