I have a problem evaluating a Grading System.

I have data from a grading system.  The problem is to determine if the 
system is fair.  There are applications to be graded and there are hundreds. 
  There are graders and there are about a dozen of them.  Exams are held 
several times a year and this process has been going on for years.  The 
applicants fall into groups depending on who they will be working for if 
they pass their exam.  The Graders groups differ in terms of the kind of 
work the applicant is applying to do.  The assignment to the grading group 
is made by an expert panel.  When an application comes in it can either be 
thrown out or scored.  In either case I know which happens.  If it's scored 
then I have the score.  If the score is really bad the applicant can apply 
for a later exam.  An applicant can come back up to several times.  I know 
which ones come back and can link the records.  Graders are pretty stable 
over time but they have their careers with start and end times.  My 
questions; Are  applicants treated differently by the graders depending on 
which type of work they applying for?  Is there any way to distinguish a 
poor group of applicants from a bias against an applicant group in the 
graders?  Does it matter how many times the applicant comes back for follow 
up exams.  The scores generally get better but there is no guarantee.
  The scores are not normally distributed and I have tried several 
approaches.  I have looked at the risk ratio for the initial turn-down.  I 
have looked at the Mann Whitney Wilcoxon for detecting shifts in the rank of 
the scores.  I don't know if the variance of the scores is constant across 
applicant groups and I know that is confounded with the wilcoxon test.  If 
the risk ratio isn't at least 1.8:1 people won't notice a difference.  How 
do I estimate and correct for the lack of power in that test.  The examiners 
want to use average scores to evaluate the grading system but I don't know 
how to properly test the averages considering I don't know the 
distributions.  Are there standard ways to approach this problem?  Problems 
like this come up in cases of employment discrimination,  What sort of model 
should I use?
Stan

_________________________________________________________________
Send and receive Hotmail on your mobile device: http://mobile.msn.com

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to