Hi Stata Forum
I have an issue with a tobit regression predicting too high values.

As a part of me and my colleague’s master’s thesis, we are trying to use a Tobit regression to predict the proportion of Special Items (an accounting post) that can be considered as opportunistic.

Inspired by an American research paper we are using Compustat data for all public Nordic companies as the basis for our predictions. We have recreated the variables from the American paper and our coefficients resemble the original when running the tobit-regression.

The company observations are grouped by industry and year, and we have created the IndYear variable accounting for this - IndYear are dropped if they have less than 30 observations.

We run a forvalues loop for each IndYear and try to make IndYear specific predictions, but the predicted SI (PredSI) is higher than the actual SI 75 pct. of the time - for the American paper it predicted too high 7 pct. of the time.
The model should predict the “real” SI and the residual represents the opportunistic part, so the residuals can’t be negative.

We are using the following command for the regression-loop:

gen PredSI = .
forvalues i = 1/62 {
tempvar test
tobit SI $cntrl if IndYear ==`i', vce(cluster CompanyName)
predict `test', ystar(0,.5)
replace PredSI = `test' if IndYear == `i'
}

PredSI = Predicted Special Items
SI = Special Items
$cntrl holds 14 variables

All variables have been winsorized by IndYear using winsor2 at a (1 99) or (0 99) if they have a lower boundary.

Hope somebody can help!

Best regards
Mathias