Re: st: Interactions and multiple-imputation

[Richard Goldstein <richgold@ix.netcom.com>]

take a look at White, IR, Royston, P and Wood, AM (2011), "Multiple
imputation using chained equations: issues and guidance for practice,"
_Statistics in Medicine_, 30: 377-399

Rich

On 3/23/11 8:34 PM, Nic wrote:
> Hello all,
> 
> In Alan C. Acock’s “A Gentle Introduction to Stata” (2010:367), it is
> recommended  to create interaction terms in the original dataset before
> doing the multiple-imputation stage. That’s how I’ve proceeded thus far,
> but I’m curious if I should in fact be doing so. I'll explain why below.
> 
> My survey dataset contains multiple measures of the same construct. For
> example, 5 questions are used to measure the extent of childhood
> physical abuse. In my non-multiply-imputed dataset I have created a
> single "physical abuse" scale that is the average of the 5 component
> variables. I have a small number of cases in which all 5 component
> variables are missing. I have other cases in which the respondent
> answered some but not all of the 5 component questions. For these cases
> it seems as though I should be imputing the missing values for the
> component variables and *then* creating the final scale by averaging the
> complete sets of 5 questions. Otherwise, I will end up with some cases
> in which the scale is completed but is based on averaging less than the
> 5 component questions and will not receive the benefit of imputation.
> 
> However, my interaction terms are the products of these types of scales
> (like "physical abuse"). And as I mentioned at the beginning of this
> email, the best advice according to Acock is to create interaction terms
> in the original dataset and then impute the missing interaction terms.
> 
> So I cannot do it both ways. I can either:
> 1. Create my interaction terms in the original dataset based on
> component variables which may themselves be comprised of missing values
> and then impute the missing interaction terms.
> or
> 2. Impute missing values in the original dataset with no scales or
> interaction terms created. Then, with the multiply-imputed dataset,
> create scales and then create interaction terms.
> 
> Option 2 seems to make more sense to me, but I thought it was a good
> idea to post here before I defy the advice found in Acock's book. I also
> suspect that the proper solution may be more complex than I realise.
> 
> With thanks,
> Nic
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/