Historical Designed Experiments

View Article - PDF

Title: Some Comments on "Historical" Designed Experiments
Author: Keith M. Bower
Publication Source: American Society for Quality, Six Sigma Forum, November 2004

Abstract

This article addresses the concept of applying techniques used for analyzing designed experiments, with historical data.

Notes

This paper was inspired by several conversations in which quality practitioners had been instructed to use the 2^k factorial design procedure on historically collected data. A key reason for this was that the practitioners required "proof" that they had done a designed experiment as part of certification during a quality improvement project. Of course, they had done nothing of the sort.

I wanted to summarize some of the key reasons when and why such an approach would be statistically invalid, and this was the resulting paper. The concept of removing datapoints to "get the software to work" was something I was very keen to address in the paper - i.e. it is wholly inappropriate.