06-12-2015 10:27 AM
we do have a distance matrix with "1 minus jaccard" as distances in a triangle shape.
To estimate the coordinates we use proc mds.
What level to we have to assume? Is it level=absolut as given in the example with the flying mileages?
When using the level=absolut option we get a fit plot where the data points differ very much from the diagonal.
And they don't differ randomly but in a certain pattern.
Does this mean, that the estimated coordinates aren't good or interpretable?
09-14-2015 03:25 AM
Hello all together,
today we can share our experiences on MDS with you.
Our data is like this: matrix of distances with more than 500 rows and columns, distances between 0 and 1, small distances are more important than bigger ones.
We got the best results with level=ordinal, which fits an ordinal MDS
To estimate the smaller distances with higher accuracy we used the following weights: weight=(observed distances)**(-5)
Especially increasing the number of levels of the MDS (up to 15 or 20) gave better results.
Better results were measured as a lower Stress value, better fit plot with randomly scattered points around the diagonal, better Shepard-Plot (observes versus estimated distances).
Our Shepard-Plot showed less variance for smaller distances than for bigger ones due to our weights.
For further information see P. Groenen, I. Borg (2005), "Modern Multidimensional Scaling", Springer