I have used growth curve modeling based on raw scores. I wonder if I can somehow incorporate multiple scores by multiple reporters on the same construct. Using standardized scores at each measurement wouldn't make sense. What would be the best way, if any, that I can look at the growth curves of the scores that "combine" multiple reporters? May I use factor scores based on confirmatory factor analyses at each occasion? Thanks very much!
Researcher posted on Tuesday, December 20, 2005 - 7:32 am
My model is like the one in Example 6.14. Is it okay to let the residuals of the factor indicators to correlate (=three factor indicators, two indicators/ factor)? Otherwise the RMSEA estimate is too large.
Also, why the growth curve examples in the manual are described in terms of random effects of the factor indicators? I see many researchers using the program by defining intercept and trend factors with the usual "BY" command.
bmuthen posted on Tuesday, December 20, 2005 - 8:18 am
Yes, if the residual correlations for the factor indicators make substantive sense.
The term random effects is used by statisticians for what behavioral scientists now tend to call growth factors - same thing. You can use the BY statement or the newer growth language using the bar ( | ) statement.
It's awesome to know that I can use Mplus to test a "multiple indicator growth model" that looks like a second order factor analysis model. Thanks for your helpful answers and guidance. Merry Christmas!!
I would like to use factor scores from CFA in a growth curve model (I do not have enough power for a multiple indicator growth curve model). Is there a way that I can use the mean and SD from the first measurement occasion for the repeated measures so that they represent changes in SD units across occasions?
I would like to use saved factor scores from CFAs across several time points in a growth model. Are the saved factor scores standardized within the sample and, if so, would this affect my ability to detect change over time?
Thank you in advance for any input you can provide.
Thank you for your response. I have a few follow-up questions:
Perhaps standardized is not the correct term, but are the factor scores calculated in reference to others in the sample (i.e., indicate position within the sample, rather than a raw score)?
Re: doing the analysis in 1 step, I have 36 factor indicators. Would estimating all of the latent factors across time and using the resulting latent variables in the growth analysis in a single input file overcome issues caused by using saved factor scores in a growth model?
If CFA model is identified by fixing variance of factors to 1 and then you save factor scores, how are they not standardized? I ask as well given reports in the literature (e.g., Wade et al., JAMA PSY, 2018) where saved factor scores are used to model growth traj and in those processes the growth appears to be tracked by the variation around a mean of zero (within group) but it is less clear how one would compare means between groups if mean slope is zero to begin with.
The factor scores are not standardized in the sense that after they are computed they are standardized; they are not. I agree with your implication that it is improper to do CFA at each of several occasions, then computing factor scores, and then using those factor scores to do growth modeling. You need measurement invariance over time for that, allowing the factor means and variances to vary across time as they are expected to do in a growth model.
A) If one does growth modeling with factor scores (assuming MI for the moment), the growth would be at the interindividual level (between) and one could predict deviation from this if the i and s factor variances are significant. Here b/c wide, one could first test for MI and then model the growth.
On the other hand, with multilevel CFA, one could truly look at intraindivudal change (within) with the drawback being that MI is 'assumed' for all intents and purposes.
Would the above be an accurate take?
B)Would it make a difference to the growth part if (after establishing configural, metric, and scalar MI on indicators) one scales the factors by fixing variances to 1 as opposed to marker indicator approach?
Thanks so much Bengt. Are any of the submitted 2018 papers (referenced in the SMEP address slides) available as preprints? Judging from their titles they address precisely the issues I am wrapping my head around.
I am running a series of growth/autoregressive models, and was originally planning on just using the observed, mean score for each scale (the complexity of the model prevents me from using latent factors for each measure).
Someone recommended that I use the factor scores instead of the mean score, to better account for the fact that items load onto the factor differently.
When conducting the CFAs, I set the factor score mean to zero and variance to 1 to set the scale. However, as noted in a comment above, I then realized this was problematic for the growth models (as now all of the means were identical).
Would simply using a different scaling method (e.g., using a marker variable) to estimate the means for each factor solve this issue? Thanks in advance for your help.
I would not go there. To make sure you get the factor scores in a comparable metric, you would have to analyze all time points together imposing measurement invariance - and then you are basically doing as much work as the longitudinal model with factors.
Thank you for your response. One clarification, the problem is not so much that the items load differently across waves, but instead the reviewer noted that within each wave, the items load with varying strengths (i.e, item 1 = .55, item 2 = .89, etc). Her argument was that when you take a scale average, you are assuming each item contributes equally, which the CFA suggests is not the case.
Her solution was to export factor scores and use those, rather than mean scores. However, because the mean of the factor scores is zero, when I run the growth models the slope is not significant (which is not accurate based on the mean scale data). I'm not quite following how the measurement invariance would correct for this?
I understand the reviewer concern about different loadings within time point. My answer concerns avoiding the problem you mention about factor score means being zero if you do them one time point at a time. To avoid that, you have to do what I mentioned using measurement invariance.