Hi Rysia, Doug, I'll just chime in that a fundamental decision you have to make is whether you want a statistical model that includes interaction terms. If you have 1 group factor (GRP) and 1 continuous covariate (COVAR1), a DODS model is essentially (using SAS shorthand): Y = GRP GRP*COVAR1
while a DOSS model is essentially: Y = GRP COVAR1 (i.e., main effects of GRP and COVAR1 only) Modeling an interaction is a fundamentally different model, and may lead to different results for the main effect of GRP. I think of the issue of centering or not as a separate issue. [By proper setting of the values in the contrast vector, it should be possible to equate the statistical results of a model with centering to one without, as noted by Doug (although obviously you are testing different null hypotheses when you use different contrast vectors for any given model)]. As a related note, when using QDEC with 2 discrete factors, QDEC by default models the interaction between the two factors. If you want a model with only main effects, that must be analyzed with "glmfit" outside of QDEC (at least currently). Again, the modeling of interaction terms is a different model from one having only main effects, so the results may be quite different. Ultimately it comes down to a model selection issue, and whether the selected model is "appropriate" or not. And that will depend on your particular data and the questions you want to ask. Hope that helps, Mike H. On Thu, 2009-01-22 at 15:48 -0500, Douglas N Greve wrote: > Hi Rysia, I've been re-thinking the issue of demeaning variables > (continuous covariates). First, demeaning has no effect on the value or > significance of the slope of the variable, it only has an effect on the > offset/intercept value and significance. When you have a continuous > covariate in the model but your only interested in the intercept/offset, > you still have to chose a value of the covariate at which the test of > the intercept is done. By default, this test is done at the covariate=0. > When your covariate is demeaned, then the test is implicitly done at the > mean of the covariate. So, if you have age as a (demeaned) variable, > then the test is done at an age equal to the mean of the sample. This > yields identical results to when you do not demean but set up your > contrast to be [1 Mean(age)]. > > So the question of demeaning is really a question of at what value of > the covariate do you want to do your intercept test? At first glance, it > makes sense to test it at the mean of your sample because that's where > you data are (instead of extrapolating back to 0). On the other hand, if > you add or remove subjects, you test will change because the mean age > will change. If you or someone else tries to replicate with a different > sample, then the results might not be the same because the test will be > different. And why the mean instead of the median or the mode? I'm not > sure what the right answer is, probably testing at the mean age of the > population (instead of the sample of the population you have drawn). > This would make the test in the range of your acquired data without > being dependent on the actual sample. > > Anyone else want to chime in? > > doug > > Burmicz, Ryzarda wrote: > > Thanks for your comments Michael. > > > > Does non mean centered DODS extrapolate back to zero? I thought it appeared > > to be extrapolating back to the beginning of the dataset (i.e. the smallest > > value of the covariate of interest, in this case age). > > > > Mean centering appears to reduce the significance map to a DODS analysis, I > > think (although I find this all very confusing) because in DOSS one is just > > looking at differences between groups at mean age? I am assuming that it is > > not a wild coincidence.. and I understand that DOSS constrains groups to be > > of the same slope. > > > > Another observation is that the 'group differences' significance map of > > DODS, without mean centering the age, is highly similar to the > > 'thickness-age correlation difference' map of mean centered age DODS > > analysis. This also makes sense to me, however I > > > > It is clear that mean-centering changes the meaning of the analysis. > > However I have read some old mailing list threads in which it is not > > recommended. However, if I just want to remove possible age interactions, I > > would have thought that mean-centering is the best approach. Incidentally, > > both groups are highly similar in mean, median range and s.d. of age. The > > age range being 63-23 is some of the reason for my cautious approach and > > given that foreach of my groups n=15 I do not have much freedom in terms of > > looking at smaller age groups. > > > > I have a final question: the DOSS with age covaried (and DODS with > > mean-centred age) comparison between patients and controls is almost > > identical to a straight (no covariates considered) analysis between > > patients and controls. Can I necessarily conclude that age is not a > > significant factor in this analysis? > > > > Any comments and corrections welcomed. > > > > Many thanks, > > > > Rysia > > > > -----Original Message----- > > From: Michael Harms [mailto:mha...@conte.wustl.edu] > > Sent: 13 January 2009 15:01 > > To: Burmicz, Ryzarda > > Cc: freesurfer@nmr.mgh.harvard.edu > > Subject: Re: [Freesurfer] DODS/DOSS choice in patient/control study > > > > > > Hello Rysia, > > > > I'll add a comment by calling your attention to one statistical > > "subtlety". When using a DODS ("separate slopes") model in the presence > > of a continuous covariate, and then examining the significance of the > > group effect, the interpretation (and results) depend critically on the > > meaning of "zero" of the continuous covariate (e.g., whether the > > covariate is centered or not). In your case, I see that you are using > > age as the covariate (a very common covariate). If you do not center > > the age variable, the resulting test of the group effect is in effect > > whether the estimated intercepts, extrapolated back to age 0, differ > > significantly. If you center the age variable (say around the mean > > age), the test of the group effect is still whether the "intercept" > > differs between groups, but now the intercept is centered at the mean > > age along the x-axis. > > > > If you experiment with different ages as the reference point (e.g., > > centered around 20, 40, 60 years), you will find that the DODS results > > for the group effect can change dramatically. > > > > Incidentally, when using a DOSS model, centering is irrelevant, because > > the regression lines for the two groups have (by design) the same slope, > > and thus a constant vertical difference, regardless of the location of > > "zero" along the x-axis. > > > > cheers, > > Mike H. > > > > > > On Tue, 2009-01-13 at 14:16 +0000, Burmicz, Ryzarda wrote: > > > >> Hello, > >> > >> > >> > >> We are looking at patients versus control group analysis. We have done > >> a straight analysis and are now looking at covariates. DOSS produces > >> similar results to a straight patient vs control analysis, in that > >> some blobs are located in the same cortical regions. In the case of > >> DODS, there is no similarity whatsoever to the straight analysis > >> maps... which is why I am so stuck in deciding, because it really does > >> impact on the results quite drastically. > >> > >> > >> > >> I understand the difference between DODS and DOSS – and I know the > >> following question is not an easy one to answer, however we have > >> sample sizes of 15 patient vs 15 control subjects and would like to > >> know if that is great enough such that we are better off going for > >> DODS. > >> > >> > >> > >> There is no reason to say that the patient cortical thickness – age > >> relationship is any different in slope from that of normal ageing > >> generally; however in certain cortical regions patients exhibit a > >> thickening of the cortex, and thinning in other (albeit fewer) > >> regions. > >> > >> > >> > >> Do you have any comments? I would have thought that our sample sizes > >> are rather small, but would appreciate your input. > >> > >> > >> > >> I attach the comparison of patient vs control covaried for age under > >> both DODS and DOSS to illustrate the differences. > >> > >> > >> > >> Thanks! > >> > >> > >> > >> Rysia > >> > >> > >> > >> Ryszarda Burmicz > >> > >> Research Assistant > >> > >> Centre for Neuroimaging Sciences > >> > >> PO89 > >> > >> Institute of Psychiatry > >> > >> > >> > >> > >> _______________________________________________ > >> 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 > > > > > > > > _______________________________________________ > 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
