> Thank you, Rich, for your message.
> You are correct, 12 scales out 20 in my study are skewed. So, does that mean
> that I need to transform the data before imputing?
"equal interval" is the important characteristic for data for models and tests.
I always considered whether variables needed transformations - though, at
times, the main investigator ruled that out because there was too much (bad)
conventional practice to justify a minor gain. If you don't have equal intervals,
you can't expect homogeneity of variance ... putting the testing into question.
And how do you construct a reasonable linear model when "one unit" is not the
same at both ends of the scoring?
Still, "scales", if these are summative rating scales, are somewhat less prone to
needing transformations that those measures in natural units (which some
people naively take as "equal") like blood titers, distances, times, or counts.
But with low "counts" in a scale, you might want square root (as with counts).
The bigger treatment of nonlinearity for scales is the IRT approach, which uses
the logistic transformation since the scales are bounded at both ends.
Skewness and outliers or other non-normality are warning signs that you may
not have equal intervals. If you mention the actual variables and their distributions,
readers here would have more to advise from.
> Can you share a bit more how I can compute and recode them to the range?