[Rasch] Sparse rating scales
Mike Linacre (RMT)
rmt at rasch.org
Mon Jan 30 07:18:00 EST 2006
Quote: "Things become even worse when people run Rasch analysis using 50
cases when their rating scales range from 1-100. They do not even have 1
person per scale point. By the way, Mike, how would you comment on that?"
Since they are approximating distributions, they could use a Rasch model
which smooths over the scale, for instance Andrich's "Guttman components"
approach, or a continuous Rasch model using a polynomial or Fourier
transform to define the rating scale structure. A direct approach would be
to impose a set of linearly increasing Rasch-Andrich thresholds: 0, 1, 2,
3, 4,....,99. And then adjust the logit scaling (USCALE= in Winsteps) until
the overall mean-squares approximate 1: www.rasch.org/rmt/rmt142n.htm
More information about the Rasch