I estimated a growth curve model with 5 waves 2 latent predictors (observed indicators for these latent constructs are continuous, categorical and nominal) and 5 observed controls. I would like to plot trajectories for latent construct values. Is there anyway I can get/calculate wave specific estimates for Low vs High values of the latent constructs for example?
There is not an automatic way to plot this in Mplus. But you can compute from the estimated model the estimated means and variances for the latent constructs at each time point. Then you can choose a low/high value for the construct at each time point as say -1SD/+1SD off the mean. For those values you can then compute the observed variable mean at each time point and plot that. You can do this for example in Excel.
Thanks Bengt that was helpful. I have another question regarding assessment of model fit. As mentioned in my previous note, my model contains 2 latent predictors where the observed indicators are a mix of continuous, categorical and nominal observed measures. Using MLR, MPLUS does not produce the standard fit measures that I usually use to assess model fit. Any suggestions on what to use in this case to show/evidence model fit.
You can do difference testing using the loglikelihood. With MLR, you need to use the scaling correction factor provided in the output. This is described on the website under Chi-Square Difference Test for MLM and MLR.
Yangjun Liu posted on Tuesday, July 02, 2019 - 12:21 pm
Hi, I am establishing a LCM model with binary outcome and one latent factor as independent variable. I found that by defacult Mplus did not correlate the latetn factor with other covariates. Should I specify that they are correlated? And I also have some difficulty in interpretating the coefficient of the latent factor. I know it represents the log odds ratio. But what is the unit of the latent factor? if it has no unit, how to interpret the results? I am a beginner in Mplus and LCM, maybe my question is very basic, but thank you for your kind answers in advance.
I would regress the factor on the covariates. The coefficient for the factor is a log odds ratio if you are using the ML estimator which defaults to logit parameterization, not if you are using the WLSMV or Bayes estimator. The unit of the factor is its standard deviation (the mean is zero) which is the square root of its estimated variance. You can also look at the STD solution in which case the factor SD is 1.
Yangjun Liu posted on Wednesday, July 03, 2019 - 7:01 am
Thanks a lot for your answers. Further question base on the above reply: (1) Could you explain a bit more about why regress the latent factor on the covariates instead of correlating them? They all independent predictors of the latent intercept(and/or slope). If regress the latent factor on covariates, and latent intercept also regress on both latent factor and covariates. Then it seems that latent factor works as mediating covariable (latent intercept -- latent factor -- covariates). (2) When I interpret the coefficients of the predictors, is it only the latent factor that I should refer to the STD result? But the the loading for the intercept and slope also be standardized, does the standardization of other variables affect the interpretation of the latent factor? And for other observed covariates, is it right that I refer to the unstandardized results? (3) another new question, if the independent variable assessed by questionnaire, and the questionnaire has two dimensions, one is method factor, should this method factor be included in the latent curve model? e.g regress the intercept/ slope on the method factor. Many thanks for your further reply.
(1) When you correlate the factor with a covariate, you add a normality assumption for the covariate which may not be suitable, e.g. or a binary covariate. And often covariates describe subjects backgrounds so that having it as a predictor of the factor makes sense. If you add a direct effect from the covariate, the mediation situation is not restrictive.
(2)Use STDYX for all coefficients.
(3) This is a general analysis strategy question which is suitable for SEMNET.
Yangjun Liu posted on Thursday, July 04, 2019 - 1:14 am
Thank you very much for your clear and helpful reply. In terms of the interpretation of the coefficient, I still feel a little bit confused. My model look like this:
USEVARIABLES ARE y1 y2 y3 sex age a b c d e; CATEGORICAL ARE y1 y2 y3; Model: x by a b c d e; d with e; x ON sex age; I S | y1@0y2@1y3@2; I WITH S; I on sex age x;
Analysis: ESTIMATOR = ML;
(1) Does it mean that for the coefficients of the independent variables I refer to the STDYX. And I also found that the p value of some coefficient is not significant in unstandardized results but is significant in standardized result. Is this the right case? (2) How about the mean of slope, which represents the log odds ratio of one unit change in time, should I still refer to the unstandardized result since the loadings of slope have been changed in the STDYX. (3) Regarding the correlation between intercept and slope, which result should I look at? unstandardized or STDYX? Thanks a lot for your very kind reply again.
(1) You should look up the entry Standardized in the UG and read about our recommendations.
For (2) and (3) you should consult SEMNET since this is really general growth modeling learning topics. But I'll say this:
(2) You can talk about how a change in s of 1 SD units without looking at the standardized solution. For instance, if the variances estimate is 9, you consider a 3-unit change in s. See also our Topic 4 video and handout for the schizophrenia example.
(3) Unstand'd gives their covariance, standardized their correlation.
Yangjun Liu posted on Saturday, July 06, 2019 - 12:37 am
Thanks a lot, Bengt. You really gave me helpful guidance on my way of learning LCM and Mplus!