I am running a series of unconditional and conditional growth models using the TYPE = RANDOM MISSING command statement. I am also using the TSCORE command in conjunction with the previous code. MPlus does not calculate standardized estimates for this combination of code, at least according to the output files that I have evaluated. Is there a way to get these values via MPlus or otherwise? Thanks in advance!
I would like to follow-up with a question about missing data. I am again using the TYPE = RANDOM MISSING command along with TSCORE. My assumption was that the missing command would use all possible data in the analysis. So, for example, if I was missing data on the last time-point for a participant but they had data for the previous 3 time-points, this person would still be included in the analysis. My sample goes from 1300 to 584 because I am missing cases on the X-variables according to the MPlus output file. I have examined all the missing data patterns and the program is basically doing listwise deletion. Do I need to use the PATTERN command to avoid this, or something else?
Individuals with missing on x variables - which includes Tscore variables - will be deleted. The solution is to not include such variables in the Missing = statement - this gives the correct ML estimation if whenever there is missing on such a variable for a given time point, there is also missing for the corresponding outcome (check your data carefully to see that this is the case). Such data points are not contributing to the ML estimation due to the missing data handling, so the missing data on x has no effect. This should give you the sample size you want.
Thank you for your quick and thoughtful response. I have a follow-up question if you do not mind. You mentioned that I should not include such variables on my Missing = statement. I do not. In the analysis section I simply have:
Type = Random Missing;
Does this assume by default that these variables are included?
The Missing = statement is in the VARIABLE command, not the ANALYSIS command.
C. Lechner posted on Thursday, September 12, 2013 - 5:40 am
Just a quick question: When p-values of the significance tests for unstandardized vs. standardized parameters in a growth model differ - which of the two should one rely on? Is there any clear rationale favoring one over the other? Thanks!
This happens because the parameters have different sampling distributions. I would tend to go with the model estimated unstandardized parameters. I would also be conservative given that many tests are usually performed at the same time and a Bonferroni type correction should probably be made.
C. Lechner posted on Thursday, September 12, 2013 - 7:36 am
Thank you, Linda. Allow me a follow-up question: Suppose I have a set of three dummy variables that, together with the reference group, represent four levels of the same (ordinal) variable. I am interested in the effect of these dummies on the latent growth factors (i s q) in a quadratic growth model. (1) Is it correct that, rather than looking at the z-tests of the single dummies, I would evaluate the influence of the whole set of dummies via a chi-square difference test comparing a model with the paths from the dummies to the growth factors fixed to zero with an unconstrained model? (2) Is it correct that with such a set of dummy variables (where all dummies code for the same construct) one must not interpret the standardized parameters of the paths from each dummy to the latent growth factors? I am interested in whether, at least descriptively, there is a clear dose-response relationship for the levels of the variable represented by the dummies – I would look at the unstandardized, rather than standardized, coefficients of each dummy, right? Thanks!