Caio Moretto It is very difficult to say exactly why this happens in your application without seeing more details on the data and models that you ran (observed covariance/correlation matrix, model specification, fit statistics, parameter estimates, etc.). In particular, it would be important to know what exactly the specification of the SEM with latent variables and a multiple-indicator measurement model looked like relative to the manifest regression and path models. If you could provide more details, that is likely to be helpful, and you might then get some better and more concrete advice / answers from the various experts here on Researchgate.

One thing that always comes to mind when the signs of regression coefficients change are suppression effects (effects of high collinearity of predictor variables). In SEM with latent variables, the collinearity effect could be stronger due to the correction for measurement error (which leads to higher correlations relative to the manifest analyses), leading to strongly correlated factors. Are your factors much more strongly correlated than your manifest composite scores? (You could examine this through a CFA model in which all factors are allowed to freely correlated with one another.) This could happen, for example, when items within a given factor are rather heterogeneous (not so strongly correlated, e.g., due to low reliability or heterogeneous content). Since a factor represents only what is common to all items (their covariance), the correlation of this factor with other factors may be much stronger than the correlations among manifest composite scores when items are heterogeneous.

Another potential explanation may be model misspecification in the SEM. Maybe the measurement model is incorrect, for example, due to the indicators not being unidimensional. This could lead to bias in the estimation of the regression coefficients.

Typically, the differences in results between a regression or path model based on manifest item sum or average (composite) scores and an SEM that uses the same items as multiple indicators of latent variables should not be extremely large (because the composite scores are also more reliable than the individual items), unless you have quite heterogeneous indicators. Therefore, I wonder whether your SEM may be misspecified, potentially causing parameter estimate bias. What does the fit (chi-square) of the SEM look like?

Caio Moretto My apologies, I just now saw that you did post a path diagram with the estimates of the models as well. It looks like you have a lot of correlated errors in your latent SEM model. Are those error correlations all theory-driven? These error correlation parameters could be part of the problem. How did you arrive at this specification with so many correlated error variables?

Also, are you looking at standardized estimates in these pictures or unstandardized estimates? (Some of the error covariance estimates seem to be > 1.0, which would be an improper estimate in the standardized solution since a correlation cannot be > 1.0 by definition.)

My suggestion is, before focusing on the weight sign, we need to check the statistical significance of the corresponding parameter. If the p-value is greater than 0.05 for example, we do not need to provide an explanation for the corresponding correlation. However, if the p-value is less than 0.05, we need to provide an explanation for the negative signal from the literature or data consistency.

Dear Adel Bessadok and Christian Geiser , thank you very much for your suggestions and reflections, they helped me a lot. Indeed, some of the errors were not theory-driven, which impacted the estimates, which were fixed later. Besides that, the p-value didn't show statistical significance for the items. After several tries and analysis, we concluded that there was something problematic with the questionnaire items that only becomes salient in more sophisticated analyses such as SEM. Thank you again for your time and attention!