Message/Author |
|
Jon Heron posted on Monday, October 04, 2004 - 9:02 am
|
|
|
Hi, having just read: "Studying Multivariate Change Using Multilevel Models and Latent Curve Models" MacCallum, Kim, Malarkey and Kiecolt-Glaser, Multivariate Behavioural Research 1997, 32, 215-253. I thought I might be able to gain a greater understanding of both disciplines by fitting a comparable(ish) model in MLwiN and Mplus. My model is height measured at 5 equidistant, time points (from birth to 1 year) + I have no missing data. I have two questions: 1) It seems to me that the latent factor created in LGM bears no relationship at all to the fixed and random effects of MLM, contrary to the paper's claim that most models could be fitted both ways. Is this true? This observation comes partly from the fact that in LGM I don't appear to be able to restrict higher order age terms to having just a fixed effect. 2) In my MLM, there is a large random effect of quadratic age and also a fixed effect of cubic age whereas in the LGM I get silly correlations and a dodgy residual covariance matrix presumably due to the high correlation between s and q. I think perhaps that my confusion re (2) stems from my ignorance re (1). Please can anyone help? thanks Jon |
|
Jon Heron posted on Monday, October 04, 2004 - 9:05 am
|
|
|
I should clarify (2) by saying that i'm not fitting a cubic in LGM. Merely that Mplus appears to be telling me that anything higher than linear is no good due to collinearity, whereas I have found a sig. cubic in MLM. |
|
bmuthen posted on Monday, October 04, 2004 - 3:14 pm
|
|
|
MLM and LCGM give the same results when setting up the models to be the same and using the same time scores. In Mplus LCGM you can fix the variance of any of your growth factors (any of the age terms) to zero. Often problems arise with models having quadratic terms if the time scores don't subtract the average time, i.e. use age-agebar and (age-agebar)^2, where agebar is the average age in the sample. If this doesn't help, as a licensed user you can send your input, output and data - plus your MLwiN output - to support@statmodel.com. |
|
Jon Heron posted on Wednesday, October 06, 2004 - 7:32 am
|
|
|
I think I'd better do that. Jon |
|
Jon Heron posted on Thursday, October 07, 2004 - 3:03 am
|
|
|
I have realised (perhaps in error) that the fixed and random effects from MLM are effectively the same as the means and variances of the growth factors. I don't know why that didn't occur to me before. The next step in my MLM would be to add some level-2 covariates either fixed or random. I assume that the 'equivalent' in LCGM is to regress the growth factors on these covariates but that here the results will no longer be directly comparable? |
|
|
You are correct in your first paragraph. Adding level 2 covariates still gives equivalent results between MLM and LCGM. |
|
Jon Heron posted on Thursday, October 14, 2004 - 2:27 am
|
|
|
So, if I knew how to do it, I could take the multiple estimates obtained from regressing the intercept, slope and quadratic factors on my covariate, and combine them in such a way to give me the single fixed-effect estimate given by MLwiN? I'm not asking how one might do that, but it'd be good to know that that was the case. many thanks (again!!) Jon |
|
bmuthen posted on Thursday, October 14, 2004 - 11:38 am
|
|
|
I don't see that your situation would give a single fixed-effect estimate from MLwiN. If you have i, s, and q as growth factors and a single x, this x generates 3 fixed effects (the slopes) and that is what you find in MLwiN as well. |
|
Jon Heron posted on Friday, October 15, 2004 - 1:54 am
|
|
|
Oh dear, back to the drawing board |
|
|
Why don't you send your MLwiN and Mplus outputs along with your data to support@statmodel.com and we will show you how to set the MLwiN model up in Mplus. |
|
Back to top |