I am hoping to test an interaction effect, though being new to statistics/quant research in general, I am a little unsure as to which test would be most appropriate.

I have read around using regression and a two-way ANOVA to test interaction, but I feel as if they may not work with my variables.

My IV variable is nominal with 5 categories My moderating variable is ordinal with either two or three groups (undecided whether I do high/low or high/medium/low) Whilst my DV is a scale variable

Any help regarding what statistical test would be most appropriate would be greatly appreciated! Also, if you know of any suggested reading in the area, do please let me know.

  • $\begingroup$ Have you tried fitting a linear model with interaction and checking if the corresponding interaction coefficients are significatively distinct from 0? $\endgroup$ – David Mar 15 at 9:57
  • $\begingroup$ Hi @asdf thanks for getting back to me. Yes I have tried a linear model using my moderator as a scale, rather than an ordinal value. When I do so, there is a significant moderation, and it is particularly strong for 2 of my 5 IV categories. However, I am not sure that a linear model tells the full story, as when I split the moderator into high/low or high/medium/low there are even more differences that do not appear in a linear model. I am just a little confused as to how I should proceed $\endgroup$ – enoon Mar 15 at 10:10
  • $\begingroup$ But did you include an interaction term into the linear model? I insist on the linear model because ANOVA/ANCOVA and all the other AN[a-z]VA are just particular cases of the general linear model, so if they are useful, so is linear regression $\endgroup$ – David Mar 15 at 10:18
  • $\begingroup$ I believe I did @asdf I used PROCESS for the linear model which used dummy variables for the catergorical IV, and it found that the interaction between two of the five IVs and the moderator were significant. I would not be unhappy to just use this (as the findings are indeed interesting), but a straight linear model does not account for the difference between high, medium, and low. eg. for one group variable, the linear model shows a direct positive relationship between that and the DV. However, when the moderator is split into three groups, the medium score is higher than the high $\endgroup$ – enoon Mar 15 at 10:44
  • $\begingroup$ Hmmm... interesting. You could add extra terms with powers of the "weird-behaiving variables" to reflect that relationship in the model. However, you may lose overall accuracy and interpretability, so I am not really sure. So a good approach may be to use linear regression only to check with interaction effect and significant, then move back to what you were doing. $\endgroup$ – David Mar 15 at 12:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.