Logit model

[klo alikaj]

Hello everybody
I'm trying to estimate the loss given default of a borrower.
Since this variable takes values between [0,1] is it apropriate to use logistic regression ???
Thank you in advance.

[Carlo Lazzaro]

Klo:
it depends on your dependent variable.
If your dependent variable is the (yes/no) borrower default, go -logit-; if your dependent variable is the amount of the loss experienced by the lender due to the borrower default, go -regress-.
-

[klo alikaj]

HI Carlo,
First of all, thank you for your answer.
The variable I'm trying to model is expressed as a percentage of the exposure, simplified as follows:
loss given default=(exposure at default -recovery )/exposure at default, those the values are continuous numbers between [0,1}.
if we think about loss given default as the probability of total recover(loss given default=0) against the probability of total loss(loss given default=1), I think we can model it as a logistic regression, but i'm not totally sure.
All the BEST,
Klodiana,

[Brad Anderson]

If this is a percentage or probability ranging from 0 - 1 I'd suggest taking a look at fractional regression or possibly beta regression. Both are available in Stata 14. Fractional regression can be estimated using glm. with logit link and binomial distribution.

[klo alikaj]

thank you Brad,
The problem is I'm not familiar with these regressions, but I will try to find something,
Thank you for your suggestion.

[Brad Anderson]

Here's a reference. The authors are either Papke and Wooldridge or Wooldrige and Papke. There's a fracreg command in Stata (I believe it became available with Stata 14). I'm sure the manual offers some explanation and references. Or just search fractional response regression and Stata and you'll get lots of hits. The coefficients have little intrinsic meaning so you may want to use the margins command to make more sense out of results. I'm not sure why one would choose fractional response regression over beta regression, or vice versa. I don't think beta regression works if there are observations at the lower limit of 0 or upper limit of 1. The fractional regression model can be estimated easily with glm if you don't have a recent version of Stata. Her's the reference: “Econometric Methods for Fractional Response Variables with an Application to 401(k) Plan Participation Rates” (with L.E. Papke), Journal of Applied Econometrics 11, 619-632, NovemberDecember, 1996.