Can logit/probit marginal effects be compared to linear probability model coefficients?

[Tom Martin]

I am using a logit model to calculate the marginal effect of my independent variables on a dichotomous dummy variable. In testing the robustness of the logit model, I am comparing the logit marginal effects to the marginal effects obtained from using a probit model. These appear to be quite similar. I also decided to estimate a linear probability model using OLS. Are the marginal effects from the logit/probit models comparable to the OLS coefficients from the linear probability model? That is, can I conclude that obtaining linear probability model coefficients that are similar to logit/probit marginal effects increases the robustness of the main logit model?

[Richard Williams]

My experience is that all 3 (logit, probit, lpm) tend to produce similar AMEs, at least if you don't have things like interactions and squared terms in the model. I don't think similarities would make the case for logit more robust. If anything people may do the opposite, claiming LPM is ok because of logit results. Paul Allison has a blog entry about this coming out soon. In the meantime you can look at

http://statisticalhorizons.com/in-de...f-logit-part-1