Drawbacks of screening:
- No simultaneous assessment of
allEs andCs. - No guarantee final model
contains all “relevant” variables
General screening situation:
nsubjects,kpredictors (Xi,
i¼1,...,k)
k“large enough” to require screening
Method 0:
Consider predictors one-at-a-
time
Screen-out thoseXinot
significantly associated withD
Questions about Method 0:
Q1. Any criticism?
Q2. Depends on types ofXs?
Use if severalEs andCs?
Use if oneEand several
Cs?
Use if onlyEs?
Q3. How largekcompared ton?
k¼10,n¼50: 20%?
k¼10,n¼100: 10%?
k¼10,n¼200: 5%?
Q4. Other ways than Method 0?
Q5. Collinearity and/or
screening?
The two primary drawbacks of screening are:
- Does not accomplish simultaneous assess-
ment of all exposure and control variables
recommended from the literature or conceptu-
alization of one’s research question. - No guarantee that one’s final model con-
tains all the relevant variables of interest,
although there is no such guarantee for any
modeling strategy.
Consider the followinggeneral screening situa-
tion: Your dataset containsnsubjects andk
predictors, and you decidekis large enough
to warrant some kind of screening procedure
to reduce the number of predictors in your
initial model.
A typical approach (let’s call itMethod 0)isto
screen-out (i.e., remove from one’s initial
model) those variables that are not individually
significantly associated with the (binary) out-
come.
Q1. Is there anything that can be criticized
about Method 0?
Q2. Should the use of Method 0 depend on
types of predictors? E.g., whether your
predictors are a mixture ofEs andCs, involve
oneEand severalCs, or only involveEs?
Q3. How large doeskhave to berelative to nin
order to justify screening?
Q4. Are there other ways (i.e., Methods A, B,
C,...) to carry out (one-at-a-time) screening
and when, if at all, should they be preferred to
the typical approach?
Q5. Where does collinearity assessment fit in
with this problem?
Presentation: III. Screening Variables 265