Dear Colleagues,
When performing logistic regression with standard software such
as SPSS, the output often includes a classification table. This
compares the observed binary responses with those predicted by
the fitted model, based on a user-defined threshold for the
estimated probability of success.
Clearly there is some bias because the training data are also used
as test data, but this table nevertheless gives a useful indication of
the model's performance. At least twice though, I have obtained a
strange result whereby the predicted responses tend to be all the
same (either 0 or 1). This occurred despite using a threshold of 0.5
and having a reasonable balance of observed 0 and 1 responses in
the training data.
I would be grateful to hear from anybody else who has encountered
this anomaly, particularly if you have a good explanation for it.
Many thanks,
_______________________________________________________
Dr David F. Percy,
Centre for Operational Research and Applied Statistics,
University of Salford, Greater Manchester, M5 4WT.
tel (0161) 295 4710, fax (0161) 295 4947
_______________________________________________________
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|