|
I have just performed my first multiply imputed multiple regression
analyses (using Schafer's freestanding version of NORM for Windows), and the
results have brought up a question for me that I'm hoping listmembers will have
some thoughts about.
The multiple regression analyses all involved the same set of 6
predictors with a number of different dependent variables, using a dataset with
a sample size of 613 cases. I conducted each regression model five
times using the five imputed datasets that I generated with NORM. I should note
that most of the missing data in these analyses were in the dependent variables
and not the predictors. For these variables, data were available from only 569
to 578 participants.
What was most surprising to me was the huge variability in the degrees of
freedom generated by the analyses. For example, age was one of the predictors,
and df associated with this predictor varied from 21 to 7012 for different
dependent variables. The "missing information" statistic for age was similarly
variable. Neither the df nor the missing information statistic seemed
to correspond to the actual percentage of missing values in the
predictor or the DV.
I'd be grateful if folks on this list could help me interpret such results.
For example, what does it mean that the missing information statistic can vary
so widely for a predictor when the actual % of missing values is constant among
DVs? Thanks in advance for your thoughts on what is probably a very basic
question!
Best
Jon
__________________________________
Jonathan Mohr, Ph.D.
Assistant Professor Department of Psychology Loyola College 4501 North Charles Street Baltimore, MD 21210-2699 |
