Specifically we were interested in modeling
case-control status as a function of several patient
covariates including serum creatinine which in our
data ranges from 0.11 to 1.98.
Because of skewness and to make the odds ratio
independent of the units measurement, we decided to
log-transform the creatinine values before entering
them into our logistic model. However the reviewer
wrote "Using a log-transform for creatine is absurd
because a 1-unit increase in ln(x) is equivalent to
increasing x by a factor of 2.718 which is in the
realm of impossibility"
Is he correct? Isn't the coefficient estimated such
that the predicted values are within the range of the
data and this only a problem if you attempt to
extrapolate beyond the data range? What I am missing?
While I am not an expert in creatinine (whatever that is) I am inclined to
agree with you. You can always plug in implausible/impossible numbers and
come up with a prediction, e.g. how much would somebody make if they had
-2,000,000 years of education? I've never heard of a rule which says that
x = 1 has to be a plausible or even possible value. For presentation
purposes, you might want to scale your variables in ways which make them
easier to understand and present (e.g. measure income in thousands of
dollars rather than in dollars) but it is not essential. There may be
other good reasons for not doing what you are doing, but the reason given
seems odd to me, unless maybe it violates some sort of convention in your
field. If you want to make this reviewer happy, maybe you could measure
creatinine in milligrams instead of grams or whatever happens to be
reasonable so that a 1 unit increase in x is possible.
