A quote from Dawson (2014; Moderation in Management Research: What, Why, When, and How. Journal of Business and Psychology, March, 2014, Vol.29(1), p.1(19)

Should I Center My Variables?

Different  authors  have  made  different  recommendations regarding the centering of independent variables and moderators.  [...]  In truth, with the  exception  of  cases  of  extreme  multicollinearity,  the decision does not make any difference to the testing of the interaction term; the p-value for the interaction term and the subsequent  interaction  plot  should  be  identical  whicheverway it is done (Dalal and Zickar2012; Kromrey and Foster-Johnson 1998). Centering does, however, make a difference to the estimation and significance of the other terms in the model: something we use to our advantage in the indirect form  of  the  simple  slope  test,  as  the  interpretation  of  the X coefficient is the relationship between X and Y when Z=0.  Therefore,  unless  the  value  0  is intrinsically  meaningful  for  an  independent  variable  or moderator (e.g., in the case of a binary variable), I recommend  that  these  variables  are  either  mean-centered  or z-standardized before the computation of the interaction term.

Dear Arjun,

please find attached two articles covering this topic. It is not necessary to center the predictor variables in a moderated regression, because this will not solve multicollinearity problems. On the other hand, centered variables are more straight forward to interpret, because after centering 0 is a meaningful value, i.e. the mean value.

If you are dealing with observed variables in the model, it is advisable to center the predictor as well as outcome variables.

However, if you are dealing with latent constructs (with multiple indicators could involve higher order constructs), the issue does not arise.

A little off topic, but I'm hoping someone can answer. When we center the variables (we are using PROCESS) some of our interactions go from being nonsignificant to highly significant. My impression is that this should not happen? any ideas what might be going on here?

It is absolutely possible. Centering may help to reduce non-essential multicollinearity. When you compare the centered and uncentered VIF or Tol values (indicators for multicollinearity), I am pretty sure that VIF will be reduced in case of centering. Multicollinearity causes increased standard errors for the zero (lower) order predictors, but not(!) the interaction term (a myth that prevailed quite a time), and does not affect the statistical power of the interaction. Therefore, it is not necessary if you are only interested in the interaction, but not the lower order predictors, although it will help interpretability.

This can be done, but the reliability indicators will decrease

Centeration is important to avoid multi collinearity

Rainer Duesing diving into mean centering for a study i have the following question. When applying mean centering it should be applied (as I understand) to the DV's and IV's. However, it is not clear to me weather the CV also need to be centered? Furthermore I was wondering how to deal with a interaction term (event-study) when applying mean centering? Thanku for the help in advance!

Noor Noij There is no need to center the DV, only the IV and CV, this will help to interpret the results, since the regression coefficients show the effect, when the other predictors have the value 0. You do not need to center, if the IV/CV have already a meaningful 0 value.

What do you mean with "event-study". You can calculate the interaction term from the centered variables and use it as normal predictors. Do not(!) calculate the interaction term from the uncentered variables and center the interaction itself. That would not help.

Rainer Duesing Thank you for the fast reply! It is help full.

With "-event-study" i meant to explain a study over e.g. time period 2008-2015 where I would like to measure the impact of event X that occurred in 2012.

The last means that when using STATA first the new centered variable must be created before the interaction terms are constructed?

Do you might happen to know if any clear STATA guides are available online for application of centering?

Still dont know the event-study thing, but this seems not important I guess.

And yes, you first calculate the centered variables and then the interaction term from the centered variables, simple as that. Simply dont use the original variables.

For STATA: sorry, never worked with it, I have not the slightest clue, just used SPSS, R and MATLAB. But if you have the possibility in STATA to calculate new variables with formulas, it is simply variable1_centered = variable1 - mean_variable1, where mean_variable1 is simply the mean value of the whole variable1. But be aware in case of missing data! Check that the mean of variable1_centered is really 0 in your final model!

Rainer Duesing thank you for the explanation. I was also getting very different results using Andrew Hayes process model 1 and MMR (without centering). After centering they were identical. Can you guide me to (if there exists) to a journal or link that explains this in more details.

Somak Banerjee I think Andrew Hayes Book about Mediation, Moderation and Conditional Process Analysis is quite good on this topic, but also the regression book by Darlington & Hayes.

A nice overview when and how to center gives this article:

https://journals.sagepub.com/doi/full/10.1177/1094428111430540?casa_token=mVZlqmVI1y8AAAAA%3AfQfI_0oE9d6vw0f6J5c8EBG5KSh0kDbbEkjQ-wcvv9krI9ze6luzqly8xfckc7TvQi9VR4mZJWagRVw

It is simply centering the independent and the moderator variable. There after get the interaction to fit in the model. this centering is to reduce the multicollinearity.

But also what is confusing me it that if the moderator variable has no significant relationship with the DV, what you do?

It is not a necessary condition that the moderator is correlated with the DV, just test it yourself in R:

library(stargazer)

set.seed(1235)

n=100

x1

See output

Thanks Rainer for this wondeful comment.

I have understood

Mugarura Alex you are welcome!

Did you publish your results? I am doing similar moderation testing.

The moderation effect of subjective well-being on the relationship between workplace spirituality and job satisfaction &

The moderation effect of positive character strength on the relationship between workplace spiruituality an job satisfaction.