EXAMPLE
MRSA example – Initial Model:
Logit PðXÞ¼aþb 1 E 1 þb 2 E 2 þg 1 V 1
þg 2 V 2 þd 11 E 1 W 1
þd 12 E 1 W 2 þd 21 E 2 W 1
þd 22 E 2 W 2 þd*E 1 E 2 ;
where
D¼MRSA statusð 0 ; 1 Þ;
E 1 ¼PREVHOSP
E 2 ¼PAMU;V 1 ¼W 1 ¼AGE;
V 2 ¼W 2 ¼GENDER
At least one collinearity problem,
i.e., involves PREVHOSP and
PREVHOSP × AGE Drop
Another possible collinearity problem
(CNI¼34.3)
Two alternatives at this point:
Stop further collinearity
assessment
Drop PAMUAGE and continue
We now illustrate the use of collinearity diag-
nostics for the MRSA dataset we have
described earlier. We consider the initial
model shown at the left, which contains two
Es, twoVs, 4EWs, and a singleEE.
Using the collinearity macro introduced above,
we obtain the (edited) collinearity diagnostic
output shown at the left. From this table, we
see that the highest CNI is 45.6, which is consid-
erably higher than 30, and there are two VDPs
greater than 0.5, corresponding to the variables
PREVHOSP (VDP¼0.95) and the product term
PREVHOSPAGE (VDP¼0.85).
Based on these results, we decide that there is
at least one collinearity problem associated
with the highest CNI and that this problem
involves the two variables PREVHOSP and
PREVHOSPAGE.
Proceeding sequentially, we would now drop
the product term from the model and reassess
collinearity for the resulting reduced model.
The results are shown at the left. From this
table, we see that the highest CNI is now 34.3,
which is slightly higher than 30, and there are
two VDPs greater than 0.5, corresponding to
the variables AGE (VDP¼0.74) and the prod-
uct term PAMUAGE (VDP¼0.75).
Since the highest CNI here (34.3) is only
slightly above 30, we might decide that this
value is not high enough to proceed further to
assess collinearity. Alternatively, proceeding
conservatively, we could drop the product
PAMUAGE and further assess collinearity.
274 8. Additional Modeling Strategy Issues