help > F test and observed statistics problem? (2x2x2 cell means design)
Showing 1-7 of 7 posts
Oct 27, 2020 04:10 PM | Alex Karagiorgis - Aristotle University of Thessaloniki
F test and observed statistics problem? (2x2x2 cell means design)
Dear experts,
I have encountered something that troubles me when using an F test with the NBS method. I have a 2x2x2 design (Group: Young/Old, Time: Pre/Post-training, Condition: Standard/Deviant stimuli) with 15 subects in each Group. I have used the cell means approach for the design matrix, with 8 predictors.
The ordering of my data is this:
Subject1 Old Post Dev
Subject1 Old Post Std
Subject1 Old Pre Dev
Subject1 Old Pre Std
Subject2 Old Post Dev
Subject2 Old Post Std
.
.
.
Subject30 Young Pre Std
The design matrix is this:
(order of the columns: OldPreStd, OldPreDev, OldPostStd, OldPostDev, YoungPreStd, YoungPreDev, YoungPostStd, YoungPostDev)
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
0 1 0 0 0 0 0 0
1 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
.
.
.
0 0 0 0 1 0 0 0
I have also used exchange blocks as [1;2;3;4;1;2;3;4;1;2;3;4 etc. till the end]
The problem I encountered is that when I run an F test I get the same results (statistic value of each connection) depending on which predictors I use, irrespective of the weights in the contrast (1 or -1). For example, the main effect of Group (contrast: [-1 -1 -1 -1 1 1 1 1]) gives the same observed statistics with the main effect of Time [-1 -1 1 1 -1 -1 1 1] or Condition [-1 1 -1 1 -1 1 -1 1]. I have also separated factors with the same outcome. The effect of Time in only the Old group [-1 -1 1 1 0 0 0 0] gives the same result with the effect of Condition in only Old group [-1 1 -1 1 0 0 0 0] and the same result with their interaction [-1 1 1 -1 0 0 0 0]. Same thing happens with Time, Condition and their interaction in only Young, as well as with all the possible separations of the factors. It's as if the contrast weights play no role at all.
Taking a look at the code of NBSglm, I saw that in the case of F test the GLM.contrast is not anyhow used, except for simply regressing out the predictors denoted with a 0 in the contrast matrix. Is this normal?
So what is it that goes off here? Is it because of the cell means approach? But why? Are the exchange block correct?
Thank you in advance!
ps. I also tried another design like this:
order of the columns: All 1's, Group, Time, Condition
1 1 1 1
1 1 1 -1
1 1 -1 1
1 1 -1 -1
1 1 1 1
1 1 1 -1
.
.
.
1 -1 -1 -1
with contrasts:
[0 1 0 0] for main effect of Group and so on, but I prefer the cell means approach because I can test any combination of effects/interactions.
I have encountered something that troubles me when using an F test with the NBS method. I have a 2x2x2 design (Group: Young/Old, Time: Pre/Post-training, Condition: Standard/Deviant stimuli) with 15 subects in each Group. I have used the cell means approach for the design matrix, with 8 predictors.
The ordering of my data is this:
Subject1 Old Post Dev
Subject1 Old Post Std
Subject1 Old Pre Dev
Subject1 Old Pre Std
Subject2 Old Post Dev
Subject2 Old Post Std
.
.
.
Subject30 Young Pre Std
The design matrix is this:
(order of the columns: OldPreStd, OldPreDev, OldPostStd, OldPostDev, YoungPreStd, YoungPreDev, YoungPostStd, YoungPostDev)
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
0 1 0 0 0 0 0 0
1 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
.
.
.
0 0 0 0 1 0 0 0
I have also used exchange blocks as [1;2;3;4;1;2;3;4;1;2;3;4 etc. till the end]
The problem I encountered is that when I run an F test I get the same results (statistic value of each connection) depending on which predictors I use, irrespective of the weights in the contrast (1 or -1). For example, the main effect of Group (contrast: [-1 -1 -1 -1 1 1 1 1]) gives the same observed statistics with the main effect of Time [-1 -1 1 1 -1 -1 1 1] or Condition [-1 1 -1 1 -1 1 -1 1]. I have also separated factors with the same outcome. The effect of Time in only the Old group [-1 -1 1 1 0 0 0 0] gives the same result with the effect of Condition in only Old group [-1 1 -1 1 0 0 0 0] and the same result with their interaction [-1 1 1 -1 0 0 0 0]. Same thing happens with Time, Condition and their interaction in only Young, as well as with all the possible separations of the factors. It's as if the contrast weights play no role at all.
Taking a look at the code of NBSglm, I saw that in the case of F test the GLM.contrast is not anyhow used, except for simply regressing out the predictors denoted with a 0 in the contrast matrix. Is this normal?
So what is it that goes off here? Is it because of the cell means approach? But why? Are the exchange block correct?
Thank you in advance!
ps. I also tried another design like this:
order of the columns: All 1's, Group, Time, Condition
1 1 1 1
1 1 1 -1
1 1 -1 1
1 1 -1 -1
1 1 1 1
1 1 1 -1
.
.
.
1 -1 -1 -1
with contrasts:
[0 1 0 0] for main effect of Group and so on, but I prefer the cell means approach because I can test any combination of effects/interactions.
Oct 27, 2020 10:10 PM | Andrew Zalesky
RE: F test and observed statistics problem? (2x2x2 cell means design)
Hi Alex,
The formulation of your design matrix is not correct.
Represent each of your three main effects with a single column in your design matrix.
You should have 3 columns for the main effects + 1 column for the grand mean (if appropriate) + interaction effects.
The NBS uses the same design matrix formulation used by FSL.
If you are using exchange blocks to model within-subject means, use: [1 1 1 1 2 2 2 2 3 3 3 3....]
Further details of an F-test can be found in the help section as well as the example provided on the FSL site: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM
Best wishes,
Andrew
Originally posted by Alex Karagiorgis:
The formulation of your design matrix is not correct.
Represent each of your three main effects with a single column in your design matrix.
You should have 3 columns for the main effects + 1 column for the grand mean (if appropriate) + interaction effects.
The NBS uses the same design matrix formulation used by FSL.
If you are using exchange blocks to model within-subject means, use: [1 1 1 1 2 2 2 2 3 3 3 3....]
Further details of an F-test can be found in the help section as well as the example provided on the FSL site: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM
Best wishes,
Andrew
Originally posted by Alex Karagiorgis:
Dear experts,
I have encountered something that troubles me when using an F test with the NBS method. I have a 2x2x2 design (Group: Young/Old, Time: Pre/Post-training, Condition: Standard/Deviant stimuli) with 15 subects in each Group. I have used the cell means approach for the design matrix, with 8 predictors.
The ordering of my data is this:
Subject1 Old Post Dev
Subject1 Old Post Std
Subject1 Old Pre Dev
Subject1 Old Pre Std
Subject2 Old Post Dev
Subject2 Old Post Std
.
.
.
Subject30 Young Pre Std
The design matrix is this:
(order of the columns: OldPreStd, OldPreDev, OldPostStd, OldPostDev, YoungPreStd, YoungPreDev, YoungPostStd, YoungPostDev)
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
0 1 0 0 0 0 0 0
1 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
.
.
.
0 0 0 0 1 0 0 0
I have also used exchange blocks as [1;2;3;4;1;2;3;4;1;2;3;4 etc. till the end]
The problem I encountered is that when I run an F test I get the same results (statistic value of each connection) depending on which predictors I use, irrespective of the weights in the contrast (1 or -1). For example, the main effect of Group (contrast: [-1 -1 -1 -1 1 1 1 1]) gives the same observed statistics with the main effect of Time [-1 -1 1 1 -1 -1 1 1] or Condition [-1 1 -1 1 -1 1 -1 1]. I have also separated factors with the same outcome. The effect of Time in only the Old group [-1 -1 1 1 0 0 0 0] gives the same result with the effect of Condition in only Old group [-1 1 -1 1 0 0 0 0] and the same result with their interaction [-1 1 1 -1 0 0 0 0]. Same thing happens with Time, Condition and their interaction in only Young, as well as with all the possible separations of the factors. It's as if the contrast weights play no role at all.
Taking a look at the code of NBSglm, I saw that in the case of F test the GLM.contrast is not anyhow used, except for simply regressing out the predictors denoted with a 0 in the contrast matrix. Is this normal?
So what is it that goes off here? Is it because of the cell means approach? But why? Are the exchange block correct?
Thank you in advance!
ps. I also tried another design like this:
order of the columns: All 1's, Group, Time, Condition
1 1 1 1
1 1 1 -1
1 1 -1 1
1 1 -1 -1
1 1 1 1
1 1 1 -1
.
.
.
1 -1 -1 -1
with contrasts:
[0 1 0 0] for main effect of Group and so on, but I prefer the cell means approach because I can test any combination of effects/interactions.
I have encountered something that troubles me when using an F test with the NBS method. I have a 2x2x2 design (Group: Young/Old, Time: Pre/Post-training, Condition: Standard/Deviant stimuli) with 15 subects in each Group. I have used the cell means approach for the design matrix, with 8 predictors.
The ordering of my data is this:
Subject1 Old Post Dev
Subject1 Old Post Std
Subject1 Old Pre Dev
Subject1 Old Pre Std
Subject2 Old Post Dev
Subject2 Old Post Std
.
.
.
Subject30 Young Pre Std
The design matrix is this:
(order of the columns: OldPreStd, OldPreDev, OldPostStd, OldPostDev, YoungPreStd, YoungPreDev, YoungPostStd, YoungPostDev)
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
0 1 0 0 0 0 0 0
1 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0
.
.
.
0 0 0 0 1 0 0 0
I have also used exchange blocks as [1;2;3;4;1;2;3;4;1;2;3;4 etc. till the end]
The problem I encountered is that when I run an F test I get the same results (statistic value of each connection) depending on which predictors I use, irrespective of the weights in the contrast (1 or -1). For example, the main effect of Group (contrast: [-1 -1 -1 -1 1 1 1 1]) gives the same observed statistics with the main effect of Time [-1 -1 1 1 -1 -1 1 1] or Condition [-1 1 -1 1 -1 1 -1 1]. I have also separated factors with the same outcome. The effect of Time in only the Old group [-1 -1 1 1 0 0 0 0] gives the same result with the effect of Condition in only Old group [-1 1 -1 1 0 0 0 0] and the same result with their interaction [-1 1 1 -1 0 0 0 0]. Same thing happens with Time, Condition and their interaction in only Young, as well as with all the possible separations of the factors. It's as if the contrast weights play no role at all.
Taking a look at the code of NBSglm, I saw that in the case of F test the GLM.contrast is not anyhow used, except for simply regressing out the predictors denoted with a 0 in the contrast matrix. Is this normal?
So what is it that goes off here? Is it because of the cell means approach? But why? Are the exchange block correct?
Thank you in advance!
ps. I also tried another design like this:
order of the columns: All 1's, Group, Time, Condition
1 1 1 1
1 1 1 -1
1 1 -1 1
1 1 -1 -1
1 1 1 1
1 1 1 -1
.
.
.
1 -1 -1 -1
with contrasts:
[0 1 0 0] for main effect of Group and so on, but I prefer the cell means approach because I can test any combination of effects/interactions.
Oct 28, 2020 10:10 PM | Alex Karagiorgis - Aristotle University of Thessaloniki
RE: F test and observed statistics problem? (2x2x2 cell means design)
Dear Andrew,
Thanks a lot for your quick reply! I read the examples at the FSL site and took a look at the forum here and they were all enlightening. However I found no clues about how to split a factor (eg. search for Time effect only in one group). I have tried three methods discussed below. Now just for confirmation:
For the main effect of Group I averaged the 4 within subject conditions, set up a design with 1s or 0s and ran a simple 2-samples t-test with contrasts [1 -1] and [-1 1].
For the within subject effects (Time and Condition), I set up a design with 1 column (1 or -1) for Time, another for Condition (1 or -1) as well as a column for each of the interaction: TimeXCondition, GroupXTime and GroupXCondition and Triple Interaction. Also added a column with 1s for each subject and exchange blocks as you specified (thanks, I had it wrong!). Contrast for f-test of Time effect: [1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
So, hopefully up until now I have valid designs for the main effects and interactions. What I need to do now is to split my factors e.g. the Group factor, to search for Time effect only in Group1 or Group2 but I am not sure which is the best way to do that. For simplicity let's stick to this example.
I have already tried three methods:
1. Using the same design as the one for the within subject effects, but filling with zeros the values in the Time column that correspond to the unwanted group. This way the Time column contains 1 and -1 for the one group (for Pre and Post) and 0 for the other group. I'm not sure if this is correct, I haven't seen this in other examples. I ran a t-test with contrast same as above: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
2. Omitting the data of one group and setting up a new design with 60 connectivity matrices (half of the total 120 observations). This way I end up with a column for Time (1 or -1), a column for Condition (1 or -1), a column for their interaction and 15 columns for the subjects. I ran a t-test with contrast: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
3. As n.2 above, but since I am not interested in the Condition effect, I averaged it so I ended up with 30 connectivity matrices, the design contains a column for the Time effect (1 or -1) and 15 columns for the subjects and ran a paired-samples t-test with contrast [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
The results of all three methods are similar and they all give significant results, with n.3 having the least power and 1. and 2. being very close with somewhat larger statistic values. However I don't feel very confident about any of those. Is there some other method wich is optimal for this scenario?
Any insight would be greatly appreciated! Thanks in advance
All the best,
Alex Karagiorgis
edit: I made a mistake in describing n.3 and corrected it.
Thanks a lot for your quick reply! I read the examples at the FSL site and took a look at the forum here and they were all enlightening. However I found no clues about how to split a factor (eg. search for Time effect only in one group). I have tried three methods discussed below. Now just for confirmation:
For the main effect of Group I averaged the 4 within subject conditions, set up a design with 1s or 0s and ran a simple 2-samples t-test with contrasts [1 -1] and [-1 1].
For the within subject effects (Time and Condition), I set up a design with 1 column (1 or -1) for Time, another for Condition (1 or -1) as well as a column for each of the interaction: TimeXCondition, GroupXTime and GroupXCondition and Triple Interaction. Also added a column with 1s for each subject and exchange blocks as you specified (thanks, I had it wrong!). Contrast for f-test of Time effect: [1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
So, hopefully up until now I have valid designs for the main effects and interactions. What I need to do now is to split my factors e.g. the Group factor, to search for Time effect only in Group1 or Group2 but I am not sure which is the best way to do that. For simplicity let's stick to this example.
I have already tried three methods:
1. Using the same design as the one for the within subject effects, but filling with zeros the values in the Time column that correspond to the unwanted group. This way the Time column contains 1 and -1 for the one group (for Pre and Post) and 0 for the other group. I'm not sure if this is correct, I haven't seen this in other examples. I ran a t-test with contrast same as above: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
2. Omitting the data of one group and setting up a new design with 60 connectivity matrices (half of the total 120 observations). This way I end up with a column for Time (1 or -1), a column for Condition (1 or -1), a column for their interaction and 15 columns for the subjects. I ran a t-test with contrast: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
3. As n.2 above, but since I am not interested in the Condition effect, I averaged it so I ended up with 30 connectivity matrices, the design contains a column for the Time effect (1 or -1) and 15 columns for the subjects and ran a paired-samples t-test with contrast [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
The results of all three methods are similar and they all give significant results, with n.3 having the least power and 1. and 2. being very close with somewhat larger statistic values. However I don't feel very confident about any of those. Is there some other method wich is optimal for this scenario?
Any insight would be greatly appreciated! Thanks in advance
All the best,
Alex Karagiorgis
edit: I made a mistake in describing n.3 and corrected it.
Oct 29, 2020 11:10 PM | Andrew Zalesky
RE: F test and observed statistics problem? (2x2x2 cell means design)
Hi Alex,
I am not quite sure if I understand your comment about splitting a factor and searching for the effect of time in one group. Including an interaction between time and group should model this effect (although I am not sure if I understand your query).
You should not need to "split" the group factor to test the effect of time in one of the groups. The group x time interaction should model this effect and is a very typical approach.
I would avoid three-way and higher-order interactions, particularly for continuous variables, since these can be challenging to interpret.
Andrew
Originally posted by Alex Karagiorgis:
I am not quite sure if I understand your comment about splitting a factor and searching for the effect of time in one group. Including an interaction between time and group should model this effect (although I am not sure if I understand your query).
You should not need to "split" the group factor to test the effect of time in one of the groups. The group x time interaction should model this effect and is a very typical approach.
I would avoid three-way and higher-order interactions, particularly for continuous variables, since these can be challenging to interpret.
Andrew
Originally posted by Alex Karagiorgis:
Dear Andrew,
Thanks a lot for your quick reply! I read the examples at the FSL site and took a look at the forum here and they were all enlightening. However I found no clues about how to split a factor (eg. search for Time effect only in one group). I have tried three methods discussed below. Now just for confirmation:
For the main effect of Group I averaged the 4 within subject conditions, set up a design with 1s or 0s and ran a simple 2-samples t-test with contrasts [1 -1] and [-1 1].
For the within subject effects (Time and Condition), I set up a design with 1 column (1 or -1) for Time, another for Condition (1 or -1) as well as a column for each of the interaction: TimeXCondition, GroupXTime and GroupXCondition and Triple Interaction. Also added a column with 1s for each subject and exchange blocks as you specified (thanks, I had it wrong!). Contrast for f-test of Time effect: [1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
So, hopefully up until now I have valid designs for the main effects and interactions. What I need to do now is to split my factors e.g. the Group factor, to search for Time effect only in Group1 or Group2 but I am not sure which is the best way to do that. For simplicity let's stick to this example.
I have already tried three methods:
1. Using the same design as the one for the within subject effects, but filling with zeros the values in the Time column that correspond to the unwanted group. This way the Time column contains 1 and -1 for the one group (for Pre and Post) and 0 for the other group. I'm not sure if this is correct, I haven't seen this in other examples. I ran a t-test with contrast same as above: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
2. Omitting the data of one group and setting up a new design with 60 connectivity matrices (half of the total 120 observations). This way I end up with a column for Time (1 or -1), a column for Condition (1 or -1), a column for their interaction and 15 columns for the subjects. I ran a t-test with contrast: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
3. As n.2 above, but since I am not interested in the Condition effect, I averaged it so I ended up with 30 connectivity matrices, the design contains a column for the Time effect (1 or -1) and 15 columns for the subjects and ran a paired-samples t-test with contrast [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
The results of all three methods are similar and they all give significant results, with n.3 having the least power and 1. and 2. being very close with somewhat larger statistic values. However I don't feel very confident about any of those. Is there some other method wich is optimal for this scenario?
Any insight would be greatly appreciated! Thanks in advance
All the best,
Alex Karagiorgis
edit: I made a mistake in describing n.3 and corrected it.
Thanks a lot for your quick reply! I read the examples at the FSL site and took a look at the forum here and they were all enlightening. However I found no clues about how to split a factor (eg. search for Time effect only in one group). I have tried three methods discussed below. Now just for confirmation:
For the main effect of Group I averaged the 4 within subject conditions, set up a design with 1s or 0s and ran a simple 2-samples t-test with contrasts [1 -1] and [-1 1].
For the within subject effects (Time and Condition), I set up a design with 1 column (1 or -1) for Time, another for Condition (1 or -1) as well as a column for each of the interaction: TimeXCondition, GroupXTime and GroupXCondition and Triple Interaction. Also added a column with 1s for each subject and exchange blocks as you specified (thanks, I had it wrong!). Contrast for f-test of Time effect: [1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
So, hopefully up until now I have valid designs for the main effects and interactions. What I need to do now is to split my factors e.g. the Group factor, to search for Time effect only in Group1 or Group2 but I am not sure which is the best way to do that. For simplicity let's stick to this example.
I have already tried three methods:
1. Using the same design as the one for the within subject effects, but filling with zeros the values in the Time column that correspond to the unwanted group. This way the Time column contains 1 and -1 for the one group (for Pre and Post) and 0 for the other group. I'm not sure if this is correct, I haven't seen this in other examples. I ran a t-test with contrast same as above: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
2. Omitting the data of one group and setting up a new design with 60 connectivity matrices (half of the total 120 observations). This way I end up with a column for Time (1 or -1), a column for Condition (1 or -1), a column for their interaction and 15 columns for the subjects. I ran a t-test with contrast: [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
3. As n.2 above, but since I am not interested in the Condition effect, I averaged it so I ended up with 30 connectivity matrices, the design contains a column for the Time effect (1 or -1) and 15 columns for the subjects and ran a paired-samples t-test with contrast [1/-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0].
The results of all three methods are similar and they all give significant results, with n.3 having the least power and 1. and 2. being very close with somewhat larger statistic values. However I don't feel very confident about any of those. Is there some other method wich is optimal for this scenario?
Any insight would be greatly appreciated! Thanks in advance
All the best,
Alex Karagiorgis
edit: I made a mistake in describing n.3 and corrected it.
Oct 30, 2020 04:10 PM | Alex Karagiorgis - Aristotle University of Thessaloniki
RE: F test and observed statistics problem? (2x2x2 cell means design)
Dear Andrew,
I wanted to split the group factor because the group x time interaction only tells us that the time effect (post-pre) in one group is greater than the time effect in the other group. This could mean that there is an increase with time only in one group while the other remains the same, or that there is an increase in one group and a decrease in the other, or that both groups increase but one has a greater slope than the other. All those cases are covered by the interaction effect. So in order to know what is the case, I need to run t-tests for the time effect on each group separately. Is it clear now?
Best regards,
Alex
I wanted to split the group factor because the group x time interaction only tells us that the time effect (post-pre) in one group is greater than the time effect in the other group. This could mean that there is an increase with time only in one group while the other remains the same, or that there is an increase in one group and a decrease in the other, or that both groups increase but one has a greater slope than the other. All those cases are covered by the interaction effect. So in order to know what is the case, I need to run t-tests for the time effect on each group separately. Is it clear now?
Best regards,
Alex
Oct 31, 2020 01:10 AM | Andrew Zalesky
RE: F test and observed statistics problem? (2x2x2 cell means design)
Hi Alex,
Thanks for the clarification. To understand the nature of the interaction, I suggest performing a post-hoc analysis, if the interaction effect turns out to be significant.
Connectivity can be averaged over the connections comprising the subnetwork associated with the interaction effect. This will yield a single (averaged) connectivity value for each subject (see Manual for instructions and example). Post-hoc analyses (e.g. t-tests, or simply plotting as a function of time and group) can then be applied to the single connectivity value.
Of course, post-hoc analyses can be performed on each connection separately, but that could be cumbersome if the subnetwork is large.
Also note that the sign of the contrast (-1 versus 1) on the interaction effect will provide some insight into the type of interaction. Depending on how the interaction effect is modelled in the design matrix, a +1 contrast specifically tests for an increase in one group, while -1 tests for an increase in the other.
I hope that helps.
Andrew
Originally posted by Alex Karagiorgis:
Thanks for the clarification. To understand the nature of the interaction, I suggest performing a post-hoc analysis, if the interaction effect turns out to be significant.
Connectivity can be averaged over the connections comprising the subnetwork associated with the interaction effect. This will yield a single (averaged) connectivity value for each subject (see Manual for instructions and example). Post-hoc analyses (e.g. t-tests, or simply plotting as a function of time and group) can then be applied to the single connectivity value.
Of course, post-hoc analyses can be performed on each connection separately, but that could be cumbersome if the subnetwork is large.
Also note that the sign of the contrast (-1 versus 1) on the interaction effect will provide some insight into the type of interaction. Depending on how the interaction effect is modelled in the design matrix, a +1 contrast specifically tests for an increase in one group, while -1 tests for an increase in the other.
I hope that helps.
Andrew
Originally posted by Alex Karagiorgis:
Dear Andrew,
I wanted to split the group factor because the group x time interaction only tells us that the time effect (post-pre) in one group is greater than the time effect in the other group. This could mean that there is an increase with time only in one group while the other remains the same, or that there is an increase in one group and a decrease in the other, or that both groups increase but one has a greater slope than the other. All those cases are covered by the interaction effect. So in order to know what is the case, I need to run t-tests for the time effect on each group separately. Is it clear now?
Best regards,
Alex
I wanted to split the group factor because the group x time interaction only tells us that the time effect (post-pre) in one group is greater than the time effect in the other group. This could mean that there is an increase with time only in one group while the other remains the same, or that there is an increase in one group and a decrease in the other, or that both groups increase but one has a greater slope than the other. All those cases are covered by the interaction effect. So in order to know what is the case, I need to run t-tests for the time effect on each group separately. Is it clear now?
Best regards,
Alex
Oct 31, 2020 10:10 AM | Alex Karagiorgis - Aristotle University of Thessaloniki
RE: F test and observed statistics problem? (2x2x2 cell means design)
Dear Andrew,
Yes, the interaction was significant and the t-test of the interaction showed that the Young group had a greater time effect than the Old. I will try the post hoc approach that you suggested.
Thank you for the precious comments!
All the best,
Alex Karagiorgis
Yes, the interaction was significant and the t-test of the interaction showed that the Young group had a greater time effect than the Old. I will try the post hoc approach that you suggested.
Thank you for the precious comments!
All the best,
Alex Karagiorgis