Whoops, I Did It Again

Bill Parr

This note abstracts a general pattern I have seen repeatedly in management of processes. Not a good pattern, but a common one.

It goes something like this - every time.

Suddenly, a yield crash is noted in a critical product. Analysis quickly (or perhaps slowly!) isolates, using some solid technical knowledge plus available control charts, a particular process as the candidate culprit. The responsible engineer begins to fidget. It turns out that there is a control chart on a key output of their process. And, when one examines that control chart, it becomes clear that about 1 month ago the mean on that chart dropped, noticeably. In fact, it appears that the mean shifted to a value which was about 3*sigma below the old mean. The control chart almost immediately signalled the shift, and shows the shift has been sustained.

So, why no response to the shift?

"Things have been really busy. And I didn't really know that variable mattered."

Further questioning turns up a pattern of not responding to the signals sent by the control chart. One of failing to learn the lessons the process is trying to teach us through the chart. And one of putting process and product at risk as a result.

What are the possible lessons?

1) There is always value in plotting meaningful measurements in time order, with (preferably) control limits to help us in deciding what sorts of changes are true signals.

2) That value won't be realized unless we actually use the chart - seeking to understand the root cause of any out of control signals. The control chart is not like some sort of garlic one might wear around their neck to protect themself from a mythical vampire - it requires conscious and deliberate action on the part of the process manager to turn the information into a tool to guide action.

3) Shifts in a variable may have effects which will only be seen further down the line. Knowing these causal links will require a combination of good solid process knowledge plus the statistical tools - either the process knowledge or the statistical tools alone will in the long run prove to be insufficient.

So - let's not sing along with Britney next time.