An interesting discussion - and I may be suffering delayed after-effects of the festive season, but I don't think that all the comments here about pooled error analyses are right. I think there are 2 issues: In probably rather simple terms - in a pooled error test, if done correctly, the df are greater but are valid. They are not 'inflated'. The error is also greater - being pooled from what would be different partitioned error cells. If sphericity holds (or is corrected for), this should be just as valid as a partitioned error test. I understand it to involve a pooled *estimate* of an error variance that may (or may not) be the same for all cells - and this estimate is more powerful, with more df, if assumptions hold. If many contrasts per subject are taken to the 2nd level as in the 3 subjects/ 12 levels example, and subject effects are not modelled (discounted from the error term) as the Flexible Factorial allows, then I think Roberto is right to suggest that the inference is not to the population, as between- and within- subject effect estimates are mixed up (mean square effect contains both, incorrectly). But I don't think this is because there is a problem with the 'pooled' df per se. If someone (a proper statistician?) can clarify in more precise language, or correct me, that would be great! I recall a fairly recent discussion on this in which some differing opinion about the validity of this approach was left somewhat in the air... cheers Alexa On 5 January 2011 16:22, Roberto Viviani <[log in to unmask]> wrote: > Right: texts say you can't use the pooled method unless certain >> interaction terms disappear (ie, are not significant), I think. Though it >> would be interesting to see exactly why it's wrong if you use the pooled >> variance when textbooks say you can't. I can't recall seeing a precise >> exposition on that, other than that you have a larger error term but also >> more dof's, which work in different directions. >> > > Yes, additivity must hold, but I hadn't connected that to the df problem. > Maybe I misunderstood your point: I thought you meant that the dfs are not > right for inference on the population. In fact, I looked it up, finding that > the old texts say it isn't inference on the population: under the null and > additivity, "... the expectation of the treatment mean square is equal to > the expectation of the error mean square. It should be noted that the > expectation is not with respect to some infinite population of repetitions > of the experiment but over the possible randomizations of the experiment." > (Kempthorne, 1973 edition, p. 129). > > Unsurprisingly, if you have 3 subjects and 12 levels, say, and inflate dfs > at the second level by taking all effects estimates there, the inference is > no longer on the population. It's Fisher-style with RFT used as a shorthand > for permutation. Therefore, the standard claim of using second-level > estimation to account for subjects as a random effect and conduct inference > on the population is strictly speaking no longer valid in this setup. > > I think the discussion on where the interaction goes tend to conclude it > inflates the treatment mean square, but this issue is made somewhat obscure > by controversies as to whether this interaction would be random or not. > > >> Your second point: I'd have to think about it. :-) >> > > I empathize, as I tend to feel my brain to run out of steam when I try to > figure out what happens with F tests at the second level. As an > afterthought, using permutation would cure these doubts of mine as well; > furthermore, it is explicitly inference on the randomization, not > repetitions of the experiment. > > Best wishes, > Roberto Viviani > Dept. of Psychiatry, University of Ulm, Germany > > > > ________________________________________ >> From: [log in to unmask] [[log in to unmask]] >> Sent: Tuesday, January 04, 2011 5:12 AM >> To: Fromm, Stephen (NIH/NIMH) [C] >> Cc: [log in to unmask] >> Subject: Re: Design matrix for each or all subject(s)? >> >> ... >> >>> Part of my reluctance is related to my disagreement with the way >>> repeated measures are handled by SPM, which is a separate topic. As >>> outlined in "ANOVAs and SPM" >>> link http://www.fil.ion.ucl.ac.uk/~wpenny/publications/rik_anova.pdf<http://www.fil.ion.ucl.ac.uk/%7Ewpenny/publications/rik_anova.pdf> >>> there's the partitioned variance method and pooled variance method. >>> IMHO the pooled variance method (the one commonly used by the SPM >>> community) is incorrect (because it gets df counting wrong), though >>> that appears to be a minority opinion. >>> >> Well this is an interesting point, but one that would also apply to F >> tests conducted in analogous designs in textbook univariate >> situations. I'd expect there should be something on this in that >> well-researched (indeed by now dated) literature. >> >> >> On the other hand, if I recall correctly, there was a thread on the >>> listserv devoted to the topic of the main effect of group which >>> implicitly showed that the pooled variance method was indeed faulty. >>> >> One thing I'd like to know, where was ever shown that the smoothness >> of F maps can be estimated from residuals? That is, irrespective of >> the numerator df's? That does not seem intuitive to me. Given that >> residuals are good to estimate smoothness of t maps (numerator df = >> 1), it does not follow they are good for higher df's. When I look at F >> maps, they seem different from t maps. This seems relevant to the >> pooled error idea, which relies on F testing. >> >> Best wishes, >> Roberto Viviani >> Dept. of Psychiatry, University of Ulm, Germany >> > > -- Dr. Alexa Morcom RCUK Academic Fellow, University of Edinburgh http://www.ccns.sbms.mvm.ed.ac.uk/people/academic/morcom.html The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336