I've got an idea how to implement continuous improvement. Every year, fire your bottom 10% producers, and hire others to replace them. On an ongoing basis, this will solve your personnel problems. Cool idea, huh?

It won't work, of course. And the main problem won't be the ensuing morale problems, or even turnover. The problem is that very few of the fired employees contributed to the root cause of the company's problems. Very few of them contributed to the company's bottleneck. Sadder, some of the fired employees might be among the company's best. How can that be?

Everything has variation. Some of the variation has a cause, and some is statistically insignificant "random noise". With extreme amounts of employee variation, all the variation could be random noise. In other words, the worst guy this year could be the best guy next year, in spite of the fact that nobody changed what they were doing.

To find out whether a particular person's performance is statistically significant, you do some statistic evaluations on the performance figures to obtain an Upper Control Limit (UCL) and a Lower Control Limit (LCL). All performances between the upper and lower control limits are considered normal performance, not subject to any discipline, correction or awards. Anyone above the UCL should be evaluated to find out what he's doing right, so everyone can do the same thing. Anyone falling below the LCL should be, for lack of a better word, troubleshot, to find out what is going wrong with his performance. There are cases where this statistical analysis revealed that the employee needed a new eyeglass prescription, and in fact once the new eyeglasses were obtained, the problem vanished. Below is a cartoon, based on a control chart, lampooning those who would "grade on the curve" in a corporation:

The gunslinger is making a fool of himself because all employees, including Barnes, are within the statistical limits, thus none of them is the cause of any company problems, and the company cannot be improved by getting rid of, or tampering with, any one of them. If the gunslinger wants to improve productivity, he must look at the production system common to all those employees.

The process of assigning causation through statistical analysis is called Statistical Process Control (SPC). This was W. Edwards Deming's starting point. It's an extremely powerful tool. SPC isn't used only to evaluated the performance of people. It can evaluate anything that can be expressed numerically. Basically, when any given "thing" falls outside of the UCL/LCL borders, it should be investigated and either fixed or propagated, as appropriate. Notice that this bestows the opportunity to answer the question "why was this month so good" (of course after determining that the months figures were above the UCL). That answer can then be used to make permanent improvements.

So SPC first alerts us to problems, and then gives us some tools to diagnose those problems. Note that at any point in the investigation we can move from SPC to a diagnostic process like the Universal Troubleshooting Process, Jim Roach's Diagnosis, the Six Step Loop, Root Cause Analysis, or Theory of Constraints, to finish getting down to the root cause.

SPC is deeply rooted in the quality movement. One hallmark of "quality" is the reduction of variation. As variation decreases, the UCL and LCL approach each other. In the case of the people in the preceding cartoon, the reduction of variation would come not from hiring new people, but from creating a system that allows them more consistent (and hopefully higher) production. This would be equally true if the preceding control chart were ball bearing diameters instead of human performance.

In the vernacular of SPC, anything outside the UCL/LCL limits warrants investigation and is called an "assignable cause", or sometimes a "special cause". Anything within those limits is termed a "chance cause", a variation born of random noise needing no investigation. If one wants to reduce the variation or raise the numbers en-masse, the system common to all the numbers must be investigated.

There are many books on the subject. The book I own is called "Understanding Statistical Process Control by Donald J. Wheeler and David S. Chambers. This book contains all the equations you need to calculate the UCL and LCL, as well as the math to diagnose numerous problems that can be evaluated statistically. As is obvious from this article, I haven't even scratched the surface of the information contained in this book.