16 Feb
2007
16 Feb
'07
2:39 p.m.
On Thu, 15 Feb 2007, Jan Teorell wrote:
Hi Amelia-folks, I have a few questions regarding
variable transformations
in the imputation models:
1. Do you I need to use the same variable transformations in my analysis
model as in the imputation model? If, for example, I use the
quadratic root transformation to normalize a couple of event count
variables, do I have to use the same root transformed variables in my
analysis model or could I use the "raw" variables?
in theory yes. you want at least as much info in the imputation model as
the analysis model. in practice, a linear approximation may be good
enough for the imputation model, at least sometimes.
2. I have a couple of variables (e.g., oil exports as percentage of GDP)
that are highly positively skewed, containing many zeros but with a
long tail. What kind of transformation would be advisable in that
case (I gather the natural log might not be a good idea since there
are so many zeros; moreover, the results look weird if I log these
variables)?
you could try the square root. (you can't log zero of course)
3. I have a variable (a democracy index) which ranges from 0-10 and is
pretty severely bimodal, that is, with most of the cases falling at
either low or high values. What kind of transformation could help
normalize this variable?
i'd probably leave it as is, or dichotomize it.
4. Relatedly, my dependent variable is the annual change (first
difference) of this democracy index. The imputation model does an
extremely poor job at predicting this variable (judging from the
overimputation plot), however, which makes substantively sense in
light of my analytical results. Would a better idea then be to impute
the level and lagged level of this variable, then compute the annual
change variable from these imputed variables, and run the analysis?
often imputing the components of an index is better, yes. levels are more
predictable than differences often. not sure it will help with the
confidence intervals on your quantities of interest, but it might.
Best of luck with your research.
Gary
All the best,
Jan Teorell
Docent, Associate Professor of Political Science
Department of Political Science, Lund University
Box 52, SE-221 00 Lund, SWEDEN
Ph: +46 46 222 8093 Email: jan.teorell(a)svet.lu.se
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=amelia