A good resource with SPSS example data and syntax on a CD-ROM is Cohen, Cohen, West, & Aiken (2003):

Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Applied multiple regression/correlation analysis for the behavioral sciences (3rd ed.). Lawrence Erlbaum Associates Publishers.

Book Applied Multiple Regression/Correlation Analysis For The Beh...

The SPSS files are also supposed to be accessible from the website above, but I could not access them today.

If you want to calculate the conditional effects of the focal predictor at -1SD, mean, and +1SD of the moderator, you would:

Hello, I am not a proponent to "overs-use" PROCESS, but if you know what you are doing, it is much less error prone (e.g. decimal places while calculating effects, missing data on some, but not all variables....) and it gives you all simple/conditional effetcts you want. So why not using it, if you know it already?

Besides that, I would recommend to use 16th, 50th and 84th percentice instead of +/- 1 SD. The latter may use implausible predictor values, if the distribution is skewed (no there is no assumption that the predictors need to be normally distributed, in case you wonder), whereas the former always uses real values of the predictor. In case of (nearly) normally distributed variables, both approaches will come to (nearly) the same conclusion and values, therefore, the percentile methods seems to be more "robust" in a sense.

Shalom Dinesh Kumar !

Thanks for you suggestion but it's very problematic. First of all, to the best of my knowledge, PROCESS calculates the conditional effects on the basis of the entire sample. Therefore, it is not the same thing as calculating anything on three smaller samples. Second, PROCESS calculates the effect AT the -1SD or 16th percentile. What you suggest would be to calculate the effect for three groups - one RANGING from 1SD and higher, one for -1SD to +1SD, and one for -1SD and lower - not the same thing at all. Third, I was in contact with Hayes and he said that this is not the way that PROCESS calculates things. So, your suggestion may give an approximation of conditional effect sizes but not something that is reportable.

Shalom Rainer Duesing!

The reason that I cannot use PROCESS (I am a user but hope that I'm not an over-user :) ) is because PROCESS doesn't know how to work with multiple imputation data sets. Many SPSS procedures do know how to do that (they analyze each imputed data set and also give a pooled result).

Thanks for the recommendation about using the percentiles in contrast to the SDs.

Shalom Prof. Lazar,

Thank you for your feedback. I am curious to know what method worked for you.

Aryeh Lazar Ok, you cannot use PROCESS because of multiple imputations, but in principle, it is what you want?

I think what Dinesh Kumar wrote is correct, but maybe not clearly explained. you do not actually create "subsets", but you shift the scale of the moderator variable, similar to centering. For the simple slope analysis you pick a point, e.g. 16th percentile or -1 SD, and "re-center" your moderator on this point and calcualte the regression model for inference.

For further explanations look at the Cohen et. al book Christian mentioned, or in Hayes book on moderation/mediation. In the current 3rd edition you find it in chapter 7.4, page 263 ff. "Probing an Interaction". Or Darlington & Hayes (2016), page 422 ff.

Darlington, R. B., & Hayes, A. F. (2016). Regression analysis and linear models: Concepts, applications, and implementation. Guilford Publications.

P.S.: and since you are doing nothing different, it is just the basic multiple regression with different (shifted) variables, you can use the SPSS dialogue and hence the multiple imputations.

My dear Rainer Duesing !

Right on the button!!! I looked it up in Hayes book (page 266 to be exact!).

I will spell it out for others:

To probe an interaction at any level (percentile, SD or whatever) of the moderator variable:

1. Calculate the value of the desired level of the moderator (e.g., 16th percentile of M is at 2.4).

2. Center the moderator using this value (e.g., M16th=m-2.4)

3. Create the interaction term by multiplying the X by the centered M (e.g., INT16=X*M16th).

4. Enter it all into hierarchical regression:

REGRESSION

/MISSING LISTWISE

/STATISTICS COEFF OUTS R ANOVA CHANGE

/CRITERIA=PIN(.05) POUT(.10)

/NOORIGIN

/DEPENDENT Y

/METHOD=ENTER X M16th

/METHOD=ENTER INT16.

In the Coefficients table, look at the B and / or Beta and associated significance for the X variable in Model 2 (which includes the interaction term).

If you want to perform the entire process using mean centered variables, then first mean center the X and the moderator variables and then proceed to step 1 (calculate the value of the centered variable at its percentile or SD).

I compared my results using this method (raw and centered) and the results were the same as for PROCESS.

Thank you so much Rainer and you Andrew Hayes!

Hey glad I was able to help. One small remark concerning die calculation of percentiles in SPSS and R. There are several formulas/algorithms and the default setting is different for SPSS and R, therefore, check this before analyzing (although the differences should be typically very small, but I was bothered that they were not the same). In R you can use the "type" option within the quantile command to sepcify which formula to use. "6" is equal to the SPSS default setting, for example:

quantile(x, prob =c(.16, .84), type = 6)

does a level of significance on 0.1 in the table of conditional effects accepted?