[amelia] Re: to impute or not to impute?

Gary, I'm stumped. Can I ask what you would do in the following situation? I have an index composed of 4 opinion items. For each of the 4 items, I have DK responses from 5-20% of respondents, but typically the DK is on only 1 of the 4 items (i.e. DK responses are not highly correlated across the items of the index). Do I: a. delete listwise and lose 20% of my respondents, losing efficiency and introducing massive bias (my DK respondents are poor, female, low educ, low political interest, etc.) -- clearly not my preferred option b. Impute the missing data for each item and construct the index using the imputed values. I wouldn't normally want to impute for an opinion variable, but if the respondent was able to answer 3 other closely related questions, why should I believe that s/he couldn't also answer the fourth? But does EMis still work if I've transformed the imputed data eg. by smushing it into an index? c. impute the index score. Again, some of these DKs are legitimate, but others aren't, and this would get around the issue of transforming the imputed data. But it would also mean throwing out information from the three items that the respondent DID answer. (OK, not throwing out, because I'd use that info to impute, but still...) d. compute the index score for respondents with one DK out of 4 items by setting the value of missing items at the mean of the remaining items. This seems to me to be taking less than full advantage of the other information in the dataset about how these people might have responded. e. treat each item separately as a categorical variable. Messy and not nearly as much fun as working with this index. I'm not crazy about any of these choices. What do you think? Any advice much appreciated... Julie

Julia Lynch Assistant Professor Department of Political Science University of Pennsylvania 202 Stiteler Hall Philadelphia, PA 19104 tel 215 898 4240 fax 215 573 2073 email jflynch(a)sas.upenn.edu