Moderated Multiple Regression versus Factorial ANOVA.

#1
I am analyzing an interaction effect between two continuous independent variables (Z and X) on a continuous dependent variable (Y).

I have conducted both, a moderated multiple regression analysis and a two-way ANOVA (spliting each predictor at its median), obtaining different results. Using the two-way ANOVA, the interaction term (Z x X) was significant (p = .29), whereas using the moderated regression analysis the interaction term was non-significant (p = .183). Main effects (Z and X) are signiifcant with both procedures (p < .01). Analysis were conducted with SPSS. Sample size: n = 100.

In performing the moderated regression analysis with SPSS, I first center de predictors to create the product term: (Z - mean of Z) * (X - mean of X). Then I carry out a stepwise regression analysis entering both predictors and the product term as independent variables: Z, X and Z x X; and Y as the dependent variable.

Is there anything wrong? Could someone explain me these differences? Any suggestion to continue with the analysis?
Thank you all in advance.
 
#2
Using the two-way ANOVA, the interaction term (Z x X) was significant (p = .29), whereas using the moderated regression analysis the interaction term was non-significant (p = .183)
Excuse me, I meant "Using the two-way ANOVA, the interaction term (Z x X) was significant (p = .029), whereas using the moderated regression analysis the interaction term was non-significant (p = .183)"

I know that using median splits should be avoided because of the reduction in statistical power. ¿But what if the two-way ANOVA is the only procedure that leads to significant effects? ¿Would it be legitimate to use median splits then?
 

spunky

Doesn't actually exist
#3
I know that using median splits should be avoided because of the reduction in statistical power. ¿But what if the two-way ANOVA is the only procedure that leads to significant effects? ¿Would it be legitimate to use median splits then?
i still yet have to see a good reason as for why median splits should be used. too much ink has already been spilled both in analytical an simulation results showing why it is a bad idea.

ps- median splits not only reduce your power. they also increase your type 1 error rate. if my memory serves me right, i think Cohen et. al. devote a very good section in Chapter 7 of their book on multiple regression which explains the various ways in which both power and type 1 error rates are affected by median splits.

so i guess my first assumption would be that the significance of your ANOVA could very well be just a statistical artifact of your median split.
 

spunky

Doesn't actually exist
#5
most certainly... if the variables are continuous i think there's merit in keeping them continuous and analyzing them as such...