I meant to post some more on the MDS diagram and NNet from the last day or two.
First up, here’s a map of the Lake Eyre Basin, the area where Karnic languages were spoken.
The map doesn’t have quite all the varieties I used in my dataset. Nhirrpi is in the SE corner of the area labelled “Yandruwandha” and Yanda is up near Guwa in the Northeast. The varieties with REU after them were recorded in the 19th Century by Johannes Reuther and were treated separately because (a) they aren’t always the same dialect as the later recorded languages (e.g. Diyari and Yandruwandha) and (b) because they show some influence from Arabana that appears to be the result of collecting the data at Killalpaninna mission (in Arabana-Wangkangurru country).
The MDS 2-D pic from yesterday corresponds fairly closely to the actual geographical layout of the languages, but with two main exceptions. If the diagram is oriented the way I have it, the West-East axis is fine, but the North-South is skewed. (a) Warluwarric ends up a long way east of where it ‘should’ be: closer to Wangkumara; and (b) Pitta-Pitta and Wangkayutyuru are south of where they should be. (But if we just flip that axis, Yanda, Guwa, etc would be south of where they should be).
Here’s what I think is happening: There is a Central Karnic innovative area (identified by Peter Austin in his 1990 “Classification of Lake Eyre languages” paper). Peter didn’t include Arabana-Wangkangurru but the Reuther data contains enough loans that AW and Diyari end up looking quite similar. Note the relative placement of Arabana, ArabanaREU, DiyariPA (Peter’s Diyari data) and DiyariREU. This gets us a grouping of Diyari, Arabana, Yarluyandi, Ngamini, and Yandruwandha (and associated dialects).
Now, in my NNet work with different subsets of the data, I think it’s fairly clear that Pitta-Pitta and Wangkumara share a number of archaisms (I know that Breen 2007 didn’t find many, but my data has Pitta-Pitta grouping variably with Arabana, Wangkumara or ‘Karnic other than Arabana’ depending on the subset). So I suspect that Pitta-Pitta and Wangkatyuru are being ‘pulled’ towards Wangkumara because of the shared archaisms. Or rather, what we see here in the entral Karnic cluster, and then the other languages arranged around the edges.
David asked for 3D MDS. Here it is.
This looks much more similar to the actual geographical layout. The first MDS was monotone MDS in 2D. I don’t know how the formulae differ so I’m not sure why these two are so different.