Hi Alexander, great video! I just had a question about whether this can be extended to repeated measures factorial ANOVAs with 4 independent variables. I have a study with 4 IVs and 1 DV where, where the number of levels within the IV can be represented as 3x2x2x2. So essentially, each subject faced 24 different conditions. Does this imply that there are 24 groups or are there 24 measurements? I'm not sure which option under G power 3.0 to consider for such a design Thanks in advance!
7:40 - are you sure that you should divide this by 2 (for the 2 groups)? Seems pretty unintuitive for the creators of G*Power to expect users to do this. Haven't been able to see it in their instructions manual. Could you clarify this? Because to me, "sample size" in practice mostly refers to "number of participants".
Thanks for the video! To design a 3-intervention 3-period crossover design, do you think that the ANOVA repeated measures within-factor is enough? I am thinking to set number of groups to 6 to include all possible combinations of exposure to treatments, and number of measurements to 3. Do you think either between-within factor or MANOVA can be more accurate?
I’d say that you should focus on your smallest effect, whatever that may be. Trying to do a fancy power analysis might steer you in the wrong direction, thinking your sample size is enough. I would also choose the smallest between-subjects effect, too, as within-subs effects are inherently more powerful.
Thanks for this video Alexander, great stuff. I had one other question on this. If you look at the bottom of the screen, there is an "options" box, which lets you select a particular effect size specification. The default is GPower 3.0, but can be switched to "as in GPower 3.0 with implicit rho" "as in SPSS" and "as in Cohen (1988) - recommended". Do you know what these options mean and why this options box only appears for repeated measures F tests?
Thank you so much for your video! If I am conducting two different statistical analysis, to answer my two hypothesis, do I also need to conduct two power analyses?
Hi, Alexander, in your video 2:38 if I prefer to effect size from means how should I write the SD within each group? for example group 1 SD=1,50 and group 2 SD=1,08 thank you for your help.🙏 for example It does not accept (1,50-1,08)
Were you using periods or commas? I don’t know if that makes a difference depending on the language your gpower is set to, but it stuck out to me. If that’s not the case, could you elaborate on “it does not accept”?
thank you for this video! I was wondering about the number of measurements. N reflects the TOTAL number of measurements right? isn't the 'number of measurements' the TOTAL number of measurements instead of measurements per group?
Hello, I need to do a GPower to determine the sample size for my intervention. I was wondering if you could confirm whether I'm calculating it right. I have 4 modalities of intervention (virtual, in-person, active control and passive control), and will assess participants before and after the intervention, and at 6-month and 12-month follow-ups. I also have 7 dependable variables. Therefore, I asked for a Manova: Repeated measures, within-between interaction, with number of groups = 4 and number of measurements = 4. Have I perfomed the test as I should? Thank you!!
Dear Alexander, what test should I use for a mixed three-way AΝΟVA with two within and one between subject variables? All have 2 levels. Thanks in advance!
Awesome question! You will want to do the power analysis for your smallest effect. So if your smallest effect size is with one of your w/s variables, choose that option in G*Power. If the smallest is with the b/s variable, choose that one. If they are all relatively close to one another and it is seemingly hard to choose, go with the conventional wisdom: w/s are more powerful in general, so it is easier to detect effects. In that case, compute your sample size based on the weaker of the three variables: the b/s one! Good luck!
@@AlexanderSwan perhaps it is a very stupid question, but how do I know which variable has the smallest effect size? Also, does this mean I should not use the “ Fixed effects, special, main effects and interactions” ?
@@konstantinosd.9427 not a stupid question! It's perhaps the most important thing, so asking is perfect. You find your effect size by looking at the literature or computing it in gpower yourself. The smallest number (closest to zero) is your smallest effect. As for your second question, if you're interested in the interaction, then you could use that one. But again, only of the interactions are smallest effect you're trying to find.
@@AlexanderSwan Actually I am interested in both the interactions and main effects, that’s why I chose to do a three way mixed ANOVA. In a similar study I did, someone told me to use the effect size of 0.2 by just filling this in in g power. I thought of doing the same now, or would that be wrong? I guess I can use the repeated measures, between factors then? With number of groups 2 (= levels in BS factor) and number of measurements 2*2=4 ? (Because participants will be in both levels of two WS factors). What do you think?
Thank you so much for your very helpful video. Just to confirm my understanding, for the power analysis for anova within-between interaction, whatever the sample size we calculate for the interaction effect (given the alpha, the effect size, etc), that sample size should also work for detecting the between and within effects, in addition to the interaction effects right? My recollection from grad school years is that interaction effects require more power to detect than either of those effects alone. Thank you for any insights you can share with me on this.
Yes, interactions require more power, BUT, whatever the smallest effect is that you're looking for is going to require the largest sample size. For example (this is completely off the top of my head), if the interaction effect is found to be a f = .30 but the between effect is f = .15, then look for the between effect sample size. The interaction in this case requires less power than the b/s effect. On top of all of that, you want to make sure you are getting the sample size for the smallest, MOST IMPORTANT effect. There's a lot to be said about having a strong theoretical justification for your sample size rather than only relying on statistical effect sizes in complex models with multiple effects.
Hello! I am planning a 2x3 within-subjects ANOVA. If I run the within-subjects power test, would I leave group at 1 and measurements at 3 since that's the factor with the larger number of measurements? Or would it be all possible measurements (i.e., 6)? Does this also account for interactions? Thank you!
Look at each of these effects in the literature and base your sample size needs on the smallest predicted effects. Either that or focus on the effect you are most invested in (main effects or interactions). The power analysis should be geared toward one of those, and the rest of the effects will come along for the ride.
Hi Alexander, thanks for your video, it's very clear, and I know how to work on G*power now! I have a couple of questions about "Number of measurements" In 2*2 mixed design, you type "2" in Number of measurements. Does it mean the level of the within-subjects factor? So if I have a 3*4 mixed design, the number of measurement would be 4. Is that right? Besides, if I have a 2*2*3 mixed design, I mean I have two within-subject factors, what's the number for the measurements? and how do I get the number? Many thanks!
Very good questions! I'll tackle your second one first: when dealing with three-way or higher order interactions, you'll want simplify your power analysis. Anything larger than a two-way interaction would not be useful for a PA. Number of measurements generally does apply to within-Ss variable, if you're only getting one measurement per manipulation. In your example, if the W/S variable is the 4-level variable, then yes, that would be the number of measurements.
Thank you so much for the video! I wondered how I can get the Nonspericity correction and the Corr, among the rep. measures in a post-hoc analysis of my SPSS output? Can I calculate it?
Sphericity correction should really be left as the default value if you believe you won't violate the assumption. The correlation among variables can be done using the correlation function in SPSS
is g*power enough for calculating 2 (between) x 2 (within) x 2 (within) mixed ANOVA? I have been struggling with sample size determination for this design, but haven't ended up yet, do I need to use different methods (using morepower software etc.)
Hi, Thank you for providing the tutorial. But I have a couple of questions. I want to calculate the number of participants required. The study has two groups and is longitudinal in nature. Data is collected at two time points. The experiment has three conditions and a control condition. There will be other measures like proficiency, etc. (background information). How do I calculate the number of participants required in a group for a large effect size? Also, what exactly is the number of measurements?
Number of measurements is 2 for your two time periods. Number of groups is 4. I would determine sample size using the latter because it is harder to detect effects regardless of effect size estimates for between-subject variables. That will tell you how many people per group you want to recruit and your within-subject time variable should be sufficiently powered.
@@AlexanderSwan Hi, Thank you for the response. I'm sorry for not getting back to you sooner. I have a couple more questions. 1. Why did you say that the number of groups is 4? 2. I am interested in understanding the interaction between different variables in the study (change in performance within the experimental conditions over time; difference between the groups across time points). Hence, while determining the sample size, I have to select within-between interactions, right? 3. How can I determine effect size? Do I need to compute it? Or can I put 0.7-0.8 since that is considered a large effect size? 4. Also, how do I get a correlation among repeated measures? Can I assume 0.5? It would be great if you could clarify these doubts. Thank you once again.
@@chirstinajohnson4342 Found a person with similar questions😅! What should be the course of action for longitudinal studies with 2 groups of participants compared at 3 time points (pre, in-between and post)? How do you determine the number of measurements and groups? Kindly advise!!
@@AlexanderSwan Hi, It would be great if you could provide an explanation for these questions. I am trying to understand the whole concept of power analysis and using GPower.
Hi there, these options change what value of effect size you're using to calculate. So if you use "same as SPSS" it wants you to put in an effect size of f(U), which is different from other specifications.
Hello, I am running a repeated measures MANOVA and have conducted my power analysis but am unsure on how to write it up correctly, please could you advise me? I had one group, power was 80%, medium effect was 0.25, correlation among repeated measures was 0.3 and alpha was 0.05. is this written correctlty: The required sample size for a Multivariate Analysis of Variance (MANOVA) with one group using an alpha of 0.05, a power of 0.80, and a medium effect size (0.25) and a correlation among repeated measures of 0.3 revealed N = 46.
I was wondering about the follow-up measurements as well. Im designing a study with 8 different treatment arms, where every group is measured several times (baseline, week 4, 8, 12, 16, 20 and 24 > repeated measurements). a follow-up measurement is done at 12 months. is the follow-up measurement included in the number of measurements?
Yes, that's week 52. If it is part of the same variable, then it's a level/condition just the same. In this case, time is the variable, measured in weeks. Any same measurement of time can be measured in weeks. Easy peasy
I want to test the following three null hypotheses: H01: the dependent variable scores are the same for each level in factor 1 (ignoring factor 2). H02: the dependent variable scores are the same for each level in factor 2 (ignoring factor 1). H03: the two factors are independent or that interaction effect is not present.
@@AlexanderSwan even like that it is still fast and hard to follow...at least for me who I m not a native english speaker. And for me, whose teacher is useless cause he can t explain a damn thing about statistics so I have to watch videos on the internet...
Thank you so much! This was really helpful and thanks to you I managed to calculate my sample size for my UG dissertation! :D
Awesome! That's the goal and I'm glad it helped :)
Hi Alexander, great video! I just had a question about whether this can be extended to repeated measures factorial ANOVAs with 4 independent variables. I have a study with 4 IVs and 1 DV where, where the number of levels within the IV can be represented as 3x2x2x2. So essentially, each subject faced 24 different conditions. Does this imply that there are 24 groups or are there 24 measurements? I'm not sure which option under G power 3.0 to consider for such a design
Thanks in advance!
From my understanding, if all 4 IVs are RM IVs, then you have 24 measurements.
7:40 - are you sure that you should divide this by 2 (for the 2 groups)? Seems pretty unintuitive for the creators of G*Power to expect users to do this. Haven't been able to see it in their instructions manual. Could you clarify this? Because to me, "sample size" in practice mostly refers to "number of participants".
This is my understanding, because I have not been able to find it in the manual either. I have not contacted the developers myself however.
Thanks for the video! To design a 3-intervention 3-period crossover design, do you think that the ANOVA repeated measures within-factor is enough? I am thinking to set number of groups to 6 to include all possible combinations of exposure to treatments, and number of measurements to 3. Do you think either between-within factor or MANOVA can be more accurate?
I’d say that you should focus on your smallest effect, whatever that may be. Trying to do a fancy power analysis might steer you in the wrong direction, thinking your sample size is enough. I would also choose the smallest between-subjects effect, too, as within-subs effects are inherently more powerful.
Thanks for this video Alexander, great stuff. I had one other question on this. If you look at the bottom of the screen, there is an "options" box, which lets you select a particular effect size specification. The default is GPower 3.0, but can be switched to "as in GPower 3.0 with implicit rho" "as in SPSS" and "as in Cohen (1988) - recommended". Do you know what these options mean and why this options box only appears for repeated measures F tests?
These are different effect size values. So the default is Cohen's f. But you can change it to other effect size calcs.
Thank you so much for your video! If I am conducting two different statistical analysis, to answer my two hypothesis, do I also need to conduct two power analyses?
If they aren’t part of the same method, yes. If they are combined with the same sample, then only look for your smallest effect
@@AlexanderSwan Thank you for your quick reply! :)
Thank you a lot!!!!!!
Hi, Alexander,
in your video 2:38
if I prefer to effect size from means
how should I write the SD within each group? for example group 1 SD=1,50 and group 2 SD=1,08
thank you for your help.🙏
for example It does not accept (1,50-1,08)
Were you using periods or commas? I don’t know if that makes a difference depending on the language your gpower is set to, but it stuck out to me.
If that’s not the case, could you elaborate on “it does not accept”?
thank you for this video! I was wondering about the number of measurements. N reflects the TOTAL number of measurements right? isn't the 'number of measurements' the TOTAL number of measurements instead of measurements per group?
Hello, I need to do a GPower to determine the sample size for my intervention. I was wondering if you could confirm whether I'm calculating it right.
I have 4 modalities of intervention (virtual, in-person, active control and passive control), and will assess participants before and after the intervention, and at 6-month and 12-month follow-ups. I also have 7 dependable variables.
Therefore, I asked for a Manova: Repeated measures, within-between interaction, with number of groups = 4 and number of measurements = 4. Have I perfomed the test as I should?
Thank you!!
This is a good start. I would also run global effects MANOVA calculation for sample size. Compare the two and go with the larger sample size number.
@@AlexanderSwan Thank you so much for answering and for your videos!
Dear Alexander, what test should I use for a mixed three-way AΝΟVA with two within and one between subject variables? All have 2 levels. Thanks in advance!
???
Awesome question! You will want to do the power analysis for your smallest effect. So if your smallest effect size is with one of your w/s variables, choose that option in G*Power. If the smallest is with the b/s variable, choose that one. If they are all relatively close to one another and it is seemingly hard to choose, go with the conventional wisdom: w/s are more powerful in general, so it is easier to detect effects. In that case, compute your sample size based on the weaker of the three variables: the b/s one! Good luck!
@@AlexanderSwan perhaps it is a very stupid question, but how do I know which variable has the smallest effect size? Also, does this mean I should not use the “ Fixed effects, special, main effects and interactions” ?
@@konstantinosd.9427 not a stupid question! It's perhaps the most important thing, so asking is perfect. You find your effect size by looking at the literature or computing it in gpower yourself. The smallest number (closest to zero) is your smallest effect.
As for your second question, if you're interested in the interaction, then you could use that one. But again, only of the interactions are smallest effect you're trying to find.
@@AlexanderSwan Actually I am interested in both the interactions and main effects, that’s why I chose to do a three way mixed ANOVA. In a similar study I did, someone told me to use the effect size of 0.2 by just filling this in in g power. I thought of doing the same now, or would that be wrong? I guess I can use the repeated measures, between factors then? With number of groups 2 (= levels in BS factor) and number of measurements 2*2=4 ? (Because participants will be in both levels of two WS factors). What do you think?
Thank you so much for your very helpful video. Just to confirm my understanding, for the power analysis for anova within-between interaction, whatever the sample size we calculate for the interaction effect (given the alpha, the effect size, etc), that sample size should also work for detecting the between and within effects, in addition to the interaction effects right? My recollection from grad school years is that interaction effects require more power to detect than either of those effects alone. Thank you for any insights you can share with me on this.
Yes, interactions require more power, BUT, whatever the smallest effect is that you're looking for is going to require the largest sample size. For example (this is completely off the top of my head), if the interaction effect is found to be a f = .30 but the between effect is f = .15, then look for the between effect sample size. The interaction in this case requires less power than the b/s effect. On top of all of that, you want to make sure you are getting the sample size for the smallest, MOST IMPORTANT effect. There's a lot to be said about having a strong theoretical justification for your sample size rather than only relying on statistical effect sizes in complex models with multiple effects.
@@AlexanderSwan yes, I think your advice sounds right. Thank you so much for your time and thoughts on this.
Hello! I am planning a 2x3 within-subjects ANOVA. If I run the within-subjects power test, would I leave group at 1 and measurements at 3 since that's the factor with the larger number of measurements? Or would it be all possible measurements (i.e., 6)? Does this also account for interactions? Thank you!
Look at each of these effects in the literature and base your sample size needs on the smallest predicted effects. Either that or focus on the effect you are most invested in (main effects or interactions). The power analysis should be geared toward one of those, and the rest of the effects will come along for the ride.
@@AlexanderSwan Thanks so much!
Hi Alexander,
thanks for your video, it's very clear, and I know how to work on G*power now!
I have a couple of questions about "Number of measurements"
In 2*2 mixed design, you type "2" in Number of measurements. Does it mean the level of the within-subjects factor?
So if I have a 3*4 mixed design, the number of measurement would be 4.
Is that right?
Besides,
if I have a 2*2*3 mixed design, I mean I have two within-subject factors, what's the number for the measurements? and how do I get the number?
Many thanks!
Very good questions! I'll tackle your second one first: when dealing with three-way or higher order interactions, you'll want simplify your power analysis. Anything larger than a two-way interaction would not be useful for a PA.
Number of measurements generally does apply to within-Ss variable, if you're only getting one measurement per manipulation. In your example, if the W/S variable is the 4-level variable, then yes, that would be the number of measurements.
@@AlexanderSwan Thank you so much! it's very helpful!
Thank you so much for the video! I wondered how I can get the Nonspericity correction and the Corr, among the rep. measures in a post-hoc analysis of my SPSS output? Can I calculate it?
Sphericity correction should really be left as the default value if you believe you won't violate the assumption. The correlation among variables can be done using the correlation function in SPSS
is g*power enough for calculating 2 (between) x 2 (within) x 2 (within) mixed ANOVA? I have been struggling with sample size determination for this design, but haven't ended up yet, do I need to use different methods (using morepower software etc.)
I would use this test and focus on the between-subjects effect, as this is the one that will require the most power.
Hi,
Thank you for providing the tutorial. But I have a couple of questions. I want to calculate the number of participants required. The study has two groups and is longitudinal in nature. Data is collected at two time points. The experiment has three conditions and a control condition. There will be other measures like proficiency, etc. (background information). How do I calculate the number of participants required in a group for a large effect size? Also, what exactly is the number of measurements?
Number of measurements is 2 for your two time periods. Number of groups is 4. I would determine sample size using the latter because it is harder to detect effects regardless of effect size estimates for between-subject variables. That will tell you how many people per group you want to recruit and your within-subject time variable should be sufficiently powered.
@@AlexanderSwan Hi, Thank you for the response. I'm sorry for not getting back to you sooner. I have a couple more questions.
1. Why did you say that the number of groups is 4?
2. I am interested in understanding the interaction between different variables in the study (change in performance within the experimental conditions over time; difference between the groups across time points). Hence, while determining the sample size, I have to select within-between interactions, right?
3. How can I determine effect size? Do I need to compute it? Or can I put 0.7-0.8 since that is considered a large effect size?
4. Also, how do I get a correlation among repeated measures? Can I assume 0.5?
It would be great if you could clarify these doubts. Thank you once again.
@@chirstinajohnson4342 Found a person with similar questions😅!
What should be the course of action for longitudinal studies with 2 groups of participants compared at 3 time points (pre, in-between and post)? How do you determine the number of measurements and groups? Kindly advise!!
@@AlexanderSwan Hi, It would be great if you could provide an explanation for these questions. I am trying to understand the whole concept of power analysis and using GPower.
@@riya_r28 I am trying to understand how GPower works. I am not sure how well I can answer your questions!
What happens if we click on option tab in case of repeated measures within, between and select "same as SPSS
Hi there, these options change what value of effect size you're using to calculate. So if you use "same as SPSS" it wants you to put in an effect size of f(U), which is different from other specifications.
Hello, I am running a repeated measures MANOVA and have conducted my power analysis but am unsure on how to write it up correctly, please could you advise me? I had one group, power was 80%, medium effect was 0.25, correlation among repeated measures was 0.3 and alpha was 0.05. is this written correctlty:
The required sample size for a Multivariate Analysis of Variance (MANOVA) with one group using an alpha of 0.05, a power of 0.80, and a medium effect size (0.25) and a correlation among repeated measures of 0.3 revealed N = 46.
That sounds pretty good. Don't think too much on it, there's really no standard for reporting power analyses to my knowledge
@@AlexanderSwan thank you!
I was wondering about the follow-up measurements as well. Im designing a study with 8 different treatment arms, where every group is measured several times (baseline, week 4, 8, 12, 16, 20 and 24 > repeated measurements). a follow-up measurement is done at 12 months. is the follow-up measurement included in the number of measurements?
Yes, that's week 52. If it is part of the same variable, then it's a level/condition just the same. In this case, time is the variable, measured in weeks. Any same measurement of time can be measured in weeks. Easy peasy
hello sir. I would like to ask what I should use if my test is TWO WAY REPEATED MEASURES ANOVA.
I want to test the following three null hypotheses:
H01: the dependent variable scores are the same for each level in factor 1 (ignoring factor 2).
H02: the dependent variable scores are the same for each level in factor 2 (ignoring factor 1).
H03: the two factors are independent or that interaction effect is not present.
It sounds like repeated measures ANOVA within-subjects measure is your power analysis test
you speak WAY TOO FAST to be understandable... damn. It s a freaking hard course to understand...
Use the video controls to slow me down to .75 :)
@@AlexanderSwan even like that it is still fast and hard to follow...at least for me who I m not a native english speaker. And for me, whose teacher is useless cause he can t explain a damn thing about statistics so I have to watch videos on the internet...
@@suscaionutz3722 Well, at least you can watch a bit and pause! Can you slow me down to .5? I don't know if that's a thing...