Moderated multiple regression and mixed factorial ANOVA

I conducted a moderated multiple regression and a mixed factorial ANOVA on the same variables and got two different results (using moderated multiple regression-the interaction term was significant and a using a mixed factorial ANOVA-the interaction term was non-significant). I would like to find reliable information and/or peer reviewed journals that talks more in-depth about each method of analysis and its justification in using one over the other.

Re: Moderated multiple regression and mixed factorial ANOVA

could you please tell us a little bit more about your data and how you handled it? if you only have continuous data which you sub-divided into groups you could've lost power for the ANOVA or you could've run into coding issues with regression... i'd just like to rule those out first before engaging in i guess a more theoretical argument here because the sum-of-squares in both cases should be the same, however, i have seen before that interactions are handled differently in each one... also, which program did you use? SPSS is notorious for sometimes using what theoretically wouldn't be the most appropriate standard errors to test for paramter significance...

Re: Moderated multiple regression and mixed factorial ANOVA

So my study wanted to see whether an intervention had an effect in reducing stigma. Participants were given a survey before the intervention. The participants were then split into two groups (experimental and control). The experiment group was given a new type of intervention, while the control group was given a traditional intervention. After the intervention, the same survey was given to see if stigma was reduced. The survey consisted of likert scale (1-9). I want to see whether the intervention had an effect in reducing stigma, and if it was much greater effect depending what group you are in. (I used SPSS)

I am not sure, but I think I can do either a mixed factorial ANOVA or a moderated multiple regression.

when I did the moderated multiple regression, I standardized my continuous variable. I then proceeded in multiplying the standardized continuous variable with the categorical variable to get the interaction term. Proceeding with the analysis, I placed the categorical variable, the standardized continuous variable (i.e. pre-intervention), and the interaction term as independent variables and post intervention as my dependent variable. The results indicated a significant relationship between pre-intervention and post-intervention, controlling for the other variables; indicating that the intervention had an effect. Furthermore, the interaction term was significant; indicating that the intervention had an effect in reducing stigma, and it had much greater effect depending what group you are in.

when i did a mixed factorial ANOVA (Pre-intervention and post intervention as my Within subjects variables; and conditions as my between subjects variable), there was significant difference between the pre-intervention and post-intervention, but the interaction was not significant.

Which result is more correct? Did I do either one correctly? How can I justify either one I use from past theories?

Re: Moderated multiple regression and mixed factorial ANOVA

aha! i knew something had to be wrong somewhere if you were getting different results. creating cross-product terms for interactions **only** works if you have continous predictors. in your case, you have categorical predictors (groups, so to speak, because if you're a member of the experimental group versus a member of the control group "predicts" whether stigma was reduced or not). to do interactions with categorical predictors you need to use contrast coding and the contrasts are what give you the interaction, not the product term. besides you also have the time variable going on here (pre- and post- intervention) which requires yet another type of coding... although doable, in my opinion, using linear regression to analyse experimental settings can be a pain in the #@$@! because of how meticulous one has to be in tracking all of the contrasts being done.

my guess is you did the split-plot (between- and within- groups) ANOVA right because it only requires you to enter data correctly into SPSS and to do a few clicks here and there, so i'd go with those results. but the regression version of it was definitely not done right.

you can justify either/or because, if you do both correctly, they should give you the same results, since both are instantiations of the general linear model.

Re: Moderated multiple regression and mixed factorial ANOVA

thank you! this is really helpful. I have two other questions for you. first, i want to see whether stigma was reduced regardless of what condition people were in. Would this be a simple paired samples t-test or repeated measures analysis (controlling for condition). In other words, do I have to control this categorical variable (i.e. condition) to see whether the mean differed between pre-intervention and post-intervention, regardless of anything else.

second, this study is much more complex. both groups were given two vignettes (person with depression and person with schizophrenia). i noticed that the stigma differed between both illnesses (i.e. higher stigma toward someone with schizophrenia than someone with depression). i noticed after the intervention, the gap narrowed. In other words, both had similar means and was non-significant. I want to show this phenomenon. Can i simply conduct a paired samples t-test comparing the mean difference of the pre-vignettes (Pre:schizo-dep) with the mean difference of the post-vignettes (Post:schizo-dep)?

Re: Moderated multiple regression and mixed factorial ANOVA

Hi folks, i need help regarding multiple moderated regresison analysis in SPSS.

I have one independent variable (X), one Dependent variable (Y) and six moderator variables (Z1, Z2......Z6).

My Dependent varibale is measured on 10-point semantic diffirential scale. i measured it by asking two questions in my survey. So now i took the mean of both the questions which made it approximately (Ratio) continous variable as far as i think.

My Independent varibale is also measured on 10-point semantic differential scale and i used the same techniqu as above to make it (Ratio) or close to continous.

My all moderators are measured on 7-point likert scale. Therefore i multiplied my dependent and independent variables with 0.7 to reduce them to the scale of all moderators for analyses in SPSS.

X has a strong positive linear relationship with Y and i am looking for the potential moderators from Z1.....Z6 which moderates the relationship between X and Y.

I have created all the interaction terms ie Z1*X....Z6*X. The question is.

1) Can i build different models from testing each moderator? for example model 1 with X, Z1 and X*Z1 and see if Z1 moderates the X Y relationship and then build another model with X, Z2, X*Z2 to see if Z2 moderates the relation between X Y .....so on upto 6.

OR

2) Do i have to build only One model with X, (Z1....Z6) and Z1*X....Z6*X. to see which all potential moderators moderates the relation between X and Y.

Now the problem i have is, when i build saperate models for each moderator variable with interaction terms, i get some good results as expected but when i build only one model with all the independent variables, moderator variables and interaction terms i do not get the results as expected.

I just want to ask if 1) is the right way or you can say..is not the wrong way for doing analysis in this case.