If you have several hypothesis some contain just main effects and some contain interaction effects - is it possible to perform several power analysis with GPower and just add the sample size up? Or is there even a way to combine that in one Power analysis? And additional question would be whether performing a test for a 3-way interaction would work the same as shown in the video? Thanks in advance!
It depends. If you test all hypotheses with the same data, you probably would calculate the needed sample size for each hypothesis and then take the highest number. If, however, you test hypotheses in different subsamples (e.g., one about male participants and one about female participants), then you add up the resulting sample sizes for the subsamples. If you test for a 3-way interaction you have 7 predictors (3 + 3 two-way interactions + 1 three-way-interaction). The rest of the power calculation stays the same. However, you should take into account that in most cases the effect for a 3-way interaction will be quite small. Because in a 3-way interaction any decreased reliability of the measures will multiply (of course, that is not a problem if you run an experiment and all 3 predictors are based on experimental conditions because then the assignment to the experimental conditions has perfect reliability).
Thank you for the clear video! Could this also be used for post-hoc power analysis when running Hayes PROCESS, or would you recommend something different?
One could use GPower for a post-hoc power analysis, using the r2-change from PROCESS model 1 output to calculate the necessary effect size. However, there is a debate whether post-hoc power analyses should even be used (and I am more on the side that does not want to run post-hoc power analyses). Some sources for that: Gelman, A. (2018, September 24th). Don’t calculate post-hoc power using observed estimate of effect size. Statistical Modeling, Causal Inference, and Social Science. statmodeling.stat.columbia.edu/2018/09/24/dont-calculate-post-hoc-power-using-observed-estimate-effect-size/ Hoenig, J. M., & Heisey, D. M. (2001). The abuse of power: The pervasive fallacy of power calculations for data analysis. The American Statistician, 55(1), 19-24.
This video is incredibly helpful, thank you for posting. One quick question, I am working on a research project that uses a Moderated Moderation Model (Model 3 in Hayes Process Analysis). I believe there are 8 total predictors in the model as I am using one covariate as well (IV, MOD1, MOD2, MOD1xIV, MOD2xIV, MOD1xMOD2, MOD1xMOD2xIV, COV1). In addition, I have 6 hypothesis which corresponds to 6 of the paths from the predictors to the DV (all paths except the MOD1xMOD2 path, and the COV1 path). Would my number of tested predictors be 6 and my number of total predictors be 8?
If you want to test 6 hypotheses then you have to run 6 power analyses (you can't realistically expect all your effects to have the same effect size). So, I think you would have 6 power calculations with 1-8 (tested-total) each.
Clear video! I have a question though. I am currently writing my master thesis and i have a 2x3 between subject design ; 1 independent variable: reward type (nominal) with 3 levels, 1 independent variable / moderator: brand loyalty (ordinal) with 2 levels and 1 dependent variable: customer retention (ordinal, likert scale). I want to use process 1. Do you think i should use this analysis as well? or should i use ANCOVA, or another one? Thank you in advance!
This is so incredibly helpful in helping me prepare for my thesis defense! Quick question: I've followed your instructions to conduct a power analysis using the F test for moderation with 5 predictors. When I enter .02 (small) as the effect size compared to .15 (medium), the projected sample size jumps from 395 to 55. Do you know what might explain such a drastic change? Also, I've been trained to consider a moderate effect threshold as the standard rather than small as you suggest. In case I need to defend this, why would you say assuming a small effect is the best choice? Thank you so much!
Such a drastic increase is quite normal when changing the effect size from medium to small, not only for regression. (E..g. for an independent samples t-test the size jumps from 102 to 620 if you change the effect size from medium to small there.) There is one reason why it is more difficult to find a medium size effect in a moderation analysis: Unless you use measures with (almost) perfect reliability (such as age, gender) in most of the cases you are using scales values. And a reliability less than 1.00 reduces the measured effect size. In moderation analysis you are interested in the interaction, which is the product of two variables. Therefore, here the (reduced) reliabilities of both measures work together to reduce the maximum effect size you could find.
Brilliant video, thank you. Quick question; I'm analysing the moderation of gender between social identity and psychological health. I will gather my data via questionnaires / surveys. In this video you suggested a small effect size, but would this mean my study lacks power? Also, I couldn't find literature to support that. Also, by doing so, I'd need a sample of 395, vs running a R2 deviation from zero with medium effect size which would would require a sample of 77... thoughts?
In your case maybe you could get away with a smaller sample. The reason why for a moderation in general you need larger samples than for other regressions is reliability. Reduced reliability attenuates the power of a test. And in a product term (interaction) you have the attenuating effect of two variables, IV and MOD, combined/multiplicated. For that reason it's in general more difficult to get significant results in moderation analyses. However, in your case that might be different because one of the variables for your product term, gender (assuming it is binary), could be measured with perfect reliability.
@@RegorzStatistik wow thank you for the speedy and amazing reply. It’s given me a lot to think on. I will definitely be using the medium effect size therefore and going with the smaller sample size. I doubt many, if any, of my sample would sit outside the binary (I am looking at new parents within the last three years) and putting it out to cloud research. It would be interesting if I did have data outside male/female but doubt it’ll be enough for this thesis.
Hi Sir, thanks for sharing! May I know if this estimated power is only for the interaction term? or the whole moderation model (including IV, moderator, and interaction term)? Thanks! I am looking forward to your generous reply!
"F: R2 increase" tests the additional effect of the interaction (against a model only with IV and moderator), the t-test for a simple regression coefficent tests that effect as well. The power for the whole model (which in most cases isn't very interesting) would be a different power calculation, "F: R2 deviation from zero".
@@RegorzStatistik Thanks for the prompt reply! Let me try to summarize your reply: 1) So does it mean that the sample size 395 is in power for testing the moderating interaction term? 2) if I want to estimate a sample size for a whole model, I should choose "F test, Linear multiple regression: Fixed model, R2 deviation from zero", am I right?
@@cailawrance8369 I'd say: 1 yes, 2 yes (however, that is almost never what we want - because I haven't seen yet in any journal article a hypothesis that IV, MOD and interaction *together* explain significant variance).
@@RegorzStatistik Hi sir, one quick follow-up question: what if I want to test the sample size for the IV in the moderation model, how should I do? Should I also indicate "1" in the tested predictor and "3" in the total predictors? It seems a bit weird to me. What do you think? Thank you very much!
Thank you for your video. If i have 4 hypothesis: the relaitonship between X and Y will be moderated by Z, the relationship between X and Y will be moderated by T, the relationship between X and W will be moderated by Z and the relationship between X and W will be moderated by T, what's the number of tested and total predictors?
You would run 4 power analyses, one for each hypothesis, and then take the largest resulting sample size. Tested predictors will be one for each hypothesis, total predictors depend on how you test these hypotheses (all in one model, 4 seperate models, etc.).
Thank you so much for this video. I want to test moderation effects in a multilevel model with the following interaction term for moderation: symptoms × condition × time2 In the overall model the lower interaction terms are also included which are: Time (T1 vs T2 vs T3) Time2 (2, 4, 9) Condition (experimental vs control), Condition × time2 Symptoms Condition × symptoms symptoms x time2 To calculate the power (using F-test) is it correct if I put number of tested predictors at 1, and number of of total predictors at 8 (i.e., lower interaction terms (7) + interaction term (1)
I found this video really helpful, thank you so much! I am pretty bad at these things so I have a quick question: is this suitable for my model? I have a categorial independent variable with two levels and a continuous dependent variable, I also have a categorical moderator with two levels. Thus, my experiment is a 2 (IV: one option vs. the other) x 2 (Moderator: one option vs. the other) between-subjects design. Do I need to follow the same steps as in the video to calculate my sample size? Or what steps are needed for my design? Thank you so much in advance!
Hi, Regorz! I am a new viewer and your videos are very helpful (the big S button has been pressed!). I am wondering whether it matters for the GPower analysis if the moderator is categorical or continuous?
@@RegorzStatistik Thank you! I have assumer small effect sizes to stay on the safe side no matter what (since I have no way of knowing any real estimate).
This is very helpful. I do have a question and would really appreciate your help. If I have two moderators (thus two interaction terms) will the total number of predictors be 5 then? Also, will the number of tested predictors be 2, since I have two moderators? Lastly, do you conduct different power analyses for each hypothesis you have? Thank you in advance for your reply and for this helpful video! :)
If you don't want to test a three-way interaction (i.e., both moderators should moderate the effect independently), then you have in total 5 predictors (IV, MOD1, MOD2, INT1, INT2). Number of tested predictors depends on your hypotheses. If you put in 2 (and chose multiple regression, R² increase), then you would test whether both moderations taken together explain additional variance. If you want to test each moderation seperately, I would use "multiple regression, single regression coefficient" instead. And yes, you should conduct a seperate power analysis for each hypothesis and then choose the largest resulting sample size - in that case your sample size is large enought to achieve (at least) the necessary power for all your hypotheses.
thank you, this is very helpful. i would like to ask tho, if i have two moderators with three hypotheses (one is the IV is negatively associated with DV, one is MOD1 moderates the interaction between them, and another one is MOD2) then how should i run the test for sample size? there have been a research with similar variables, so is that okay if i input the effect size to be 0.5, type I error 0.05, and type II error 0.95? with the total number of predictors to be 5 and tested predictor 4..? thank you
You have to run three power analyses (one for each hypothesis) and then take the largest resulting sample size. If you have 1 IV, 2 MOD, 2 Interactions, then you have 5 predictors. But for each hypothesis, I believe, you test just one regression weight, so the number of tested predictors is 1 for each power calculation. With the effect size I can't help you since I don't know your research area.
thank you so much, this was really helpful. but in description you wrote Hayes' PROCESS macro (model 1), but as I know moderation is either model 2 or 3, is this a mistake?
Model 1 is a simple moderation, model 2 is a moderation with two moderators (independent of each other) and model 3 is a moderated moderation (i.e. a moderation with a three-way interaction between IV, MOD1 and MOD2).
Question: I’m doing a Moderation Malaysia and I need to know how much the total sample size is. So this is basically how u calculate the total sample size for a moderation analysis? And what is “number of predictors” I didn’t get this. Can you please tell me what this is?
Yes, this is about calculating the total sample size you need. For a power analysis in the context of a multiple regression, and a moderation is tested in a multiple regression, you need the total number of predictor variables in the model. This is at least 3: IV, MOD, and interaction IV-MOD. If you have additional covariates, then their number is added to that.
Thank you so much, amazingly helpful. Swift question. I have three independent variables, one moderator, and one dependent variable. However, the first independent variable has three reflective sub-indicators; the second independent variable has its four reflective sub-indicators and the third independent variable has two sub-indicators. The moderator is single-facet and the dependent variable has three reflective sub-indicators. Under F-tests and R square increase as the video showed, when filling in the blank of the number of tested predictors and the total number of predictors, what numbers can I fill in, please? get confused..thank you so much in advance!
If I understand you correctly, you are running a moderation within a SEM framework? (because there you would have reflective sub-indicators). I am afraid G*Power does not help you for a power calculation with a latent interaction.
This may be a dumb question, but to ask: what is the difference between *linear multiple regression: fixed model, single regression coefficient* and *linear multiple regression: fixed model, r squared increase* ? which should I use for moderation analysis and for mediation analysis? I have 2 IVs, 1 mediator, 2 moderators, 4 DVs
"single regression coefficient" tests, whether one coefficent in a multiple regression is significantly different from zero. "r squared increase" tests, whether one step in a hierarchical regression explains significant additional variance. For a moderation you can use either of them, the first for testing the interaction, the second for testing, whether including the interaction in a hierarchical regression explains significantly more than a model without the interaction. Unfortunately, the last question is not so easy to answert. With that many variables you probably run a path model and power calculation for path models is not the purpose of GPower.
Wie berechne ich die Poweranalyse, wenn ich zwei Moderationsvariablen habe? In meiner Studie habe ich drei AVs und eine UV (Gruppe). Ich will für jede AV die beiden Moderatoren anschauen. Ich wäre sehr dankbar über Hilfe! Danke für die hilfreichen Videos!
Thank you so much for the video! I have a moderated mediation model (model 59 PROCESS), what would be the number of tested predictors and total predictors? Thank you so much in advance!
heyy, thank u for ur video !! my study has one iv one moderator that has two levels and a dv or two levels as well, does that mean my number of tested predictor is 1 and total number of predictors is 3 ? thanks
@@mandytan5205 Whether you have a binary or continous predictor (or a binary or contious moderator) that does not change the number of predictors for the power calculation (so: tested 1, total 3) for each of your two models (one model per dv).
An alternative would be Monte Carlo simulation, but I don't have any experience with that. (And I believe there are alternatives to GPower, e.g. an r package).
That depends. Do you want to run 2 moderation analyses, e.g., 2 times PROCESS model 1? Or do you want to run 1 analysis testing both moderators , e.g., PROCESS model 2?
@@horizon_954 With PROCESS model 2 you have to options: You could look at a test whether both moderators together moderate the relationsship (then number of tested predictors = 2). Or you could look at the two tests for the two moderators (then number of tested predictors =1). Since you have seperate hypotheses I think this second option would be the right one.
Unfortunately, that question is much too complex to fully answer in a comment. Short version: t-test tests one predictor, F-test tests a model or a model step. If you include one additional predictor (e.g. the interaction), then the F-test for that step should give the same p-value as a (two-tailed) t-test for that predictor.
@@marcogiancola2667 If you have a directional hypothesis, then you could perform a one-tailed t-test (and for that you would have to use a t-test, because there is no one-sided F-test). However, you don't have to; I guess the majority of papers with directional hypotheses use two-tailed tests.
With PROCESS model 3 you have (at least) 7 predictors IV, MOD1, MOD2, three 2-way-interactions and one 3-way-interaction. So you could calculate the power for a single regression coefficent (e.g., the 3-way-interaction, if that is the basis for your hypothesis) in a multiple regression with 7 predictors (+ the number of covariates if you have any in your model).
Very helkpful. Thank you!
If you have several hypothesis some contain just main effects and some contain interaction effects - is it possible to perform several power analysis with GPower and just add the sample size up? Or is there even a way to combine that in one Power analysis?
And additional question would be whether performing a test for a 3-way interaction would work the same as shown in the video?
Thanks in advance!
It depends. If you test all hypotheses with the same data, you probably would calculate the needed sample size for each hypothesis and then take the highest number. If, however, you test hypotheses in different subsamples (e.g., one about male participants and one about female participants), then you add up the resulting sample sizes for the subsamples.
If you test for a 3-way interaction you have 7 predictors (3 + 3 two-way interactions + 1 three-way-interaction). The rest of the power calculation stays the same. However, you should take into account that in most cases the effect for a 3-way interaction will be quite small. Because in a 3-way interaction any decreased reliability of the measures will multiply (of course, that is not a problem if you run an experiment and all 3 predictors are based on experimental conditions because then the assignment to the experimental conditions has perfect reliability).
Thank you for the clear video! Could this also be used for post-hoc power analysis when running Hayes PROCESS, or would you recommend something different?
One could use GPower for a post-hoc power analysis, using the r2-change from PROCESS model 1 output to calculate the necessary effect size.
However, there is a debate whether post-hoc power analyses should even be used (and I am more on the side that does not want to run post-hoc power analyses).
Some sources for that:
Gelman, A. (2018, September 24th). Don’t calculate post-hoc power using observed estimate of effect size. Statistical Modeling, Causal Inference, and Social Science. statmodeling.stat.columbia.edu/2018/09/24/dont-calculate-post-hoc-power-using-observed-estimate-effect-size/
Hoenig, J. M., & Heisey, D. M. (2001). The abuse of power: The pervasive fallacy of power calculations for data analysis. The American Statistician, 55(1), 19-24.
@@RegorzStatistik thanks for the quick reply!
This video is incredibly helpful, thank you for posting. One quick question, I am working on a research project that uses a Moderated Moderation Model (Model 3 in Hayes Process Analysis). I believe there are 8 total predictors in the model as I am using one covariate as well (IV, MOD1, MOD2, MOD1xIV, MOD2xIV, MOD1xMOD2, MOD1xMOD2xIV, COV1). In addition, I have 6 hypothesis which corresponds to 6 of the paths from the predictors to the DV (all paths except the MOD1xMOD2 path, and the COV1 path). Would my number of tested predictors be 6 and my number of total predictors be 8?
If you want to test 6 hypotheses then you have to run 6 power analyses (you can't realistically expect all your effects to have the same effect size).
So, I think you would have 6 power calculations with 1-8 (tested-total) each.
Clear video! I have a question though. I am currently writing my master thesis and i have a 2x3 between subject design ; 1 independent variable: reward type (nominal) with 3 levels, 1 independent variable / moderator: brand loyalty (ordinal) with 2 levels and 1 dependent variable: customer retention (ordinal, likert scale). I want to use process 1. Do you think i should use this analysis as well? or should i use ANCOVA, or another one? Thank you in advance!
Unfortunately, experiments are not my core area of expertise. I am more used to correlational designs.
This is so incredibly helpful in helping me prepare for my thesis defense! Quick question: I've followed your instructions to conduct a power analysis using the F test for moderation with 5 predictors. When I enter .02 (small) as the effect size compared to .15 (medium), the projected sample size jumps from 395 to 55. Do you know what might explain such a drastic change? Also, I've been trained to consider a moderate effect threshold as the standard rather than small as you suggest. In case I need to defend this, why would you say assuming a small effect is the best choice? Thank you so much!
Such a drastic increase is quite normal when changing the effect size from medium to small, not only for regression. (E..g. for an independent samples t-test the size jumps from 102 to 620 if you change the effect size from medium to small there.)
There is one reason why it is more difficult to find a medium size effect in a moderation analysis: Unless you use measures with (almost) perfect reliability (such as age, gender) in most of the cases you are using scales values. And a reliability less than 1.00 reduces the measured effect size. In moderation analysis you are interested in the interaction, which is the product of two variables. Therefore, here the (reduced) reliabilities of both measures work together to reduce the maximum effect size you could find.
Brilliant video, thank you. Quick question; I'm analysing the moderation of gender between social identity and psychological health. I will gather my data via questionnaires / surveys. In this video you suggested a small effect size, but would this mean my study lacks power? Also, I couldn't find literature to support that. Also, by doing so, I'd need a sample of 395, vs running a R2 deviation from zero with medium effect size which would would require a sample of 77... thoughts?
In your case maybe you could get away with a smaller sample.
The reason why for a moderation in general you need larger samples than for other regressions is reliability. Reduced reliability attenuates the power of a test. And in a product term (interaction) you have the attenuating effect of two variables, IV and MOD, combined/multiplicated. For that reason it's in general more difficult to get significant results in moderation analyses.
However, in your case that might be different because one of the variables for your product term, gender (assuming it is binary), could be measured with perfect reliability.
@@RegorzStatistik wow thank you for the speedy and amazing reply. It’s given me a lot to think on. I will definitely be using the medium effect size therefore and going with the smaller sample size. I doubt many, if any, of my sample would sit outside the binary (I am looking at new parents within the last three years) and putting it out to cloud research. It would be interesting if I did have data outside male/female but doubt it’ll be enough for this thesis.
Hi Sir, thanks for sharing!
May I know if this estimated power is only for the interaction term? or the whole moderation model (including IV, moderator, and interaction term)?
Thanks! I am looking forward to your generous reply!
"F: R2 increase" tests the additional effect of the interaction (against a model only with IV and moderator), the t-test for a simple regression coefficent tests that effect as well.
The power for the whole model (which in most cases isn't very interesting) would be a different power calculation, "F: R2 deviation from zero".
@@RegorzStatistik Thanks for the prompt reply! Let me try to summarize your reply:
1) So does it mean that the sample size 395 is in power for testing the moderating interaction term?
2) if I want to estimate a sample size for a whole model, I should choose "F test, Linear multiple regression: Fixed model, R2 deviation from zero", am I right?
@@cailawrance8369 I'd say: 1 yes, 2 yes (however, that is almost never what we want - because I haven't seen yet in any journal article a hypothesis that IV, MOD and interaction *together* explain significant variance).
@@RegorzStatistik Cool, thanks! I do think so! Thank you very much for your kind replies! I really appreciate!
@@RegorzStatistik Hi sir, one quick follow-up question: what if I want to test the sample size for the IV in the moderation model, how should I do? Should I also indicate "1" in the tested predictor and "3" in the total predictors? It seems a bit weird to me. What do you think? Thank you very much!
Thank you for your video. If i have 4 hypothesis: the relaitonship between X and Y will be moderated by Z, the relationship between X and Y will be moderated by T, the relationship between X and W will be moderated by Z and the relationship between X and W will be moderated by T, what's the number of tested and total predictors?
You would run 4 power analyses, one for each hypothesis, and then take the largest resulting sample size.
Tested predictors will be one for each hypothesis, total predictors depend on how you test these hypotheses (all in one model, 4 seperate models, etc.).
Thank you so much for this video. I want to test moderation effects in a multilevel model with the following interaction term for moderation: symptoms × condition × time2
In the overall model the lower interaction terms are also included which are:
Time (T1 vs T2 vs T3)
Time2 (2, 4, 9)
Condition (experimental vs control),
Condition × time2
Symptoms
Condition × symptoms
symptoms x time2
To calculate the power (using F-test) is it correct if I put number of tested predictors at 1, and number of of total predictors at 8 (i.e., lower interaction terms (7) + interaction term (1)
Unfortunately, I don't know how to run a power analysis for a multilevel model.
I found this video really helpful, thank you so much! I am pretty bad at these things so I have a quick question: is this suitable for my model? I have a categorial independent variable with two levels and a continuous dependent variable, I also have a categorical moderator with two levels. Thus, my experiment is a 2 (IV: one option vs. the other) x 2 (Moderator: one option vs. the other) between-subjects design. Do I need to follow the same steps as in the video to calculate my sample size? Or what steps are needed for my design? Thank you so much in advance!
In principle yes, I think. However, for that design I am not so sure what realistic effect size measures are.
Hi, Regorz! I am a new viewer and your videos are very helpful (the big S button has been pressed!). I am wondering whether it matters for the GPower analysis if the moderator is categorical or continuous?
Whether the moderator is binary or continous does not change the power calculation (however, it could change the realistic effect size).
@@RegorzStatistik Thank you! I have assumer small effect sizes to stay on the safe side no matter what (since I have no way of knowing any real estimate).
This is very helpful. I do have a question and would really appreciate your help. If I have two moderators (thus two interaction terms) will the total number of predictors be 5 then? Also, will the number of tested predictors be 2, since I have two moderators? Lastly, do you conduct different power analyses for each hypothesis you have? Thank you in advance for your reply and for this helpful video! :)
If you don't want to test a three-way interaction (i.e., both moderators should moderate the effect independently), then you have in total 5 predictors (IV, MOD1, MOD2, INT1, INT2).
Number of tested predictors depends on your hypotheses. If you put in 2 (and chose multiple regression, R² increase), then you would test whether both moderations taken together explain additional variance. If you want to test each moderation seperately, I would use "multiple regression, single regression coefficient" instead.
And yes, you should conduct a seperate power analysis for each hypothesis and then choose the largest resulting sample size - in that case your sample size is large enought to achieve (at least) the necessary power for all your hypotheses.
thank you, this is very helpful. i would like to ask tho, if i have two moderators with three hypotheses (one is the IV is negatively associated with DV, one is MOD1 moderates the interaction between them, and another one is MOD2) then how should i run the test for sample size? there have been a research with similar variables, so is that okay if i input the effect size to be 0.5, type I error 0.05, and type II error 0.95? with the total number of predictors to be 5 and tested predictor 4..? thank you
You have to run three power analyses (one for each hypothesis) and then take the largest resulting sample size.
If you have 1 IV, 2 MOD, 2 Interactions, then you have 5 predictors. But for each hypothesis, I believe, you test just one regression weight, so the number of tested predictors is 1 for each power calculation.
With the effect size I can't help you since I don't know your research area.
@@RegorzStatistik i see, thank you!!! i will read more about it. have a nice day!
thank you so much, this was really helpful.
but in description you wrote Hayes' PROCESS macro (model 1), but as I know moderation is either model 2 or 3, is this a mistake?
Model 1 is a simple moderation, model 2 is a moderation with two moderators (independent of each other) and model 3 is a moderated moderation (i.e. a moderation with a three-way interaction between IV, MOD1 and MOD2).
Thank you so much for this helpful video! Would you be able to advise how we would calculate GPower for a moderation analysis using Hayes Model 2?
Maybe my answer to the comment for this video by Maya Elshafei can help you there.
Question: I’m doing a Moderation Malaysia and I need to know how much the total sample size is. So this is basically how u calculate the total sample size for a moderation analysis? And what is “number of predictors” I didn’t get this. Can you please tell me what this is?
Yes, this is about calculating the total sample size you need.
For a power analysis in the context of a multiple regression, and a moderation is tested in a multiple regression, you need the total number of predictor variables in the model. This is at least 3: IV, MOD, and interaction IV-MOD. If you have additional covariates, then their number is added to that.
Thank you so much, amazingly helpful. Swift question. I have three independent variables, one moderator, and one dependent variable. However, the first independent variable has three reflective sub-indicators; the second independent variable has its four reflective sub-indicators and the third independent variable has two sub-indicators. The moderator is single-facet and the dependent variable has three reflective sub-indicators. Under F-tests and R square increase as the video showed, when filling in the blank of the number of tested predictors and the total number of predictors, what numbers can I fill in, please? get confused..thank you so much in advance!
If I understand you correctly, you are running a moderation within a SEM framework? (because there you would have reflective sub-indicators). I am afraid G*Power does not help you for a power calculation with a latent interaction.
@@RegorzStatistik got it, thank you, Dr.
This may be a dumb question, but to ask: what is the difference between *linear multiple regression: fixed model, single regression coefficient* and *linear multiple regression: fixed model, r squared increase* ? which should I use for moderation analysis and for mediation analysis? I have 2 IVs, 1 mediator, 2 moderators, 4 DVs
"single regression coefficient" tests, whether one coefficent in a multiple regression is significantly different from zero.
"r squared increase" tests, whether one step in a hierarchical regression explains significant additional variance.
For a moderation you can use either of them, the first for testing the interaction, the second for testing, whether including the interaction in a hierarchical regression explains significantly more than a model without the interaction.
Unfortunately, the last question is not so easy to answert. With that many variables you probably run a path model and power calculation for path models is not the purpose of GPower.
@@RegorzStatistik now I understand it much better. Thanks!
Wie berechne ich die Poweranalyse, wenn ich zwei Moderationsvariablen habe? In meiner Studie habe ich drei AVs und eine UV (Gruppe). Ich will für jede AV die beiden Moderatoren anschauen. Ich wäre sehr dankbar über Hilfe! Danke für die hilfreichen Videos!
Hier ist ein Video für die Power bei zwei Moderatoren (Modell 2):
ruclips.net/video/60ACU9DtjrI/видео.html
Thank you so much for the video! I have a moderated mediation model (model 59 PROCESS), what would be the number of tested predictors and total predictors? Thank you so much in advance!
I don't think you can use this approach to calcualte the power for a moderated mediation, unfortunately.
@@RegorzStatistik Ok, thank you so much for your response!
heyy, thank u for ur video !! my study has one iv one moderator that has two levels and a dv or two levels as well, does that mean my number of tested predictor is 1 and total number of predictors is 3 ? thanks
If the dv has two levels then you would need a power analysis for a logistic regression, not for a linear regression.
@@RegorzStatistik sorry i meant to say that i have two dvs
@@mandytan5205 Whether you have a binary or continous predictor (or a binary or contious moderator) that does not change the number of predictors for the power calculation (so: tested 1, total 3) for each of your two models (one model per dv).
Is there an alternative method for computing the a priori sample size for moderation? thanks
An alternative would be Monte Carlo simulation, but I don't have any experience with that. (And I believe there are alternatives to GPower, e.g. an r package).
If I have 2 Hypothesis, each with different Moderators, makes that Number of Tested Predictors = 2 ?
I would be very glad if you could help me :S
That depends. Do you want to run 2 moderation analyses, e.g., 2 times PROCESS model 1? Or do you want to run 1 analysis testing both moderators , e.g., PROCESS model 2?
@@RegorzStatistik First, thanks for your answer ;) I think i will chose Model Nr. 2, so how many tested predictors would that be?
@@horizon_954 With PROCESS model 2 you have to options:
You could look at a test whether both moderators together moderate the relationsship (then number of tested predictors = 2).
Or you could look at the two tests for the two moderators (then number of tested predictors =1). Since you have seperate hypotheses I think this second option would be the right one.
@@RegorzStatistik Yes, that makes sense, thank you so much 👍
are the t test and f test equivalent? If not, which are the differences?
Unfortunately, that question is much too complex to fully answer in a comment.
Short version: t-test tests one predictor, F-test tests a model or a model step. If you include one additional predictor (e.g. the interaction), then the F-test for that step should give the same p-value as a (two-tailed) t-test for that predictor.
@@RegorzStatistik Thank you for your kind answer. However, if I have a directional hypothesis I should performe a one-tailed t test or not?
@@marcogiancola2667 If you have a directional hypothesis, then you could perform a one-tailed t-test (and for that you would have to use a t-test, because there is no one-sided F-test). However, you don't have to; I guess the majority of papers with directional hypotheses use two-tailed tests.
Model 3 (moderated moderation), how can i calculate sample size?
With PROCESS model 3 you have (at least) 7 predictors IV, MOD1, MOD2, three 2-way-interactions and one 3-way-interaction. So you could calculate the power for a single regression coefficent (e.g., the 3-way-interaction, if that is the basis for your hypothesis) in a multiple regression with 7 predictors (+ the number of covariates if you have any in your model).
@@RegorzStatistik Sir, could you share this paper with us? I am going to cite it in my paper. Thank you!
@@cailawrance8369 I don't have a paper for that.
@@RegorzStatistik Ok, thank you anyway!