I am estimating a growth curve model with couple as the unit of analysis. I have data from husbands and wives over 10 annual assessments, and some data are missing. The model runs just fine, but I have 3 related questions.
First, I am not sure how to interpret the section of the SAMPSTAT output where 'estimated sample statistics' are presented. I have ANALYSIS: TYPE= MEANSTRUCTURE MISSING H1; From p. 112 of the user's guide, it appears that the means presented here are 'unrestricted' means, but I am not sure how they were calculated and how they should be interpreted. I want to plot the means over time for each spouse using the best estimate of the means given information from the total sample (as opposed to plotting means from either the subsample with complete data at all assessments or the means from only the valid cases at each assessment). Are the estimated means giving me the information I am looking for?
Second, I would like to obtain a parallel set of means controlling for follow-up status over time (e.g., divorced, withdrew from the study, contributed adat at all assessments). This would be analogous to obtaining means adjusted for covariates. Are such means obtainable?
Finally, when I obtain intercepts, slopes, and quadratic efefcts via orthogonal polynomial contrasts, it is clear that the effects are not really orthogonal. I assume, as in regular multiple regression, this is because the number of observations varies at each assessment. I know that Kirk in his book on experimental design has an appendix on how to construct orthogonal coefficients with unequal ns. Are these procedures relevant in the latent growth curve (or multilevel) context? I can interpret the linear and quadratic effects via the standard hierarchical method, but was wondering if there was a way to construct a set of contrasts for the intercept, slope, and quadratic effect for each spouse so that the effects are truly orthogonal.
The sample statistics that you obtain with TYPE=MISSING MEANSTRUCTUE H1 can be interpreted as regular sample statistics. They are estimated in the same way that the model is estimated using ML and MAR. They would be the values that you want.
I'm not sure I understand your question about means adjusted for covariates. But if you want means, for example for those who withdrew from the study, you could subset on these individuals and do the analysis in the same way as for the complete sample.
I don't know of any reason that the principles in Kirk would not apply here. But I wouldn't know how to do it.
i develop a growth model which has 3 dummy variables. whereas d1=1(high income countries) d2=1 (middle income countries) d3=1(low income countries).how to interprete these dummies?how do i know that the model is robust?is it because these dummies are not significant when i regress the model?
Hi Linda, is there a way to obtain the significance of the correlations in the ESTIMATED SAMPLE STATISTICS section of Mplus output?
The correlations in this section of Mplus output are exactly what we need, yet we don't know whether the correlations displayed are significant or not.
I think we'd prefer to get the correlations from this default matrix rather than from a model created in the MODEL line because we have some categorical variables, and when we put those in a so called "agnostic" model (not a causal model, using only WITH commands), we get the message "*** ERROR in MODEL command. Covariances for categorical, censored, count or nominal variables with other observed variables are not defined."
SAS does give signficance of correlations between categorical and continuous variables, and among categorical variables. There is a tetrachoric option in SAS.
What if we want to use TYPE=COMPLEX so that we can apply weights and estimate the correlations for select subgroups? This requires the STRATIFICATION command and TYPE=COMPLEX.
We could alternatively choose to not apply sampling weights at this stage of our analyses, and apply the weights later when we move on to causal modeling. If we choose this, we could just specify TYPE=BASIC and use the USEOBSERVATIONS line to select our groups.
I just wonder if we can obtain the SEs for the correlations with TYPE=COMPLEX, used for subpopulation weighting.
Jen posted on Wednesday, October 01, 2014 - 7:52 pm
This is probably obvious, but I cannot figure out why the standard errors in the "SAMPLE STATISTICS FOR ESTIMATED FACTOR SCORES" are so large. I think they must not be standard errors, but I am uncertain what else they would be. For example, I have an intercept and slope. In the "MODEL RESULTS" section, they have M (SE) .581 (.023) and -.222 (.014), respectively. Under "SAMPLE STATISTICS," I = .581 and I_SE = .566; S = -.222 and S_SE = .345. What are these _SE values?
(I am using a latent change score model and with trajectories fit to change scores, and I would love to interpret the means of the change scores to show the slope of my quadratic trajectory at each time point, thus my interest in these values; the I & S are just easy examples.)
Two things. First, I assume you have an unconditional growth model (no covariates). Second, when you look at the sample stats for the estimated factor scores, perhaps you are looking at the mean section where "I_SE" would be the average SE estimate for the factor scores over subjects.
If that doesn't help, you have to send a full output to support with license number so we can see exactly what you are referring to.