contradictory odds ratios--a problem with the equation or the interpretation?

Hi jonathan,

Another idea that may possibly help. Suppose you break down the problem
into smaller pieces.

1) start with only language as predictor (I think I would use Mandarin as
reference). It's interesting to see what happens when one of the other
predictors is added. Which other predictor, after adding, makes the
significant language difference (if present) dissapear?

2) analyse each word separately.  This means a simple regression per word.
You probably have 53 observations then? Maybe entering all 6 predictors is
too much, so you could again start with languages only, etc. On the other
hand, entering all 6 predictors could show that for some words languages
differ in the way you expect, and for some words they do not.

Good luck, Ben.