Thanks, all this discussion made me realize that i don't wish to start arguing with him back and forth and start showing him why i am right. He suggested i use factor analysis instead, so to appease him, and since to me it doesn't really matter as the neighbourhoods were already chosen, i decided to "surrender". Maybe when i am "old and wise" (this is only my fourth attempt to submit something), i will rebut or argue. Thanks to all of you who helped. Hilit
On Tue, Mar 9, 2010 at 5:41 PM, Burton Shank <burtonsh...@gmail.com> wrote: > Hi Hilit, > > I haven't seen a really useful response to your posting s don'to thought I > would chime in. > > You don't mention explicitly but I assume you're using nMDS for this > ordination. If so, the thing to keep in mind is that nMDS does its best to > accurately portray the (dis)similarities in your dataset while compressing > the data into a limited number of spatial dimensions. We're used to seeing > 2d MDS plots in publications because they are easily visualized on a flat > page. This does not mean that two dimensions are most appropriate for > portraying your dataset. Also keep in mind that the first vs second axis > are arbitrary in nMDS, as apposed to CCA or PCA. It's just points in space, > you can rotate and look at them from any angle with equal validity. > > I suggest one of two approaches: run the MDS with only one dimension and > see how the indicators fall out. If you can clearly see the segregation > between the social status in just one dimension, then you can be fairly > confident that your social indicators are strongly aligned with the > strongest gradient in your dataset. This is mostly for your peace of mind, > as plotting data on a one-dimensional plot would confuse a lot or readers. > > For the publication, I'd first run a varimax rotation on the resulting MDS > scores if possible. This actually aligns the results optimally with your > input variables (i.e it gives your axes some meaning in the context of your > input data). Check, your software may do this by default. Second, include > a biplot array in the plot, which shows how each of your variables > correlates with the MDS axes. If you can't generate a biplot directly, > simply run a correlation between each of your variables and the MDS scores > for your samples and discuss the strongest correlations (i.e. MDS axis 1 is > strongly correlated with var 1, 3, and 6; etc.). > > Hope this helps. > > Burton > -- Hilit Finkler PhD student Zoology department The George S. Wise Life sciences faculty Tel Aviv University Israel