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

Erroneous data

+

Correct if possible

Drop if not correctable

Correct data

+

Decision to delete up to researcher

Report and interpret in“discussion”

Inappropriate model?

+

Difficult to decide if due to

influential observation

Summary:

Be careful about deleting.

Conservative approach:

drop subject only if

uncorrectable error

Certainly, if a subject’s data is erroneous, it

should be corrected if possible. If such an

error is not clearly correctable, then the subject

may be dropped from the analysis.

However, if the data on an influential subject is

not erroneous, the researcher has to decide

whether the subject should be dropped. For

example, if the subject is much older than most

subjects (i.e., an outlier), the researcher may

have to decide whether the age range initially

allowed needs to be modified. Instead of deleting

such an individual, the researcher may wish to

report and interpret the presence of influential

subjects in the “discussion” of results.

It is typically difficult to determine whether a

large influence statistic results from an inappro-

priate model. Since the initial model is rarely

one’s final (i.e., best) model, a final decision as

to whether a given subject is influential should

wait until one’s final model is determined.

In summary, the researcher must be careful

when considering whether or not to delete an

observation. A very conservative approach is to

only delete an observation if it is obviously in

error and cannot be corrected.

Presentation: V. Influential Observations 279