In classical regression equations you input the independent variables that you may observe and measure; these are so called explicit or open (not hidden) variables - you are able to monitor them because you are aware that they modulate the variability of your dependent variable. In numerous fields (f.i. sociology, psychology, marketing, polithology, etc) you may also have a considerable impact of the so called latent variable. They are not directly measured in your research, but rather are deduced from a theory; you are aware that they are important and contribute significantly to shape your dependent variable up. The latest may be endo- or exogenic (in- or outside of your model). Based on the above you may attribute each of the variables in your model to one of four categories: (a) explicit exogenic, (b) explicit endogenic, (c) latent exogenic and (d) latent endogenic variables. The appropriate classification of these variables in your model is decisive in the appropriate solution of your equation. To even more complicate the above, there are also residual variable, that exist and have some impact in your model but you do not include them in your equation. Please, also refer f.i. to Loehlin, 1987: Latent variable models to read much more.

In classical regression equations you input the independent variables that you may observe and measure; these are so called explicit or open (not hidden) variables - you are able to monitor them because you are aware that they modulate the variability of your dependent variable. In numerous fields (f.i. sociology, psychology, marketing, polithology, etc) you may also have a considerable impact of the so called latent variable. They are not directly measured in your research, but rather are deduced from a theory; you are aware that they are important and contribute significantly to shape your dependent variable up. The latest may be endo- or exogenic (in- or outside of your model). Based on the above you may attribute each of the variables in your model to one of four categories: (a) explicit exogenic, (b) explicit endogenic, (c) latent exogenic and (d) latent endogenic variables. The appropriate classification of these variables in your model is decisive in the appropriate solution of your equation. To even more complicate the above, there are also residual variable, that exist and have some impact in your model but you do not include them in your equation. Please, also refer f.i. to Loehlin, 1987: Latent variable models to read much more.

36

To add to Prof Watala's excellent reply: one reason for using SEM is where a construct is measured by a number of variables. If we use regression, then we are asking what is the effect of each predictor variable when the effects of the others are accounted for. This can result in some strange effects. For example, if you put education and social class into a model, you are asking what is the effect of a 1-unit change in education when social class is adjusted for (in other words, what is the effect of being less or more educated than the expected level for your social class) – that is, you are focussing on an education/social class discordance.

With multiple measures of the same construct, too, the intercorrelations make the model standard errors inflate.

With SEM, on the other hand, the information from both variables can be combined into a 'social status' variable. And since we are actually using the communialities of the variables to construct the latent variable, their intercorrelations are a strength.

So: one variable per construct with minimal correlation - regression. Several variables per construct - SEM.

Hello Sofia,

Just a small addition to Cezary (33 years I lived in Lodz prior migrating, so Hello Cezary) One could say that SEM is extension of regression as SEM is a combination of factor analysis (CFA and EFA) and regression analyses (paths). 

Also you have mentioned causality and even SEM does not allow for that if the data was obtained at one specific time point (cross sectional). 

However, please bear in mind that whatever you are going to apply (regression or SEM) it should be theory driven decision. Perhaps it would be a good idea to draw your model first and then decide upon the technical analysis you are going to use. 

Best wishes

Just to add to be above interesting anwsers

SEM; as an example of second generation of multivariate analysis, which differ greatly from first-generation techniques such as factor analysis, discriminant analysis or regression; is a statistical technique for simultaneously testing and estimating a set of hypothesised relationships among multiple independent and dependent variables (Gefen et al., 2000). SEM helps the researcher to answers a set of interrelated research questions in a single, systematic, and comprehensive analysis. SEM is mostly used to generate theories and concepts. One of the most important steps in structural equation modelling is assessing whether a specified model ‘fits’ the collected data or not (Yuan, 2005). SEM has the ability to model theoretical constructs that are hard or impossible to measure directly.

SEM is more appropriate than other statistical technique when one exogenous (dependent) variable becomes an endogenous (independent) variable. Hence, the proposed research model will be tested simultaneously. However, a large number of multiple analyses would be required when using first generation statistical tool.

You also use SEM when your model is relatively complex, and thus scarifies the parsimony. Using first generation statistical tools are not applicable to test complex modelling whereas SEM is more valuable when testing complex mathematical models.

Good luck

I agree with all above comments on application of classic regression analysis and SEM. I have some additional thoughts-classic regression analysis gives the associations between dependent variable(s) with independent co-variables. The association between one co-variable with the dependent can be considered as holding the others as the constant.

By using SEM, you are able to measure the direct and indirect effects of a list of variables on the dependent, their individual effect and how an individual variable may also have the effect on the dependent via the others. This may give more insights on how the exposures may affect the study outcome.