2

I have an SEM that fits well according to several metrics and returns meaningful results. However, I'm not sure if the structure is valid. I have 3 between subjects conditions (2 treatments and one control), one potential continuous moderator, and two continuous DVs. One DV has three subscales and the other DV has 4 subscales.

My goals are (1) to determine if the conditions can predict the value of the each DV subscale and (2) to determine if the proposed moderator moderates the effect of the conditions on the DVs.

My question concerns the validity of the SEM structure to achieve my goals. Specifically, I would like to know if I have setup the IVs appropriately to determine the presence of moderation. I have attached an image of my current SEM with boxes with blue text that explains the nature of each variable (standardized vs not, etc.) Any help would be greatly appreciated!

SEM structure

N Matthews
  • 21

1 Answers1

0

Welcome to CrossValidated.

This looks correct to me.

It's not clear to me that it's necessary to set this up as an SEM though. The only thing you're testing is whether the 12 (I think) residual covariances that aren't in the model are equal to zero. That's not a very interesting thing to test - if they are equal to zero, your results are equal to a multivariate (multiple outcome) regression, if they're not equal to zero, your parameter estimates will be inaccurate (even if the chi-square test suggests they're not).

You might consider treating your outcome variables as two latent variables, instead of 7 measured variables.

Jeremy Miles
  • 17,812
  • Thank you for your help! The idea was to provide the same results as a multi-variate regression to test for moderation effects. The SEM isn't necessary but (I think) can detect moderation with less power relative to a regression. My sample is around 180 so I'm not quite powered enough for regression.

    I had considered making the DVs another latent variable. I will do that and report back.

    Also, you indicated that this tests residual covariance, given the goals I mentioned in my first post, do you think that this design addresses these goals validly?

     – N Matthews Nov 23 '18 at 17:00
  • I don't think you'll get any more power from the SEM - I think the results should be the same as the multivariate regression (if you allow the residuals to correlate.) I think the design addresses the goals. – Jeremy Miles Nov 23 '18 at 18:35
  • Thanks again, Jeremy. You might be right about power issues. I do, however, have another reason why I'm using SEM. Because multiple regression (MR) requires one DV, I cannot run an MR that includes both DVs simultaneously unless I average them together. – N Matthews Nov 27 '18 at 20:38
  • You can run a multivariate regression, which has two outcomes. – Jeremy Miles Nov 27 '18 at 23:05
  • https://stats.stackexchange.com/questions/11127/multivariate-multiple-regression-in-r – Jeremy Miles Nov 27 '18 at 23:06