>I am using SPSS for my analysis. > >I have both MCAR and non-random missing data. The non-random missing data >comes from questions about religion where some people have just refused to >answer any of the questions. > >I am planning to use EM to impute the missing data for the items with MCAR >data. The non-random missing data is a bit trickier. I don't want to dump >the items with non-random missing data because they are important. I can't >dump the people who didn't answer the items because that would introduce >bias. Although the SPSS manual suggests that EM is not suitable for data >which is not missing at random I see that as my best option.
I think it's good that you're accounting for those participants, because there is no way to "not treat" missing data - deleting those cases assumes that variable relationshis in those participants are the same as for those who are included. So no matter what, you're doing "something", and it's good to take your best shot. I think Multiple Imputation is a viable option here, and easy to work with, using Schafer's program (free download from methcenter.psu.edu). First, even with MCAR data, there are problems with using EM to estimate your parameters, namely, standard errors aren't correct because you don't have a good estimate of sample size. MI will take care of the standard error problem presented by the use of EM. However, there still exists the problem of the nonignorable mechanism, which I think MI can do a fair job with. My attitude toward nonignorable mechanisms is that *a portion* of this mechanism is actually accessible - often more than we think (see the sensitivity analysis presented in Graham, Hofer, et al. (1997)) - and making an attempt to account for as much of this as possible often yields sensible results. For instance, as Ahmad suggested, you can use other data sources to account for some of this mechanism. Too, you likely have data which will account for factors which predict this mechanism - perhaps ethnicity? I would explore the data and see what variables you have which are associated with missingness in that question. Sure, it's not perfect, but unless you're able to model that nonresponse mechanism, nothing's going to be perfect, so you need to do the best you can with it. So, for what it's worth, that's what I think. Comments are welcome! Hope it helped... Jeff Wayman, M.S., Ph.D., Tri-Ethnic Center for Prevention Research Colorado State University Ft. Collins, CO 80521 phone: (970) 491-6969 email: [email protected]
