OK, here's a further question, then: if I interact my term for each
treatment (a 0 or 1, effectively), when the treatment value is zero, should
I force a zero in the product of the treatment and varX when the interaction
term is missing (since 0 * varX always =) and only impute where the
interaction is 1 * varX (since I don't know what varX is in that instance)?
Or should I impute both interaction terms?
On Wed, Mar 2, 2011 at 11:49 PM, Matt Blackwell <blackwel(a)fas.harvard.edu>wrote;wrote:
Hi Donald,
There's no real way to combine imputations from separate imputation
models, since they would have completely different underlying
parameter value draws. I'm not sure what separate imputations would
accomplish, though. It would still estimate separate mean values of
the outcome for each treatment group, but would also "interact" the
treatment with the all the independent variables. That is, it allows
the relationships between the independent variables and the dependent
variables to vary across the treatment groups. You can accomplish the
same effect by interacting the treatment group binary variables with
each of the independent variable. With this, you can impute everything
all together.
Cheers,
matt.
On Wed, Mar 2, 2011 at 11:13 PM, Donald Braman <dbraman(a)law.gwu.edu>
wrote:
I have data from an experiment with a control and
four treatments. There
are several independent and dependent variables, and all have missing
data
(which, for the sake of argument, assume occurs
randomly). The
independent
variables can each be imputed from other
independent data *across* the
entire experiment. But the dependent variables, I take it, should be
imputed only from other data *within* each condition. Is that right? Is
there some way to specify this within Amelia? If not, do you recommend
imputing all the independent variables across conditions, then merging
them