 |
 |
Dear Julie, entering more than one covariate (no matter if of interest or not) might become problematic. It is absolutely not a problem to use a multiple regressions model when you have a general question like: How well can my regressors predict the BOLD contrast? In this case all of your regressors would be of interest and you may be able to answer your quesiton with a multiple regression and a "suitable" F-contrast (so called "effects-of-interest"). In any case where you want to disentangle somehow the effects of different regressors, you should check if regressors are correlated. Suppose you have - as you mentioned in your mail - a variable A of interest and another variable B that should serve as covariate of no interest. If both variables correlate perfectly with each other, any contrast testing one of the variables alone (either [1 0], [0 1]) leads to a null result. Null results might of course be "true" but in this case results are not interpretable. If both variables are orthogonal (correlation exactly zero), you have no problems at all and contrasts are interpretable. The absolute values of correlations are of course in the range between >0 and <1 (at least in most cases) and usually one assumes A and B to be independent and any departure from a null-correlation is random and therefore low. It is up to you to decide how low is low enough, but the higher the (absolute) correlation of A and B, the less would be the variance in the BOLD contrast accounted for by A (A variable of interest, B covariate). But in any case such a model gives you an answer to the question: How much variance in the BOLD contrast can be explained by A which can not be explained by B? Hope this helps... Thilo On Tuesday 23 September 2008 18:29, Julie McEntee wrote: > I have a question re: design when there are both covariates of "interest" > AND of "no interest". For ex., in this particular experiment, if one had > reason to believe number of years of tobacco use may contribute to BOLD > contrast (post-pre treatment) and wanted to determine the effect of TD that > cannot be explained by length of tobacco use. Would you include both (i.e., > TD and years of use) in a multiple regression, to make sure years of use is > not related to BOLD contrast, or can you include it in a regression as a > covariate (of no interest)? > > Thanks for anyone's thoughts- > > Julie > > Julie E. McEntee, M.A., C.C.R.P. > Senior Research Program Coordinator > Department of Psychiatry and Behavioral Sciences > Division of Psychiatric Neuroimaging > Johns Hopkins University School of Medicine > 600 N. Wolfe St./ Phipps 300 (office: room 317) > Baltimore, MD 21287 > Phone: 410-502-0468 > Fax: 410-614-3676 > > > -----Original Message----- > From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On > Behalf Of Thilo Kellermann > Sent: Tuesday, September 23, 2008 10:24 AM > To: [log in to unmask] > Subject: Re: [SPM] design problem > > Dear Vlad, > > as far as I understand covariates and paired t-tests, it does not make much > sense to include a covariate per subject unless you have two different > assessments of that covariate for each time point (pre and post drug > treatment). Put differently, a paired t-test is the same as testing the > difference of pre and post against zero, which is the same as a one sample > t-test. When you then include the covariate and this does not change from > pre > to post within each subject, then you would enter a column of zeros as > covariate, which his not what you want. > > By the way, including TD as a covariate in a one sample t-test for > instance, > > does not show you the impact of TD on the BOLD contrast. It would show you > where in the brain the BOLD contrast deviates from zero, which canNOT be > explained by TD. Covariates are always of no interest. > > What you can do is to calculate a simple regression, which is actually the > way > how calculate "covariates of interest". When you have only one TD value per > subject you can define three simple regressions: > 1) TD and pre-treatment scan > 2) TD and post-treatment scan > 3) TD and difference of pre- and post-treatment scans (using e.g. imcalc) > > The third one would be the model you want to examine. > > In case you have different TD values for pre- and post-scans you can do > more > > or less the same: > 1) TDpre and pre-scans > 2) TDpost and post-scans > 3) TDpost-TDpre and (post-scans)-(pre-scans) > > Good luck, > Thilo > > On Tuesday 23 September 2008 15:19, Vlad Kushnir wrote: > > Hello, > > > > > > > > I have encountered a problem and would greatly appreciate some direction > > in > > > resolving it. > > > > > > In my study design there are two drug conditions, Pre and Post-drug > > scans. During each scan subjects are presented with a task that shows > > drug-related > > > and neutral cues, as well as a rest fixation cross in a blocked design. > > As I have one subject group and the subjects vary in their level of > > tobacco dependence (TD), my goal is to create a model that examines how > > TD influences differential activation between the two scan sessions. > > Therefore, in essence I would like to create a paired-samples t-test with > > TD as a covariate of interest. > > > > > > So far I have tried the following: > > > > To best look at how drug cues elicit brain activation in my population of > > interest, for every subject during each of the two scans, I have created > > a 1st level drug cue > neutral cue contrast. These contrasts were later > > taken > > > into a 2nd level paired t-test design (with TD as a covariate for each > > pair), however the job would not compute. I'm wondering if there is a way > > to look at the effect of TD on the difference between the two scans, or > > as a covariate of interest. > > > > Then I tried creating a full factorial design with one factor (Drug) that > > has two levels (pre and post). There was no independence between the two > > levels. TD was the covariate, but it was entered twice repeatedly, once > > for > > > the pre-drug level and once for the post-drug. I am not sure if this is > > the > > > most appropriate design. > > > > In an alternative method, I took every subject's pre and post drug scans > > into 1st level and created a drug cue pre > drug cue post (and vice > > versa) contrast. I then took these contrast images into a 2nd level > > multiple regression analysis, where TD was the covariate. I am also not > > sure if > > this > > > is the most appropriate design, especially since neutral cues are not > > taken > > > into context. > > > > Please advise. > > > > Thank you in advance, > > > > Vlad Kushnir -- Thilo Kellermann Department of Psychiatry and Psychotherapy RWTH Aachen University Pauwelsstr. 30 52074 Aachen Tel.: +49 (0)241 / 8089977 Fax.: +49 (0)241 / 8082401 E-Mail: [log in to unmask]