Eric: the fact that the imputations differ from the observed values is not necessarily an argument against imputation, if the imputation model is reasonable. Here it seems that the second wave variable is a good predictor of first wave -- I'd expect this religiosity variable to quite stable over time. So imputation holds some promise, based the wave two religiosity and any other variables in current and other waves that predict the missing variable in a complete case regression. I recommend multiple imputation to propagate imputation error, using something like IVEware. Rod Little
Quoting Eric Nauenberg <[email protected]>: > > > Dear Impute listserv members: > > We have a survey instrument called the Canadian COmmunity Health Survey > (CCHS-similar to the national health interview survey in the United States). > We are using the survey to study the interaction between aging, religiousity, > and health services utilization. > > The problem we are having is that the first wave of the survey in 2001 > (the only wave appropriate for our questions) has a sample size of > approximately 133,000 but the module on religiousity was only given to a > random subset of approximately 3,000. We feel that we have some good > predictors of church attendance in some of our behavioral variables such as > smoking and alcohol consumption; however, Wilcoxon tests of our imputations > with the distributions of the actual values from the sample subset > force us to > reject the hypothesis that they are from the same distribution. > > The same question on religiousity was also used in a second wave of the > survey instrument in late 2001 in which all 37,000 in the second wave were > asked the question we are interested in. Unfortunately, the dependent > variables were not available in this wave. What is useful here in that the > distribution of responses to the question of interest (frequency of church > attendance) is statistically not different than the distribution for the same > question in the first wave (according to the results of the Wilcoxon text). > > Given this information, can you suggest anything we might try which might > have a shot at imputing values for such a large percentage of a sample and > offer a test of accuracy of the imputation other than the Wilcoxon test? > > Any help you can provide will be much appreciated. > > > > > -- > Eric Nauenberg, Ph.D. > Associate Professor of Health Economics > Department of Health Policy, Management and Evaluation > Faculty of Medicine, University of Toronto > Health Sciences Building > 155 College Street Suite 425 > Toronto, ON Canada M5T 3M6 > (416) 212-6109 > ----- End forwarded message ----- > > > -- > Eric Nauenberg, Ph.D. > Associate Professor of Health Economics > Department of Health Policy, Management and Evaluation > Faculty of Medicine, University of Toronto > Health Sciences Building > 155 College Street Suite 425 > Toronto, ON Canada M5T 3M6 > (416) 212-6109 > ----- End forwarded message ----- > > > -- > Eric Nauenberg, Ph.D. > Associate Professor of Health Economics > Department of Health Policy, Management and Evaluation > Faculty of Medicine, University of Toronto > Health Sciences Building > 155 College Street Suite 425 > Toronto, ON Canada M5T 3M6 > (416) 212-6109 > > _______________________________________________ > Impute mailing list > [email protected] > http://lists.utsouthwestern.edu/mailman/listinfo/impute > > >
