Lee_50 posted on Friday, December 30, 2005 - 12:46 pm
I use the Mplus User's Guide and then have some questions.
1. Some papers write down the status facotr, it is the same the intercept growth factor, isn't it? ( I think it is same but I want to make sure.)
2. Fix the model, consist of two basic random-effects growth factors: an overall status factor and a linear growth facotr...so overall status factor is also the same in (1).
3. Want to use: Full-information maximum-likelihood estimation under missingness, so I search it but the option='estimator', it is not find...could you tell me? just default or ML( is the FIML)
4. Want to use: satorra-bentler corrections to chi-square tests of model fit and parameter standard errors, but if I just use the Multilevel model, I got chi-square test value. is this same??
5. Use the other kind of modelin, I got the P-value from output, but the mulitlevel model result didn't give this part(p-value) how to calcuate p-value? or it is possible in Mplus?
Some questions is not difficult. But I really want to make sure. Cuz, each book and paper..they use the different word even if the same meanning..this kinds of thing it is not good me..confuse. Thank you so much for helping. It is too help to me.
1-2. When the slope growth factor has the zero time score at the first timepoint, the intercept growth factor describes initial status. If the zero time score is at the last timepoint, the intercept growth factor describes final status. Perhaps this is what you mean.
3. Any maximum likelihood estimator will give you full-information maximum likelihood.
4. MLM gives the Satorra-Bentler chi-square.
5. I think you mean you are not getting chi-square and the p-value for chi-square. When means, variances, and covariances are not sufficient statistics for model estimation, chi-square is not available.
I have run a latent growth curve model to analyze the development of ADH problems. I have analyzed whether the development of these problems is different in the control and intervention condition, by regressing the intercept and the slope of the growth curve model on the intervention status (cond). There is a significant effect of the intervention status on the slope (not on the intercept). QUESTION: I would like to visualize the growth of ADH problems in the control and intervention condition. Can this be done in Mplus?
Hi Linda, I am running a latent growth model with 4 time points, and received this message: 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 SLOPE.
Do you see anything wrong with my code below? I tried loosening starting values for loadings to slope, and tried providing a positive starting value for the variance of SLOPE, since the message suggested that there may be a negative variance for this factor--I did indeed see this in the TECH4. Can you give me suggestions, or may I send you my data? Thank you!
Part of the problem may be that you are not specifying the growth model correctly. If you use BY, you must fix the intercepts to zero and free the growth factor means. It would be easier if you use the growth language:
Thanks, Linda! It worked! I have one more question: I saw both variance of latent intercept and slope, and RESIDUAL variance for latent intercept and slope reported in Bengt Muthen's writeup of a similar analysis, where he modeled the effect of predictors on latent intercept and latent slope: http://gseis.ucla.edu/faculty/muthen/Papers/Article_080.pdf
In my output, I do see the residual variances, including for latent intercept and latent slope, but not the variances of latent intercept and slope themselves.
Where do I find variance of latent intercept and latent slope in the Mplus output (not residual variances)?
Thanks, Bengt. Can you tell me what I am doing wrong here (see code below)?
Is it possible to estimate both the variance and residual variance of a latent variable using the delta method? I get the message: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING PARAMETER 14. THE CONDITION NUMBER IS 0.000D+00.
Bengt, I apologize for any oversight I am making, but I can't think of how to define the variances of latent intercept and latent slope using other terms in the model, except R2 + residual variance (i.e., variance explained plus variance unexplained). However, R2 and residual variance are not have explicitly named terms in the model; these are just statistics produced automatically by Mplus.
I apologize: In my previous post, I meant to write, "However, R2 and residual variance are not explicitly named terms in the model; these are just statistics produced automatically by Mplus." So, I am just not sure how to write the model constraints.
TECH4 does not have standard errors. For a well-fitting model, you can obtain standard errors for the variances by running the unconditional model where variances are estimated. This is probably the best solution in your case where residual variances at not model parameters.
If a predictor GENDER with females coded as 1 affects a latent growth factor negatively, is it more appropriate to say that females showed LESS growth from the starting point (reflected by score on Initial Status) or showed SLOWER growth from the starting point?
I have heard velocity of latent growth. How does one measure velocity of latent growth? Is this something different than what is reflected by the loadings onto the latent growth factor, or the effect of a predictor such as gender on the latent growth factor?
Linda, my quesrtion is: How does one measure velocity of latent growth? Is this something different than what is reflected by the loadings onto the latent growth factor, or the effect of a predictor such as gender on the latent growth factor? Is it reflected in how much freely estimated loadings onto the latent growth factor jump from one time point to another? For example, if you have several loadings: .00, *, *, *, *, 1.00 (with * meaning it is freely estimated), and the jump between loadings between time 1 and time 2 is a bigger jump than the loadings between time 4 and time 5, could you say that the first part of the trajectory has greater VELOCITY of growth? Is this velocity quantifiable?
Hi, Dr. Muthen! I think Lisa M. Yarnell may be referring to the linear term in a growth model. I've heard the linear term referred to as velocity and the quadratic term as acceleration. I think the mean of the linear term in the LGM is what she's looking for.
Drs. Muthen: If in a 4 time point LGM, I set error variance to be the same across time points (which I see demonstrated in the Mplus manual), is it also possible to correlate errors for subsequent time points? Given that it is the same scale administered at multiple time points, it makes sense that scores would be correlated. However, does this introduce some sort of dependency issue in the model--either due to the errors being set to be equal across the time points, or otherwise? The model below runs well without the correlated errors, but crashes when I correlate them in either of the patterns shown below (in the correlations between YEAR1-YEAR4).
The fact that the same item is repeatedly measured is what the growth models takes into account. There may also be a need for residual covariances across time. You can look at modification indices to check this. You can hold the residual variances equal across time or not. There are generally held equal in multilevel modeling programs but that is not necessary in Mplus. If you want me to look at your output, please send it and your license number to email@example.com.
I am running a conditional LGCM with multiple indicators at each time point; each indicator is an ordered categorical. In my model I have time-varying (TVC) and time-invariant covariates (TIC). TVC and TIC are included as sets of dummies, as they are all categorical too. The model runs well, however I loose 50% of the cases compared with the unconditional LGCM.
Then I tried to run the same conditional model after restating the names of the TIC and TVC, as we do for other models when we want Mplus to use all the available cases (full-information ML), for example: construct1991 0n educ1 educ2; educ1 educ2; ...
In this second case I only loose around 15% of cases, but unfortunately the model does not converge (message: number of iterations exceeded).
Please, can I do anything to avoid dropping all those cases and get my model to run?
If not, would it be correct to compare results, e.g. growth factors' estimates, between the unconditional and conditional models if the number of cases varies so much between the two?
Here's a trick that we suggest when you have missing on tics and the missingness for the tic at time t implies that the outcomes y at time t are missing. Score the tics at values not designated as missing data symbols. Then subjects with missing on tics will be included but not affect the likelihood due to missing on outcomes.
ywang posted on Wednesday, March 26, 2014 - 12:56 pm
Dear Drs. Muthen,
We worked on a latent class growth model. The outcome variable is clearly very skewed. One reviewer mentioned that it has implications for erroneously identifying latent classes in a population when we treated it as a continuous variable. What can we do for the skewed outcome variable in latent class growth modeling (and the parallel latent growth modeling linked to another outcome)? Thanks a lot!
Two issues are relevant in this situation. One is whether the classes make substantive sense based on theory. The other is whether external validity can be demonstrated using a distal outcome. See the following paper which is available on the website:
Muthén, B. (2003). Statistical and substantive checking in growth mixture modeling. Psychological Methods, 8, 369-377.
I have developed a quadratic LGCM with 5 time points (varying times of observation).
I notice that when I add a time invariant covariate (dichotomous) to the model, it will not converge, stating that there is a "non-positive definite fisher information matrix". The output also states: PROBLEM INVOLVING PARAMETER 9. Using TECH 1 I see that this parameter is the correlation between my covariate and the intercept. By chance I have managed to resolve the problem by defining time differently (time = time/24).
1. However, I am unclear as to what the initial problem was. Are you able to tell me what this might have been?
2. Also, unfortunately, this new time scale does not make much sense in my model, is there any other way of resolving this problem without changing the time scale?
Thank you very much for your reply. I believe I have solved this problem. However, I have a remaining issue, which is that the Log-likelhood of the unrestricted model (H1) appears to be more negative than the restricted model (H0 - my model). This appears to imply that the my model fits the data better than this unrestricted model, which doesn't seem to make sense. This also provides a negative chi-square when comparing the model. I am wondering if you have any idea why this might be the case, and what I might be able to do to rectify this?
Dear Linda or Bengt! I have run a latent growth model where I regress the intercept and slope growth factors on a continuous predictor X. X has a negative effect on the intercept and a positive effect on the slope while the correlation between the intercept and the slope is negative. The suspicion arises that the positive effect of X on the slope might be confounded by the negative association between X and the intercept (those high in X tend to have lower starting values and therefore more room to increase in the outcome). Would it make sense to let the slope regress on both X and the intercept? Would this give me the effect of X on the slope while controlling for the intercept?
I am trying to estimate a conditional model for a piecewise latent growth curve model using the following truncated syntax which does not show the syntax for estimating the latent intercepts (alphas) and slopes (betas):
Alpha_P with Beta1_P Beta2_P; Beta1_P with Beta2_P;
Beta1_P Beta2_P on Alpha_C; Beta1_C with Beta2_C;
Alpha_C with Beta1_C Beta2_C;
Beta1_P with Beta1_C; Beta2_P with Beta2_C; Panic_M with Alpha_P Beta1_P Beta2_P; Crit_M with Alpha_C Beta1_C Beta2_C;
I'm running a 2-group latent growth model of longitudinal model to find support for invariance (the slope loadings and residual variances at each time point) between the two treatment groups in my sample. I want to justify collapsing both groups into one sample to run subsequent analysis.
When I first ran the two groups together, the model of random linear and random quadratic slope fit well, but the variance of each growth factor was not significant..
When running the model as a 2-group, the variance of each growth factor within the each group is now significant!
Can this be due to increased power with a 2-group approach?
I am running a quadratic latent growth model with four time points, and the intercept fixed at time 1. The unconditional model yieled a non-significant negative residual variance for the linear slope term, so this term was fixed to 0. When various predictors of growth parameters are added in the conditional model, I am receiving an error that there is a non-positive definite first-order derivative product matrix, and the error indicates that the problematic parameter is the variance of the latent intercept in the psi matrix. I have tried a number of solutions, including fixing the residual variance, changing the starting values, etc., but nothing seems to solve the problem. An additional issue seems to be that the conditional model has worse fit than the unconditional model
Do you have any suggestions for how to proceed? Are the estimates for prediction of the growth parameters from the conditional model interpretable?
Thanks for the quick response! There are not any binary covariates, but there are some Likert-type covariates that are positively or negatively skewed for which the means are estimated. Could this result in the same situation as the binary covariate?