Hi everyone, I apologyze if my question is a rather basic one.
I'm running a multiple linear regression in spss to analyze how several factors contribute to the performance in a theory of mind task ("reading the mind in the eyes task" - continuous), including gender (0 = male, 1=female, dichotomous) and the performance in an author recognition test (ART - continuous). Since I'm new to multiple linear regression, I'm not sure how to codify and enter the interaction in the model. Specifically, I have the following questions:
1) Should I create a dummy variable by multiplying ART scores by gender values (0, 1)?
2) What method should I choose for entering the variables (gender, ART score and their interaction)? Forced entry? Stepwise?
3) Should I include the interaction in a second block, and include all main effects (gender, ART, etc.) in the first one?
Thank you all for yor attention and patience!
Just fit the gender*ART "dummy variable" along with gender and ART.
1. If the interaction is not significant, remove it.
2. If the interaction is significant, keep all three variables in the model.
It is possible that ART only has an effect on one gender, hence you keep all variables to prove this.
Thank you so much for your prompt answer! I have a couple of follow up questions to clarify:
1) should the interaction dummy variable be calculated as the product of ART * gender? should I use ART mean to do this?
2) Are there any method preferences to introduce the variables in the model? like stepwise, for instance?
3) Should I include main effects and the interactions in separate blocks?
I do suspect ART has a significant effect on one gender only.
Thanks again!
1) The product, no averaging.
2) I've found stepwise works fine and (in the proper circumstances) even out-performs the fancier methods that the literature claims is better. But you will have to obey your advisor.
3) If you think it only works on one gender, you could try performing separate analyses by gender.
As an add on: it can be a good decision to include ART as measuring the deviation of each observation's ART from the mean, i.e. de-meaning the variable. This makes the interpretation of the gender-main effects more meaningful, since they can then be interpreted as the gender-effect for observations with mean (instead of zero) ART.
Thank you both for your useful answers and suggestions, I will definitely substract the mean from each value of ART and use a stepwise method to enter the variables.
I wasn't sure about running separate analysis for each gender, would it be better than including the interaction in the model?
Basically, the advantage of the interaction term compared to running the regression of the dependent variable on ART separately for each gender is that the interaction also tells you whether the different impact of ART on the outcome variable for each gender is significantly different. If you just run both regressions for each gender, you get the regression coefficients, not the results of a test whether both coefficients differ.
I think this would be the only difference, but I am not completely sure. Here is a possibly more reliable answer: http://stats.stackexchange.com/questions/33692/joint-model-with-interaction-terms-vs-separate-regressions-for-a-group-comparis
If the author suspects ART does not matter for one gender, then a separate analysis will test for whether ART = 0 for that gender.
If a pooled analysis is desired, then contrast tests can be done.
Thank you all again for your comments, I realize I should have pointed out that sample size is unequal between men (n=61) and women (n=122). Would this affect the analysis in anyway?
Lastly: would it be correct to run a pooled multiple regression first, and then if the interaction is significant, run separate regressions on each gender?
Do as you describe. When you run separate regressions, you are testing for whether ART has a significant association in that gender. As long as you have decent sample sizes for both genders, they do not have to be equal.
- Badges
- Science topic