A control chart with a regular pattern demands some analysis. Regardless of the pattern that emerges when data is charted using SPC software, it should trigger an alarm and generate efforts to gather more information about the process. Learn what steps you should take to discover why a non-random pattern is emerging.
If you’re decorating your family room, you may be looking for regular, repetitive patterns, to tie the look together. But using control charts for statistical process control is a different story altogether. Keep those regular patterns out of your charting life, if you want to get the most out of your data analysis.
The object of control charts is to determine stability in a system and to reduce variation. “Stable” does not mean “beautifully, evenly patterned,” however; a chart with such a regular pattern should elicit some attention and a need to take steps to discover why it is emerging, since every system and every process will have variation.
That variation will always be reflected in the charting of the data. Tampering with a system, of course, can render the chart invalid—even though it may seem to have a terrific, predictable pattern. Deming’s funnel experiment demonstrates this loss of validity, but there are also everyday examples.
Think about a meeting room, for example, where temperature data is collected every morning and every evening. Every night, the maintenance crew sets the thermostat to 63 degrees. In the morning, those who use the room find this to be too chilly, so they turn the thermostat to 73. Within minutes, the temperature reaches a point close to that level. But later in the day, the thermostat is once again set for 63, and the environmental temperature plummets accordingly. If one were to collect data twice a day—once in the early morning, and once in the evening—a regular pattern would clearly emerge. But a chart that illustrates this data would indicate nothing about the system itself. It would only indicate that the room temperature had been altered twice each day.
What do you do if you see a pattern like the one below forming on your SPC chart? For those who value order, this chart seems to represent a repetitive, orderly, and perhaps predictable process. Besides, it’s a nice pattern, you might say.
This is what is known as a “sawtooth pattern” in the data. The data points alternate above and below the mean, in a regular and somewhat predictable pattern. For some reason, alternate subgroups have alternating greater and smaller averages.
Though it may produce an attractive chart with some level of predictability, this chart is all a little too neat. It represents what is known as a nonrandom pattern. In a normal distribution—that is, when a system is in control or stable—one could imagine tilting the chart on one end and watching as all the points slip down to form a normal curve. About half the points would fall above and half below the centerline. One could expect to find about two thirds of the total points in the middle regions and no repeatable patterns in the data. This would not be the case with the chart above.
But what does one do about it, anyway? Patterns that form from data are by definition not random, and therefore demand further investigation. Examples of typical patterns include:
- Data points too close to the average
- Data points too far from the average
- 2 of 3 points beyond 2 sigma
- 4 of 5 points beyond 1 sigma
- And, of course, the sawtooth
Stratifying or splitting the data by key variables may assist in analyzing this problem. This may occur if alternate samples from two machines or production lines are used.
Regular, predictable patterns can be caused by a number of circumstances, including:
- Edited data
- Reduced variability without recalculation of control limits
Whenever a pattern emerges from the data, the most important step is to try to establish the reason. Is this apparent improvement genuine? Can the improvement be maintained? If the improvement can be maintained, then the control limits need to be recalculated. Although the data looks more stable than normal, this condition is referred to statistically (and paradoxically) as unstable.
Each nonrandom pattern offers unique opportunity and demands slightly different analysis. When patterns such as cycles occur, for example, the data rises and falls in a rhythmic pattern. The pattern is definitely not random. This could be caused by some regular, periodic change in the system.
A positive aspect of cycles, if there is one, is that they tend to indicate that there is one major cause of variation, which will typically be changing in a similar cyclic fashion. If the cause of the cycle can be established and reduced, this should result in a major improvement to the process.
Regardless of the pattern that emerges when data is charted, it should trigger an alarm and generate efforts to gather more information about the process.
After all, it’s stability in the process, not neat patterns in the charts, that will produce healthy outcomes in the system. Randomness is your friend, if this is your objective.
Now, back to that family room project…