I want to build a model that determines whether to pre-screen my widgets for defects. If I do pre-screen, it costs a fixed amount per check and I resolve the problem 100% of the time. If I don't pre-screen, its cost will be the cost of fixing the problem, plus the cost of materials for expired inventory. I have a whole bunch of variables that impact whether a defect will occur in each particular widget, but if it does occur and I didn't pre-screen it, it will definitely result in a fixed repair cost and a variable expired inventory cost based on cost of materials. For each widget, I need to know if the expected cost of the defect exceeds or is less than the cost of pre-screening.

Things I know:

  • A - Cost of pre-screening
  • B - Model to Predict Whether it will fail - based on test data - of ones it says it won't have a defect, it is right 98.5% of the time (false negatives).
  • C -Model to Predict expected cost of Raw Materials for each widget produced
  • D - % of repairs that will be successful and result in no expired inventory cost
  • E - Fixed Cost of Repairs

Therefore, I should screen if:

A - B*(E+C*D) > 0

I have all of these pieces, but I don't know how to build this into a model.

Anyone with R or Stata skills, I would appreciate being pointed in the direction of a good package to use for building this model.

  • $\begingroup$ Please let me know if this answered your question or if you have any further concerns. $\endgroup$ – Jack Ryan Apr 30 '14 at 11:25

Considering my predilection for R, I am going to do something unlikely: I would advise you not to use R or any other program. If you agree to the principle that the production should be done correctly, I think you would agree with the principle that determining the inspection criteria and the inspection itself should be as simple as possible (but you can call it a "decision tree" to justify your title):

Question 1: Is the supply of your part stable?

A1 == NO. Considering a type of widget: if the supply of this widget is not in a state of control, strongly consider 100% inspection (Deming, 1986). DONE.

A1 == YES. OK. . .

Question 2: Does the cost of testing outweigh the cost of not testing?

If the supply of this part is in control, you can adopt an all-or-nothing testing strategy (Deming, 1986), given:

p   = [historic] average percent defective [from] incoming lots of parts 
      (e.g. received 100 widgets Friday, 3 defective; p = 3%).  
      Remember: stable process exists! 
k   = cost to inspect one part
l   = cost to dismantle, repair, reassemble, and test an assembly that 
      fails because a defective part was put into the production line.
k/l = break even quality, or break-even point (always between 0 and 1)

A2 == NO. If p > k/l inspect 100%. You will save money with each inspection by preventing expensive mistakes further down the production line. DONE.

A2 == YES. If p < k/l inspect 0%. You will lose money with each inspection by spending valuable time inspecting parts that are most likely acceptable. DONE.

The statistical proof of this is "exceedingly simple," but involves a chart which I might attempt to reproduce later if I'm feeling ambitious. I would encourage you to find a copy of the text which I created this answer from:

Deming, W. E. (1986). Out of the Crisis (Ch 15). Cambridge, MA: MIT Center for Advanced Engineering Study.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.