Ordinary least squares is the regression subset of the General Linear Model. The GLM is a beautiful statistical structure unlike any other in our discipline. It provides a consistent theory and methods for regression, Analysis of Variance, Analysis of Covariance, and--through its use to generate results for other analyses--many other statistical methods (David Booth notes that SEM is one of these). Simple results such as t-test and F-tests are just special cases. In addition, it provides a wealth of diagnostic tools that apply across the board to all of these methods. Leverage, influence measures, partial regression plots, and residual analyses all make it more likely that a carefully performed GLM analysis will be valid and intelligible. Methods that rely on computations that are too complex to be visible can't do that. You are stuck with a P-value and some coefficients and no insight into what might really be going on in your data.
My advice:
* Plot your data
* Re-express if relationships are not linear
* Remove outliers, split your data into subgroups if they are not a consistent whole.
* Fit OLS models carefully. Be willing to entertain alternative models--there often is not a single "best" model and alternative models can be informative.
* Diagnose your models: examine leverage, influence measures, plots of residuals, partial regression plots for the coefficients of interest, and other related tools.
* Then--and only then--if everything looks good and you think there is still more to learn about your data, try one of the other methods.
The Gauss-Markov theorem and the properties of a normal distribution. For details see a linear statistical models book. SEM is just a dressed up version of least squares as is PLS-SEM. Best wishes, David Booth. PS in the old days i.e. when I was a student all of the SEM and Path Analysis calculations were done with ordinary least squares regression - no special programs. See Oscar Kempthorne's book, An Introduction to Genetic Statistics to see how path analysis was originally done.
Hayes, A. F. (2009). Beyond Baron and Kenny: Statistical mediation analysis in the New Millennium.
Hayes, A. F., & Rockwood, N. J. (2017). Regression-based statistical mediation and moderation analysis in clinical research: Observations, recommendations, and implementation.
Ordinary least squares is the regression subset of the General Linear Model. The GLM is a beautiful statistical structure unlike any other in our discipline. It provides a consistent theory and methods for regression, Analysis of Variance, Analysis of Covariance, and--through its use to generate results for other analyses--many other statistical methods (David Booth notes that SEM is one of these). Simple results such as t-test and F-tests are just special cases. In addition, it provides a wealth of diagnostic tools that apply across the board to all of these methods. Leverage, influence measures, partial regression plots, and residual analyses all make it more likely that a carefully performed GLM analysis will be valid and intelligible. Methods that rely on computations that are too complex to be visible can't do that. You are stuck with a P-value and some coefficients and no insight into what might really be going on in your data.
My advice:
* Plot your data
* Re-express if relationships are not linear
* Remove outliers, split your data into subgroups if they are not a consistent whole.
* Fit OLS models carefully. Be willing to entertain alternative models--there often is not a single "best" model and alternative models can be informative.
* Diagnose your models: examine leverage, influence measures, plots of residuals, partial regression plots for the coefficients of interest, and other related tools.
* Then--and only then--if everything looks good and you think there is still more to learn about your data, try one of the other methods.