Re: [Freesurfer] Linear Mixed Models in FS?
Yes there is numerical instability when p-values becomes extremely low and your solution is OK. You are just being conservative. Your actual p-value might have been 1e-25 but you couldn't observe it exactly (got zero instead) because of numerical limitations of the Matlab's fcdf function. I'll fix this issue on github:https://github.com/NeuroStats this weekend. Best -Jorge De: Lars M. Rimol lari...@gmail.com Para: jorge luis jbernal0...@yahoo.es CC: Freesurfer support list freesurfer@nmr.mgh.harvard.edu Enviado: Jueves 19 de junio de 2014 5:01 Asunto: Re: [Freesurfer] Linear Mixed Models in FS? Hi Jorge, Thank you! Yes this makes sense, because this confirms that the second covariate in fact tests for an effect across all groups, just as I expected. Now, there's another problem with these analyses: Please consider the attached figure lh_model1_01000_interaction_nopatch_lateral.tif, which shows a p-map (cortical area, smoothed with 30 mm fwhm) with two gray regions in the temporal lobe and the insula/IFG in the middle of highly significant regions. There is no darker blue transition into the non-significant regions. These gray regions appear to be artifacts based on eye balling of the maps. I checked the output of the significance testing in those regions and it appears that the output of this formula (in lme_mass_f.m) is extremely low: pval(i) = 1-fcdf(Fstat,szC,m); I assume there may be numerical instability when pval becomes extremely low? Could that explain this? I changed the code so that if 1-fcdf(Fstat,szC,m) is zero or less than 1e-15, then pval(i) = max(1-fcdf(Fstat,szC,m),1e-15); (or, if it's negative pval(i) = min(1-fcdf(Fstat,szC,m),-1e-15) ) If it's not, then the old code applies: pval(i) = 1-fcdf(Fstat,szC,m); This seems to have fixed the problem as the figure model1_area_lh_01000_winteraction_wpatch_lateral.tif shows. Again, based on eye-balling the maps. I have seen this problem in several data sets, both cortical and subcortical data. In all cases using a lower limit for 1-fcdf(Fstat,szC,m) - either 1e-15 or 1e-20 - seems to fix the problem. Do you concur that the problem is numerical instability and is this a good way to fix it? Thank you! yours, LMR On Wed, Jun 18, 2014 at 3:49 PM, jorge luis jbernal0...@yahoo.es wrote: Hi LMR If the interaction term is not statistically significant then there is no evidence of the existence of two different groups in your sample (as far as the longitudinal trajectory is concerned they are all controls, the groups might be different at baseline though). This is why main effects are only tested after the interactions have been previously tested. In your model a common “base time slope” is assumed for both groups (the second coefficient) but you are also explicitly modeling the possibility of the case-group slope being exceeding the control/common base slope by an extra quantity. That quantity is the interaction term. Hope this makes sense Best -Jorge De: Lars M. Rimol lari...@gmail.com Para: FS maling list freesurfer@nmr.mgh.harvard.edu Enviado: Miércoles 18 de junio de 2014 8:57 Asunto: Re: [Freesurfer] Linear Mixed Models in FS? Hi Jorge, Thank you for your reply! Again considering the same model from before intercept(random effect) + centered age + group + group x centered age + sex I think what is confusing me is that I think of the [centered age] covariate as a column vector which will contain the centered age of both the control- and the case group. This is how it would be seen in a GLM using the same design matrix. Therefore it is difficult for me to understand how the contrast [0 1 0 0 0] can inform us about the control group alone. To me it would seem obvious that this contrast tells me something about the effect of [centered age] on the whole of the sample, regardless of the group each subject belongs to. On the other hand, I agree with you that the interaction term could tell us something about the effect of [centered age] on the case-group by considering the contrast vector [0 0 0 1 0]. Just for the sake of argument, please consider the following model intercept(random effect) + (1-group) x centered age + group + group x centered age + sex and compare to the one presented above. Here (1-group) is a column vector which is 1 where the [group] vector is 0, and vice versa. This difference ensures that the second term only includes numbers from the control-group. Applying the contrast [0 1 0 0 0] to this model, would this not be more appropriate for consider the effect of [centered age] on the control-group alone? Given your previous answers I suspect I'm missing something here, but I would greatly appreciate if you could please take the time to explain to me how I've gone wrong. Thanks! LMR --- Hi
Re: [Freesurfer] Linear Mixed Models in FS?
Ey! Dice Lilla que estas currando en el Childrens? Juan Eugenio Iglesias Postdoctoral researcher BCBL www.jeiglesias.com www.bcbl.eu Legal disclaimer/Aviso legal/Lege-oharra: www.bcbl.eu/legal-disclaimer - Original Message - From: jorge luis jbernal0...@yahoo.es To: Lars M. Rimol lari...@gmail.com Cc: Freesurfer support list freesurfer@nmr.mgh.harvard.edu Sent: Thursday, June 19, 2014 8:56:38 AM Subject: Re: [Freesurfer] Linear Mixed Models in FS? Yes there is numerical instability when p-values becomes extremely low and y our solution is OK. You are just being conservative. Your actual p-value might have been 1e-25 but you couldn't observe it exactly (got zero instead) because of numerical limitations of the Matlab's fcdf function. I'll fix this issue on github: https://github.com/NeuroStats this weekend. Best -Jorge De: Lars M. Rimol lari...@gmail.com Para: jorge luis jbernal0...@yahoo.es CC: Freesurfer support list freesurfer@nmr.mgh.harvard.edu Enviado: Jueves 19 de junio de 2014 5:01 Asunto: Re: [Freesurfer] Linear Mixed Models in FS? Hi Jorge, Thank you! Yes this makes sense, because this confirms that the second covariate in fact tests for an effect across all groups, just as I expected. Now, there's another problem with these analyses: Please consider the attached figure lh_model1_01000_interaction_nopatch_lateral.tif , which shows a p-map (cortical area, smoothed with 30 mm fwhm) with two gray regions in the temporal lobe and the insula/IFG in the middle of highly significant regions. There is no darker blue transition into the non-significant regions. These gray regions appear to be artifacts based on eye balling of the maps. I checked the output of the significance testing in those regions and it appears that the output of this formula (in lme_mass_f.m) is extremely low: pval(i) = 1-fcdf(Fstat,szC,m); I assume there may be numerical instability when pval becomes extremely low? Could that explain this? I changed the code so that if 1-fcdf(Fstat,szC,m) is zero or less than 1e-15, then pval(i) = max(1-fcdf(Fstat,szC,m),1e-15); (or, if it's negative pval(i) = min(1-fcdf(Fstat,szC,m),-1e-15) ) If it's not, then the old code applies: pval(i) = 1-fcdf(Fstat,szC,m); This seems to have fixed the problem as the figure model1_area_lh_01000_winteraction_wpatch_lateral.tif shows. Again, based on eye-balling the maps. I have seen this problem in several data sets, both cortical and subcortical data. In all cases using a lower limit for 1-fcdf(Fstat,szC,m) - either 1e-15 or 1e-20 - seems to fix the problem. Do you concur that the problem is numerical instability and is this a good way to fix it? Thank you! yours, LMR On Wed, Jun 18, 2014 at 3:49 PM, jorge luis jbernal0...@yahoo.es wrote: Hi LMR If the interaction term is not statistically significant then there is no evidence of the existence of two different groups in your sample (as far as the longitudinal trajectory is concerned they are all controls, the groups might be different at baseline though). This is why main effects are only tested after the interactions have been previously tested. In your model a common “base time slope” is assumed for both groups (the second coefficient) but you are also explicitly modeling the possibility of the case-group slope being exceeding the control/common base slope by an extra quantity. That quantity is the interaction term. Hope this makes sense Best -Jorge De: Lars M. Rimol lari...@gmail.com Para: FS maling list freesurfer@nmr.mgh.harvard.edu Enviado: Miércoles 18 de junio de 2014 8:57 Asunto: Re: [Freesurfer] Linear Mixed Models in FS? Hi Jorge, Thank you for your reply! Again considering the same model from before intercept(random effect) + centered age + group + group x centered age + sex I think what is confusing me is that I think of the [centered age] covariate as a column vector which will contain the centered age of both the control- and the case group. This is how it would be seen in a GLM using the same design matrix. Therefore it is difficult for me to understand how the contrast [0 1 0 0 0] can inform us about the control group alone. To me it would seem obvious that this contrast tells me something about the effect of [centered age] on the whole of the sample, regardless of the group each subject belongs to. On the other hand, I agree with you that the interaction term could tell us something about the effect of [centered age] on the case-group by considering the contrast vector [0 0 0 1 0]. Just for the sake of argument, please consider the following model intercept(random effect) + (1-group) x centered age + group + group x centered age + sex and compare to the one presented above. Here (1-group) is a column vector which is 1 where the [group] vector is 0, and vice versa. This difference ensures that the second term only
Re: [Freesurfer] Linear Mixed Models in FS?
Hi Jorge, Thank you for your reply! Again considering the same model from before intercept(random effect) + centered age + group + group x centered age + sex I think what is confusing me is that I think of the [centered age] covariate as a column vector which will contain the centered age of both the control- and the case group. This is how it would be seen in a GLM using the same design matrix. Therefore it is difficult for me to understand how the contrast [0 1 0 0 0] can inform us about the control group alone. To me it would seem obvious that this contrast tells me something about the effect of [centered age] on the whole of the sample, regardless of the group each subject belongs to. On the other hand, I agree with you that the interaction term could tell us something about the effect of [centered age] on the case-group by considering the contrast vector [0 0 0 1 0]. Just for the sake of argument, please consider the following model intercept(random effect) + (1-group) x centered age + group + group x centered age + sex and compare to the one presented above. Here (1-group) is a column vector which is 1 where the [group] vector is 0, and vice versa. This difference ensures that the second term only includes numbers from the control-group. Applying the contrast [0 1 0 0 0] to this model, would this not be more appropriate for consider the effect of [centered age] on the control-group alone? Given your previous answers I suspect I'm missing something here, but I would greatly appreciate if you could please take the time to explain to me how I've gone wrong. Thanks! LMR --- Hi LMR 1) Yes, you should use n-1 (0/1) covariates to model n groups. Eg. (Controls, Case 1 and Case 2) the model would be: intercept(random effect) + centered age?(might be a random effect too)?+ ?Case1 + Case1 x centered age + Case2 + Case2 x centered age + sex 2)In model: intercept(random effect) + centered age + group + group x centered age + sex the fourth coefficient is the interaction term that represents the difference in slope between the patient and control groups. This is easy to see from your Question 1 equations. It's also easy to see from those equations that [0 1 0 0 0] tests the effect of time in the control group since the group-specific slope is only equal to the coefficient of the time covariate (the second covariate) when the group covariate is zero (i.e for the controls). Hope this makes sense. Best -Jorge -- yours, Lars M. Rimol, PhD St. Olavs Hospital Trondheim, Norway ___ 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] Linear Mixed Models in FS?
Hi LMR If the interaction term is not statistically significant then there is no evidence of the existence of two different groups in your sample (as far as the longitudinal trajectory is concerned they are all controls, the groups might be different at baseline though). This is why main effects are only tested after the interactions have been previously tested. In your model a common “base time slope” is assumed for both groups (the second coefficient) but you are also explicitly modeling the possibility of the case-group slope being exceeding the control/common base slope by an extra quantity. That quantity is the interaction term. Hope this makes sense Best -Jorge De: Lars M. Rimol lari...@gmail.com Para: FS maling list freesurfer@nmr.mgh.harvard.edu Enviado: Miércoles 18 de junio de 2014 8:57 Asunto: Re: [Freesurfer] Linear Mixed Models in FS? Hi Jorge, Thank you for your reply! Again considering the same model from before intercept(random effect) + centered age + group + group x centered age + sex I think what is confusing me is that I think of the [centered age] covariate as a column vector which will contain the centered age of both the control- and the case group. This is how it would be seen in a GLM using the same design matrix. Therefore it is difficult for me to understand how the contrast [0 1 0 0 0] can inform us about the control group alone. To me it would seem obvious that this contrast tells me something about the effect of [centered age] on the whole of the sample, regardless of the group each subject belongs to. On the other hand, I agree with you that the interaction term could tell us something about the effect of [centered age] on the case-group by considering the contrast vector [0 0 0 1 0]. Just for the sake of argument, please consider the following model intercept(random effect) + (1-group) x centered age + group + group x centered age + sex and compare to the one presented above. Here (1-group) is a column vector which is 1 where the [group] vector is 0, and vice versa. This difference ensures that the second term only includes numbers from the control-group. Applying the contrast [0 1 0 0 0] to this model, would this not be more appropriate for consider the effect of [centered age] on the control-group alone? Given your previous answers I suspect I'm missing something here, but I would greatly appreciate if you could please take the time to explain to me how I've gone wrong. Thanks! LMR --- Hi LMR 1) Yes, you should use n-1 (0/1) covariates to model n groups. Eg. (Controls, Case 1 and Case 2) the model would be: intercept(random effect) + centered age?(might be a random effect too)?+ ?Case1 + Case1 x centered age + Case2 + Case2 x centered age + sex 2)In model: intercept(random effect) + centered age + group + group x centered age + sex the fourth coefficient is the interaction term that represents the difference in slope between the patient and control groups. This is easy to see from your Question 1 equations. It's also easy to see from those equations that [0 1 0 0 0] tests the effect of time in the control group since the group-specific slope is only equal to the coefficient of the time covariate (the second covariate) when the group covariate is zero (i.e for the controls). Hope this makes sense. Best -Jorge -- yours, Lars M. Rimol, PhD St. Olavs Hospital Trondheim, Norway ___ 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.