Regresión logística binaria: el problema de la SEPARACION COMPLETA

Extraños sucesos tienen lugar cuando una variable por sí misma puede separar completamente la presencia de eventos de su ausencia.

Eso es la separación completa. Se verá un ejemplo de separación completa en el video adjunto, y se mostrará cómo los intervalos de confianza del Odds Ratio (expB) nos ayudan a detectarla.

Y si el SPSS da un mensaje de no convergencia, no se debe confiar en los coeficientes de las distintas variables:

In general, you should not use the results from a logistic regression unless the parameter estimates have converged. You can increase the number of iterations allowed to see if they will converge, though usually, if the estimation doesn't converge within the default number (20 for the LOGISTIC REGRESSION procedure), it's quite possible convergence isn't going to be achieved by allowing more iterations. Inspection of the iterations history table may show that some estimates are growing larger though the log likelihood is stable, in which case a complete or quasi-complete separation is likely the issue. If this is the case, while the individual estimates are not valid (because finite estimates do not exist for such models), the predicted probabilities produced by the entire model can be used


https://drive.google.com/open?id=0B9C7VyfotFyJV2ZkX0tBUjlJZE0




Material suplementario de UCLA:
http://stats.idre.ucla.edu/other/mult-pkg/faq/general/faqwhat-is-complete-or-quasi-complete-separation-in-logisticprobit-regression-and-how-do-we-deal-with-them/