2
$\begingroup$

In this discussion on probit vs. OLS models 'Conjugate Prior' mentions that probit and all GLMs allow to specify a DV distribution.

How is the GLM improved by specifying a DV distribution and what is the implied DV distribution for OLS then?

$\endgroup$
2
$\begingroup$

GLMs aren't improved by specifying a DV distribution; the specification of a distribution is part of forming a GLM. For example, if we perform a logistic regression, which is a kind of GLM, we are saying the dependent variable is distributed Bernoulli conditional upon the values of the regressors (independent variables.) Where the improvement comes is that by specifying a GLM we are able to use maximum likelihood to estimate the parameters, which, assuming the model is correct, have lots of nice properties as the sample size gets large.

OLS minimizes the sum of squared errors; this doesn't imply any conditional distribution for the dependent variable, it's just an objective function. If, however, we assume that the errors are i.i.d. Normal, then the OLS procedure not only minimizes the sum of squared errors, it also maximizes the likelihood.

$\endgroup$

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.