Re: range issues in ANCOVA

Isabelle <isabellesup@xxxxxxxxxxx> wrote in
news:2384821.1180230250542.JavaMail.jakarta@xxxxxxxxxxxxxxxxxxxxxx:

A research scientist told me that he wants me to study some of the
ranges issues that may exist with the predictor variables in his
ANCOVA or Linear Model Analysis.

His understanding of statistics is intuitive, rather than rigorous.

The only range issue I can think of are that which arise when a
categorical var has the values 1 and 2 (say man=1, woman=2) instead of
the values 0 and 1. If one wants the gender coefficient to be
meaningful, one would rather use a 0/1 variable than the 1/2 variable.

Is that really the case? Won't the coefficient be determined by the
difference between the numeric values assigned to the two groups? You might
run into interpretability problems if you were working with interactions,
but in such situations you should always be plugging sample case values
into the full estimated model, rather than looking at the separate
coefficients anyway.

A potential problem, although I am not sure it is called a "range effect",
may arise when the range of a covariate is fairly narrow in the sample
compared with its range "in the wild". You may see a higher ratio of the
dependent variable variance to the covariate range, causing wide confidence
intervals (low power to identify effects).

--
David Winsemius
.

Loading