Thanks Jorge! In our case, at step 4, an F-test does show significant evidence for a quadratic term, but not across the whole brain. We have a strong hypothesis that there should be a time effect (linear or quadratic) across most of the brain, and it makes sense that some regions will show linear and others will show quadratic effects.
We thought that a top-down approach makes sense, whereby we drop the quadratic term from the model to investigate significant linear time effects. This results in overlap in regions showing a significant quadratic effect (first model) and a significant linear effect (second model). 1. Do you have any thoughts on this approach, and on how to decide whether there are linear or quadratic effects across different regions of the brain? 2. We also want to look at group differences. We find no significant group x quadratic time effects, but we do find significant group x linear time effects (in our second model). The significant regions overlap with those regions where we found significant quadratic effects for the whole group (first model). While this scenario makes sense to me statistically, we have had some researchers tell us that this scenario is not possible/doesn't make sense. Do you have any advice for looking at group differences in our situation? Thank you for your time/help. Sarah ________________________________ From: jorge luis [jbernal0...@yahoo.es] Sent: Tuesday, 21 April 2015 2:37 AM To: Sarah Whittle Cc: freesurfer@nmr.mgh.harvard.edu Subject: Re: advice/question regarding your LME toolbox for freesurfer Hi Sara For those analysis I went through the following steps: First considered a full model for both the mean and the covariance and selected the best model for the covariance: 1- Fitted a full model (model1) including intercept, time and time squared as both fixed effects and random effects. 2- Separately fitted a model (model2) including intercept, time and time squared as fixed effects but only including intercept and time as random effects. 3- In order to compare the previous models I then applied the likelihood ratio test vertex-wise and corrected for multiple comparisons using FDR. The result was that over 80% of the vertices showed no significant results for the previous test after correcting for multiple comparisons. So the model with three random effects wasn't significantly better than the model with two random effects and thus I considered model2 a better fit for the data. Once the best model for the covariance was selected I proceeded to select the best model for the mean: 4- I tested the null hypothesis of no quadratic term in the model for the mean using an F-test on model2. After correction for multiple comparisons there was no significant evidence for a quadratic term so I dropped it from the model for the mean. So the final model was one including intercept and time as both fixed and random effects. No quadratic term included either in the model for the mean or the model for the covariance. Hope that helps -Jorge ________________________________ De: Sarah Whittle <swhit...@unimelb.edu.au> Para: "jber...@nmr.mgh.harvard.edu" <jber...@nmr.mgh.harvard.edu>; "jbernal0...@yahoo.es" <jbernal0...@yahoo.es> Enviado: Lunes 20 de abril de 2015 2:15 Asunto: advice/question regarding your LME toolbox for freesurfer Dear Dr. Bernal-Rusiel, I was hoping that you could please give me some advice about model selection using the LME tools that you have developed for freesurfer. My colleague has posted to the mailing list about this but hasn't got a response. I am wondering specifically about the best approach for identifying quadratic versus linear age effects for mass-univariate analysis of longitudinal data, in addition to group differences in linear versus quadratic age effects. I have read the following in your 2013 paper: "After correcting for multiple comparisons, over 80% of the cortex vertices included both the intercept and time, and not time squared, as the optimal set of random effects. Hence, these two random effects were included in the final model for all remaining analyses and time squared (the quadratic term) was not included as a random effect. We then tested the null hypothesis of no group differences in the quadratic term (i.e., the coefficient of the “time squared” fixed effect) and no vertex exhibited a statistically significant association after multiple comparisons correction. Therefore, we dropped the quadratic term from the model." I was wondering how you did this more specifically? In the model where you found >80% vertices included time but not time squared, was this using using two contrasts from the same 'full' model? Or did you run one model with only linear age effects, and then a second with linear + quadratic effects? I have a situation where I have both significant linear and quadratic effects (after multiple comparison correction), but neither fit >80% vertices. Some regions have both linear and quadratic effects. Do you have any advice/thoughts on representing these results? Thank you for your time and I look forward to hearing your thoughts. Sincerely, Sarah Whittle, PhD Senior Research Fellow Melbourne Neuropsychiatry Centre The University of Melboourne 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.
