It does not matter, unless you aim to do sequential F-tests. Even then: you can always do any sequence of tests on any sub-models.

Hi Esther Curtin

sounds like you have already devised a strategy based on the expected effects. If you work that into your text as well as into the regression model itself, it will make it easier for your readership to understand, but in the end the order of the blocks does not matter.

Best

Marel

Hello Esther,

If testing for changes in R-squared from a smaller model (fewer IVs) to a larger model (more IVs) is what you're after, then yes: the order of block/step entry matters (that's what Jochen Wilhelm refers to as sequential tests).

The choice of order should be driven by your specific research questions in conjunction with whatever relevant theory is guiding your work (as implied by Marcel Grieger's response). Yes, different folks might use different sequences, but that's in keeping with what questions they seek to address by the analysis.

In the final/full model, it makes no difference as to when a given IV was entered; the estimated regression coefficients are conditional based on all other IVs and the relationships in your data set, and not order of entry. (Both previous responses noted this as well.)

Good luck with your work.

Everything you did sounds plausible. Covariates should be entered into the first step of the hierarchical regression model if you want to examine the effects of your predictors (IVs and interactions) adjusted for those covariates. After this, your predictors are entered into the next step. This can be done by changing "blocks" on SPSS or fitting the models and calculating the R-squared change between the models in R (you can pull r.squared from the summary table of the regression).

As long as your predictors are entered into the same step, the order of the predictors doesn't matter. However, for interpretation, it would be beneficial to enter the IVs first, and then the interaction terms later.

On a side note, conducting simple slope analysis and testing the differences between simple slopes could be beneficial as well as interpreting the significant main and interaction effects. Robinson et al., 2013 compared simple slope comparison to interaction analysis and found advantages of simple slope comparison (more power and less risk of Type II error).

Good luck with your analysis!

Robinson, C. D., Tomek, S., & Schumacker, R. E. (2013). Tests of moderation effects: Difference in simple slopes versus the interaction term. Multiple Linear Regression Viewpoints, 39(1), 16-24.

Hi Hanjoo Kim thanks for the response.

In Robinson et al's paper, they mention that simple slope analyses can be used to further probe the interactions if they appear significant - is this plausible or should I choose one or the other?

Also, from my understanding, interaction terms lend themselves better for continuous variables, and simple slopes for categorical - is this correct?

From this previous thread (https://www.researchgate.net/post/How-to-do-a-simple-slope-analysis-with-a-continuous-predictor-and-a-categorical-moderator-4-categories) I gather that I should create dummy variables for my categorical moderators (e.g., a dummy for female and male as the reference) and then create interaction terms for the dummy variable instead of the original sex variable?

Thanks a lot,

Esther

Dear Esther Curtin ,

I think your task is a bit ahead of your understanding. Instead of reading papers about special applications (which are often using suboptimal analysis strategies) I'd recommend to read a book (or two) about regression modelling. Giving you a useful advice for your problem would require to give half a course (or possibly even a full course) on regression modelling, what is not feasable through RG.

These books might be a start:

ISBN: 9781107023987

ISBN: 0521867061

ISBN: 9780471754985

Great, thanks a lot!

Hi Esther Curtin

These are good questions.

In Robinson et al's paper, they mention that simple slope analyses can be used to further probe the interactions if they appear significant - is this plausible or should I choose one or the other?

- I recommend presenting the results from both. Begin with interaction analysis and do the simple slope analysis as a follow-up. They are not mutually exclusive and you'll get the same results in most cases.

Also, from my understanding, interaction terms lend themselves better for continuous variables, and simple slopes for categorical - is this correct?

- Yes, in order to conduct a simple slope analysis, you'll need categories. Researchers do simple slope analysis for continuous variables by transforming those into categorical variables (e.g., high vs. low group rather than a continuous predictor.) This is just for the simple slope analysis, and you can keep the continuous variables as it is for the regression analysis.

From this previous thread (https://www.researchgate.net/post/How-to-do-a-simple-slope-analysis-with-a-continuous-predictor-and-a-categorical-moderator-4-categories) I gather that I should create dummy variables for my categorical moderators (e.g., a dummy for female and male as the reference) and then create interaction terms for the dummy variable instead of the original sex variable?

- Yes, you're right. A categorical moderator should be dummy coded since otherwise the statistical software would recognize it as a continuous variable. A group with the highest frequency is typically considered a reference group, and it is coded 0 across all dummies. Then, the effect of the reference group will appear as a constant in your regression model, and you'll see the main effects and interaction effects below. Sometimes, you don't need to dummy code your predictors in SPSS when there are built-in posthoc analyses, such as ANOVAs or univariate analyses. However, dummy coding matters when running a linear regression.

If the sex variable is binary and already coded 0 and 1, you don't need to do anything else about it but enter it with the associated interaction terms (IV x sex) into the model.

I hope this helps!

Hi Hanjoo Kim

Thanks very much for your detailed answers, it helps a lot!

One of my key moderators has well-recognised norms to be able to split participants into high/low categories, however another does not. Would you therefore advise to split participants by the mean/SD or median? I assume the latter would make equal sized groups, therefore I would have to find another plausible way to decide on which subgroup to use as my reference category.

Many thanks again,

Esther

Hi Esther Curtin, Good question. Unless you have a certain cutoff standard validated by other studies, "median split" is the way to go.