0

I'm working on an SEM problem, and I could use some second opinions! The diagram below shows the proposed SEM model, where I have a 4 latent constructs which are estimated themselves from 2 second-order constructs, and these second order constructs predict 3 Dependent Variables. there is an additional latent construct which is believed to act directly on one of the second order constructs. This all seems valid, except that some of the coefficients predicting the Dependent Variables have the wrong sign. They are negative, for instance, when we believe they should be positive.

Some useful information about the data: the exogenous variables are binary. The endogenous variables (dependent variables) are measured as 7-point Likert Scales.

Any thoughts and suggestions are appreciated. Happy to answer any questions as well, obviously! Thank you!

enter image description here

logisticregress
  • 127
  • 9
  • How's the fit of the measurement model? Perhaps there's a specification error? Could also be a suppressor situation. See this: https://stats.stackexchange.com/questions/73869/suppression-effect-in-regression-definition-and-visual-explanation-depiction#:~:text=So%2C%20a%20suppressor%20mostly%20%22suppresses,complements%22%20to%20the%20coefficients). – Rick Hass Apr 23 '21 at 16:25
  • Good question - the fit of the Measurement model is good. we're now extending it to predict outcomes. I hadn't considered the suppressor option. Thanks! I'll need to think about how to avoid that, as the stakeholders want all Latent variables in the model (of course they do!) – logisticregress Apr 23 '21 at 16:28
  • Haha! of course... but you can always test the difference in fit between nested models if they want some convincing! – Rick Hass Apr 23 '21 at 16:32
  • FWIW - if I take away the second order Latent Constructs (which are there because the first order ones are highly correlated), it makes for a really complex path to the three DVs. It's seemed too complex... so I tried to simplify it by taking advantage of know correlations between the first order LCs. – logisticregress Apr 23 '21 at 16:51
  • That sounds like it's a good way to go. Good luck! – Rick Hass Apr 23 '21 at 18:00

0 Answers0