|
|
Re: off-line change detection & clustering in a time series of
multidimensional data
Posted:
May 16, 2012 1:02 AM
|
|
Hmm I'm not completely following. Wikipedia agrees that ranking is one example of non-parametric statistics, and that it plays a central role in many approaches, but I wouldn't think tests like Kolmogorov-Smirnov that estimate the distribution would have much to do with rank ordering. I can see how permutation tests might. But even then, why would you expect rank ordering to give "normal" residuals? This viewpoint is somewhat opposite to the introductory material I've seen about non-parametric tests where advantages are usually given (robustness, fewer assumptions, etc.). Do you have a reference I could look up on your arguments? In any case, I could re-consider ANOVAs, though I don't expect my data to have "normal" residuals (which I thought was needed for ANOVA). I also might have issues with outliers with non-robust stats: the system doesn't always properly measure the location of the markers.
|
|
