1
$\begingroup$

I am running a second-order latent growth curve model with seven repeated measures. I have 2 time-invariant covariates (TIC, country and sex) and 2 time-varying covariates (TVC, both are latent variables (5 items each), seven repeated measures). All of the variables were measured at the same time.

I have already run an unconditional linear growth curve with my focal variable, let's say, A. The fit is good for this model. Later I ran a conditional model with the TICs. The fit decreased but was still good.

However, when I added the TVC, variables B and C, the model did not fit. (I did not add both at the same time. One by one as separate models) I understand that it means that both of the variables shifted their corresponding value of variable A in such a way that it does not fit a line anymore.

My first question is, if the aim of the analysis for the model with TVC is to see how much they shift the value of A in a certain direction, is it plausible to make an argument that "the model no longer fits a linear trajectory owing to the effects of the TVCs" and not respecify the model to make it fit? (or find a better-fitting model)

My second question is, does the worse fit for a linear model indicate that the effects of the TVC on variable A are inconsistent? I am assuming that if the effects of the TVC on variable A are constant, a linear trajectory should still fit.

I have a wald test code in my syntax to test this (I saw it in a textbook) but am not really sure how it works or if it is applicable here. Below is the syntax. (I use Mplus)

    AE1 ON AS1(P1);
    AE2 ON AS2(P2);
    AE3 ON AS3(P3);
 model test:
   P1=P2; P2=P3; 

Note. AE is my focal variable and AS is one of my TVC. The numbers following them indicate wave of data collection.

Thanks to everyone in advance who gave a thought to this post.

$\endgroup$

1 Answer 1

1
$\begingroup$

The misfit when you add additional variables with multiple indicators can have many reasons (especially in a complex model like yours), including problems with the measurement model (e.g., correlated error terms within or across constructs, cross-loadings). It is impossible to say what the reasons are without conducting a thorough analysis of your model, data, and results. Examining model residuals and/or modification indices can sometimes be helpful to locate the sources of misfit.

$\endgroup$
2
  • $\begingroup$ I see... this tells me that I need a closer look at my results. Thank you for pointing out some points to look at. Does your answer mean that I cannot interpret the model with bad fit (as is) when my aim is to see the effects of the TVC on each time point? $\endgroup$
    – EmH
    Feb 7 at 21:28
  • $\begingroup$ I would be hesitant to interpret any parameter estimates in a misspecified (not well fitting) model because the estimates may be biased. $\endgroup$ Feb 7 at 21:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.