In an SEM with latent variables, you account for measurement error (random noise) in the observed variables related to the exogeneous (independent, predictor) variables, whereas in standard OLS regression you don't. This alone may explain the differences.

Also, an OLS regression model is saturated (reproduces the observed data perfectly), whereas SEMs with latent variables tend to be overidentified. Overidentification means there can be misfit between your model and data. This also (or in addition) could explain discrepancies between your SEM and regression results.

Another issue is statistical power, which could differ between OLS regression and SEM, depending on various factors. Differences in power may explain differences regarding tests of significance between regression and SEM.

Structural equation modeling (SEM) and linear regression are both statistical techniques used for analyzing relationships between variables. However, they differ in their underlying assumptions and the types of models they can handle. These differences can lead to variations in the significance and impact levels of variables.

1. Assumptions: Linear regression assumes a linear relationship between the independent and dependent variables, as well as independence of observations. On the other hand, SEM allows for more complex relationships, including non-linear and mediated relationships, and can handle non-independent observations.

2. Model complexity: SEM is a more comprehensive framework that allows for the analysis of both observed and latent variables. It incorporates measurement models (relating observed variables to latent constructs) and structural models (relating latent constructs to each other and to observed variables). Linear regression, on the other hand, focuses on the relationship between observed variables only.

3. Error term considerations: In linear regression, the error term is assumed to be independent and identically distributed with constant variance. In SEM, the error term can be correlated and have different variances across variables, allowing for a more flexible modeling of complex relationships.

4. Control for measurement error: SEM explicitly models and accounts for measurement error in the measurement models, while linear regression does not consider measurement error unless it is explicitly addressed through methods like instrument variables.

Given these differences, it is possible for the significance and impact levels of variables to differ between SEM and linear regression. SEM provides a more comprehensive analysis by considering the relationships among latent and observed variables, measurement error, and complex model specifications. This can lead to a better understanding of the underlying structure and relationships in the data. However, it is important to select the appropriate modeling technique based on the research question and data characteristics to obtain valid and reliable results.

Thanks Christian Geiser for your answer