Mary,
your messages touches upon two different subjects: (1) why the classification table does not change from one step to the next (in a stepwise method, I presume), and (2) whether you may change the cutoff point for the classification table, using another value instead of the default SPSS cutoff value (0.50).
Regarding (1), it is perfectly possible that two successive steps give probabilities for the various cases that determine equal numbers to each side of the divide, i.e. the same number of predicted "Yes" and "No" each time, even if the probabilities themselves have somewhat changes from one step to the next. As long as the probability of a case remains below the cutoff point, it is predicted that it will not experience the outcome. As long as it remains above the cutoff, it is predicted to undergo the event. It is only sometimes that a case jumps the divide as a result of an iteration.
Regarding (2) there are many opinions. The most authorized voices call for an inspection of the curve to determine the best cutoff point. However, as with any cutoff point, it will be to some degree an arbitrary decision.
The idea of changing the cutoff point "to reflect the observed prevalence in your sample" may be based on the reasoning that the observed prevalence reflects the "average probability of the event" for a random individual, so that any individual with a probability above that average should be predicted to undergo the event, and anyone below would be predicted not to undergo it. But the foundations for such a reasoning are rather shaky. If the prevalence is 0.10, having a probability of 0.11 does not constitute adequate grounds to put you among those earmarked to undergo the event: after all, such a subject has near 90% chances of NOT undergoing the event. However, using that shaky reasoning will closely reproduce your observed prevalence, if that is what you wish. Nonetheless, I prefer to think of Logistic Regression as a tool for assessing probabilities in populations, not for classifying individuals in discrete classes.
Hector
----- Mensaje original -----
De: Mary Nolan <
[hidden email]>
Fecha: Martes, Abril 24, 2007 4:08 pm
Asunto: Logistic regression c;assification tables
> Hi
> I am doing logistic regression and have noticed in my SPSS output
> that my
> classification tables for Step 0 and Step 1 are exactly the same.
> I am
> wondering if anyone has had experience of this and if so what does
> it mean
> for my findings. Perhaps there is something I am doing incoreectly
> and can
> correct for? I recently read somewhere that I can change the cut
> off to
> reflect the prevalence of my predicted outcome in my sample, but
> it was not
> from a credible source.
> Many thanks in advance,
> Mary
>
Locked
