[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: RE: Questions on multiple imputation
--- Mark Horowitz wrote:
> I have a variable that records the year an event occurred, but about 10% of
> the observations are missing. When I run -ice-, some of the imputed values
> are future years... However, the mean imputed values for each missing
> observation are actually ok (<2007). Are the coefficients, standard errors,
> etc still ok? If not, how do you limit the imputed values to < 2007? I
> don't see how the -ice- syntax allows for this.
Multiple imputation involves replacing missing values by reasonable guesses.
-ice- is a program that uses your model (e.g. -regress year x1 x2 x3- to
estimate such reasonable guesses. It is up to you to convince your readers
that your model is reasonable, -ice- can't do that for you. If you can
convince your audience that your imputations are reasonable enough, you are
fine, and if you can't you are in trouble.
The variable year is probably discrete, and if the number of values isn't
too large you can use -mlogit- instead of -regress-. This will ensure that
your imputation remain within the observed range. An alternative is to
specify the match option. This will also ensure that the imputations remain
within the observed range.
> Also, when I used -micombine- along with -reg- (micombine reg [varlist]),
> the ANOVA table, R-squared, RMSE, etc did not show up. Is there a
> theoretical reason why these don't appear?
You can calculate average R-squares and RMSEs over complete datasets. I
wouldn't do that for the ANOVA table, since that might encourage people
to do tests that aren't justified for Multiple Impution. A lot of this
was explained in the following post and the references therein:
Hope this helps,
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
1081 HV Amsterdam
Buitenveldertselaan 3 (Metropolitan), room Z434
+31 20 5986715
* For searches and help try: