Anonymous posted on Friday, September 03, 2004 - 11:42 am
I try to use your article Muthen & Currant (1997) "General Longitudinal Model.." as dipicted in Figure 5. I am using the following set up but it doesn' t work. S2 it is supposed to be the added growth. What changes need to be done in this program. Thanks
VARIABLE: NAMES ARE id y1 y2 y3 grp; USEVARIABLES y1-y3 grp; MISSING ARE .; GROUPING is grp(1=LGM 0=ADD); ANALYSIS: TYPE=MEANSTRUCTURE;
I am also interested in the model depicted in Muthen & Curran (1997) and not quite sure how to implement it in Mplus. In the above reply did you mean the Day 2 handout from the two-day workshop on multilevel modeling in Mplus? I can't find the needed input there. Thanks! Rick
thanks Linda! As always, you guys rock! What confuses me is that it looks like the mean of t is set to 0 in the control group [t@0] but the variance is not set to zero in the control group so doesn't this factor still exist in the control group? I am also confused as to why the t@0 and q@0 statements are included in the overall model statement, perhaps I am misunderstanding these statements as I thought they set the variance of these two growth factors equal to 0 (and I can understand why one would do this in the control group for which the Muthen & Curran Fig 5 shows that t does not exist but am not sure why one would state this in the overall model section so that it would apply for the treatment group too)
The overall part of the model sets the *residual* variance of t at zero. This setting therefore holds in both the treatment and control group and says that in the treatment group t regressed on i has R-square 1 (so t varies as a function of only i) - it just turned out that this residual variance was not needed. Since the residual variance isn't mentioned in the control group, it is still fixed at zero in the control group. There was no indication of a need for a q growth factor variance so q@0 in the overall part of the model makes q a fixed effect.
thanks I think that helps Bengt. Just to make sure, it is the "t on i@0" statement in the control model combined with the residual variance of t being set to zero in both groups that sets the variance of t to zero in the control group, right? and combined with setting its mean to equal zero in the control group makes this factor specific to the treatment group? And the t on i statement models the treatment-initial staus interaction discussed on pp. 379 - 381 in Muthen & Curran? What if one didn't model the treatment-initial status interaction - how would one make t specific to the treatment group in this case?
great! thanks again! my last question was not yes-no though. It was, if one didn't model the treatment-initial status interaction and so didn't have the t on i@0 constraint in the control model, how would one make t specific to the treatment group? That is, how does one make t vanish in the control group without the t oni@0 contraint (which wouldn't be relevant if one didn't model the treatment-initial status interaction)?
when I have t@0 in the overall model (for a similar model I am currently conducting power analyses on)I get 2 warnings that the PSI matrix is not positive definite in either group with the problem involving variable t. If I put the t@0 only in the model for controls I get the same warning for the control group. Are these warnings ignorable?