Dear Kersten,

Thanks, I find it. And I have other questions: 1. The intercepts all set as one, so in this model it doesn't separate different subjects, or can say no individual subject change rate? 2. Should we set age according to different timepoint, or just use baseline age?

Thanks, Lanbo

On Tue, Mar 13, 2018 at 3:34 PM, Diers, Kersten /DZNE <Kersten.Diers@dzne.de

wrote:

Hello Lanbo,

the univariate example data can actually be downloaded from the LME tutorial website:

Search for: "An optional sample dataset which can be used to become familiar with the LME Matlab tools can be found here". The linked tar.gz archive contains two folders, one for the univariate and one for the mass-univariate example data.

Best regards,

Kersten

On Di, 2018-03-13 at 13:38 +0100, lanbo Wang wrote:

...

Dear Experts,

Hi, There is no example detail on website of LME tutorial. I have some question about it. Could you send me the table of ADNI univariate example data?

Thanks, Lanbo

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Partners Compliance HelpLine at http://www.partners.org/complianceline . If the e-mail was sent to you in error but does not contain patient information, please contact the sender and properly dispose of the e-mail.

Dear Kersten,

Thanks a lot, it's really help. I have another question, after I got results that two group have significant, then how could I get direction?

Thanks, Lanbo

On Tue, Mar 13, 2018 at 5:40 PM, Diers, Kersten /DZNE <Kersten.Diers@dzne.de

wrote:

Hello,

On Di, 2018-03-13 at 21:48 +0100, lanbo Wang wrote:

...

Dear Kersten,

Thanks, I find it. And I have other questions:

1. The intercepts all set as one, so in this model it doesn't

separate different subjects, or can say no individual subject change rate?

If I understood correctly, the question is whether or not we can get estimates for individual slopes across time?

If so, then yes, that's possible - but I'd have to run an analysis myself and look up how to do it exactly - I'll get back on this.

...

2. Should we set age according to different timepoint, or just use

baseline age?

It's better to use age at baseline.

One of the nice things of the LME is that it can separate the cross- sectional effect of age (at baseline) and the longitudinal effect of aging (=effect of time). So the aging effect is already incorporated within the 'time since baseline' variable, which of course should also be present in the model.

Since it is difficult to estimate and interpret effects that are very redundant (such as time vs age at each timepoint), it's better to just use age at baseline for the other regressor.

Best regards,

Kersten

...

Thanks, Lanbo

On Tue, Mar 13, 2018 at 3:34 PM, Diers, Kersten /DZNE <Kersten.Diers@ dzne.demailto:Kersten.Diers@dzne.de> wrote: Hello Lanbo,

the univariate example data can actually be downloaded from the LME tutorial website:

Search for: "An optional sample dataset which can be used to become familiar with the LME Matlab tools can be found here". The linked tar.gz archive contains two folders, one for the univariate and one for the mass-univariate example data.

Best regards,

Kersten

On Di, 2018-03-13 at 13:38 +0100, lanbo Wang wrote:

...

Dear Experts,

Hi, There is no example detail on website of LME tutorial. I have some question about it. Could you send me the table of ADNI univariate example data?

Thanks, Lanbo

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Partners Compliance HelpLine at http://www.partners.org/complianceline . If the e-mail was sent to you in error but does not contain patient information, please contact the sender and properly dispose of the e-mail.

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

Hi

sorry for the delayed response, I was not able to reply during the last week.

You are right that the F-test provided in the LME toolbox initially does not provide information about the direction of the effects.

You could do the following:

First, it is always useful to plot the data to get a first impression. However, if the statistical model also contains additional covariates, this plot does not completely reflect the effects as estimated by the model. So this may not give a complete picture.

A second option is to take a look at the sign of the estimated beta values, which also reflect the direction of the effects. These beta values are stored in the output of the 'lme_fit_FS' script, e.g. 'total_hipp_vol_stats.Bhat' in the tutorial example. For proper interpretation, you'd need to relate the beta values and their corresponding columns of the design matrix, and also take your contrasts of interest into account. So this is somewhat complicated.

As an alternative, the LME toolbox also provides so-called "signed p- values". Note that these are not p-values in the conventional sense. A signed p-value is the sign (-1 or +1) of the scalar product of a row of the contrast matrix and the vector of the estimated betas, and it can give information about the direction of a specific effect.

These signed p-values can be accessed from the output of the 'lme_F' script, see e.g. the 'F_C.sgn' variable for the tutorial example.

There will be as many signed p-values as there are rows in the contrast matrix, so each signed p-value corresponds to one row of the contrast matrix.

The interpretation of this signed p-value depends on how the contrast was formulated:

In a hypothetical example (different from and simpler than the tutorial example), suppose that one row of a contrast matrix starts like

0 1 -1 ...

and the entries correspond to group1, group2, group3, ..., then a positive signed p-value will in this particular case indicate that the difference 'group2 minus group3' is greater than zero, which means that group2 has greater values than group3. On the other hand, a negative signed p-value will in this particular case indicate that the difference 'group2 minus group3' is less than zero, and hence group3 must have greater values than group2.

However, also consider the equally valid alternative that a row of a contrast matrix starts like

0 -1 1 ...

and the entries again correspond to group1, group2, group3, ..., then a positive signed p-value will in this particular case indicate that the difference 'group3 minus group2' is greater than zero, which means that group3 has greater values than group2. On the other hand, a negative signed p-value will in this particular case indicate that the difference 'group3 minus group2' is less than zero, and hence group2 must have greater values than group3.

So the second interpretation is just the opposite as the first one. This illustrates that careful attention needs to be paid to the formulation of the contrasts. Also, it's best if the interpretation gained from the signed p-values agrees with the interpretation of a simple plot of the data.

To summarize, if we are testing for simple group differences, the F- test only provides a single p-value, which indicates if there is at least one significant difference among the several groups. To get an idea which groups actually differ, and in which direction, one needs to take a closer look, for example at those effects that are reflected in the rows of the contrast matrix; for this, one option is to use the signed p-values as described above.

Hope this helps,

Kersten

On Di, 2018-03-27 at 15:31 +0200, lanbo Wang wrote:

Dear Kersten,

I have a question about LME model. After I acquired p value, could I know which group is bigger?

Thanks, Lanbo

On Fri, Mar 16, 2018 at 12:13 AM, lanbo Wang <drrambow@gmail.com<mail to:drrambow@gmail.com>> wrote: Dear Kersten,

Thanks a lot, it's really help. I have another question, after I got results that two group have significant, then how could I get direction?

Thanks, Lanbo

On Tue, Mar 13, 2018 at 5:40 PM, Diers, Kersten /DZNE <Kersten.Diers@ dzne.demailto:Kersten.Diers@dzne.de> wrote: Hello,

On Di, 2018-03-13 at 21:48 +0100, lanbo Wang wrote:

...

Dear Kersten,

Thanks, I find it. And I have other questions:

1. The intercepts all set as one, so in this model it doesn't

separate different subjects, or can say no individual subject change rate?

If I understood correctly, the question is whether or not we can get estimates for individual slopes across time?

If so, then yes, that's possible - but I'd have to run an analysis myself and look up how to do it exactly - I'll get back on this.

...

2. Should we set age according to different timepoint, or just use

baseline age?

It's better to use age at baseline.

One of the nice things of the LME is that it can separate the cross- sectional effect of age (at baseline) and the longitudinal effect of aging (=effect of time). So the aging effect is already incorporated within the 'time since baseline' variable, which of course should also be present in the model.

Since it is difficult to estimate and interpret effects that are very redundant (such as time vs age at each timepoint), it's better to just use age at baseline for the other regressor.

Best regards,

Kersten

...

Thanks, Lanbo

On Tue, Mar 13, 2018 at 3:34 PM, Diers, Kersten /DZNE <Kersten.Diers@ dzne.dehttp://dzne.de<mailto:Kersten.Diers@dzne.demailto:Kersten .Diers@dzne.de>> wrote: Hello Lanbo,

the univariate example data can actually be downloaded from the LME tutorial website:

Search for: "An optional sample dataset which can be used to become familiar with the LME Matlab tools can be found here". The linked tar.gz archive contains two folders, one for the univariate and one for the mass-univariate example data.

Best regards,

Kersten

On Di, 2018-03-13 at 13:38 +0100, lanbo Wang wrote:

...

Dear Experts,

Hi, There is no example detail on website of LME tutorial. I have some question about it. Could you send me the table of ADNI univariate example data?

Thanks, Lanbo

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.ed u<mailto:Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh. harvard.edu> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Partners Compliance HelpLine at http://www.partners.org/complianceline . If the e-mail was sent to you in error but does not contain patient information, please contact the sender and properly dispose of the e-mail.

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

Hello,

this is a little bit hard to tell from the outside. So for the moment I have only some general ideas:

Did the algorithm converge in both cases? If not, I would have less confidence in the results.

One thing to consider besides the effect size and direction is the variability of the estimate (that's in .CovBhat) - and ultimately the significance of the effect, which combines estimated effect size and variability. Therefore, a true effect size near zero and / or a high variability of the estimate may also explain your observations, i.e. the direction change of the estimate.

Also keep in mind that the interpretation of a single beta coefficient assumes that all other coefficients in the model are set to zero. This includes additional contributions to this effect from groups other than the reference group if you have modeled these in additional group*time regressors. For exploration purposes, it might be worth to choose another reference group and/or to leave out those interaction regressors.

Finally, let us know if you are using the tutorial data for this analysis; we may be able to take a closer look in that case.

Best

Kersten

On Do, 2018-04-12 at 20:50 +0200, lanbo Wang wrote: Dear Kersten,

Thanks for reply. I have new question. When I set model use total_i_vol_stats = lme_fit_FS(X,[1 2],Y(:,2),ni) and total_i_vol_stats = lme_fit_FS(X,[1],Y(:,2),ni). But Bhat of two models' year have a lot different, for example, with 2 is 1.2422 without 2 is -0.5737. The direction has already changed. I feel confuse about this. Looking for your reply.

Thanks, Lanbo

On Wed, Mar 28, 2018 at 2:05 PM, Diers, Kersten /DZNE <Kersten.Diers@dzne.demailto:Kersten.Diers@dzne.de> wrote: Hi

sorry for the delayed response, I was not able to reply during the last week.

You are right that the F-test provided in the LME toolbox initially does not provide information about the direction of the effects.

You could do the following:

First, it is always useful to plot the data to get a first impression. However, if the statistical model also contains additional covariates, this plot does not completely reflect the effects as estimated by the model. So this may not give a complete picture.

A second option is to take a look at the sign of the estimated beta values, which also reflect the direction of the effects. These beta values are stored in the output of the 'lme_fit_FS' script, e.g. 'total_hipp_vol_stats.Bhat' in the tutorial example. For proper interpretation, you'd need to relate the beta values and their corresponding columns of the design matrix, and also take your contrasts of interest into account. So this is somewhat complicated.

As an alternative, the LME toolbox also provides so-called "signed p- values". Note that these are not p-values in the conventional sense. A signed p-value is the sign (-1 or +1) of the scalar product of a row of the contrast matrix and the vector of the estimated betas, and it can give information about the direction of a specific effect.

These signed p-values can be accessed from the output of the 'lme_F' script, see e.g. the 'F_C.sgn' variable for the tutorial example.

There will be as many signed p-values as there are rows in the contrast matrix, so each signed p-value corresponds to one row of the contrast matrix.

The interpretation of this signed p-value depends on how the contrast was formulated:

In a hypothetical example (different from and simpler than the tutorial example), suppose that one row of a contrast matrix starts like

0 1 -1 ...

and the entries correspond to group1, group2, group3, ..., then a positive signed p-value will in this particular case indicate that the difference 'group2 minus group3' is greater than zero, which means that group2 has greater values than group3. On the other hand, a negative signed p-value will in this particular case indicate that the difference 'group2 minus group3' is less than zero, and hence group3 must have greater values than group2.

However, also consider the equally valid alternative that a row of a contrast matrix starts like

0 -1 1 ...

and the entries again correspond to group1, group2, group3, ..., then a positive signed p-value will in this particular case indicate that the difference 'group3 minus group2' is greater than zero, which means that group3 has greater values than group2. On the other hand, a negative signed p-value will in this particular case indicate that the difference 'group3 minus group2' is less than zero, and hence group2 must have greater values than group3.

So the second interpretation is just the opposite as the first one. This illustrates that careful attention needs to be paid to the formulation of the contrasts. Also, it's best if the interpretation gained from the signed p-values agrees with the interpretation of a simple plot of the data.

To summarize, if we are testing for simple group differences, the F- test only provides a single p-value, which indicates if there is at least one significant difference among the several groups. To get an idea which groups actually differ, and in which direction, one needs to take a closer look, for example at those effects that are reflected in the rows of the contrast matrix; for this, one option is to use the signed p-values as described above.

Hope this helps,

Kersten

On Di, 2018-03-27 at 15:31 +0200, lanbo Wang wrote:

Dear Kersten,

I have a question about LME model. After I acquired p value, could I know which group is bigger?

Thanks, Lanbo

On Fri, Mar 16, 2018 at 12:13 AM, lanbo Wang <drrambow@gmail.commailto:drrambow@gmail.com<mail to:drrambow@gmail.commailto:to%3Adrrambow@gmail.com>> wrote: Dear Kersten,

Thanks a lot, it's really help. I have another question, after I got results that two group have significant, then how could I get direction?

Thanks, Lanbo

On Tue, Mar 13, 2018 at 5:40 PM, Diers, Kersten /DZNE <Kersten.Diers@ dzne.dehttp://dzne.de<mailto:Kersten.Diers@dzne.demailto:Kersten.Diers@dzne.de>> wrote: Hello,

On Di, 2018-03-13 at 21:48 +0100, lanbo Wang wrote:

...

Dear Kersten,

Thanks, I find it. And I have other questions:

1. The intercepts all set as one, so in this model it doesn't

separate different subjects, or can say no individual subject change rate?

If I understood correctly, the question is whether or not we can get estimates for individual slopes across time?

If so, then yes, that's possible - but I'd have to run an analysis myself and look up how to do it exactly - I'll get back on this.

...

2. Should we set age according to different timepoint, or just use

baseline age?

It's better to use age at baseline.

One of the nice things of the LME is that it can separate the cross- sectional effect of age (at baseline) and the longitudinal effect of aging (=effect of time). So the aging effect is already incorporated within the 'time since baseline' variable, which of course should also be present in the model.

Since it is difficult to estimate and interpret effects that are very redundant (such as time vs age at each timepoint), it's better to just use age at baseline for the other regressor.

Best regards,

Kersten

...

Thanks, Lanbo

On Tue, Mar 13, 2018 at 3:34 PM, Diers, Kersten /DZNE <Kersten.Diers@ dzne.dehttp://dzne.dehttp://dzne.de<mailto:Kersten.Diers@dzne.demailto:Kersten.Diers@dzne.de<mailto:Kerstenmailto:Kersten .Diers@dzne.demailto:Diers@dzne.de>>> wrote: Hello Lanbo,

the univariate example data can actually be downloaded from the LME tutorial website:

Search for: "An optional sample dataset which can be used to become familiar with the LME Matlab tools can be found here". The linked tar.gz archive contains two folders, one for the univariate and one for the mass-univariate example data.

Best regards,

Kersten

On Di, 2018-03-13 at 13:38 +0100, lanbo Wang wrote:

...

Dear Experts,

Hi, There is no example detail on website of LME tutorial. I have some question about it. Could you send me the table of ADNI univariate example data?

Thanks, Lanbo

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu<mailto:Freesurfer@nmr.mgh.harvard.edmailto:Freesurfer@nmr.mgh.harvard.ed u><mailto:Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu<mailto:Freesurfer@nmr.mghmailto:Freesurfer@nmr.mgh. harvard.eduhttp://harvard.edu>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Partners Compliance HelpLine at http://www.partners.org/complianceline . If the e-mail was sent to you in error but does not contain patient information, please contact the sender and properly dispose of the e-mail.

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu<mailto:Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

_______________________________________________ Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer