Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Martin Reuter]

Hi Matthieu, 

no, sorry. Maybe you can find a local biostatistician who knows. 

A slope estimate with less time points is more uncertain than one with
more time points (obviously). Also a slope estimate with time points on
a line is more certain than one with time points values all over the
place. This information is used in the LME model. But I don't know to
what extend any potential biases are removed. For example, if one group
has only 2 time points and the other had 5, I would be very suspicious.
You can, of course, always reduce the number of time points in the
second group to match the first and see if your results remain stable. 

Best, Martin



On Fri, 2018-12-07 at 22:33 +0100, Matthieu Vanhoutte wrote:
> External Email - Use Caution
> Hi Martin,
> 
> I come back on my previous question concerning survival bias in LME.
> 
> Do you have an advice ?
> 
> Best,
> Matthieu
> 
> > Le 19 oct. 2018 à 16:30, Matthieu Vanhoutte  > l.com> a écrit :
> > 
> > Hi Martin,
> > 
> > Do you mean that high attrition (less time points with time) would
> > induce more variance on the slope estimates but this effect would
> > be compensated by LME modeling ?
> > 
> > Would there be proportion of early drop outs to respect in order to
> > compensate bias with LME ?
> > 
> > Thanks,
> > Matthieu
> > 
> > 
> > Le ven. 19 oct. 2018 à 16:19, Martin Reuter  > d.edu> a écrit :
> > > Hi Matthieu, 
> > > 
> > > in LME it is OK if subjects have differently many time points in
> > > general. You need to ask a statistician about your specific
> > > setup, but
> > > I think it might be OK (basically less time points, means more
> > > variance
> > > on the slope estimates, but that should be considered in LME).
> > > But I am
> > > not a statistician. 
> > > 
> > > Best, Martin
> > > 
> > > 
> > > On Wed, 2018-10-17 at 18:46 +0200, Matthieu Vanhoutte wrote:
> > > > External Email - Use Caution
> > > > Hi Martin,
> > > > 
> > > > Thanks for your answer.
> > > > 
> > > > I actually compare neurospychological scores at baseline
> > > between
> > > > drop-out subjects and subjects with full time-points. If I ever
> > > find
> > > > that drop-out subjects are more severely affected than the
> > > subjects
> > > > with full time-points, then there might be a bias in the
> > > results of
> > > > my LME study ?
> > > > 
> > > > How could I argue that significant patterns found in my LME
> > > study
> > > > between both groups are still valid accounting for this bias ?
> > > Is LME
> > > > method robust enough for compensating this kind of drop-out ?
> > > > 
> > > > Best,
> > > > Matthieu
> > > > 
> > > > 
> > > > Le mer. 17 oct. 2018 à 18:33, Martin Reuter  > > rvard.
> > > > edu> a écrit :
> > > > > Hi Matthieu, 
> > > > > 
> > > > > 1) survival analysis is typically used if you want to detect
> > > if the
> > > > > time to an event is longer in one group vs the other (e.g.
> > > one
> > > > > group
> > > > > gets placebo the other drug and we want to know if recurrence
> > > is
> > > > > later
> > > > > in the drug group). Not sure this is what you need. The nice
> > > thing
> > > > > is,
> > > > > it can deal with drop-outs
> > > > > 
> > > > > 2) No, you can directly test that (e.g. do more dieseased
> > > drop out
> > > > > than
> > > > > healthy, or are the dropouts on average more advanced (test-
> > > scores,
> > > > > hippo-volume etc) than the diseased at baseline... many
> > > options.
> > > > > you
> > > > > could also test interactions with age , gender etc. However,
> > > not
> > > > > finding an interaction may not mean there is no bias, it is
> > > just
> > > > > small
> > > > > enough to go undetected with your data size. 
> > > > > 
> > > > > 3) Survival analysis is a different analysis than LME. 
> > > > > 
> > > > > Best, Martin
> > > > > 
> > > > > On Tue, 2018-10-16 at 16:15 +, Matthieu Vanhoutte wrote:
> > > > > > External Email - Use Caution
> > > > > > Hi Martin,
> > > > > > 
> > > > > > It's been a long time since this discussion but I return on
> > > this
> > > > > from
> > > > > > now... The problem is that I followed longitudinal images
> > > of two
> > > > > > groups where I had mainly missing time points at the end.
> > > Than
> > > > > you
> > > > > > suggested:
> > > > > > If you have mainly missing time points at the end, this
> > > will bias
> > > > > > your analysis to some extend, as the remaining ones may be
> > > > > extremely
> > > > > > healthy, as probably the more diseased ones drop out. You
> > > may
> > > > > want to
> > > > > > do a time-to-event (or survival-analysis) which considers
> > > early
> > > > > drop-
> > > > > > out.
> > > > > > 
> > > > > > 1) I know the survival analysis toolbox on matlab, but now
> > > I
> > > > > would
> > > > > > like to know what information will this survival analysis
> > > give to
> > > > > me
> > > > > > ? 
> > > > > > 2) Will this analysis tell me if there is a bias ?
> > > > > > 3) How to consider 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

External Email - Use Caution

Hi Martin,

I come back on my previous question concerning survival bias in LME.

Do you have an advice ?

Best,
Matthieu

> Le 19 oct. 2018 à 16:30, Matthieu Vanhoutte  a 
> écrit :
> 
> Hi Martin,
> 
> Do you mean that high attrition (less time points with time) would induce 
> more variance on the slope estimates but this effect would be compensated by 
> LME modeling ?
> 
> Would there be proportion of early drop outs to respect in order to 
> compensate bias with LME ?
> 
> Thanks,
> Matthieu
> 
> 
> Le ven. 19 oct. 2018 à 16:19, Martin Reuter  > a écrit :
> Hi Matthieu, 
> 
> in LME it is OK if subjects have differently many time points in
> general. You need to ask a statistician about your specific setup, but
> I think it might be OK (basically less time points, means more variance
> on the slope estimates, but that should be considered in LME). But I am
> not a statistician. 
> 
> Best, Martin
> 
> 
> On Wed, 2018-10-17 at 18:46 +0200, Matthieu Vanhoutte wrote:
> > External Email - Use Caution
> > Hi Martin,
> > 
> > Thanks for your answer.
> > 
> > I actually compare neurospychological scores at baseline between
> > drop-out subjects and subjects with full time-points. If I ever find
> > that drop-out subjects are more severely affected than the subjects
> > with full time-points, then there might be a bias in the results of
> > my LME study ?
> > 
> > How could I argue that significant patterns found in my LME study
> > between both groups are still valid accounting for this bias ? Is LME
> > method robust enough for compensating this kind of drop-out ?
> > 
> > Best,
> > Matthieu
> > 
> > 
> > Le mer. 17 oct. 2018 à 18:33, Martin Reuter  > edu> a écrit :
> > > Hi Matthieu, 
> > > 
> > > 1) survival analysis is typically used if you want to detect if the
> > > time to an event is longer in one group vs the other (e.g. one
> > > group
> > > gets placebo the other drug and we want to know if recurrence is
> > > later
> > > in the drug group). Not sure this is what you need. The nice thing
> > > is,
> > > it can deal with drop-outs
> > > 
> > > 2) No, you can directly test that (e.g. do more dieseased drop out
> > > than
> > > healthy, or are the dropouts on average more advanced (test-scores,
> > > hippo-volume etc) than the diseased at baseline... many options.
> > > you
> > > could also test interactions with age , gender etc. However, not
> > > finding an interaction may not mean there is no bias, it is just
> > > small
> > > enough to go undetected with your data size. 
> > > 
> > > 3) Survival analysis is a different analysis than LME. 
> > > 
> > > Best, Martin
> > > 
> > > On Tue, 2018-10-16 at 16:15 +, Matthieu Vanhoutte wrote:
> > > > External Email - Use Caution
> > > > Hi Martin,
> > > > 
> > > > It's been a long time since this discussion but I return on this
> > > from
> > > > now... The problem is that I followed longitudinal images of two
> > > > groups where I had mainly missing time points at the end. Than
> > > you
> > > > suggested:
> > > > If you have mainly missing time points at the end, this will bias
> > > > your analysis to some extend, as the remaining ones may be
> > > extremely
> > > > healthy, as probably the more diseased ones drop out. You may
> > > want to
> > > > do a time-to-event (or survival-analysis) which considers early
> > > drop-
> > > > out.
> > > > 
> > > > 1) I know the survival analysis toolbox on matlab, but now I
> > > would
> > > > like to know what information will this survival analysis give to
> > > me
> > > > ? 
> > > > 2) Will this analysis tell me if there is a bias ?
> > > > 3) How to consider early drop-out with this type of analysis
> > > based on
> > > > mass-univariate LME analysis of longitudinal neuroimaging data ?
> > > > 
> > > > Thanks in advance for helping.
> > > > 
> > > > Best,
> > > > Matthieu
> > > > 
> > > > Le mer. 14 déc. 2016 à 22:14, Martin Reuter  > > ard.
> > > > edu> a écrit :
> > > > > Hi Matthieu,
> > > > > 
> > > > > 1. yes, LME needs to be done first so that values can be
> > > sampled
> > > > > from the fitted model for the SA.
> > > > > 
> > > > > 2. yes, I was talking about gradient non-linearities etc that
> > > could
> > > > > be in the image from the acquisition. We currently don’t use
> > > non-
> > > > > linear registration across time points (only rigid). 
> > > > > 
> > > > > Best, Martin
> > > > > 
> > > > > 
> > > > > > On Nov 22, 2016, at 9:31 PM, Matthieu Vanhoutte
> > >  > > > > > e...@gmail.com > wrote:
> > > > > > 
> > > > > > Hi Martin,
> > > > > > 
> > > > > > Please see inline below:
> > > > > > 
> > > > > > > Le 22 nov. 2016 à 17:04, Martin Reuter  > > vard
> > > > > > > .edu> a écrit :
> > > > > > > 
> > > > > > > Hi Matthieu, 
> > > > > > > (also inline)
> > > > > > > 
> > > > > > > > On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte
> > >  > > > > > > > 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

External Email - Use Caution

Hi Martin,

Do you mean that high attrition (less time points with time) would induce
more variance on the slope estimates but this effect would be compensated
by LME modeling ?

Would there be proportion of early drop outs to respect in order to
compensate bias with LME ?

Thanks,
Matthieu


Le ven. 19 oct. 2018 à 16:19, Martin Reuter  a
écrit :

> Hi Matthieu,
>
> in LME it is OK if subjects have differently many time points in
> general. You need to ask a statistician about your specific setup, but
> I think it might be OK (basically less time points, means more variance
> on the slope estimates, but that should be considered in LME). But I am
> not a statistician.
>
> Best, Martin
>
>
> On Wed, 2018-10-17 at 18:46 +0200, Matthieu Vanhoutte wrote:
> > External Email - Use Caution
> > Hi Martin,
> >
> > Thanks for your answer.
> >
> > I actually compare neurospychological scores at baseline between
> > drop-out subjects and subjects with full time-points. If I ever find
> > that drop-out subjects are more severely affected than the subjects
> > with full time-points, then there might be a bias in the results of
> > my LME study ?
> >
> > How could I argue that significant patterns found in my LME study
> > between both groups are still valid accounting for this bias ? Is LME
> > method robust enough for compensating this kind of drop-out ?
> >
> > Best,
> > Matthieu
> >
> >
> > Le mer. 17 oct. 2018 à 18:33, Martin Reuter  > edu> a écrit :
> > > Hi Matthieu,
> > >
> > > 1) survival analysis is typically used if you want to detect if the
> > > time to an event is longer in one group vs the other (e.g. one
> > > group
> > > gets placebo the other drug and we want to know if recurrence is
> > > later
> > > in the drug group). Not sure this is what you need. The nice thing
> > > is,
> > > it can deal with drop-outs
> > >
> > > 2) No, you can directly test that (e.g. do more dieseased drop out
> > > than
> > > healthy, or are the dropouts on average more advanced (test-scores,
> > > hippo-volume etc) than the diseased at baseline... many options.
> > > you
> > > could also test interactions with age , gender etc. However, not
> > > finding an interaction may not mean there is no bias, it is just
> > > small
> > > enough to go undetected with your data size.
> > >
> > > 3) Survival analysis is a different analysis than LME.
> > >
> > > Best, Martin
> > >
> > > On Tue, 2018-10-16 at 16:15 +, Matthieu Vanhoutte wrote:
> > > > External Email - Use Caution
> > > > Hi Martin,
> > > >
> > > > It's been a long time since this discussion but I return on this
> > > from
> > > > now... The problem is that I followed longitudinal images of two
> > > > groups where I had mainly missing time points at the end. Than
> > > you
> > > > suggested:
> > > > If you have mainly missing time points at the end, this will bias
> > > > your analysis to some extend, as the remaining ones may be
> > > extremely
> > > > healthy, as probably the more diseased ones drop out. You may
> > > want to
> > > > do a time-to-event (or survival-analysis) which considers early
> > > drop-
> > > > out.
> > > >
> > > > 1) I know the survival analysis toolbox on matlab, but now I
> > > would
> > > > like to know what information will this survival analysis give to
> > > me
> > > > ?
> > > > 2) Will this analysis tell me if there is a bias ?
> > > > 3) How to consider early drop-out with this type of analysis
> > > based on
> > > > mass-univariate LME analysis of longitudinal neuroimaging data ?
> > > >
> > > > Thanks in advance for helping.
> > > >
> > > > Best,
> > > > Matthieu
> > > >
> > > > Le mer. 14 déc. 2016 à 22:14, Martin Reuter  > > ard.
> > > > edu> a écrit :
> > > > > Hi Matthieu,
> > > > >
> > > > > 1. yes, LME needs to be done first so that values can be
> > > sampled
> > > > > from the fitted model for the SA.
> > > > >
> > > > > 2. yes, I was talking about gradient non-linearities etc that
> > > could
> > > > > be in the image from the acquisition. We currently don’t use
> > > non-
> > > > > linear registration across time points (only rigid).
> > > > >
> > > > > Best, Martin
> > > > >
> > > > >
> > > > > > On Nov 22, 2016, at 9:31 PM, Matthieu Vanhoutte
> > >  > > > > > e...@gmail.com> wrote:
> > > > > >
> > > > > > Hi Martin,
> > > > > >
> > > > > > Please see inline below:
> > > > > >
> > > > > > > Le 22 nov. 2016 à 17:04, Martin Reuter  > > vard
> > > > > > > .edu> a écrit :
> > > > > > >
> > > > > > > Hi Matthieu,
> > > > > > > (also inline)
> > > > > > >
> > > > > > > > On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte
> > >  > > > > > > > hou...@gmail.com> wrote:
> > > > > > > >
> > > > > > > > Hi Martin,
> > > > > > > >
> > > > > > > > Thanks for replying. Please see inline below:
> > > > > > > >
> > > > > > > > > Le 21 nov. 2016 à 20:26, Martin Reuter  > > .har
> > > > > > > > > vard.edu> a écrit :
> > > > > > > > >
> > > > > > > > > Hi Matthieu,
> > > > > > > > >

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Martin Reuter]

Hi Matthieu, 

in LME it is OK if subjects have differently many time points in
general. You need to ask a statistician about your specific setup, but
I think it might be OK (basically less time points, means more variance
on the slope estimates, but that should be considered in LME). But I am
not a statistician. 

Best, Martin


On Wed, 2018-10-17 at 18:46 +0200, Matthieu Vanhoutte wrote:
> External Email - Use Caution
> Hi Martin,
> 
> Thanks for your answer.
> 
> I actually compare neurospychological scores at baseline between
> drop-out subjects and subjects with full time-points. If I ever find
> that drop-out subjects are more severely affected than the subjects
> with full time-points, then there might be a bias in the results of
> my LME study ?
> 
> How could I argue that significant patterns found in my LME study
> between both groups are still valid accounting for this bias ? Is LME
> method robust enough for compensating this kind of drop-out ?
> 
> Best,
> Matthieu
> 
> 
> Le mer. 17 oct. 2018 à 18:33, Martin Reuter  edu> a écrit :
> > Hi Matthieu, 
> > 
> > 1) survival analysis is typically used if you want to detect if the
> > time to an event is longer in one group vs the other (e.g. one
> > group
> > gets placebo the other drug and we want to know if recurrence is
> > later
> > in the drug group). Not sure this is what you need. The nice thing
> > is,
> > it can deal with drop-outs
> > 
> > 2) No, you can directly test that (e.g. do more dieseased drop out
> > than
> > healthy, or are the dropouts on average more advanced (test-scores,
> > hippo-volume etc) than the diseased at baseline... many options.
> > you
> > could also test interactions with age , gender etc. However, not
> > finding an interaction may not mean there is no bias, it is just
> > small
> > enough to go undetected with your data size. 
> > 
> > 3) Survival analysis is a different analysis than LME. 
> > 
> > Best, Martin
> > 
> > On Tue, 2018-10-16 at 16:15 +, Matthieu Vanhoutte wrote:
> > > External Email - Use Caution
> > > Hi Martin,
> > > 
> > > It's been a long time since this discussion but I return on this
> > from
> > > now... The problem is that I followed longitudinal images of two
> > > groups where I had mainly missing time points at the end. Than
> > you
> > > suggested:
> > > If you have mainly missing time points at the end, this will bias
> > > your analysis to some extend, as the remaining ones may be
> > extremely
> > > healthy, as probably the more diseased ones drop out. You may
> > want to
> > > do a time-to-event (or survival-analysis) which considers early
> > drop-
> > > out.
> > > 
> > > 1) I know the survival analysis toolbox on matlab, but now I
> > would
> > > like to know what information will this survival analysis give to
> > me
> > > ? 
> > > 2) Will this analysis tell me if there is a bias ?
> > > 3) How to consider early drop-out with this type of analysis
> > based on
> > > mass-univariate LME analysis of longitudinal neuroimaging data ?
> > > 
> > > Thanks in advance for helping.
> > > 
> > > Best,
> > > Matthieu
> > > 
> > > Le mer. 14 déc. 2016 à 22:14, Martin Reuter  > ard.
> > > edu> a écrit :
> > > > Hi Matthieu,
> > > > 
> > > > 1. yes, LME needs to be done first so that values can be
> > sampled
> > > > from the fitted model for the SA.
> > > > 
> > > > 2. yes, I was talking about gradient non-linearities etc that
> > could
> > > > be in the image from the acquisition. We currently don’t use
> > non-
> > > > linear registration across time points (only rigid). 
> > > > 
> > > > Best, Martin
> > > > 
> > > > 
> > > > > On Nov 22, 2016, at 9:31 PM, Matthieu Vanhoutte
> >  > > > > e...@gmail.com> wrote:
> > > > > 
> > > > > Hi Martin,
> > > > > 
> > > > > Please see inline below:
> > > > > 
> > > > > > Le 22 nov. 2016 à 17:04, Martin Reuter  > vard
> > > > > > .edu> a écrit :
> > > > > > 
> > > > > > Hi Matthieu, 
> > > > > > (also inline)
> > > > > > 
> > > > > > > On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte
> >  > > > > > > hou...@gmail.com> wrote:
> > > > > > > 
> > > > > > > Hi Martin,
> > > > > > > 
> > > > > > > Thanks for replying. Please see inline below:
> > > > > > > 
> > > > > > > > Le 21 nov. 2016 à 20:26, Martin Reuter  > .har
> > > > > > > > vard.edu> a écrit :
> > > > > > > > 
> > > > > > > > Hi Matthieu, 
> > > > > > > > 
> > > > > > > > a few quick answers. Maybe Jorge knows more. 
> > > > > > > > Generally number of subjects / time points etc. cannot
> > be
> > > > > > > > specified generally. All depends on how noisy your data
> > is
> > > > > > > > and how large the effect is that you expect to detect.
> > You
> > > > > > > > can do a power analysis in order to figure out how many
> > > > > > > > subject / time points would be needed. There are some
> > tools
> > > > > > > > for that in the LME toolbox:
> > > > > > > > https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEf
> > fect
> > > > > > > > sModels#Poweranalysis 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

External Email - Use Caution

Hi Martin,

Thanks for your answer.

I actually compare neurospychological scores at baseline between drop-out
subjects and subjects with full time-points. If I ever find that drop-out
subjects are more severely affected than the subjects with full
time-points, then there might be a bias in the results of my LME study ?

How could I argue that significant patterns found in my LME study between
both groups are still valid accounting for this bias ? Is LME method robust
enough for compensating this kind of drop-out ?

Best,
Matthieu


Le mer. 17 oct. 2018 à 18:33, Martin Reuter  a
écrit :

> Hi Matthieu,
>
> 1) survival analysis is typically used if you want to detect if the
> time to an event is longer in one group vs the other (e.g. one group
> gets placebo the other drug and we want to know if recurrence is later
> in the drug group). Not sure this is what you need. The nice thing is,
> it can deal with drop-outs
>
> 2) No, you can directly test that (e.g. do more dieseased drop out than
> healthy, or are the dropouts on average more advanced (test-scores,
> hippo-volume etc) than the diseased at baseline... many options. you
> could also test interactions with age , gender etc. However, not
> finding an interaction may not mean there is no bias, it is just small
> enough to go undetected with your data size.
>
> 3) Survival analysis is a different analysis than LME.
>
> Best, Martin
>
> On Tue, 2018-10-16 at 16:15 +, Matthieu Vanhoutte wrote:
> > External Email - Use Caution
> > Hi Martin,
> >
> > It's been a long time since this discussion but I return on this from
> > now... The problem is that I followed longitudinal images of two
> > groups where I had mainly missing time points at the end. Than you
> > suggested:
> > If you have mainly missing time points at the end, this will bias
> > your analysis to some extend, as the remaining ones may be extremely
> > healthy, as probably the more diseased ones drop out. You may want to
> > do a time-to-event (or survival-analysis) which considers early drop-
> > out.
> >
> > 1) I know the survival analysis toolbox on matlab, but now I would
> > like to know what information will this survival analysis give to me
> > ?
> > 2) Will this analysis tell me if there is a bias ?
> > 3) How to consider early drop-out with this type of analysis based on
> > mass-univariate LME analysis of longitudinal neuroimaging data ?
> >
> > Thanks in advance for helping.
> >
> > Best,
> > Matthieu
> >
> > Le mer. 14 déc. 2016 à 22:14, Martin Reuter  > edu> a écrit :
> > > Hi Matthieu,
> > >
> > > 1. yes, LME needs to be done first so that values can be sampled
> > > from the fitted model for the SA.
> > >
> > > 2. yes, I was talking about gradient non-linearities etc that could
> > > be in the image from the acquisition. We currently don’t use non-
> > > linear registration across time points (only rigid).
> > >
> > > Best, Martin
> > >
> > >
> > > > On Nov 22, 2016, at 9:31 PM, Matthieu Vanhoutte  > > > e...@gmail.com> wrote:
> > > >
> > > > Hi Martin,
> > > >
> > > > Please see inline below:
> > > >
> > > > > Le 22 nov. 2016 à 17:04, Martin Reuter  > > > > .edu> a écrit :
> > > > >
> > > > > Hi Matthieu,
> > > > > (also inline)
> > > > >
> > > > > > On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte  > > > > > hou...@gmail.com> wrote:
> > > > > >
> > > > > > Hi Martin,
> > > > > >
> > > > > > Thanks for replying. Please see inline below:
> > > > > >
> > > > > > > Le 21 nov. 2016 à 20:26, Martin Reuter  > > > > > > vard.edu> a écrit :
> > > > > > >
> > > > > > > Hi Matthieu,
> > > > > > >
> > > > > > > a few quick answers. Maybe Jorge knows more.
> > > > > > > Generally number of subjects / time points etc. cannot be
> > > > > > > specified generally. All depends on how noisy your data is
> > > > > > > and how large the effect is that you expect to detect. You
> > > > > > > can do a power analysis in order to figure out how many
> > > > > > > subject / time points would be needed. There are some tools
> > > > > > > for that in the LME toolbox:
> > > > > > > https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffect
> > > > > > > sModels#Poweranalysis
> > > > > > >
> > > > > > > 1. see above
> > > > > > > 2. yes, also time points can miss from the middle. If you
> > > > > > > have mainly missing time points at the end, this will bias
> > > > > > > your analysis to some extend, as the remaining ones may be
> > > > > > > extremely healthy, as probably the more diseased ones drop
> > > > > > > out. You may want to do a time-to-event (or survival-
> > > > > > > analysis) which considers early drop-out.
> > > > > >
> > > > > > Is there any way to do with Freesurfer this kind of analysis
> > > > > > ?
> > > > >
> > > > > https://surfer.nmr.mgh.harvard.edu/fswiki/SurvivalAnalysis
> > > > > Yes, there is also a paper where we do this. It is a
> > > > > combination of LME and Survival Analysis (as for the SA you
> > > > > need to 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Martin Reuter]

Hi Matthieu, 

1) survival analysis is typically used if you want to detect if the
time to an event is longer in one group vs the other (e.g. one group
gets placebo the other drug and we want to know if recurrence is later
in the drug group). Not sure this is what you need. The nice thing is,
it can deal with drop-outs

2) No, you can directly test that (e.g. do more dieseased drop out than
healthy, or are the dropouts on average more advanced (test-scores,
hippo-volume etc) than the diseased at baseline... many options. you
could also test interactions with age , gender etc. However, not
finding an interaction may not mean there is no bias, it is just small
enough to go undetected with your data size. 

3) Survival analysis is a different analysis than LME. 

Best, Martin

On Tue, 2018-10-16 at 16:15 +, Matthieu Vanhoutte wrote:
> External Email - Use Caution
> Hi Martin,
> 
> It's been a long time since this discussion but I return on this from
> now... The problem is that I followed longitudinal images of two
> groups where I had mainly missing time points at the end. Than you
> suggested:
> If you have mainly missing time points at the end, this will bias
> your analysis to some extend, as the remaining ones may be extremely
> healthy, as probably the more diseased ones drop out. You may want to
> do a time-to-event (or survival-analysis) which considers early drop-
> out.
> 
> 1) I know the survival analysis toolbox on matlab, but now I would
> like to know what information will this survival analysis give to me
> ? 
> 2) Will this analysis tell me if there is a bias ?
> 3) How to consider early drop-out with this type of analysis based on
> mass-univariate LME analysis of longitudinal neuroimaging data ?
> 
> Thanks in advance for helping.
> 
> Best,
> Matthieu
> 
> Le mer. 14 déc. 2016 à 22:14, Martin Reuter  edu> a écrit :
> > Hi Matthieu,
> > 
> > 1. yes, LME needs to be done first so that values can be sampled
> > from the fitted model for the SA.
> > 
> > 2. yes, I was talking about gradient non-linearities etc that could
> > be in the image from the acquisition. We currently don’t use non-
> > linear registration across time points (only rigid). 
> > 
> > Best, Martin
> > 
> > 
> > > On Nov 22, 2016, at 9:31 PM, Matthieu Vanhoutte  > > e...@gmail.com> wrote:
> > > 
> > > Hi Martin,
> > > 
> > > Please see inline below:
> > > 
> > > > Le 22 nov. 2016 à 17:04, Martin Reuter  > > > .edu> a écrit :
> > > > 
> > > > Hi Matthieu, 
> > > > (also inline)
> > > > 
> > > > > On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte  > > > > hou...@gmail.com> wrote:
> > > > > 
> > > > > Hi Martin,
> > > > > 
> > > > > Thanks for replying. Please see inline below:
> > > > > 
> > > > > > Le 21 nov. 2016 à 20:26, Martin Reuter  > > > > > vard.edu> a écrit :
> > > > > > 
> > > > > > Hi Matthieu, 
> > > > > > 
> > > > > > a few quick answers. Maybe Jorge knows more. 
> > > > > > Generally number of subjects / time points etc. cannot be
> > > > > > specified generally. All depends on how noisy your data is
> > > > > > and how large the effect is that you expect to detect. You
> > > > > > can do a power analysis in order to figure out how many
> > > > > > subject / time points would be needed. There are some tools
> > > > > > for that in the LME toolbox:
> > > > > > https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffect
> > > > > > sModels#Poweranalysis 
> > > > > > 
> > > > > > 1. see above
> > > > > > 2. yes, also time points can miss from the middle. If you
> > > > > > have mainly missing time points at the end, this will bias
> > > > > > your analysis to some extend, as the remaining ones may be
> > > > > > extremely healthy, as probably the more diseased ones drop
> > > > > > out. You may want to do a time-to-event (or survival-
> > > > > > analysis) which considers early drop-out.
> > > > > 
> > > > > Is there any way to do with Freesurfer this kind of analysis
> > > > > ?
> > > > 
> > > > https://surfer.nmr.mgh.harvard.edu/fswiki/SurvivalAnalysis 
> > > > Yes, there is also a paper where we do this. It is a
> > > > combination of LME and Survival Analysis (as for the SA you
> > > > need to have measurements of all subjects at all time points,
> > > > so you estimate that from the LME model). 
> > > 
> > > Thank you for the link, I will take a look at. So if understand,
> > > this analysis has to be done after LME statistical analysis ?
> > > Thereafter since SA need all time points, LME model will allow me
> > > to estimate missing time points ?
> > > 
> > > > > > 3. see above (power analysis)
> > > > > > 4. GIGO means garbage in, garbage out, so the less you QC,
> > > > > > the more likely will your results be junk. The more you QC
> > > > > > the less likely will it be junk, but could still be. The FS
> > > > > > wiki has lots of tutorial information on checking
> > > > > > freesurfer recons. For longitudinal, you should
> > > > > > additionally check the surfaces in the base, the brain 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

External Email - Use Caution

Hi Martin,

It's been a long time since this discussion but I return on this from
now... The problem is that I followed longitudinal images of two groups
where I had mainly missing time points at the end. Than you suggested:
*If you have mainly missing time points at the end, this will bias your
analysis to some extend, as the remaining ones may be extremely healthy, as
probably the more diseased ones drop out. You may want to do a
time-to-event (or survival-analysis) which considers early drop-out.*

1) I know the survival analysis toolbox on matlab, but now I would like to
know what information will this survival analysis give to me ?
2) Will this analysis tell me if there is a bias ?
3) How to consider early drop-out with this type of analysis based on
mass-univariate LME analysis of longitudinal neuroimaging data ?

Thanks in advance for helping.

Best,
Matthieu

Le mer. 14 déc. 2016 à 22:14, Martin Reuter  a
écrit :

> Hi Matthieu,
>
> 1. yes, LME needs to be done first so that values can be sampled from the
> fitted model for the SA.
>
> 2. yes, I was talking about gradient non-linearities etc that could be in
> the image from the acquisition. We currently don’t use non-linear
> registration across time points (only rigid).
>
> Best, Martin
>
>
> On Nov 22, 2016, at 9:31 PM, Matthieu Vanhoutte <
> matthieuvanhou...@gmail.com> wrote:
>
> Hi Martin,
>
> Please see inline below:
>
> Le 22 nov. 2016 à 17:04, Martin Reuter  a
> écrit :
>
> Hi Matthieu,
> (also inline)
>
> On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte <
> matthieuvanhou...@gmail.com> wrote:
>
> Hi Martin,
>
> Thanks for replying. Please see inline below:
>
> Le 21 nov. 2016 à 20:26, Martin Reuter  a
> écrit :
>
> Hi Matthieu,
>
> a few quick answers. Maybe Jorge knows more.
> Generally number of subjects / time points etc. cannot be specified
> generally. All depends on how noisy your data is and how large the effect
> is that you expect to detect. You can do a power analysis in order to
> figure out how many subject / time points would be needed. There are some
> tools for that in the LME toolbox:
>
> https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels#Poweranalysis
>
>
> 1. see above
> 2. yes, also time points can miss from the middle. If you have mainly
> missing time points at the end, this will bias your analysis to some
> extend, as the remaining ones may be extremely healthy, as probably the
> more diseased ones drop out. You may want to do a time-to-event (or
> survival-analysis) which considers early drop-out.
>
>
> Is there any way to do with Freesurfer this kind of analysis ?
>
>
> https://surfer.nmr.mgh.harvard.edu/fswiki/SurvivalAnalysis
> Yes, there is also a paper where we do this. It is a combination of LME
> and Survival Analysis (as for the SA you need to have measurements of all
> subjects at all time points, so you estimate that from the LME model).
>
>
> Thank you for the link, I will take a look at. So if understand, this
> analysis has to be done after LME statistical analysis ? Thereafter since
> SA need all time points, LME model will allow me to estimate missing time
> points ?
>
>
>
> 3. see above (power analysis)
> 4. GIGO means garbage in, garbage out, so the less you QC, the more likely
> will your results be junk. The more you QC the less likely will it be junk,
> but could still be. The FS wiki has lots of tutorial information on
> checking freesurfer recons. For longitudinal, you should additionally check
> the surfaces in the base, the brain mask in the base, and the alignment of
> the time points (although there is some wiggle space for the alignment, as
> most things are allowed to evolve further for each time point).
>
>
> For the alignment of the time points, should I better comparing brainmask
> or norm.mgz ?
>
>
> It does not really matter, I would use norm.mgz. I would load images on
> top of each other and then use the opacity slider in Freeview to blend
> between them (that way the eye can pick up small motions). I would not
> worry too much about local deformations which could be caused by
> non-linearity (gradient). But if you see global misalignment (rotation,
> translation) it is a cause for concern) .
>
>
> Ok thank you. The non-linearity you are talking about are well provoked by
> MRI system and not non-linear registration between time points and template
> base, aren’t they ?
>
> Best regards,
> Matthieu
>
>
>
> In order to avoid bias by adding further time points in the model by the
> -add recon all command, is this better for each subject to take into
> account all the time points existing for it or only the ones that I will
> include in the model (three time points / subject ; if existing 6 time
> points for any subject ?)
>
>
> Usually it is recommended to run all time points in the model (so a base
> with 6 time points) and not use the - - add flag. Also, Linear Mixed
> Effects models deal well with missing time points. 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Martin Reuter]

Hi Matthieu,

1. yes, LME needs to be done first so that values can be sampled from the 
fitted model for the SA.

2. yes, I was talking about gradient non-linearities etc that could be in the 
image from the acquisition. We currently don’t use non-linear registration 
across time points (only rigid). 

Best, Martin


> On Nov 22, 2016, at 9:31 PM, Matthieu Vanhoutte  
> wrote:
> 
> Hi Martin,
> 
> Please see inline below:
> 
>> Le 22 nov. 2016 à 17:04, Martin Reuter > > a écrit :
>> 
>> Hi Matthieu, 
>> (also inline)
>> 
>>> On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte 
>>> > wrote:
>>> 
>>> Hi Martin,
>>> 
>>> Thanks for replying. Please see inline below:
>>> 
 Le 21 nov. 2016 à 20:26, Martin Reuter > a écrit :
 
 Hi Matthieu, 
 
 a few quick answers. Maybe Jorge knows more. 
 Generally number of subjects / time points etc. cannot be specified 
 generally. All depends on how noisy your data is and how large the effect 
 is that you expect to detect. You can do a power analysis in order to 
 figure out how many subject / time points would be needed. There are some 
 tools for that in the LME toolbox:
 https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels#Poweranalysis
  
 
  
 
 1. see above
 2. yes, also time points can miss from the middle. If you have mainly 
 missing time points at the end, this will bias your analysis to some 
 extend, as the remaining ones may be extremely healthy, as probably the 
 more diseased ones drop out. You may want to do a time-to-event (or 
 survival-analysis) which considers early drop-out.
>>> 
>>> Is there any way to do with Freesurfer this kind of analysis ?
>> 
>> https://surfer.nmr.mgh.harvard.edu/fswiki/SurvivalAnalysis 
>>  
>> Yes, there is also a paper where we do this. It is a combination of LME and 
>> Survival Analysis (as for the SA you need to have measurements of all 
>> subjects at all time points, so you estimate that from the LME model). 
> 
> Thank you for the link, I will take a look at. So if understand, this 
> analysis has to be done after LME statistical analysis ? Thereafter since SA 
> need all time points, LME model will allow me to estimate missing time points 
> ?
> 
>> 
>>> 
 3. see above (power analysis)
 4. GIGO means garbage in, garbage out, so the less you QC, the more likely 
 will your results be junk. The more you QC the less likely will it be 
 junk, but could still be. The FS wiki has lots of tutorial information on 
 checking freesurfer recons. For longitudinal, you should additionally 
 check the surfaces in the base, the brain mask in the base, and the 
 alignment of the time points (although there is some wiggle space for the 
 alignment, as most things are allowed to evolve further for each time 
 point). 
>>> 
>>> For the alignment of the time points, should I better comparing brainmask 
>>> or norm.mgz ?
>> 
>> It does not really matter, I would use norm.mgz. I would load images on top 
>> of each other and then use the opacity slider in Freeview to blend between 
>> them (that way the eye can pick up small motions). I would not worry too 
>> much about local deformations which could be caused by non-linearity 
>> (gradient). But if you see global misalignment (rotation, translation) it is 
>> a cause for concern) .
> 
> Ok thank you. The non-linearity you are talking about are well provoked by 
> MRI system and not non-linear registration between time points and template 
> base, aren’t they ?
> 
> Best regards,
> Matthieu
> 
>> 
>>> 
>>> In order to avoid bias by adding further time points in the model by the 
>>> -add recon all command, is this better for each subject to take into 
>>> account all the time points existing for it or only the ones that I will 
>>> include in the model (three time points / subject ; if existing 6 time 
>>> points for any subject ?)
>>> 
>> 
>> Usually it is recommended to run all time points in the model (so a base 
>> with 6 time points) and not use the - - add flag. Also, Linear Mixed Effects 
>> models deal well with missing time points. It is perfectly OK to have 
>> differently many time points per subject for that. You should still check if 
>> there is a bias (e.g. one group always has 3 time points the other 6) that 
>> would not be good. Maybe also consult with a local biostatistician if you 
>> are not comfortable with the stats. The LME tools are matlab, and so are the 
>> survival-analysis scripts. 
>> 
>> Best, Martin
>> 
>> 
>> 
>>> Best regards,
>>> Matthieu
>>> 
 
 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

Dear FS's experts,

Could you answer me about questions from below inline precedent mail ?

Best regards,
Matthieu


2016-11-22 21:31 GMT+01:00 Matthieu Vanhoutte :

> Hi Martin,
>
> Please see inline below:
>
> Le 22 nov. 2016 à 17:04, Martin Reuter  a
> écrit :
>
> Hi Matthieu,
> (also inline)
>
> On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte <
> matthieuvanhou...@gmail.com> wrote:
>
> Hi Martin,
>
> Thanks for replying. Please see inline below:
>
> Le 21 nov. 2016 à 20:26, Martin Reuter  a
> écrit :
>
> Hi Matthieu,
>
> a few quick answers. Maybe Jorge knows more.
> Generally number of subjects / time points etc. cannot be specified
> generally. All depends on how noisy your data is and how large the effect
> is that you expect to detect. You can do a power analysis in order to
> figure out how many subject / time points would be needed. There are some
> tools for that in the LME toolbox:
> https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels#
> Poweranalysis
>
> 1. see above
> 2. yes, also time points can miss from the middle. If you have mainly
> missing time points at the end, this will bias your analysis to some
> extend, as the remaining ones may be extremely healthy, as probably the
> more diseased ones drop out. You may want to do a time-to-event (or
> survival-analysis) which considers early drop-out.
>
>
> Is there any way to do with Freesurfer this kind of analysis ?
>
>
> https://surfer.nmr.mgh.harvard.edu/fswiki/SurvivalAnalysis
> Yes, there is also a paper where we do this. It is a combination of LME
> and Survival Analysis (as for the SA you need to have measurements of all
> subjects at all time points, so you estimate that from the LME model).
>
>
>
Thank you for the link, I will take a look at. So if understand, this
analysis has to be done after LME statistical analysis ? Thereafter since
SA need all time points, LME model will allow me to estimate missing time
points ?


>
>
>
> 3. see above (power analysis)
> 4. GIGO means garbage in, garbage out, so the less you QC, the more likely
> will your results be junk. The more you QC the less likely will it be junk,
> but could still be. The FS wiki has lots of tutorial information on
> checking freesurfer recons. For longitudinal, you should additionally check
> the surfaces in the base, the brain mask in the base, and the alignment of
> the time points (although there is some wiggle space for the alignment, as
> most things are allowed to evolve further for each time point).
>
>
> For the alignment of the time points, should I better comparing brainmask
> or norm.mgz ?
>
>
> It does not really matter, I would use norm.mgz. I would load images on
> top of each other and then use the opacity slider in Freeview to blend
> between them (that way the eye can pick up small motions). I would not
> worry too much about local deformations which could be caused by
> non-linearity (gradient). But if you see global misalignment (rotation,
> translation) it is a cause for concern) .
>
>
>
 Ok thank you. The non-linearity you are talking about are well provoked by
MRI system and not non-linear registration between time points and template
base, aren’t they ?


>
>
> In order to avoid bias by adding further time points in the model by the
> -add recon all command, is this better for each subject to take into
> account all the time points existing for it or only the ones that I will
> include in the model (three time points / subject ; if existing 6 time
> points for any subject ?)
>
>
> Usually it is recommended to run all time points in the model (so a base
> with 6 time points) and not use the - - add flag. Also, Linear Mixed
> Effects models deal well with missing time points. It is perfectly OK to
> have differently many time points per subject for that. You should still
> check if there is a bias (e.g. one group always has 3 time points the other
> 6) that would not be good. Maybe also consult with a local biostatistician
> if you are not comfortable with the stats. The LME tools are matlab, and so
> are the survival-analysis scripts.
>
> Best, Martin
>
>
>
> Best regards,
> Matthieu
>
>
> Best, Martin
>
> On Nov 21, 2016, at 7:07 PM, Matthieu Vanhoutte <
> matthieuvanhou...@gmail.com> wrote:
>
> Dear Freesurfer’s experts,
>
> I would have some questions regarding the LME model to be used in
> longitudinal stream:
>
> 1) Which are the ratio limits or % of missing timepoints accepted ?
> (according time, I have less and less subjects time points)
>
> 2) Is it possible to include patients that would miss the first timepoint
> but got the others ?
>
> 3) Considering a group in longitudinal study, which is the number of
> subjects minimal of this group accepted for LME modeling ?
>
> 4) Finally, concerning quality control and among a big number of total
> time points, which essential controls are necessary ? (Control of norm.mgz
> of the base, 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

Hi Martin,

Please see inline below:

> Le 22 nov. 2016 à 17:04, Martin Reuter  a écrit :
> 
> Hi Matthieu, 
> (also inline)
> 
>> On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte 
>> > wrote:
>> 
>> Hi Martin,
>> 
>> Thanks for replying. Please see inline below:
>> 
>>> Le 21 nov. 2016 à 20:26, Martin Reuter >> > a écrit :
>>> 
>>> Hi Matthieu, 
>>> 
>>> a few quick answers. Maybe Jorge knows more. 
>>> Generally number of subjects / time points etc. cannot be specified 
>>> generally. All depends on how noisy your data is and how large the effect 
>>> is that you expect to detect. You can do a power analysis in order to 
>>> figure out how many subject / time points would be needed. There are some 
>>> tools for that in the LME toolbox:
>>> https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels#Poweranalysis
>>>  
>>> 
>>>  
>>> 
>>> 1. see above
>>> 2. yes, also time points can miss from the middle. If you have mainly 
>>> missing time points at the end, this will bias your analysis to some 
>>> extend, as the remaining ones may be extremely healthy, as probably the 
>>> more diseased ones drop out. You may want to do a time-to-event (or 
>>> survival-analysis) which considers early drop-out.
>> 
>> Is there any way to do with Freesurfer this kind of analysis ?
> 
> https://surfer.nmr.mgh.harvard.edu/fswiki/SurvivalAnalysis 
>  
> Yes, there is also a paper where we do this. It is a combination of LME and 
> Survival Analysis (as for the SA you need to have measurements of all 
> subjects at all time points, so you estimate that from the LME model). 

Thank you for the link, I will take a look at. So if understand, this analysis 
has to be done after LME statistical analysis ? Thereafter since SA need all 
time points, LME model will allow me to estimate missing time points ?

> 
>> 
>>> 3. see above (power analysis)
>>> 4. GIGO means garbage in, garbage out, so the less you QC, the more likely 
>>> will your results be junk. The more you QC the less likely will it be junk, 
>>> but could still be. The FS wiki has lots of tutorial information on 
>>> checking freesurfer recons. For longitudinal, you should additionally check 
>>> the surfaces in the base, the brain mask in the base, and the alignment of 
>>> the time points (although there is some wiggle space for the alignment, as 
>>> most things are allowed to evolve further for each time point). 
>> 
>> For the alignment of the time points, should I better comparing brainmask or 
>> norm.mgz ?
> 
> It does not really matter, I would use norm.mgz. I would load images on top 
> of each other and then use the opacity slider in Freeview to blend between 
> them (that way the eye can pick up small motions). I would not worry too much 
> about local deformations which could be caused by non-linearity (gradient). 
> But if you see global misalignment (rotation, translation) it is a cause for 
> concern) .

Ok thank you. The non-linearity you are talking about are well provoked by MRI 
system and not non-linear registration between time points and template base, 
aren’t they ?

Best regards,
Matthieu

> 
>> 
>> In order to avoid bias by adding further time points in the model by the 
>> -add recon all command, is this better for each subject to take into account 
>> all the time points existing for it or only the ones that I will include in 
>> the model (three time points / subject ; if existing 6 time points for any 
>> subject ?)
>> 
> 
> Usually it is recommended to run all time points in the model (so a base with 
> 6 time points) and not use the - - add flag. Also, Linear Mixed Effects 
> models deal well with missing time points. It is perfectly OK to have 
> differently many time points per subject for that. You should still check if 
> there is a bias (e.g. one group always has 3 time points the other 6) that 
> would not be good. Maybe also consult with a local biostatistician if you are 
> not comfortable with the stats. The LME tools are matlab, and so are the 
> survival-analysis scripts. 
> 
> Best, Martin
> 
> 
> 
>> Best regards,
>> Matthieu
>> 
>>> 
>>> Best, Martin
>>> 
 On Nov 21, 2016, at 7:07 PM, Matthieu Vanhoutte 
 > wrote:
 
 Dear Freesurfer’s experts,
 
 I would have some questions regarding the LME model to be used in 
 longitudinal stream:
 
 1) Which are the ratio limits or % of missing timepoints accepted ? 
 (according time, I have less and less subjects time points)
 
 2) Is it possible to include patients that would miss the first timepoint 
 but got the others ?
 
 3) Considering a group in 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Martin Reuter]

Hi Matthieu, 
(also inline)

> On Nov 21, 2016, at 10:28 PM, Matthieu Vanhoutte 
>  wrote:
> 
> Hi Martin,
> 
> Thanks for replying. Please see inline below:
> 
>> Le 21 nov. 2016 à 20:26, Martin Reuter > > a écrit :
>> 
>> Hi Matthieu, 
>> 
>> a few quick answers. Maybe Jorge knows more. 
>> Generally number of subjects / time points etc. cannot be specified 
>> generally. All depends on how noisy your data is and how large the effect is 
>> that you expect to detect. You can do a power analysis in order to figure 
>> out how many subject / time points would be needed. There are some tools for 
>> that in the LME toolbox:
>> https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels#Poweranalysis
>>  
>> 
>>  
>> 
>> 1. see above
>> 2. yes, also time points can miss from the middle. If you have mainly 
>> missing time points at the end, this will bias your analysis to some extend, 
>> as the remaining ones may be extremely healthy, as probably the more 
>> diseased ones drop out. You may want to do a time-to-event (or 
>> survival-analysis) which considers early drop-out.
> 
> Is there any way to do with Freesurfer this kind of analysis ?

https://surfer.nmr.mgh.harvard.edu/fswiki/SurvivalAnalysis 
 
Yes, there is also a paper where we do this. It is a combination of LME and 
Survival Analysis (as for the SA you need to have measurements of all subjects 
at all time points, so you estimate that from the LME model). 

> 
>> 3. see above (power analysis)
>> 4. GIGO means garbage in, garbage out, so the less you QC, the more likely 
>> will your results be junk. The more you QC the less likely will it be junk, 
>> but could still be. The FS wiki has lots of tutorial information on checking 
>> freesurfer recons. For longitudinal, you should additionally check the 
>> surfaces in the base, the brain mask in the base, and the alignment of the 
>> time points (although there is some wiggle space for the alignment, as most 
>> things are allowed to evolve further for each time point). 
> 
> For the alignment of the time points, should I better comparing brainmask or 
> norm.mgz ?

It does not really matter, I would use norm.mgz. I would load images on top of 
each other and then use the opacity slider in Freeview to blend between them 
(that way the eye can pick up small motions). I would not worry too much about 
local deformations which could be caused by non-linearity (gradient). But if 
you see global misalignment (rotation, translation) it is a cause for concern) .

> 
> In order to avoid bias by adding further time points in the model by the -add 
> recon all command, is this better for each subject to take into account all 
> the time points existing for it or only the ones that I will include in the 
> model (three time points / subject ; if existing 6 time points for any 
> subject ?)
> 

Usually it is recommended to run all time points in the model (so a base with 6 
time points) and not use the - - add flag. Also, Linear Mixed Effects models 
deal well with missing time points. It is perfectly OK to have differently many 
time points per subject for that. You should still check if there is a bias 
(e.g. one group always has 3 time points the other 6) that would not be good. 
Maybe also consult with a local biostatistician if you are not comfortable with 
the stats. The LME tools are matlab, and so are the survival-analysis scripts. 

Best, Martin



> Best regards,
> Matthieu
> 
>> 
>> Best, Martin
>> 
>>> On Nov 21, 2016, at 7:07 PM, Matthieu Vanhoutte 
>>> > wrote:
>>> 
>>> Dear Freesurfer’s experts,
>>> 
>>> I would have some questions regarding the LME model to be used in 
>>> longitudinal stream:
>>> 
>>> 1) Which are the ratio limits or % of missing timepoints accepted ? 
>>> (according time, I have less and less subjects time points)
>>> 
>>> 2) Is it possible to include patients that would miss the first timepoint 
>>> but got the others ?
>>> 
>>> 3) Considering a group in longitudinal study, which is the number of 
>>> subjects minimal of this group accepted for LME modeling ?
>>> 
>>> 4) Finally, concerning quality control and among a big number of total time 
>>> points, which essential controls are necessary ? (Control of norm.mgz of 
>>> the base, alignment of longitudinal timepoints on base,… ?)
>>> 
>>> Best regards,
>>> Matthieu
>>> 
>>> 
>>> ___
>>> Freesurfer mailing list
>>> Freesurfer@nmr.mgh.harvard.edu 
>>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer 
>>> 
>>> 
>>> 
>> 
>> 

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

Hi Martin,

Thanks for replying. Please see inline below:

> Le 21 nov. 2016 à 20:26, Martin Reuter  a écrit :
> 
> Hi Matthieu, 
> 
> a few quick answers. Maybe Jorge knows more. 
> Generally number of subjects / time points etc. cannot be specified 
> generally. All depends on how noisy your data is and how large the effect is 
> that you expect to detect. You can do a power analysis in order to figure out 
> how many subject / time points would be needed. There are some tools for that 
> in the LME toolbox:
> https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels#Poweranalysis
>  
> 
>  
> 
> 1. see above
> 2. yes, also time points can miss from the middle. If you have mainly missing 
> time points at the end, this will bias your analysis to some extend, as the 
> remaining ones may be extremely healthy, as probably the more diseased ones 
> drop out. You may want to do a time-to-event (or survival-analysis) which 
> considers early drop-out.

Is there any way to do with Freesurfer this kind of analysis ?

> 3. see above (power analysis)
> 4. GIGO means garbage in, garbage out, so the less you QC, the more likely 
> will your results be junk. The more you QC the less likely will it be junk, 
> but could still be. The FS wiki has lots of tutorial information on checking 
> freesurfer recons. For longitudinal, you should additionally check the 
> surfaces in the base, the brain mask in the base, and the alignment of the 
> time points (although there is some wiggle space for the alignment, as most 
> things are allowed to evolve further for each time point). 

For the alignment of the time points, should I better comparing brainmask or 
norm.mgz ?

In order to avoid bias by adding further time points in the model by the -add 
recon all command, is this better for each subject to take into account all the 
time points existing for it or only the ones that I will include in the model 
(three time points / subject ; if existing 6 time points for any subject ?)

Best regards,
Matthieu

> 
> Best, Martin
> 
>> On Nov 21, 2016, at 7:07 PM, Matthieu Vanhoutte > > wrote:
>> 
>> Dear Freesurfer’s experts,
>> 
>> I would have some questions regarding the LME model to be used in 
>> longitudinal stream:
>> 
>> 1) Which are the ratio limits or % of missing timepoints accepted ? 
>> (according time, I have less and less subjects time points)
>> 
>> 2) Is it possible to include patients that would miss the first timepoint 
>> but got the others ?
>> 
>> 3) Considering a group in longitudinal study, which is the number of 
>> subjects minimal of this group accepted for LME modeling ?
>> 
>> 4) Finally, concerning quality control and among a big number of total time 
>> points, which essential controls are necessary ? (Control of norm.mgz of the 
>> base, alignment of longitudinal timepoints on base,… ?)
>> 
>> Best regards,
>> Matthieu
>> 
>> 
>> ___
>> Freesurfer mailing list
>> Freesurfer@nmr.mgh.harvard.edu 
>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
>> 
>> 
> 
> ___
> 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.

___
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.

Re: [Freesurfer] Longitudinal stream : LME and limits of model

[Martin Reuter]

Hi Matthieu, 

a few quick answers. Maybe Jorge knows more. 
Generally number of subjects / time points etc. cannot be specified generally. 
All depends on how noisy your data is and how large the effect is that you 
expect to detect. You can do a power analysis in order to figure out how many 
subject / time points would be needed. There are some tools for that in the LME 
toolbox:
https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels#Poweranalysis
 

 

1. see above
2. yes, also time points can miss from the middle. If you have mainly missing 
time points at the end, this will bias your analysis to some extend, as the 
remaining ones may be extremely healthy, as probably the more diseased ones 
drop out. You may want to do a time-to-event (or survival-analysis) which 
considers early drop-out.
3. see above (power analysis)
4. GIGO means garbage in, garbage out, so the less you QC, the more likely will 
your results be junk. The more you QC the less likely will it be junk, but 
could still be. The FS wiki has lots of tutorial information on checking 
freesurfer recons. For longitudinal, you should additionally check the surfaces 
in the base, the brain mask in the base, and the alignment of the time points 
(although there is some wiggle space for the alignment, as most things are 
allowed to evolve further for each time point). 

Best, Martin

> On Nov 21, 2016, at 7:07 PM, Matthieu Vanhoutte  
> wrote:
> 
> Dear Freesurfer’s experts,
> 
> I would have some questions regarding the LME model to be used in 
> longitudinal stream:
> 
> 1) Which are the ratio limits or % of missing timepoints accepted ? 
> (according time, I have less and less subjects time points)
> 
> 2) Is it possible to include patients that would miss the first timepoint but 
> got the others ?
> 
> 3) Considering a group in longitudinal study, which is the number of subjects 
> minimal of this group accepted for LME modeling ?
> 
> 4) Finally, concerning quality control and among a big number of total time 
> points, which essential controls are necessary ? (Control of norm.mgz of the 
> base, alignment of longitudinal timepoints on base,… ?)
> 
> Best regards,
> Matthieu
> 
> 
> ___
> Freesurfer mailing list
> Freesurfer@nmr.mgh.harvard.edu
> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
> 
> 

___
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.

[Freesurfer] Longitudinal stream : LME and limits of model

[Matthieu Vanhoutte]

Dear Freesurfer’s experts,

I would have some questions regarding the LME model to be used in longitudinal 
stream:

1) Which are the ratio limits or % of missing timepoints accepted ? (according 
time, I have less and less subjects time points)

2) Is it possible to include patients that would miss the first timepoint but 
got the others ?

3) Considering a group in longitudinal study, which is the number of subjects 
minimal of this group accepted for LME modeling ?

4) Finally, concerning quality control and among a big number of total time 
points, which essential controls are necessary ? (Control of norm.mgz of the 
base, alignment of longitudinal timepoints on base,… ?)

Best regards,
Matthieu


___
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.