 |
 |
No, I haven't said you are wrong. Nor do I think you are. But I do think that your correct statements are tangential to the question asked by the original poster. Your answer is tangential because it assumes that the original poster is working with fragments and may have stuffed up the counts of elements. By contrast, my reply to the original poster assumes that she has used the customary, cautionary procedures to arrive at the number of elements (or a proxy, such as femoral heads). If she has, then we can move on to testing the significance of difference between the two stratigraphic units. In other words I have tacitly assumed, using Occam's razor, that extraneous assumptions surrounding her question need not be made. So I have not questioned her ability to identify elements. Nor, for that matter, have I questioned her ability in another quarrelsome field - the correct identification of stratigraphic units. Returning to the statistical question (and admittedly breaching my principle of using Occam's razor) I have wondered whether her worry is in fact about chi square and the exact test merely demonstrating the bleeding obvious, in the case of large counts. Perhaps she is not really interested in testing significance at all but is interested in using some descriptive measure of strength of association (such as phi-square mentioned in Richard Meadow's post). Anyway, we have given the original poster a good run for her money. I am interested in having to make no more assumptions and learning exactly what her problem is. > >Subject: Re: [ZOOARCH] Statistics help > From: "D.C. Orton" <[log in to unmask]> > Date: Fri, 11 Jan 2008 10:42:18 +0000 > To: [log in to unmask] > >Whether the data are at the taxon or the element level, the arguments >surrounding NISP and MNE/MAU are still entirely relevant. The only >difference is that when dealing with elements you only have to worry about >one, rather than two, levels of fragmentation of the unit of analysis (the >element)- i.e. only breakage of elements, not disarticulation. You can >restrict the 'NISP' and 'MN*' labels to taxonomic comparisons if you want, >but there's still a fundamental difference between fragment counts and >derived estimates when dealing with elements. Surely? Unless I'm missing >something? > >Not that it matters enormously in the present case, because neither group >of measures is really amenable to classical hypothesis testing in any case. >I know that zooarchaeologists frequently use chi-squared etc. with both >forms, but quite simply they shouldn't, especially with raw counts. This is >equally the case whether those counts are of individuals or of elements, >UNLESS there is virtually no breakage. Some form of Watson-style DZ is >probably the best bet, but even then you have to worry about >non-independence between proximal and distal ends etc. > >With regard to the use of Kolmogorv-Smirnov tests for element profiles, I'm >not an expert on this stuff either and I have to admit I haven't read that >book in detail, but it sounds like madness to me. Even if one ignores the >quantification/sample inflation issue, K-S is designed for ordinal data, >while element profiles are nominal. The KS test statistic is based on Dmax, >the greatest difference between the cumulative frequency curves of the two >distributions. With element profiles this will differ depending on the >order in which you list the elements - an unscrupulous researcher could >even try a few different orders to get the 'best' result. > >I may well be missing something here, in which case I'd be grateful if >someone could explain where I'm off the mark. I'd love to be wrong on this. > >David > >> The OP can speak for herself, but I took her statement literally: >> "differences in element frequencies between the two units." >> >> If elements have been counted, then issues of NISP versus MNI are not >> relevant to her question. >> >> >>> >>>Subject: Re: [ZOOARCH] Statistics help >>> From: "D.C. Orton" <[log in to unmask]> >>> Date: Thu, 10 Jan 2008 22:25:17 +0000 >>> To: [log in to unmask] >>> >>>I think the most important question is what form of quantification you're >>>using. >>> >>> If it's NISP-based, you can basically forget formal hypothesis testing >>> as you're samples are almost certainly subject to sample inflation (see >>> e.g. Grayson 1984, pp.22-23 for a clear explanation of this). The only >>> exception would be if the bones are barely fragmented at all. Of course, >>> if you run the tests (chi-squared would make sense by the sound of it) >>> and find no significant pattern then that's fine, as sample inflation >>> will only ever lower your p values, but you shouldn't trust an >>> apparently significant result from this kind of data. >>> >>>To be honest, if it's minimum-number based the situation isn't great >>>either, since you're then dealing with estimates with a non-random, >>>asymmetrical error term. >>> >>>Frankly, zooarchaeological data is a bit of a nightmare statistically >>>speaking. David >>> >>>> You don't say how many elements you have, but whether it is one or more >>>> then I would have thought Chi-square and Fisher's test are entirely >>>> appropriate. Why are you worried about using them? >>>> >>>> >>>>> >>>>>Subject: [ZOOARCH] Statistics help >>>>> From: Melanie Fillios <[log in to unmask]> >>>>> Date: Thu, 10 Jan 2008 15:39:26 -0600 >>>>> To: [log in to unmask] >>>>> >>>>> Dear All, I'm hoping someone may be able to point me in the right >>>>> direction with some statistical analysis of an assemablage. In short, >>>>> I am comparing two stratigraphic units and would like to test whether >>>>> differences in element frequencies between the two units are >>>>> statistically significant. Could anyone tell me what type of test I >>>>> should be using? I've looked a using a Chi-square or Fisher's test, >>>>> but neither seem appropriate. As math is not my strong point, I may be >>>>> missing something. >>>>> >>>>>Thanks for the help! >>>>> >>>>>Melanie >>>>> >>>>>Dr. Melanie Fillios >>>>>University of Sydney >>>>>NSW, Australia 2006 >>>>>[log in to unmask] >>>> >>