Logistic Regression: A Self-learning Text, Third Edition (Statistics in the Health Sciences)

(vip2019) #1
In contrast, the model given by logit P(X)
equalsaplusbEplusg 1 V 1 plusg 2 V 2 plus the
product termsd 1 EV 1 plusd 2 EV 2 is hierarchi-
cally well formulated because the lower-order
components of each variable in the model are
also in the model. For example, the compo-
nents ofEV 1 areEandV 1 , both of which are
contained in the model.

For illustrative purposes, let us consider one
other model given by logit P(X) equalsaplus
bE plus g 1 V^21 plus g 2 V 2 plus the product
termd 1 EV^21 .Isthismodelhierarchicallywell
formulated?

The answer here can be either yes or no
depending on how the investigator wishes
to treat the variableV 12 in the model. IfV 12 is
biologically meaningful in its own right with-
out considering its componentV 1 , then the
corresponding model is hierarchically well
formulated because the variableEV 12 can be
viewed as having only two components,
namely,EandV 12 , both of which are contained
in the model. Also, if the variableV^21 is consid-
ered meaningful by itself, it can be viewed as
having no lower order components. Conse-
quently, all lower order components of each
variable are contained in the model.

On the other hand, if the variableV 12 is not
considered meaningful separately from its fun-
damental componentV 1 , then the model is not
hierarchically well formulated. This is because,
as given, the model does not containV 1 , which
is a lower order component ofV 12 andEV^21 , and
also does not contain the variableEV 1 , which is
a lower order component ofEV 12.

Now that we have defined and illustrated an
HWF model, we discuss why such a model
structure is required. The reason is that if the
model is not HWF, then tests about variables in
the model – in particular, the highest-order
terms – may give varying results depending on
the coding of variables in the model. Such tests
should beindependent of the codingof the vari-
ables in the model, and they are if the model is
hierarchically well formulated.

EXAMPLE
HWF model:
logit P


X


¼aþbEþg 1 V 1 þg 2 V 2
þd 1 EV 1 þd 2 EV 2

Components ofEV 1 :
E,V 1 both in model

EXAMPLE

logit P


X


¼aþbEþg 1 V^21
þg 2 V 2 þd 1 EV^21

HWF model?
Yes,ifV^21 is biologically meaningful
components ofEV^21 EandV^21
components ofV^21 : none

No,ifV^21 is not meaningful separately
fromV 1 :
The model does not contain
 V 1 , component ofV 12
 EV 1 , component ofEV 12

Why require HWF model?


Answer:


HWF?


Tests for highest-order
variables?

No dependent on coding
Yes independent of coding


182 6. Modeling Strategy Guidelines

Free download pdf