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

D. Modeling strategy: AllEs, noCs

Model: Logit PðXÞ¼aþ~

q

i¼ 1

biEiþ~

q

i¼ 1

~

q

i^0 ¼ 1

i 6 ¼i^0

d*ii 0 EiEi^0

Step 1: Define initial model (above)

Step 2: Assess interaction involvingEs.

Option A*: Overall chunk test forEEs,

followed by backward elimination of

EEs

Option B*: Skip chunk test forEEs; start

with backward elimination ofEEs

Skip previous Step 3

Step 4: Test for nonsignificantEs if not

components of significantEEs

E. How causal diagrams can influence choice of

initial model?

III. Screening Variables (pages 263–270)

A. Problem Focus: Model contains oneE, and a large

number ofCs andECs,butcomputer program

does not run or fitted model unreliable (“large” p)

B. Screening: Exclude someCjone-at-a-time; fit

reduced model

C. Method 0:Consider predictors one-at-a-time;

screen-out thoseXinot significantly associated

with the outcome (D)

D. Questions and Brief Answers about Method 0:

1. Any criticism? Yes: does not consider
   confounding or interaction involvingCs


2. Depends on types ofXs? Yes: use if onlyEs
   and noCs.


3. How largekcompared ton? No good answer.


4. Other ways than Method 0? Yes: evaluate
   confounding and/or interaction forCs.


5. Collinearity and/or screening? Consider
   collinearity prior to and following screening.
   E. Assessing Confounding and Interaction when
   Screening C variables.
   Confounding: Compare Logit P(X)¼aþbEwith
   Logit P(X)¼aþbEþgC
   DoesdORDE¼e
   ^b
   6 ¼dORDEjC¼e
   ^b*
   ?

Interaction: TestH 0 :d¼0 for the model Logit

P(X)¼aþbEþgCþdEC

F. How to proceed if severalEs and severalCs: It

depends!

G. How to proceed if severalEs and noCs: Use

method 0.

Detailed Outline 287