Some Scary Simulations on the Robustness of Logit

Based on my training, I was always believed that logit models were relatively robust.  Specifically, the rule-of-thumb had been that having more than 50 observations was probably ok for simple MLE models like logit.  Some sticklers prefer a floor of 100 observations.  However, some recent research suggests that the floor may be need to be higher -- much higher.


Follow the link for evidence that we may need to be skeptical of any logit model with less than 750 observations.

http://www.surefoss.org/visualisation/evolution-of-a-logistic-regression/