Good morning. I ran a LGM analysis with type=missing and estimation used ML. However, the model sample size seemed to include all of the covariates in addition to the six repeated measures. I wrote Linda on this list and she (you) suggested I use type=missing random to use all available data on the y variable. This worked, however I noticed that the estimator is MLR. What is the reason for this?
I ran the model with ML estimation and the results were slightly different (i.e., one effect that was significant at p=.05 is now significant at (p<.05). Is this problematic? Can I simply use ML instead of MLR to stay consistent?
I have complete data on y_pre (n = 177), but not on y_post (n = 137). Initial runs indicate that number of observations was 137, but then I learned through the UCLA-IDRE site (http://www.ats.ucla.edu/stat/mplus/faq/fiml_counts.htm) that if I estimated the mean of y_pre (i.e., "[y_pre];") it would include all cases. Indeed, doing say changed the number of observations to 177.
However, I have some questions about how to report this and also how to interpret the findings with regard to statistical power.
Is it most appropriate to report the n as 137 with the caveat that MPLUS made use of all available data for the sample of 177, or vice versa (n = 177 with caveat about the 137).
Also from what I can tell, the statistical power to detect effects (of x) doesn't change regardless of whether I include the y_pre mean or not. Is that correct--that inclusion of the mean for y-pre may affect the estimate for the coefficient of x but it will not increase the power to detect an effect?
Thank you for your help and any opinions you may have on how to report this. My interest is to make sure I am accurate, clear, and above all, not misleading.
Only observations without missing on y_post contribute to the estimation of the regression coefficient y_post ON y_pre irrespective of the sample size shown when you include y_pre in the model.
Mark LaVenia posted on Tuesday, February 02, 2016 - 11:30 am
Thank you Linda. After searching further on the statmodel discussion board, I do see now that you have addressed this multiple times before. Sorry to have you repeat yourself.
I have to admit however, I am still not entirely clear why--if the cases without y_post do not contribute to the regression coefficient--the estimate for the regression coefficient nevertheless changes. I suppose it is just a function of now applying the distributional assumptions on the y_pre (???).
Is there a source you can recommend that speaks further on this issue? Thanks.