It is not mandatory to meet this criterion

Mediation is just a way of specifying a set of regressions. So to the extent that your normality issues are strong enough to violate the assumptions for regression, then yes, it does matter.

Hi,

I agree with David. Process runs OLS regression and they require normality--but of the residuals not the variable (often made mistake). My intuitive take would be to check the assumptions by two separate regressions including diagnostics.

If you decide to specify a SEM to test the mediation--optimally with a model with a full medition OR the x and m variables being instrumented your data should be multinormal (this time its the variables as you use the ML estimator). However, violations can easily handeled with using a correction of the ML estimator.

I would recommend to do this. Don't use PROCESS....ever.

All the best,

--Holger

Hi everyone, thank you very much for all your replies.

Dear Vanishree Kanaka Sundar , first of all, due to Central Limit Theorem, normal distribution of data is not an issue if you have optimal set of responses because Sampling Distribution of a large (depends on area of research) data set tends to be normal. Please refer to “Many people take the ‘assumption of normality’ to mean that your data need to be normally distributed. However, that isn’t what it means. In fact, there is an awful lot of confusion about what it does mean” (Field, 2013, p. 229).

Secondly, PROCESS Macro provides you the facility to bootstrap, so normality itself should not be a problem while performing a mediation analysis in PROCESS. However, in case of latent variables, Reliability and Validity would still be an issue.

Field, A. (2013). Discovering Statistics using IBM SPSS Statistics (4th ed.): Sage.

Muhammad Zia Aslam Hi. Thank you for your answer. So do you mean that if I'm going to do mediation using PROCESS Macro, the data does not need to be normally distributed ? Please do correct me if I'm wrong.