Andrew Hayes (http://www.afhayes.com/) provides fantastic tools for testing mediation called INDIRECT and PROCESS.  You can download them for free from his website; they are add-ons from SPSS and SAS.  You need to have some basic knowledge of syntax (very basic), but this tool will inevitably make your life easier.

In addition to Sliter's suggestion, it is worth also looking at Imai and colleagues research (http://imai.princeton.edu/projects/mechanisms.html). They have a package, mediation, written in R, that uses their approach. It relates to Pearl's approach (see his recent comment on their work in Psychological Methods, and his book Causality is also excellent).

Isn't multiple regression a special case of structural equations modeling (SEM)?

So, it will depend on how you write down your model. I would say write it in a SEM way, and see if it reduces to the simple case of multiple regression. If so, use the latest since it has more devoted software, usage... If not, use the complete SEM framework.

I would also say it depends on the complexity of your model, quality of the data, etc. if you have a single mediator, then regression is fine.  For more complex mediational processes, I would use sem,  it also provides more sophisticated information concerning the significance of the indiirect effects, etc.

Dear Babar,

I suggest you to find David A. Kenny's sites (http://davidakenny.net/cm/mediate.htm) who explained in clear way about mediation effect in causal relationship.  Besides analyse mediation, SEM also capables to handle interactions (some literatures called moderation effect).

Regards,

Putu

I recommend SEM. 

SEM is far more powerful, and as a result far more easily misused. In general, I tend to dislike any multivariate methods that are based on analyses of variance, and all the most common tests are (t-tests, multiple regression, ANOVA, MANOVA, etc.). These all rely overmuch on the mean, which frequently (especially in high-dimensional spaces) is a very poor basis for distances between data points. If you have a good understanding of linear algebra and multivariable mathematics in general, then SEM is the way to go. If not, multiple regression is more likely to yield results that can be interpreted such that you don't just provide a largely meaningless table or matrix of p values, but understand how the model you developed is mathematically sound and what the values you derive really "mean".

Also, as another commentator noted, SEM includes regression models.

Multiple Regression is a special case of Structural Equation Models (SEM).  And SEM can be fitted in R with following packages:

• sem: Structural Equation Models package
• lavaan: Latent Variable Analysis pacakge
• systemfit: Estimating Systems of Simultaneous Equations pacakge
• OpenMx - Advanced Structural Equation Modeling package.

http://cran.r-project.org/web/packages/sem/index.html

http://cran.r-project.org/web/packages/lavaan/index.html

http://cran.r-project.org/web/packages/systemfit/index.html

http://openmx.psyc.virginia.edu/

If your variables are all directly measured variables, then use multiple regression. If you have latent variables (ie. factors defined by indicators), you should use SEM.

Find in detail an effective way of conducting mediating analysis attached file,

Article Making Sense of Mediating Analysis: A Marketing Perspective

it's not about which one is better. since you are asking for beginners regression would be the better choice to learn. some of the journals say this is an age old way of testing mediation effect. my suggestion is to go with PROCESS ( bootstrapping) which is used by most of the good journals these days. all the best.  

If your objective of research is to test the effect of single variable without taking into consideration constructs used to describe the model than regression is best but if you want to test constructs describing variables than SEM is better .SEM is also powerful if u want to test for moderation or mediation

Hi Baber

1. The Basic Mediation Model is,

M = β0 +β 1 X=e.

2. Indirect (mediated) effect of X on Y = a*b and Direct (unmediated) effect of X on Y = c’ .

3. To test for mediation we can examine the statistical significance of the indirect effect (ie. H0 : ab = 0).

4. There are many statistical tests of the indirect effect (see Hayes & Scharkow, 2013 for a review). Some of these tests (eg. the Sobel Z test) assume the indirect effect is normally distributed.

5. Unfortunately, the indirect effect is rarely normally distributed.

6. Non-parametric bootstrapping (with confidence intervals) has been recommended for testing mediation as it does assume the indirect effect is normally distributed and yields the most accurate results .

7. Please refer to,

Hayes, A. (2013). Introduction to mediation, moderation, and conditional process analysis. New York: Guilford Press.

Regards,

Zuhair