I am validating an inspection screen for a particular issue. I require zero
escapes from the screen, so cell (1,1) must not have any samples in it. I
can accept false alarms, however, so cell (2,2) may be non-zero. If the
number of false alarms gets too high, though, then the validity of the test
is reduced. Balanced data was collected for screen samples showing/not
showing the defect. When the actual production samples were destructively
analyzed for the defect, the following data was tabulated:
Actual
Defect NoDef SubT
Screen NoDef 0 19 19
Def 6 13 19
SubT 6 32 38
The std normal variates for cells (1,1) and (2,2) are (0-3)/1.139 = -2.634.
Here are my questions:
* Should I view this value as significant, thereby validating that
this is a successful screen?
* I believe I can calculate critical values for the null hypothesis of
independence, but how can I build a confidence interval for my results?
* Does anyone know of any references for the noncentral or extended
hypergeometric distributions?
Thank you for your assistance,
Mark Balhorn
Hardware Engineer / Industrial Statistician
Guidant Corporation
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|