I'm doing a regression about how practice of sports affect the GPA. I'm using a panel data, and I'm going to do a fixed effects approach. The thing is, I need to standardize the GPA and the sports as well somehow so that I can give some kind of weight to the students who have a lower/upper GPA, or are going more to the gym, because they are seeing more or less credits. For example, a student with a GPA of 4/5 but who's seeing 20 credits isn't quite easily comparable to one that got a GPA of 4/5 as well but just with 6 credits.

So I first thought about multiplying credits by GPA, but I think the difference is exaggerated with this approach.

Then I thought I could standardize the GPA with a formula somehow like this: (x-u)*(std/2) > with x being the observation, u the mean and std the standard deviation. My intuition here is that x-u gives positive if the credits are higher than the mean so the effect on GPA will be positive, and when x is lower than the mean, well, it will go down a bit. I think this is somehow arbitrary and I think it may be wrong for me to mess with the data like this?

Then I thought I could do an auxiliary regression like this: GPA=credits+credits^2. The intuition is that credits^2 will let me now for a fact the form of the curve, and credits alone would give me a coefficient which I can add or subtract depending on whether the GPA in that semester is over que top or lower bottom of the curve, and is in fact affected upwards or downwards by the amount of credits. In this case, with the sports practice the regression would be similar.

  • $\begingroup$ Welcome to this site. A summary of data may be incorporated in body of your question and try to be brief about issues focussing on the key objective. $\endgroup$ Feb 28 '17 at 14:27

Based on the information you provided, I do not see a reason why you need to separately standardize GPA by credits; simply incorporate it into your regression model.

From your description, it seems like you could simply add the number of credits into the regression equation as another independent variable along with the interaction. e.g. GPA = Vsports + Vcredits + Vsports*Vcredits

This would allow you estimate the effects of the practicing sports at varying levels of credits. Of course, you would need to check model assumptions and may need to transform your variables accordingly.

  • $\begingroup$ hmm Ok, I get it but, the way I see it if the student has less credits in a semester, he/she also has more time to practice sports, so that's why I didn't took into account to use credits and sports both together, because they should be correlated. So in this case the interaction shows how sports would interfere with the relation between credits and GPA, but does it solve the problem of correlation? $\endgroup$
    – yatoe
    Feb 28 '17 at 21:15
  • $\begingroup$ I'm guessing that you will have enough cases of "non sport participants" with low credits and "sport participants" with a lot of credits to cover the sample space. Ideally this would allow the credits to be "adjusted for" in the regression itself, leaving the true effect of the sports participation to be estimated. $\endgroup$
    – Underminer
    Mar 2 '17 at 13:47

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.