Using Hayes PROCESS Macro for SPSS: Assumption Testing

Подписаться 4,3 тыс.

Просмотров 27 тыс.

50% 1

How can you check whether the statistical assumptions for your analysis hold when using the PROCESS macro for SPSS?
PROCESS is based on regression models and regressions have assumptions. Those assumptions have to be checked if you do not want to come to wrong conclusions about your data.
The PROCESS macro by Hayes has become a de facto standard for testing indirect effects and interactions in regression models. With thjis tool you can test for moderation (model 1), mediation (model 4) and different kinds of moderated mediation (model 7, model 8, model 14, model 15, etc.).
0:00 Start
0:31 Regression assumptions
1:07 Comprehensive assumptions test
1:58 Interaction variable
2:09 Checking assumptions in SPSS
3:34 Only important assumptions
This video will show you options for assumption checking in a PROCESS model:
- Linearity of the relationships between IVs and DV
- Normal distribution of residuals
- Homoscedasticity of the residuals
- Uncorrelatedness of residuals
- Absence of strong multicollinearity
- Appropriate scale properties
- Absence of extreme outliers
Annotated output of the assumption check:
www.regorz-statistik.de/en/che...
My CONSULTING SERVICES for PROCESS (moderation, mediation, moderated mediation):
www.regorz-statistik.de/en/con...
Here are my videos about moderated mediation with PROCESS:
Model 7: • PROCESS Model 7: Moder...
Model 8: • Moderated Mediation wi...
Model 14: • Moderated Mediation wi... (better to use model 15 than model 14!)
Model 15: • Moderated Mediation wi...
Models 21, 22, 28, 29: • Moderated Moderated Me...

Опубликовано:

6 авг 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 88

@Q4burt0n Год назад

Hey man, nice video, thank you! Really helped to refresh my memory, keep it up!

@charlotte-bp5cj 7 месяцев назад

bro this channel is so goated i owe you my whole thesis thank you so much regorz

@charlotte-bp5cj 7 месяцев назад

i wish i could cite you in my references

@Romangojuice 4 года назад

Amazing video! Thank you!

@eviemaria 3 года назад

so helpful, thank you so much!!!!

@Amicondrous 4 года назад

Great video. Thank you so much. But how weird that PROCESS cannot run these assumption checks itself...

@stasrieznik8408 4 года назад

Thank you for your help )

@L75hatesshuffle 4 года назад

Thank you!

@andyk5802 3 года назад

It would have been better if you could've walked us through how to interpret the regression results

@stephensalbod46 11 месяцев назад

Excellent video. All your videos are great. What happened to Outliers? Thank you, Steve

@RegorzStatistik 11 месяцев назад

If you rebuild the models in SPSS you can check for outliers in the conventional way.

@sippingpoetry 3 года назад

Hi, can I also use these instructions to test for the double moderation (model 2)? I computed the interaction independent variable x mod1 and independent variable x mod2 separately, and added both of them in my linear regression model (in addition to the other variables in my model). Is that okay?

@RegorzStatistik 3 года назад

Hi, yes you can use this approach for all PROCESS models. In each model, the output starts with the one (in the case of moderations) or two (in the case of mediations and moderated mediations) regression models PROCESS has calculated. You just have to rebuild those model(s). In the case of model 2, you need a model with the IV, the MOD1, the MOD2, the interaction IV-MOD1 and the interaction IV-MOD2 (and any covariates you may have included in your model).

@diillaa3 2 года назад

Hi! Thanks for the helpful video! Do you happen to have a source available that states that in order to check the regression assumptions, one should rebuild the models that Process calculates separately in SPSS? I couldn't find any reference to this in Haye's book on moderated mediation, unfortunately.

@RegorzStatistik 2 года назад

I haven't seen a source for that, yet. But there are at least two facts that should be indisputable: 1. PROCESS works based on regression analyses. 2. Regression analyses have assumptions. If you can find another way of dealing with those two facts instead of rebuilding the models in SPSS, that should be O.K.

@elsabaptista2330 3 года назад

Great video! I am doing a moderation analysis and all assumptions were met except for multicollinearity. The moderator and interaction term showed VIF values higher than 10, should I assume strong multicollinearity? If so, what's the alternative approach to test the moderation effect?

@RegorzStatistik 3 года назад

Unfortunately, for hypothesis testing and multicollinearity there are not really that many options. Techniques such as ridge regression are not a good fit for hypothesis testing, more for prediction. If your hypothesis test is not significant, then you could discuss multicollinearity as a possible reason for that in the discussion section of your paper. If your test is significant, then maybe you would like to look into: O’brien, R. M. (2007). A caution regarding rules of thumb for variance inflation factors. Quality & quantity, 41(5), 673-690. His position is that even with VIF above 10 it could be possible to interpret the data which you could discuss in your discussion section.

@juleindahouse 3 года назад

Hi! Thank you for the great video. Unfortunately, my output did not include a "product terms key" for model 4. Do I need to create an interaction term between my IV and mediator when I recreate the model?

@RegorzStatistik 3 года назад

In model 4 you don't have any interactions, so you don't have to create interaction terms.

@juleindahouse 3 года назад

@@RegorzStatistik Thank you so much for your fast response. That makes sense :)

@user-it5qh2gi8b 4 года назад

Which interactions would I need to create manually for moderated mediation (PROCESS model 8)?

@RegorzStatistik 4 года назад

I'd like to answer in a general way so this should be helpful for other PROCESS models, too: I would run the model with PROCESS , look at the output and check there which interactions it has constructed (in the tables "product terms key"). Then those interactions I would compute in SPSS.

@xibitbo9133 4 года назад

Thanks for your videos! I am using R for my mediation analysis. Since my data is incomplete, I do multiple imputation with mice in R. I was wondering when to check the assumptions in case of missing: before imputation or after? If I should check assumptions after imputation, do you know a good instruction how to do that in R? Thank you very much!!

@RegorzStatistik 4 года назад

I think I would check the assumptions after imputation because you run the tests with the imputed data so those have to conform to the test assumptions, I guess. Unfortunately, I don't know good instructions how to do that in R.

@xibitbo9133 4 года назад

@@RegorzStatistik thanks for your feedback! Unfortunately I did not find any clues at all in the internet and forums. but I continue my research. :/

@Leonie333lol Год назад

Thank you for the great video! However, I have a problem: my variables sometimes contain the value 0, which means that the interaction variables often contain the value 0 in these places, and also that I can't always work out Mahalanobis, for example. Is there any solution to this?

@RegorzStatistik Год назад

I don't know a solution for that.

@Ccccc- 4 года назад

some videos entered the interaction variable at block 2 (instead of block 1 which was what you did), what is the main difference between them?

@RegorzStatistik 4 года назад

It depends on the purpose. If I run a moderation in SPSS without using PROCESS I enter the interaction in block 2 in order to get the R2-change (=delta R²), i.e. the additional variance explained by the interaction. But if I use PROCESS and I am using SPSS for assumption checking only I am not interested in this information from SPSS (only in the assumption checks) because I get the R2-change from PROCESS: "Test(s) of highest order unconditional interaction(s)"

@Ccccc- 4 года назад

@@RegorzStatistik that makes much more sense now! Thanks for clarifying :)

@silkegeraats305 3 года назад

Hi, I am doing a mediation analysis. My independent variable is dichotomous. I was wondering if I could use the Durbin-Watson test in PROCESS to check the independence of residuals? If not: how do I check this assumption using PROCESS? Another question I have is that in another comment you said: In mediation (model 4) PROCESS uses bootstrapping for the indirect effect, so normality and homoscedasticity are not a problem. However, you should check linearity. But for the other effects in model 4 (a, b, c', maybe c) you don't get bootstrapping by default so there you have to check those assumptions (or choose bootstrapping and HC4 and check linearity).

@RegorzStatistik 3 года назад

It is questionable whether the Durbin-Watson test is helpful for a cross-sectional design because it is made for time series - it tests first order autocorrelation (whether there is a correlation between two *successive* errors). And in a cross-sectional design there are no successive errors so the result of the Durbin-Watson test depends on the (more or less random) order of the observations in your data file. For the assumption of independence of residuals I would look at the design - if it is cross-sectional and you don't have dependencies in your subjects (no nested designs, e.g. students in classes or employees in workgroups or husband&wife couples) that assumption should be met. For the key result of model 4, the significance of the indirect effect, you have boostrapping, so normality and homoscedasticity are not that relevant. But most of the time you will want to report the results for the a-path or b-path, too. And if that is the case you have to use bootstrapping there as well or check the assumptions by hand (in SPSS).

@silkegeraats305 3 года назад

@@RegorzStatistik Thank you so much for your answer, this really helps me writing my thesis! Two last questions: I don't see a model for bootstrapping the a-path and b-path in PROCESS, so how can I do this using SPSS? Also, do you have a reference about to confirm the assumption about independence of residuals?

@RegorzStatistik 3 года назад

You can request bootstrapping for all model parameters in each PROCESS model (PROCESS dialog on the bottom left hand side). I don't have a source available for the independence assumption, I would try regression textbooks.

@aliciadonadio2597 3 года назад

Ich geb mal Feedback: Super nett und sozial, solche Erklärvideos zu machen. Der Output mit Notizen ist auch nützlich. Der Akzent ist ziemlich stark im Englischen, aber das ist natürlich Nebensache. Einzige substantielle Kritik: Für einen Einsteiger ist es wahrscheinlich verwirrend, dass hier eine Kombination aus Moderation und Mediation durchgeführt wird. Es wäre hier wahrscheinlich besser gewesen, erstmal eine "reine" Mediation durchzuführen.

@RegorzStatistik 3 года назад

Danke für das Feedback! Das Video ist primär als Begleitung zu Tutorials für die verschiedenen Modelle der moderierten Mediation entstanden, daher auch am Beispiel einer moderierten Mediation. Es stimmt, für Einsteiger wäre es einfacher gewesen, das mit einer einfachen Moderation oder Mediation zu machen.

@nono-yd1tw 2 года назад

Hi! If I'm testing for a moderated mediation, with two variables controlled. When testing for linearity, do I test for the relationship between the IV and the DV, the mediator and the DV, and the interaction and the DV? or also between the control variables and the DV? Thank you.

@RegorzStatistik 2 года назад

Including covariates into the model assumes a linear relationship between the covariates and the DV. Therefore I'd check linearity there as well.

@nono-yd1tw 2 года назад

@@RegorzStatistik Thank you so much for the quick response. What would be the next step if there is no linearity between any of the variables and the DV? Thank you.

@nono-yd1tw 2 года назад

What are the implications for this analysis if you carry it out without linearity in all IV DV relationships?

@RegorzStatistik 2 года назад

Without linearity I would use polynomial regression (outside of PROCESS, as a path model, e.g. with AMOS).

@RegorzStatistik 2 года назад

Implications: One assumption for regression analyses is that you have specified the correct model. If this assumption is not true, then you can get false results.

@user-it5qh2gi8b 4 года назад

Can I also use standardized versions of continuous variables instead of mean centering?

@RegorzStatistik 4 года назад

I think it is possible, in principle. However, the interpretation in the context of moderation models can be tricky. So I prefer mean centering.

@user-it5qh2gi8b 4 года назад

@@RegorzStatistik thank you for your answer. I presume that interpretation could still be difficult if measurement scales are not the same for different variables. Is that correct? Also, I am uncertain as to whether I should mean center all variables or only the continuous ones.

@RegorzStatistik 4 года назад

@@user-it5qh2gi8b I would mean center continous variables only. For binary variables mean centering (or z-standardisation) makes the interpretation more difficult in most cases. Different scales are a problem for the interpretation primarily if you want to compare the effects of different predictors.

@user-it5qh2gi8b 4 года назад

@@RegorzStatistik This was really helpful. Thank you very much!

@emmanuelleriehl9117 Год назад

Hi, can I use HC0, HC1, HC2, HC3 or HC4 if the assumption of homoscedasticity is violated, or can this function in Process only be used to check this assumption?

@RegorzStatistik Год назад

I would use HC3 or HC4 - then you get results that are robust against violations of homoscedasticity (in that case you will see in the output that a robust SE has been used for the statistical tests).

@emmanuelleriehl9117 Год назад

@@RegorzStatistik Thank you, that's very helpful! If both bootstrapping and HC4 were applied, which output should be reported? The one where HC4 is applied or the bootstrap results, or a combination of both?

@RegorzStatistik Год назад

@@emmanuelleriehl9117 That depends on why you have used those two methods. Bootstrapping is robust against violations of normality (and to a certain extent against hetereoscedasticity), HC4 only against heteroscedasticity. If you haven't tested the assumptions I would report bootstrapping (confidence interval) and maybe in addition HC4-results (p-value)

@emmanuelleriehl9117 Год назад

@@RegorzStatistik I tested the assumptions of homogeneity of variance and normality and both were violated. That’s why I used both methods.

@RegorzStatistik Год назад

@@emmanuelleriehl9117 Then I would report both.

@ness702 3 года назад

Hi sir, Great video but I do have a question. I am using model 3 to test moderated moderation with only dichotomous predictors. However when I am putting my independent variables in a linear regression it leaves some of the interactive terms out due to multicollinearity. Also when I use another statistical software packages like Stata the variables are ommited. Do you have any suggestions how to solve this problem or advice? Kind regards, Ernesto

@RegorzStatistik 3 года назад

Have you put exactely the same 7 predictor variables into your linear regression that are listed in the output of PROCESS model 3? (That is: IV, MOD1, MOD2, and four interactions, IVxMOD1, IVxMOD2, MOD1xMOD2, IVxMOD1xMOD2) If yes: Are all 8 possible combinations of your binary variables (IV, MOD1, MOD2) represented in your data or are there some combinations that have a n of zero?

@ness702 3 года назад

@@RegorzStatistik Yes I have done as mentioned above and all combinations are represented in the dataset (I assume you mean that all of them have values?) The one problem I can imagine is the mean centering of the dichotomous predictors, but again this should not make a difference. However, when I ran the model with the uncentred predictors and interacting there is still variables ommited. Any idea why this is happening ?

@RegorzStatistik 3 года назад

@@ness702 With all combinations I mean (with a coding of e.g. 0-1 for the IV, MOD1 and MOD2) that all these combinations are present in the data set: 0-0-0 (IV, MOD1, MOD2) 0-0-1 0-1-0 0-1-1 1-0-0 1-0-1 1-1-0 1-1-1 If that is the case I don't really know what the source of the problem is. (Mean centering does not make much sense for binary predictors, but it shouldn't be the source of the problems)

@ness702 3 года назад

@@RegorzStatistik Thank you for your answer sir but I think I do not completely understand. Do you mean that all these posible combinations are computed seperately and put in the regression model. Currently I entered the following in the regression analysis: IV, MOD1, MOD2, IVxMOD1, IVxMOD2, MOD1xMOD2, IVxMOD1xMOD2. Since variables are mean centered there are no values of 0, because even if the values would be 0*0*0 then due to the mean centering computed interaction would still not equal 0. Do you perhaps mean that I should compute all these combinations myself 0-0-0 0-0-1 0-1-0 0-1-1 1-0-0 1-0-1 1-1-0 1-1-1 And then put all of these as independent variables in my regression model?

@RegorzStatistik 3 года назад

@@ness702 What I meant is this: Are all possible combinations of the three predictors present? 0-0-0 is meant to be a person who has the lower value on the IV, the lower value on MOD1 and the lower value on MOD2, etc. 1-1-1 is meant to be a person who has the higher value on all three variables. (And no, you should not enter these combinations as independent variables).

@nono-yd1tw 2 года назад

Hi! How can I cite your work? thank you :)

@RegorzStatistik 2 года назад

According to APA 7th edition: Regorz, A. (2020, February 28). Using Hayes' PROCESS macro for SPSS: Assumption testing. RU-vid. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-LgAA_nUYEg8.html (and the title in italics)

@anoukschippers3984 Год назад

Hi, i'm using Process Macro in R, but the homoskedasticity assumption is violated. Do I have to transform the DV or is homoskedasticity not important for Process Macro? If it is important what should I do? I was not really sure about this after watching the video.

@RegorzStatistik Год назад

You could use robust standard errors (HC3 or HC4 - those are used to deal with possible heteroskedasticity; parameter: hc =3 or hc = 4) or you could use bootstrapping for all model coefficients (parameter: modelbt =1), since bootstrapping is fairly robust against heteroskedasticity as well.

@anoukschippers3984 Год назад

@@RegorzStatistik yes Process Macro Analyses makes use of Bootstrapping right? So, is the homoskedasticity assumption is violated, that is not a problem?

@RegorzStatistik Год назад

@@anoukschippers3984 As a default PROCESS uses bootstrapping only for some analyses (indirect effect in model 4, conditional indirect effects and index of moderated mediation in models 7 - ...). If you want to have robust results for the other estimates as well, then you have to request that (e.g., by setting: modelbt = 1).

@anoukschippers3984 Год назад

@@RegorzStatistik Okay thank you very much! And do you have any reference for that? Because I need to substantiate my choices😅

@RegorzStatistik Год назад

@@anoukschippers3984 Maybe you could look into Hayes' book (Introduction to Mediation, Moderation, and Conditional Process Analysis: A Regression-Based Approach).

@fayehelffer5614 4 года назад

Is this the same for mediation analysis using PROCESS?

@RegorzStatistik 4 года назад

Yes, it applies to all PROCESS models, but with slight differences depending on the model type. In mediation (model 4) PROCESS uses bootstrapping for the *indirect* effect, so normality and homoscedasticity are not a problem. However, you should check linearity. But for the other effects in model 4 (a, b, c', maybe c) you don't get bootstrapping by default so there you have to check those assumptions (or choose bootstrapping and HC4 and check linearity). (One partial exception: If you use PROCESS with a binary mediator or a binary dependent variable you get logistic regression; there the assumptions and their test are different).

@fayehelffer5614 4 года назад

@@RegorzStatistik Amazing, thank you!

@AmorFatiYT 4 года назад

@@RegorzStatistik What are the assumptions for a binary dependent variable?

@AmorFatiYT 4 года назад

@@RegorzStatistik Would be very helpful for master thesis. Best, Max

@RegorzStatistik 4 года назад

@@AmorFatiYT I haven't run such a model yet. You could use this searchstring in Google: binary logistic regression assumptions

@Cranberrycheesecake1 3 года назад

why don't you use the standardized residuals?

@RegorzStatistik 3 года назад

If I remember it correctly from the statistics lectures the studentized residuals are slightly better because of their distributional characteristics. But testing normality with the standardized residuals is also OK, I think.

@Reine111 4 года назад

Thank you so much for this. However, it isn't simplified enough for me. You did not take it step by step, in a way that one can replicate.😥😥 I just ran a model 4 and would like to check assumptions.

@RegorzStatistik 4 года назад

Unfortunately, a step by step explanation isn't feasible here because the specific steps could differ for different PROCESS models (moderation, mediation, moderated mediation). Therefore I could only show general principles here.

@LaMagica1927 4 года назад

Thank you, great video!

@septiankusuma9682 3 года назад

Thank you so much for your video. The second approach in your video, if I have a problem with normality, just click on bootstrapping, and is it clear? Once again, in the heteros test, can I choose HC0, HC1, HC2 HC3? I'm sorry for my bad English.

@RegorzStatistik 3 года назад

If your sample size is 50 or higher, normality will not be necessary as long as you are using bootstrapping, yes. And for the robust standard errors, HC0-HC2 are somewhat dated. I would use HC4 or HC3 (probably HC4).

@septiankusuma9682 3 года назад

@@RegorzStatistik Thank you for your answer. When I try to compare H0-H4, I found H0 and H1 are the best results. Can I choose one of them?

@RegorzStatistik 3 года назад

Unfortunately not. You should decide on which robust error to use in advance before looking at the data.

@septiankusuma9682 3 года назад

@@RegorzStatistik OK so I just can leave the data & create new data set. Normality & classic asumption are my big problem in secondary data. Thank you so much.