Statistics Group Report Two types of SPC There is a lot of confusion about SPC. Walter Shewhart, Don Wheeler, and the BDA booklets "Why SPC?" and "How SPC?" say one thing. Almost all other books on SPC say something else. The reason for the disagreement is not that the other books are completely wrong in everything they say. (They are wrong in some things). The main difference is that the aim of SPC as intended by Walter Shewhart, and as described by most other writers, are completely different. The two aims are: Adjustment: To detect quickly when a process needs readjust- ment or repair. Improvement: To find the most urgently needed improvement in the system itself. Adjustment is all most books cover. It has some use, because it prevents things getting worse. The point of making the adjustment is to try to keep the product "within specification". This makes the statistical methods quite complicated. To be sure that most of the individual results are within specification, when we only have measured a few samples, we have to know that the process is stable, and that the individual measurements are "normally distributed". Otherwise we can not predict the proportion of the product within specification. Unfortunately, as Henry Neave has shown by computer simulation studies, even if these assumptions are true, the predictions this method makes can be wildly wrong. What is worse, we know in practice that the assumptions never are true. Improvement is what Dr Shewhart and Dr Deming are talking about, but hardly any books mention: or if they do, they are more concerned with the first problem, and so suggest rather ineffect- ive ways of finding improvements. Adjustment is needed too, but it takes second place, instead of being the sole aim. The reason why the approach has to be different is that some "signals" that you can pick up from a control chart tell you little about the nature of the underlying reason for going out of control. For example, a slow drift in the mean can rarely point straight to a cause of change, whereas an isolated point quite unlike the points on either side of it usually tells you all you need to know. At least, it does if the process operators themselves are keeping the chart: they will usually know just what happened at that point. By comparison, a slow drift may result from something that started to go wrong long before. Naturally, if all you are going to do is to alter the controls to bring the mean back into line, you want to detect a slow drift or change as soon as possible, and put it right. This is why many books suggest such a wide range of "out of control" signals, such as runs above the mean, or runs in the same direction. On the other hand, if you want to trace the underlying cause, and do something permanent about it, these signals are usually nothing but a nuisance. The process gets readjusted before you can trace the cause. So in the Deming-Shewhart approach, the only signal worth much is the simple q3SD from the mean. The distrib- ution, normal or otherwise rarely matters at all. And of course, we do not have to start with a process that is "under control". The aim is to find out why it is not: or if it is, to see clearly the effect of experimental improvements. Instead of emphasising complicated rules for detecting drift or a change of mean, what is needed is great care to see that the information about factors which might affect the process, and knowledge of changes, is immediately available to someone who can see the connections, and can get things done. In this approach control charts on inputs to the process, such as raw materials, temperatures, pressures, and so on, are as important, or more important, than control charts on the final product. For adjust- ment, only the final product matters. Obviously improvement is better in the long run, from all points of view. If simply adjusting the process is enough to meet specifications, improvement will meet them many times over. And the general effects on the system which result from improvement will have good effects that spread far and wide. The statistical methods used in improvement are also much easier to understand and use. The drawback, in many companies, is that short-term thinking rules, and no-one has the power to change the system.