By Lloyd S. Nelson
My consulting experience with manufacturing/engineering groups has revealed that virtually all of them have one or more hand-operated processes with someone in charge of keeping the process "on target." This has invariably led to overadjusting, because:
a) most of the time there is no control chart in place to guide them; and
b) the operators (engineers or technicians) in charge of monitoring the processes inherently feel that, unless they are continually adjusting the processes, they are not doing the job they are paid to do.
To be specific, let us suppose that a process output is designed to be 100. An operator measures the output and finds it to be 97. The process is adjusted by increasing its output by 3 units to correct the deficiency. This procedure is carried out from time to time. (It is related to "Rule 2" of the "funnel experiment" described by Deming (1986, p. 327)). It is "obviously" a sensible approach. Unfortunately, as we shall see, it is often not a sensible approach at all (see MacGregor (1990) for situations in continuous process industries where frequent adjustment may be called for).
How should this problem be addressed? If you, as a consultant, report at a sufficiently high level in the organization, you can request that the process be allowed to run with no adjustment—"just to see what will happen." At a lower level of operation it will probably take quite some "selling" to justify such an action. Nevertheless, this must be tried. It has been my experience that when a process is left alone, it runs for a much longer time before it becomes obvious that some action is needed. In some such cases the process has run, essentially, indefinitely without requiring any intervention. At this point, you should have the attention of the group in charge of the process. Now is when you introduce the concept of choosing a rule for indicating when a process needs to be fixed (not simply adjusted). Now is the time to introduce the concept of the Shewhart control chart.
Line up a conference room with a blackboard or white board, or at least a flip chart. Convene a meeting to discuss what data need to be collected in the form of subgroups. Make sure that the concept of a subgroup is clear, and that they understand what information can be derived from within and among subgroups. Help them with the mechanics of gathering data on the output of their process. You will need at least a dozen (and, better, two dozen) subgroups of some appropriate (economic) size. The size should be at least two, and does not need to be more then ten.
Reconvene after you have made a control chart from the collected data. Point out that the main purpose of the chart is to show quantitatively the "natural" distribution of the process data. Illustrate how the chart can reveal ways in which the process can go "out of control." Emphasize that without some indication from the control chart that the process is out of control, "adjusting" the process is not appropriate. Convince them that such action would serve to increase the variability of the output. It can be shown mathematically that the variance will double. It can be demonstrated physically as follows.
If you have a quincunx available, you can clearly demonstrate this increase in variability. If you do not have a quincunx, then you should draw a normal distribution on its side on a separate piece of paper and move it up and down while drawing a stream of data coming from it. No leap of imagination is required to appreciate that the out-flowing data will form a distribution that is wider than the original distribution.
If a Shewhart control chart does not indicate an out-of-control situation, then it would be incorrect to "adjust" the process. If it does show an out-of-control situation (for example, a point outside a three-sigma limit, six points in a row steadily increasing or decreasing, or other tests described by Nelson (1984, 1985)), then it is reasonable to suspect that the process needs to be fixed. This will almost certainly be more complicated than simply adjusting it.
The main point is that it is often folly to adjust a process by moving it either up or down whenever the most recent measurement shows it to be either too low or too high. When such a "compensating" adjustment is made, the variance of the output will be increased. A control chart will inform you when something has happened that requires fixing. A successful study will result in a process that operates in a steady manner with a minimum of attention. It will, as the old phrase goes, be "in statistical control."
Key Words: Process Adjustment, Shewhart Control Charts.
Read Full Article (PDF, 60 KB)