Do we make inferences on marginal effects or coefficient estimates in a Probit model? I am asking because I found the discrepancy (i.e. in terms of both signs and statistical significance) between the results generated by two approaches.