Hi all,
I have a question about my analyses for a revise and resubmit I am working on.
I have a predictor and two moderators I am testing on my outcome (which has several dimensions - so multiple regressions).
For the analyses: for each outcome, I ran several moderated regression analyses.
To test the first interaction (predictor x moderator 1), I did the following:
In block 1:
predictor
moderator 1
moderator 2
And then in block 2:
added predictor and moderator 1 interaction (predictor X mod1)
To test the other interaction (predictor x moderator 2), I ran a SEPARATE linear regression.
In block 1:
predictor
moderator 1
moderator 2
And then in block 2:
added predictor and moderator 2 interaction (predictor X mod2)
One of my reviewers says that both interactions should be included in the SAME model.
When I tested this, interactions that were significant are no longer significant (because including BOTH interactions takes away explained variance).
My question is -- when testing multiple moderators, do they have to be included in the same model? By testing all moderator effects together, I lose the explained variance.
But I am not aware of what is best practice. To me, it makes sense to test in two separate models (one for each moderators).
Only one reviewer wanted this, and the other didn't mention it.
Any insight or help is much appreciated. Thanks!
Erin
It is my belief that the reviewer is correct. Presumably it is your belief that the best model of your experiment has two moderators acting at the same time but on different IVs. Thus this best model contains x1 + x1*m1 +x2 +x2*m2.+ rest of model with the regression coefficients included with all model terms.You test what you believe happens and see if that stands or falls. Because you believe m1 and m2 are working at the same time in nature. if the regression coefficient of x1*m1 or that of x2*m2 doesn't reject the hypotheses that the regression coefficient = 0 then the corresponding interaction is rejected. Whatever you think is happening that is what you test. Best wishes, David Booth
When you choose to run a moderator analysis using multiple regression, part of the process involves checking to make sure that the data you want to analyse can actually be analysed using multiple regression. You need to do this because it is only appropriate to use a moderator analysis using multiple regression in SPSS.
for example 1= There needs to be a linear relationship between the dependent variable and the independent variable for each group of the dichotomous moderator variable. Whilst there are a number of ways to check for these linear relationships, you can create a scatterplot using SPSS Statistics, and then visually inspecting this scatterplot to check for linearity. If the relationship displayed in your scatterplot is not linear, you will have to either run a non-linear regression analysis or "transform" your data, which you can do using SPSS Statistics. In our enhanced moderator guide, we show you how to create and interpret a scatterplot to check for linearity when carrying out a moderator analysis using SPSS Statistics and some of the options available to you when you do not have linearity.
example 2= Your data must not show multicollinearity, which occurs when you have two or more independent variables that are highly correlated with each other. This leads to problems with understanding which independent variable contributes to the variance explained in the dependent variable, as well as technical issues in calculating a multiple regression model. Therefore, in our enhanced moderator analysis guide, we show you: (a) how to use SPSS Statistics to detect for multicollinearity through an inspection of correlation coefficients and Tolerance/VIF values; and (b) how to interpret these correlation coefficients and Tolerance/VIF values so that you can determine whether your data meets or violates this assumption.
You may have already considered this but I'll add it here just in case. How are the two moderators related to each other? Are they statistically independent and the objective is to show that they together moderate the effect of the IV on DVs? In this case, I agree with the reviewer and the valuable comments above that both should be present in the model.
However, if the the moderators are correlated to each other and are measuring the same underlying construct, you may want to consider combining them first into a latent variable and perform the moderation analysis with the latent variable. This can be achieved by using structural equation modelling (SEM), cf: http://www.quantpsy.org/pubs/little_card_bovaird_preacher_crandall_2007.pdf
In the case of high correlation between moderators or between the moderators and the IVs, the problem of multicollinearity has also been reported as suggested by one of the comments above, cf: Article Mean centering, multicollinearity, and moderators in multipl...
Thanks, everyone! This is very helpful. Usman Rashid this is a good question about the moderators being related. They have a Pearson r correlation of .67, so they are related but I don't think enough to justify combining them into one variable. I do not believe I have issues with collinearity, according to VIF, etc. I will continue to read up, but it looks like my discussion will be re-done to focus on the main effects rather than the interaction (which has disappeared in the model with both moderators included). Thanks again! Erin
- Badges
- Science topic