Heteroscedasticity

[Dexter Hewlett]

will like to know the command for checking heteroscedasticity and how to explain the results using logistic regression analysis

[Maarten Buis]

There is no command for that, and it is a very different problem in logistic regression compared to linear regression: http://www.stata.com/statalist/archi.../msg00996.html

[Alfonso Sánchez-Peñalver]

Maarten, in the linked message you mention that the response variable is a probability, and that's why it's different. However, I have always thought of the response variable as a binary (bernoulli) variable, and thus these models estimate a mean of the response variable given some explanatory variables, which is what OLS does with a continuous variable. The main difference is then that logit and probit are non-linear models, in that they have an inverse link function around the linear expression. I believe this is what really makes detecting heteroskedasticity very much more complicated. Thanks for the reference in the link, something to read next week.

[Maarten Buis]

What is in the data is a binary variable, but what is modeled is the mean of that binary variable. The mean of a binary (coded 0,1) is the proportion of 1s, which you could see as an estimate of the probability of a 1 (in a frequentists sense).