Picking variables for a fracreg

[Jane Smith]

I have 15 variables of interest, and 15 variables that are well established in the literature as independent variables predictive of my dependent variable. How do I select which variables for a fractional regression?

[Richard Williams]

If all 15 variables have been previously found to be theoretically justified and have been found to have important effects in past work, why not use all 15 of them? If you exclude some for no apparent reason, others may ask why you did that given that you had measures of the variable in question.

There may be good reasons for narrowing the list. Maybe the sample is too small to include 15 variables. Maybe you think that the effects of some variables are indirect, and estimated direct effects will disappear once other variables are in the model. Maybe you have good theoretical reasons for believing some of these variables will have no effect in the population you are studying.

But, given that all 15 are well established, I think you would have to make the case for not using all of them. Maybe you develop a theory about how some effects are indirect or spurious, and then estimate models to support your case. Or, if sample size is small or the data have other limitations, explain why this affects the number of variables that can be looked at and why you chose the ones you did.

But again, just based on what you say, why not use all 15, or at least start with all 15? Maybe all 15 will be important for you too. Or, maybe you'll be able to identify the variables that really really really are important, since other studies may have suffered by only having a few of the measures.

If you wanted to use mindless empiricism, I suppose you could use stepwise regression. But, that is generally frowned on.

Incidentally, the fact that you are using fracreg does not seem relevant. With any statistical procedure, you have to decide which variables to look at.

[Jane Smith]

Thank you for the very helpful advice. My issue is I have 15 variables known to be associated with the outcome but I also have 15 other variables I want to test (so 30 variables in total!). And my sample size is way too small to handle that. I realise stepwise is frowned upon but I would be interested to see what that throws up. Can this be done with fracreg? I cannot seem to find any instructions

[Jane Smith]

stata tells me fracreg is not supported by stepwise. I could do betareg, but am I able to get stata to produce odds ratios using betareg?

[Richard Williams]

You can use glm instead of fracreg:

use https://www3.nd.edu/~rwilliam/statafiles/401kpart, clear
sw, pe(.05): glm prate mrate ltotemp age sole, vce(robust) link(p) family(binomial) nolog

The fact that you can do it does not change the fact that many view stepwise regression as the work of the devil.

https://www.stata.com/support/faqs/s...sion-problems/

[Jane Smith]

Thank you Richard, is glm OK when my outcome is a proportion?

[Richard Williams]

Yes, glm works when the DV is a proportion.

[Jane Smith]

.

[Jane Smith]

I think I may be better off sticking with the fracreg, but need to figure out how to reduce down 30 variables (sample size is small). Is there any method? Currently I have just been putting in and out random variables and the significant predictors keep changing. I don't know how to figure out the best model.