Moderating Effects with Seemingly Uncorrelated Variables
Jeremy Taylor | Wednesday, July 13, 2011 at 7:00AM Imagine the circumstance where you are testing whether there is an association between "number of carrots consumed" and "blood pressure". Imagine further that you have reason to believe that the association between these two variables varies by age (for this example let's make age dichotomous, i.e. old vs. young). Perhaps you expect that the more carrots someone eats, the lower their blood pressure will be (negative association), but you think this will be more true for older people than younger (i.e. age moderates the effect of carrot consumption on blood pressure).
Since you'd expect that the association will probably still be negative in both groups, but more negative in older people (if your hypothesis is accurate), you might expect to see a graph like the one below:
Video: How To Test for Mediators, Moderators, and Suppressors
Video: How To Run a Multiple Regression in SPSS
Discussion Forum: How to Create Dummy Variables for Multiple Regression
Discussion Forum: General linear model- univariate spss write up
This graph is fairly typical of a two-way interaction, where the two groups (young vs. old) have differing slopes. Since both group's slopes are negative, it isn't supriising that the overall sample slope is also negative. However, imagine a slightly different scenario where younger people's slope was actually positive for some unknown reason. In this case, your graph would look like this:
In this scenario, you would still likely have significant moderation (probably an even strong interaction effect, since the difference in slope is even larger), however you might not see a significant association between the IV (carrots) and DV (Blood Pressure) in the sample as a whole. This example highlights the danger of relying only on correlations and failing to consider/test potential moderating effects. Thanks for the great question, Ken!
Editorial Note: Stats Make Me Cry is owned and operated by Jeremy J. Taylor. The site offers many free statistical resources (e.g. a blog, SPSS video tutorials, R video tutorials, and a discussion forum), as well as fee-based statistical consulting and dissertation consulting services to individuals from a variety of disciplines all over the world.
Reader Comments (14)
Hi Jeremy,
Thanks so much for your clear and concise reply - I finally get it! Thanks again, it's much appreciated.
Ken
Great site. Keep up the good work!
Thanks, Ben!
The way you described the situation is great. As a follow-up question: what about the interaction term and its correlation with the DV? More specifically, I have seen cases where an interaction term is significant (in a regression analysis) but it is not correlated with the DV in the correlation matrix. What do you make of this?
Thanks so much for this posting! I have been searching for this information as this issue was recently raised in response to some analyses I have been working on.
I was wondering if you might be able to provide a citation that discusses this? Any information would be great! Thanks!
Bev,
It is absolutely possible for an interaction term to be significant in a regression, but not correlated with the DV in a correlation matrix. This can occur because an interaction term is really only testing what it is intended when its effects are partial estimates with the main effects of the interaction term components in the model also.
Hey Lori,
I don't know a citation for this specific issue (at least not off-hand), but the logic/rationale is based on the basic principles of a moderator analysis, for which citations like (Aiken & West, 1991) would be appropriate.
Reference:
Aiken, L. S., & West, S. G. (1991). Multiple Regression: Testing and Interpreting Interactions. Thousand Oaks, CA: Sage Publications.
Here is a link to the book on Amazon:
http://www.amazon.com/Multiple-Regression-Testing-Interpreting-Interactions/dp/0761907122
i need a quick answer... really quick response please!!! My question is : Can moderated be related to independent and dependent variable?
I have a research framework proposed by my own supervisor that takes perception of politics (POP) as an independent variable, job insecurity as dependent...and perception of support (POS) as a moderator but my argument says:As POP and POS are highly correlated.Presence of perception of politics will hamper the working of perception of support as a moderator. Therefore POS shouldn't be taken as a moderator......Please help as she isn't listening to me.Looking forward for an urgent reply!!!
Sara,
Do you believe that POS should instead be a mediator?
Can a variable be a moderator if it is significantly correlated to the IV? Let's put it another way: If i hypothesized a moderating effect that in fact doesn't work, and that the moderator is significantly correlated to the IV, is it correct to say that the moderation doesn't work BECAUSE of this high correlation?
Thank you!
Hi..Ive a question please
Iam studying the relationship between employees perception of CSR as Independent V. and Employee Engagement.. can I take Employee perception of total rewards system (monetary rewards, non-monetary rewards) as a Moderator Variable?
I will be so grateful for your reply?
Dear Ali,
If you believe that the relationship between employees perception of CSR as Independent V. and Employee Engagement varies (or differs) as a function of total rewards system (monetary rewards, non-monetary rewards), then rewards system would indeed be a moderator!
Hi,
I have a quick clarification. I'd really appreciate your help with this.
So when you run the moderation analysis. you run regression between IV and DV, then Moderator and DV, and then the interaction and the DV.
What if the Moderator and DV regression is not significant, but when you run the interaction (including plots) it is significant.
Would you still consider this moderation or no?
Thank you soooo much!