Jackknifing & Bayesian rmANOVA
Hello!
I am running an analysis on event-related-potential latency measures that were obtained using a jackknifing approach (see Kiesel et al., 2008, Psychophysiology). With a frequentist repeated-measures ANOVA, the analysis of such subsample scores requires the correction of the resulting F-values; that is because jackknifing artificially reduces the error variances in the ANOVA and consequentially, the F-values are too large (see Ulrich & Miller, 2001, Psychophysiology).
However, I now want to run a bayesian rmANOVA - is jackknifed data suitable for a bayesian analysis or does it require some kind of correction as well? I appreciate any thoughts on this.
Thanks!
Laura
Comments
Hi Laura
The jackknife is a resampling technique, right? When you're doing a Bayesian analysis that is not needed. But I guess I should look up the Kiesel paper...
Cheers,
E.J.
Dear E.J.,
yes, that's right, jackknifing is a resampling technique. When you say, that is not needed when doing a Bayesian analysis, do you mean the correction (i.e. analogous to the correction of an F-value in a frequentist analysis) is not necessary or are you referring to the resampling itself?
Maybe to clarify what I am doing: I am interested in when a certain event-related potential (ERP) component occurs and I expect differences between my experimental conditions. EEG latency measures are really sensitive to noise - that's why it is usually recommended to use jackknifing. Essentially, that means rather than obtaining one measure per subject from the ERP belonging to that one subject, I create a series of average waveforms, each of which is missing one subject (leave-one-out-averages). Those leave-one-out-averages will only be slightly different from each other, reflecting the subject who was omitted from the average. Because the waveforms (and the resulting latency measures) are very similar, the error variance has been artifically reduced. Thus, in a convential ANOVA I divide the F-value by (N-1)^2 to correct for this reduction in variance. Does that make sense?
Best,
Laura
Dear Laura,
I am referring to resampling itself. I do realize the resampling is handy in the present context, because otherwise you need some sort of complicated time series model (which will almost certainly be misspecified). This is definitely interesting but I cannot provide a confident answer at this point. But I am confident that it would be really nice to figure this out!
Cheers,
E.J.
Thanks anyway for your time to look into this!
Best,
Laura