RE: REQUEST: Predictive Modeling and Tampering

[Steven_S_Prevette@rl.gov]

The key on predictive modeling (and this is a big thing for us Operations
Researchers) with respect to Deming is to look at the prediction from a
System point of view, not an individual datum point.  

If you compare the most recent result to the predicted value (or worse, some
"standard") you are going to end up tampering with the process.  Review the
Funnel Experiment for implications on this. 

Also, if I (mis)use my best prediction as a goal, or an action setpoint, 50%
of the time I will be over the prediction, and 50% under, assuming it was a
good prediction.  I will continually be oscillating on taking actions.  50%
of the time I will be on the "wrong" side of the prediction.

The best use for predictive tools is when you have a stable system and you
have decided you need to improve its results.  Predictive models could help
assess different alternatives during the PLAN phase of PDSA.  

But the key is not whether the prediction was "right", but did you get a
positive change (special cause variation) after implementation (Do).  You
determine if you got the positive change using SPC during the Study phase.


Steve Prevette
Site Technical Authority for Statistical Trending
Environment, Safety and Health
Fluor Hanford, A Fluor Global Services Company
ASQ Certified Quality Engineer
steven_s_prevette@rl.gov
509-373-9371