Abstract:
Conditional logistic regression models have been
extensively used in the field of medicine and mainly applied
in matched case control studies. However, none of the major
statistical packages, i.e. SAS, MINITAB, SPSS provide
diagnostics to assess the goodness-of-fit of these models.
In addition the freely downloadable package R provides no
functions for this purpose. The objectives of this study are
to review the available diagnostics for software for testing
goodness-of-fit of conditional logistic regression models by
the development of a computer programme and to test this
programme which implements some of the reviewed methods
on real data. The computer programme is implemented using
Visual Basic for Applications (VBA) for Microsoft Excel and
connected to the Statistical Analysis Software (SAS) version 9.1
using the Object Linking and Embedding (OLE) automation.
The software thus developed is tested on a matched case
control study on endometrial cancer. A conditional logistic
regression model is fitted to these data and the risk factors for
endometrial cancer are identified. MINITAB and SPSS are
incapable of doing conditional logistic regression. For testing
goodness of the fitted model Proc Logistic in SAS is only
capable of giving delta-beta plots which explain the influence of
each observation on the parameters of the model. Besides, plots
obtained from the developed computer programme, in addition
provide information on stratum specific lack-of-fit statistics.
These plots were very successful in identifying 3 outlying
strata which were quite different from the other strata. In these
3 strata the case had not received estrogen whereas one or more
control had received estrogen.