Vous êtes sur la page 1sur 4

# How Many Values Do I Need for Limits?

Donald J. Wheeler
What happens when we use only a few values to create an XmR Chart? There are three aspects to the behavior of the chart that are of interest here. What happens to the number of a false alarms? What happens to the ability of the chart to detect a change? And what happens when the baseline data contain a signal? THE LIKELIHOOD OF A FALSE ALARM Mathematical theory tells us that the risk of a false alarm whenever we place a point on an X-Chart is approximately 0.27%. But what happens to this risk when only k values have been used to compute the limits? Simulation studies provide the answers shown in Table 1 and Figure 1.

1.5%

## k even Limits Based on Median Moving Ranges

1.0% 0.5%

k odd

Theoretical 0.27% 0.0% k 5 Limits Based on Average Moving Ranges 10 15 20 25 30 Number of Values Used to Compute Limits 35 40

Figure 1: The Effective Alpha-Levels for the Computed Limits The bottom curve in Figure 1 shows the effective alpha-levels for the limits based on the Average Moving Range. These alpha-levels rapidly approach the theoretical value and then plateau. The two upper curves are for limits based on a Median Moving Range. Both curves show that the effective alpha-level drops down toward the theoretical value as k increases. Moreover, while these limits do have larger effective alpha-levels than limits based on the Average Moving Range (as was expected), these effective alpha-levels do not exceed one percent except when k is 4, 6, 8, or 10. Since an alpha-level of 1.7% is still considered to be a conservative alpha-level, we do not loose the conservative nature of the charts even when the limits are based on few data. Based on Figure 1 it should be clear that we can construct an XmR Chart using as few as four original values without an undue risk of a false alarm. Therefore, the conservative nature of the process behavior chart is maintained even when small amounts of data are used to construct an XmR Chart. This means that when you get a point outside the limits on the X Chart, you can safely interpret that point as a signal, regardless of how few data have been used in constructing the limits.

## Using an XmR Chart with Few Data

Donald J. Wheeler

Table 1: Effective Alpha Levels for X Charts Based on k Data Limits based on an Average Moving Range Limits based on a Median Moving Range k alpha k alpha k alpha k alpha 4 0 14 0.0023 4 .0167 14 .0080 5 0.0008 15 0.0023 5 .0086 15 .0072 6 0.0021 16 0.0023 6 .0145 16 .0076 7 0.0021 17 0.0025 7 .0087 17 .0072 8 0.0021 18 0.0026 8 .0123 18 .0071 9 0.0021 19 0.0024 9 .0088 19 .0061 10 0.0021 20 0.0027 10 .0104 20 .0065 11 0.0022 25 0.0025 11 .0080 25 .0055 12 0.0025 30 0.0026 12 .0090 30 .0053 13 0.0023 40 0.0028 13 .0080 40 .0046 100 0.0024 100 .0030 DETECTING SIGNALS FOLLOWING THE BASELINE PERIOD Having established that we are not likely to have spurious signals, how sensitive is a chart based on few data when a real signal occurs after the baseline period? Here we look at that proportion of the time that the X Chart will detect a three-sigma shift in location that occurs on the first point following the baseline period. Both intuition and theory tell us that an X Chart should detect such a shift half the time. If three-sigma shifts are detected on the first point more than one-half the time, then the chart is more sensitive than theory predicts. If three-sigma shifts are detected on the first point less than one-half the time, then the chart is less sensitive than theory predicts. Table 2: The Proportion of the Time that a Three-Sigma Shift Occurring After the Baseline Period Was Detected on the First Observation Following the Shift by X Charts Based on k Data Limits based on an Average Moving Range k proportion k proportion 4 0.531 9 0.510 5 0.523 10 0.510 6 0.518 11 0.508 7 0.515 12 0.506 8 0.513 13 0.504 14 0.507 Limits based on a Median Moving Range k proportion k proportion 4 0.475 9 0.474 5 0.459 10 0.482 6 0.482 11 0.479 7 0.472 12 0.482 8 0.480 13 0.478 14 0.486

To empirically determine the sensitivity of the XmR Chart, limits were first computed using k observations generated from a standard normal probability model. Next an additional observation that was shifted by three sigma units was compared with these limits. By repeating this exercise 600,000 times for each value of k and noting the proportion of times that the shifted value was outside the computed limits I obtained the values in Table 2 and Figure 2. The theoretical value of 0.500 is shown as a horizontal reference line in Figure 2. The upper curve shows the probability of detecting a three-sigma shift when the limits are based on the Average Moving Range. As k increases these probabilities drop down toward the theoretical value. The middle curve is for limits based on the Median Moving Range when k is an even number. The bottom curve is for limits based on the Median Moving Range when k is an odd value. They both move toward the theoretical value as k increases.

## Using an XmR Chart with Few Data

Donald J. Wheeler

## Probability of Detecting a Three Sigma Shift

0.54 0.52 0.50 0.48 0.46 Limits Based on Median Moving Range Limits Based on Average Moving Range Theoretical Probability of 0.50

## 4 5 6 7 8 9 10 11 12 13 14 Number of Values, k, Used for Limits of XmR Chart

Figure 2: How the Number of Baseline Values, k, Affects the Likelihood of Detecting a Shift that Occurs After the Baseline Period Figure 2 shows that limits based on the Average Moving Range are slightly more sensitive than theory predicts, while limits based on the Median Moving Range are slightly less sensitive than theory predicts. However, when using the X-Chart, you will not really be able to detect the differences in sensitivity shown by the departures in Figure 2. This is why hundreds of thousands of charts were needed to quantify the differences shown. All of these departures from the theoretical value are too small to have any real impact in practice. Therefore, the ability of the X Chart to detect changes that occur after the baseline period is not dramatically affected by the number of values used in the baseline period. THE ABILITY TO DETECT A SIGNAL WITHIN THE BASELINE PERIOD When the baseline data contain values that occur both before and after a change in process location, it is inevitable that the change will influence the limits. When the baseline period contains a large number of values, the effect of any shift that occurs within the baseline period will be diluted. However, when the baseline period contains few data, any shift will have a pronounced impact upon the limits. Thus, it is inevitable that the XmR Chart will become less sensitive to a baseline shift as fewer data are used in the baseline period. This section will quantify this loss in sensitivity. Table 3: Size of Shifts Within the Baseline Period Detected Half the Time with X Charts Based on k Data Limits based on an Average Moving Range Limits based on a Median Moving Range k shift k shift k shift k shift k shift k shift 4 11 4.18 18 3.53 4 5.95 11 3.85 18 3.42 5 15.23 12 3.99 19 3.54 5 5.32 12 3.69 19 3.45 6 7.55 13 3.89 20 3.47 6 4.77 13 3.69 20 3.37 7 5.83 14 3.78 25 3.37 7 4.49 14 3.61 25 3.31 8 5.03 15 3.73 30 3.29 8 4.12 15 3.58 30 3.26 9 4.59 16 3.68 40 3.22 9 4.01 16 3.53 40 3.20 10 4.33 17 3.62 100 3.07 10 3.88 17 3.49 100 3.07 For purposes of the comparisons being made here I chose a specific type of change where the last value, out of k values, was shifted by some known amount. (Other types of signals might have greater or

## Using an XmR Chart with Few Data

Donald J. Wheeler

lesser impacts upon the computed limits, but this signal will be sufficient for our purposes.) For different numbers of baseline values, k, the smallest shift that was detected in at least 5000 out of 10,000 X Charts is given in Table 3 and Figure 3. These empirical values show the expected increase in the lack of sensitivity as k gets smaller.

Size of Baseline Shift in Sigma Units That Was Detected Half the Time

8 Limits Based on the Average Moving Range 7 6 5 4 3 5 10 15 20 25 30 40 Number of Baseline Values, k, Used to Compute Lmits Figure 3: How Shifts Within the Baseline Period Desensitize the X Chart Limits Based on the Median Moving Range

The upper curve in Figure 3 shows what happens when the limits are based on the Average Moving Range. When k is 12 or smaller these limits become very insensitive to shifts that occur within the baseline period. The lower curve in Figure 3 shows what happens when the limits are based on the Median Moving Range. Here we see that these limits display a greater ability to detect changes within the baseline period for small values of k than do limits based on the Average Moving Range. However, when k is less than 9 even these limits become very insensitive. These losses of sensitivity suggest that you should exercise reasonable care when choosing a baseline period to make sure that it is free from known changes. While this is good advice for any value of k, it is imperative with short baseline periods. In summary, we have seen that, in practice, the number of data used to compute the limits does not have an appreciable impact upon either the false alarm rate or the ability of the charts to detect changes which occur following the baseline period. As long as we exercise reasonable care to avoid known or obvious changes within our baseline period, we can construct an XmR Chart using as few as 4 to 6 original data. Remember, the objective is not to compute the right limits, but to use the limits to characterize the behavior of the process. This characterization does not require high precision in our computations, but clear thinking in how we use and interpret the data.