7

In my master thesis I have drawn a few hypotheses. I have answered them all with linear regression. In these linear regressions, I took control variables into account.

My question is: do I have to run a mediation analysis? Or is it also possible to report the regressions of all relations separately (for example: X -> Y, X -> M, M -> Y and A -> Y)?

Here is how my model look like:

enter image description here

My main hypothesis is about the relation between X and Y. I hope my question is clear.

Thank you in advance!

chl
  • 50,972
  • 18
  • 205
  • 364
Ann
  • 71
  • 1
  • 2

2 Answers2

8

You can run regression models separately, if you follow the Baron-Kenny approach. As far as I know, there are two general approaches to test for mediation: (1) path models (and, SEM, of course) and (2) the Baron-and-Kenny approach (see item (a)). I use Mplus to run my mediation models which is very handy (+ bootstrapped standard errors).

Unfortunately, you did not tell us what software package you are using to do your analysis. You have a couple of options:

(a) You might be interested in D Kenny's website on mediation . He gives a very clear description how to proceed in order to test for a mediation effect (see "Baron and Kenny Steps").

(b) If you happen to use Stata or R for your analysis, you could check out the ATA website on Stata Frequently Asked Questions (search for 'mediation') or the R package mediation. If you use SPSS, you will like this website. Kenny's website also offers a couple of tips for different software packages, e.g. how to get bootstrapped standard errors in SPSS or SAS.

Bernd Weiss
  • 7,044
  • 28
  • 40
  • Thank you for your answer! My software package is SPSS (19). If I run my model in linear regression with all my independent variables at once: only one predictor is significant. When I run them separately with the dependent variable, all the predictors are significant. Can I choose for this method (running the regressions separately instead of running the whole model at once)? Thank you! – Ann Jun 16 '11 at 09:29
  • 1
    @Bernd (+1) Very good response. I pointed to similar material in an [earlier response](http://stats.stackexchange.com/questions/9605/calculating-the-mackinnon-empirical-distribution-test-to-test-mediation) of mine :-) – chl Jun 16 '11 at 09:32
  • 1
    @Ann I found some resources for mediation analysis with SPSS (including Sobel's test and bootstraped estimates of path coefficients) on [Andrew F. Hayes](http://www.afhayes.com/spss-sas-and-mplus-macros-and-code.html)'s website. – chl Jun 16 '11 at 09:36
  • Thank you both! :-) (I can not vote up because I need 15 reputation) Well, the only problem left is the following: When you have a dependent variable with two or more independent variables, you conduct a multiple regression analysis. In my case, I have three independent variables (X, M and A). Where do I leave variable A in my mediation analysis? I conducted the mediation with variables X, M and Y. But variable A is left out. Is that a problem? Thank you! – Ann Jun 16 '11 at 09:43
  • 1
    @Ann I do not know your data (or your theoretical model). So, it is hard to comment on this behaviour (which is not uncommon though). Regarding "running the whole model", see chapter "Covariates" on Kenny's website. – Bernd Weiss Jun 16 '11 at 09:45
  • @chl Thanks! I really should have searched for older answers before I post my answer. – Bernd Weiss Jun 16 '11 at 09:49
  • @Bernd Thank you! I will run the model with A as a covariate. – Ann Jun 16 '11 at 10:49
1

Whether you control for A depends on what you are trying to accomplish. Including A in the model will increase your R-squared.
But if A is uncorrelated with X or M (as your diagram indicates) then inclusion/non-Inclusion of A will not affect coefficients or p-values for X or M.

rolando2
  • 11,645
  • 1
  • 39
  • 60