Anonymous posted on Saturday, May 14, 2005 - 12:18 pm
Hi I am new to multilevel modeling/growth curve modeling and am in need of a VERY basic answer to a simple (and probably silly) question. Could you please help me understand the basic differences between HLM and MPlus? I have a sample of parent-child interactions over 4 time points and want to look at linear change in positive interaction patterns over time. I have some specific child factors (i.e. temperament, medical history) that I want to use as predictors of slope and intercept. Can I do this in either HLM or MPlus or is one better suited than the other? Thank you.
bmuthen posted on Saturday, May 14, 2005 - 12:33 pm
In terms of standard growth modeling, Mplus and HLM largely overlap. Mplus has advantages of allowing more general modeling. For example, if you want your child factors as predictors of slope and intercept, in HLM you would have to enter a score such as the sum of the items as a proxy for the factor, whereas with Mplus you could have the factor model with its multiple indicators as part of the growth model, avoiding the bias of unreliable scores. Growth mixture modeling is another advantage of Mplus (see for instance Muthén, B. & Muthén, L. (2000). Integrating person-centered and variable-centered analyses: Growth mixture modeling with latent trajectory classes. Alcoholism: Clinical and Experimental Research, 24, 882-891. Contact the first author; mention paper #85).
Anonymous posted on Sunday, May 15, 2005 - 8:29 am
another basic question: In standard LGM (no covariates, only one growth factor) the correlation of the intercept and the slope is identical to a direct effect of s ON i (regarding model fit, covariances etc.). Although this effect is simply a direct function of the choice of the position of i, I am interested in this effect. In GMM, with varying means, var and cov across 2 classes, this does not seem to hold true anymore, why? More precisely: %c#1% i s | t1@0t2@1 t3* t4* t5*; i* s*; i WITH s*; %c#2% i s | t1@0t2@1 t3* t4* t5*; i* s*; i WITH s*; converges without problems and yields reasonable results, whereas %c#1% i s | t1@0t2@1 t3* t4* t5*; i* s*; s ON i; i WITH s@0; %c#2% i s | t1@0t2@1 t3* t4* t5*; i* s*; s ON i; i WITH s@0; does not converge and gives me a couple of error messages such as: THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO AN ILL-CONDITIONED FISHER INFORMATION MATRIX. Clearly, I am missing something here...what is it? Thank you very much, Manuel
It's difficult to say what is happening from the information provided. First of all, if you are not using Version 3.12, you should download it from Product Support. If you are, please send your input/output, data, and license number to firstname.lastname@example.org.
Anonymous posted on Monday, May 16, 2005 - 9:30 am
thank you very much for your quick reply! I was using version 3.11 and just updated to 3.12. Now the same syntax runs without any problems - problem resolved:-)) Manuel
Anonymous posted on Tuesday, June 21, 2005 - 2:29 pm
Here's real dumb one - Do all growth models have latent variables? Is that the nature of growth models?
If so, what would be my latent variable if I was modeling the growth over time of kids reading levels? (Putting aside any classes they are taking, etc.)
rich jones posted on Tuesday, June 21, 2005 - 7:02 pm
Here's my perspective: There are several general approaches to modeling change over time. Many involve modeling the idea that individual persons differ from one another in their starting level and in the pace of change over time. Sometimes these individual differences are called 'random effects'. Sometimes they are called (and modeled as) latent variables. Unless you're doing something fancy, it dosen't really matter what you call them or how you model them. Convince yourself of the equivalence of various approaches by viewing worked examples on Patrick Curran's web page (http://www.unc.edu/%7Ecurran/example.html).
So the situation you describe could be approached from a latent growth modeling persective, and the latent variables might represent individual differences in the starting value and rate of change over time.
Anonymous posted on Wednesday, June 22, 2005 - 11:59 am
I see, I was just letting the terminology throw me! I appreciate your help!
I am running a CFA with 40 items and 6 factors using Mplus 4. I got the following warning and was trying to locate TECH4 Output for more information. Thanks so much for your help.
THE MODEL ESTIMATION TERMINATED NORMALLY
WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR A LATENT VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO LATENT VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES. CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE F6.
Is there any problem having latent growth curve parameters predict two correlated continous manisfest variables using an MLR estimator. I don't think there are any problems, but could be wrong. The model is specified below:
Hi, I am trying to run growth model with longitudinal data. I have a file for all children. I want to use the same file for all the analyses. So I was thinking to use Useobservations and separate kids according to grade. I was writing (to take only 5th and 6th graders) : Useobservations = grade eq 5 or grade eq 6; But it doesn’t want to run. Am I doing something wrong?
I am a PhD student working with LCA for a continuous variable. I have your book, and took the course at University of Maryland in May 2006, but I am struggling to work with model fit, and deciding the number of classes. For example, what represents an acceptable log likelihood, what targets for BIC, AIC are desirable.
Are there a few seminal papers that use LCA that really walk the beginning student through that process of understanding these decision processes. LCA is a difficult technique for the new student to gain entry.
You want to look at several things to decide on the number of classes including whether the classes make sense. Bengt has a paper in the book edited by Kaplan. You can find this paper on the website under Recent Papers. I believe it looks at a growth model but the steps are the same. You are not looking for absolute loglikelihood and BIC values. You are looking for the highest loglikelihood and the lowest BIC when comparing 2, 3, and 4 classes for example.
Wei Chun posted on Wednesday, November 12, 2008 - 9:07 pm
What's wrong with my data? Mplus gives the following message:
*** ERROR Unexpected end of file reached in data file.
Muthén, B. & Asparouhov, T. (2009). Beyond multilevel regression modeling: Multilevel analysis in a general latent variable framework. To appear in The Handbook of Advanced Multilevel Analysis. J. Hox & J.K Roberts (eds). Taylor and Francis.
which is on our web site under Papers, Multilevel SEM.
The slope-intercept correlation is available in all multilevel programs.
Anne Chan posted on Sunday, January 17, 2010 - 11:24 am
Thanks! I have another basic question. I was following the mplus online video course, in which I learnt that the "intercept" of LGM can be interpreted as the "initial level" of the growth curve. But why does the value of the intercept not equal to the value of the mean of time one? Is that "initial level" not equal to "time one mean"?
The intercept growth factor is defined by the time point where the time score is zero. If this is the first time point, the mean of the intercept growth factor is equal to the model estimated mean of the first measurement not the observed mean. They are the same only if the model fits perfectly.
No Mplus uses maximum likelihood estimation, not GEE. The default covariance structure when doing growth modeling in the wide, single-level format of UG chapter 6 is that each time point has its own residual variance, where the residuals are uncorrelated across time. Many other residual correlation forms are possible in Mplus using WITH statements for residual covariances. For specification of auto-correlated residuals, see UG ex 6.17.
Emil Coman posted on Thursday, May 17, 2012 - 9:37 am
I prepared a response to reviewers on GEE vs GLVM in Mplus, summarized here [I may be wrong on some things, like 2!]: The advantages of generalized latent variable modeling (GLVM, Skrondal & Rabe-Hesketh 2004) over GEE come down to: 1. GEE is primarily a regression modeling approach that cannot be used for simultaneous regressions estimation, and therefore GEE does not model links between DVs or IV variables or regression parameters for that matter, across different regressions (including indirect effects; see e.g. xtgee Stata command, a panel data 1-DV-at-a-time modeling). 2. It is also clear that GLVM is more flexible in modeling relatively few time points (‘short panel’), where cross-lag causal links are known to vary dramatically by lag, especially when an intervention is conducted, while models like xtgee may be more appropriate for ‘long panels’ (Cameron & Trivedi, 2009, Ch. 8, p. 265). 3. The robust estimators used in Mplus, especially for categorical and count variables, WLSMV = robust weighted least squares estimator with diagonal matrix, perform as well or better than GEE (Muthén, du Toit, & Spisic, 1997) for one-regression models. Furthermore, GEE “treats the between-subjects correlations as nuisance parameters” Cameron McIntosh, SEMNET (for clustered data, for repeated measures the within-subject correlations are treated as nuisance).
Emil Coman posted on Thursday, May 17, 2012 - 9:37 am
-part 2- 4. The intra-individual clustering (the only clustering here due to repeated measures component of the model) is modeled specifically rather than ‘corrected for’ as in GEE. The repeated measurement non-independence in the data is flexibly modeled in GLVM framework however through various means, e.g. correlated errors of repeated measures (of lag one, or possibly more). 5. Model fit/adequacy with GEE is not easy to evaluate, see Skrondal, & Rabe-Hesketh, 2004 (p. 199).
I am looking at change in x over time on an outcome variable. I am doing latent growth curve modelling. I understand that I can look at how the intercept (initial level of x) and slope (change in x over time) may relate to the outcome. But I am wondering whether I could also regress the last repeated measurement point of x onto the outcome to also look at whether the latter level of x uniquely impacts the outcome variable? That way, I get information about initial level, change, and later level of x on the outcome. Would this work?