The Problem with Tampering
When a process is centered on target and is in state of statistical control, any adjustments to the process only increase variation. Adjusting a process that is in control is referred to as tampering.
The Funnel Experiment and Deming’s Four Rules
The classic analysis of the effects of tampering is Deming’s Funnel Experiment
. In this experiment, participants drop marbles through a funnel suspended over a target. The funnel represents the process, the marble drop location is the feature being produced, and the target is the customer specification.
Deming described four approaches—also referred to as rules
—that encompass the typical ways in which the experiment participants tamper with the funnel (Out of Crisis,
1986, p. 328).
Rule 1: No adjustment
The optimal approach is to leave the funnel fixed and aimed at the target, without making any adjustments. When a process is stable, centered, and shows only the inherent variation, there is no reason to make an adjustment.
The takeaway: Before attempting any process adjustment, you must gather enough data to make sure you understand the normal behavior of the process. Use a control chart to track variations, and then adjust the process only when special variations occur.
Rule 2: Adjustment from last position
Sometimes referred to as the “human nature” approach, some participants move the funnel after each drop, to try and compensate for the previous drop’s variation. In this approach, the funnel is moved the exact negative distance of the drop. Compensating for the “error” of the drop, might improve the on-target average but doubles the variation.
The takeaway: When participants compensate for error, the variation doubles—and remember, variation is the true issue. This problem is prevalent in gauge calibration when manufacturers adjust a gauge after taking one standard measurement.
Rule 3: Adjustment from target
Participants trying to take a “logical” approach also move the funnel to try to compensate for the previous drop. But in this instance, the funnel is moved not based on its last location, but on its distance from the target. For example, if the measurement of the previous drop was 5 units above the target, participants move the funnel 5 units below the target.
The takeaway: Although this approach seems logical, it results in an oscillating process.
Rule 4: Adjustment from last drop
In this approach, participants move the funnel to point at the previous drop rather than the target. In other words, at drop n, they set the funnel over the location of the n-1 drop. As you might expect, this approach creates a pattern that moves steadily away from the target.
The takeaway: Believe it or not, this approach occurs in calibration scenarios when one product is used to set up for the next production. This issue is typical in workplaces where on-the-job training is prevalent.