Originally posted by Aishwarya Gupta
View Post
Earlier, I said that if you thought you could meet the assumptions of censored linear regression (i.e. Tobit regression), you could try running the model with the lower limit set at 0, and the upper limit at 10, i.e.
Code:
xttobit happinessindex lgva unemployementrateaged16 mentalhealth obesityqofprevalence17 smokingattributable deathsfromhea violentcrimeincludingsexualviole socialisolation lifeexpectancypca, ll(0) ul(10)
I further cautioned that I thought that censored linear regression was a wrong model to begin with. I was trying to think of a model that was clearly superior to -xtreg-. This was the only alternative model that came to mind at the time, but given your description, this is not a case where censoring applies - I tried to make this clear, but I fear I have caused confusion, for which I apologize. Let me be clear: the clearest improvement would be to get the original data and use ordered logistic models (albeit you have to deal with the fact that you don't have repeat measures on the same individuals; I'm not quite sure how to do this).
You actually have a statistically significant, albeit small, effect of log gvpa on average satisfaction. It is what it is, and moreover, you said that you were expecting this from prior literature. I don't really see any major improvements to be made here (unless you can get the individual data).
Comment