Technometrics

ISSN: 0040-1706 (Print) 1537-2723 (Online) Journal homepage: https://www.tandfonline.com/loi/utch20

A Robust Standard Deviation Control Chart

Marit Schoonhoven & Ronald J.M.M. Does

To cite this article: Marit Schoonhoven & Ronald J.M.M. Does (2012) A Robust Standard
Deviation Control Chart, Technometrics, 54:1, 73-82, DOI: 10.1080/00401706.2012.648869

To link to this article: https://doi.org/10.1080/00401706.2012.648869

View supplementary material

Accepted author version posted online: 20

Dec 2011.
Published online: 02 Mar 2012.

Submit your article to this journal

Article views: 875

View related articles

Citing articles: 33 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=utch20
 Supplementary materials for this article are available online. Please go to http://www.tandfonline.com/r/TECH

A Robust Standard Deviation Control Chart

Marit SCHOONHOVEN and Ronald J.M.M. DOES

Institute for Business and Industrial Statistics, University of Amsterdam

1018 TV Amsterdam, The Netherlands
([email protected]; [email protected])

This article studies the robustness of Phase I estimators for the standard deviation control chart. A Phase
I estimator should be efficient in the absence of contaminations and resistant to disturbances. Most of the
robust estimators proposed in the literature are robust against either diffuse disturbances, that is, outliers
spread over the subgroups, or localized disturbances, which affect an entire subgroup. In this article, we
compare various robust standard deviation estimators and propose an algorithm that is robust against both
types of disturbances. The algorithm is intuitive and is the best estimator in terms of overall performance.
We also study the effect of using robust estimators from Phase I on Phase II control chart performance.
Additional results for this article are available online as Supplementary Material.

KEY WORDS: Average run length; Mean squared error; Phase I; Phase II; Statistical process control.

√
1. INTRODUCTION by 1 − p/p. It is common practice to use p = 0.0027 and so
ARL = 370.4 and SDRL = 369.9 when σ is known.
The performance of a process depends on the stability of
When the standard deviation is estimated, the conditional run
its location and dispersion parameters and any change in these
length—the run length given an estimate of σ —has a geometric
parameters should be detected as soon as possible. To monitor
distribution. However, the unconditional RL distribution—the
these parameters, Shewhart (1931) introduced the idea of control
run length distribution averaged over all possible values of the
charts in the 1920s. The dispersion parameter is controlled first,
estimated σ —is not geometric. Quesenberry (1993) showed that
followed by the location parameter. The present article focuses
for the X and X control charts, the unconditional ARL as well
on control charts for monitoring the standard deviation.
as the unconditional p are higher than in the (µ,σ )-known case.
We assume that in the design of such charts, the in-control
Chen (1998) studied the unconditional run length distribution of
standard deviation (σ ) is unknown. Therefore, σ must be esti-
the standard deviation control chart and showed that the situation
mated from subgroups taken when the process is assumed to be
is somewhat better than for the X control chart. To achieve the
in control. This stage in the control charting process is denoted
intended unconditional in-control performance when the limits
as Phase I (cf. Woodall and Montgomery 1999). Control limits
are estimated, one could derive Un and Ln by controlling either
are calculated from the estimated σ to monitor the process
the in-control p or ARL, or a percentile point of the in-control
standard deviation in Phase II. The Phase I and Phase II data
RL distribution. An advantage of using the ARL is its intuitive
are arranged in subgroups indexed by i. We denote by Xij ,
interpretation. A drawback, however, is that the ARL is strongly
i = 1, 2, . . . , k and j = 1, 2, . . . , n, the Phase I data and by Yij ,
determined by the occurrence of extremely long runs. Hillier
i = 1, 2, . . . and j = 1, 2, . . . , n, the Phase II data. The Xij ’s
(1969) and Yang and Hillier (1970) derived correction factors
are assumed to be independent and N (µ, σ 2 ) distributed and
for the range (R) and standard deviation (S) control charts by
the Yij ’s are assumed to be independent and N(µ, (λσ )2 ) dis-
controlling p.
tributed, where λ is a constant. When λ = 1, the standard
Jensen et al. (2006) conducted a literature survey of the ef-
deviation is in control; otherwise it has changed. Let σ̂ be an un-
fects of parameter estimation on control chart properties and
biased estimate of σ based on the Xij ’s, and let σ̂i be an unbiased
identified issues for future research. Their suggestion on page
estimate of λσ based on the ith subgroup Yij , j = 1, 2, . . . , n.
360 is the subject of the present article. More specifically, we
The process standard deviation can be monitored in Phase II by
will find robust estimators for Phase I data and we will study the
plotting σ̂i on a Shewhart-type control chart with limits
performance of these robust estimators during Phase II moni-
toring.
UCL
= Ln σ̂ ,

= Un σ̂ , LCL (1) Rocke (1989) proposed standard deviation control charts
based on the mean or the trimmed mean of the subgroup ranges
where Un and Ln are chosen so that the desired control chart or subgroup interquartile ranges. Moreover, he studied a two-
behavior is achieved when the process is in control. When stage procedure whereby the initial chart is constructed first
σ̂i falls within the control limits, the process is deemed to be in and then groups that seem to be out of control are excluded.
control. We define Fi as the event that σ̂i falls beyond the limits, The control limits are recomputed from the remaining sub-
P (Fi ) = p as the probability of that event and RL as the run groups. Rocke (1992) provided the practical details for the
length, that is, the number of subgroups until the first σ̂i falls
beyond the limits. When the limits are known, Fs and Ft (s
= t)
© 2012 American Statistical Association and
are independent and therefore RL is geometrically distributed the American Society for Quality
with parameter p. Hence, the average run length (ARL) is given TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1
by 1/p and the standard deviation of the run length (SDRL) DOI: 10.1080/00401706.2012.648869

73
74 MARIT SCHOONHOVEN AND RONALD J.M.M. DOES

construction of these charts. Tatum (1997) explained the dif- The first estimator of σ is based on the pooled subgroup
ference between diffuse and localized disturbances: diffuse dis- standard deviation
turbances are equally likely to perturb any observation, whereas   k 1/2
1
localized disturbances affect all observations in a subgroup. He S̃ = Si2 , (2)
proposed a method, constructed around a variant of the biweight k i=1
A estimator, that is resistant to both diffuse and localized distur- where Si is the ith subgroup standard deviation defined by
bances. Finally, Davis and Adams (2005) proposed a diagnostic
 1/2
1 
n
technique for monitoring data that might be contaminated with
Si = (Xij − X̄i ) 2
.
outliers to react to signals that indicate a true process shift only. n − 1 j =1
In this article, we investigate robust Phase I estimators for the
subgroup standard deviation control chart. The estimators con- An unbiased estimator is given by S̃/c4 (k(n − 1) + 1), where
sidered are the pooled standard deviation, the robust biweight c4 (m) is defined by
A estimator of Tatum (1997), and several adaptive trimmers.  1/2
Additionally, we look at an adaptive trimmer based on the mean 2 (m/2)
c4 (m) = .
deviation from the median, a statistic more resistant to diffuse m−1 ((m − 1)/2)
outliers (cf. Schoonhoven, Riaz, and Does 2011). For diffuse This estimator provides a basis for comparison under normality
outliers, we think that a control chart for individual observa- when no contaminations are present. Mahmoud et al. (2010)
tions would detect outliers more quickly. We therefore include showed that this estimator is more efficient than the mean of
an estimator based on the individuals chart. To measure the the subgroup standard deviations and the mean of the subgroup
variability within and not between subgroups, we correct for ranges when the data are normally distributed.
differences in the location between the subgroups. Finally, we We also evaluate a robust estimator proposed by Tatum
present an algorithm that combines the last two approaches. The (1997). This approach is applicable for n ≥ 4. The method
performance of the estimators is evaluated by assessing their begins by calculating the residuals in each subgroup, which
mean squared error (MSE) under normality and in the presence involves subtracting the subgroup median from each value:
of several types of contaminations. Moreover, we derive factors resij = Xij − Mi . If n is odd, then in each subgroup, one of
for the Phase II limits of the standard deviation control chart the residuals will be zero and is dropped. As a result, the
and assess the performance of the control charts by means of a total number of residuals is m = nk when n is even and
simulation study. m = (n − 1)k when n is odd. Tatum’s estimator is given by
The article is structured as follows. The next section intro-  k    
2 4 1/2
duces the standard deviation estimators, demonstrates the im- m 
i=1 j :|u |<1 res2
ij 1 − uij
Sc∗ =      ,
ij
plementation of the estimators by means of a real-world ex- (m − 1)1/2  k i=1 j :|uij |<1 1 − u2 1 − 5u2 
ij ij
ample, and assesses their MSE. Next, we present the design
schemes for the standard deviation control chart and derive (3)
the Phase II control limits. We then describe the simulation where uij = hi resij /(cM ∗ ), M ∗ is the median of the absolute
procedure and simulation results. The article ends with some values all residuals,
concluding remarks. ⎧
⎨1 Ei ≤ 4.5,
hi = Ei − 3.5 4.5 < Ei ≤ 7.5,
⎩
c Ei > 7.5,
2. PROPOSED PHASE I ESTIMATORS and Ei = IQRi /M ∗ . IQRi is the interquartile range of subgroup
In practice, the same statistic is generally used to estimate i and is defined as the difference between the second-smallest
both the in-control standard deviation σ in Phase I and the stan- and the second-largest observation for 4 ≤ n ≤ 7, and as the
dard deviation λσ in Phase II. Since the requirements for the difference between the third-smallest and the third-largest ob-
estimators differ between the two phases, this is not always the servation for 8 ≤ n ≤ 11. The constant c is a tuning constant.
best choice. In Phase I, an estimator should be efficient in uncon- Each value of c leads to a different estimator. Tatum studied
taminated situations and robust against disturbances, whereas in the behavior of the estimator for c = 7 and c = 10 and showed
Phase II, the estimator should be sensitive to disturbances (cf. that c = 7 gives an estimator that loses some efficiency in the
Jensen et al. 2006). In this section, we present six Phase I es- absence of disturbances but gains efficiency in the presence of
timators, demonstrate the implementation of the estimators by disturbances. We apply this value of c in our simulation study.
means of a real data example, and assess the efficiency of the Note that we have h(i) = Ei − 3.5 for 4.5 < Ei ≤ 7.5 instead
estimators in terms of their MSE. of h(i) = Ei − 4.5, as presented by Tatum (1997, p. 129). This
was a typographical error, resulting in too much weight on local-
ized disturbances and thus an overestimation of σ . An unbiased
estimator of σ is given by Sc∗ /d ∗ (c, n, k), where d ∗ (c, n, k) is
2.1 Estimators of the Standard Deviation a normalizing constant. During the implementation of the es-
timator, we discovered that for odd values of n, the values of
Recall that Xij , i = 1, 2, . . . , k and j = 1, 2, . . . , n, denotes d ∗ (c, n, k) given by Table 1 in Tatum (1997) are incorrect. We
the Phase I data with n the subgroup size and k the number use the corrected values, which are presented in Table 1. The
of subgroups. resulting estimator is denoted by D7, as in Tatum (1997).
TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1
 A ROBUST STANDARD DEVIATION CONTROL CHART 75

Table 1. Normalizing constants d ∗ (c, n, k) for Tatum’s estimator (Sc∗ ) An unbiased estimator of σ is MD/t2 (n), where t2 (n) equals
E(MD/σ ). Since it is difficult to obtain E(MD) analytically, it
c=7 c = 10
is obtained by simulation. Extensive tables for t2 (n) can be found
n k = 20 k = 30 k = 40 k = 20 k = 30 k = 40 in Riaz and Saghir (2009). The advantage of this estimator is
that it is less sensitive to outliers than R (cf. Schoonhoven et al.
s
5 1.070 1.069 1.068 1.054 1.053 1.053 2011). The resulting estimator is denoted by MD . The values
7 1.057 1.056 1.056 1.041 1.040 1.040 used for the Phase I control limits and the constants necessary
9 1.052 1.051 1.050 1.034 1.034 1.033 to obtain unbiased estimates from the screened data are given in
11 1.047 1.046 1.046 1.029 1.029 1.028
Table 2. Both are obtained by simulation.
13 1.044 1.044 1.043 1.026 1.025 1.025
15 1.041 1.041 1.041 1.023 1.023 1.023
For subgroup control charts, only adaptive trimming meth-
ods based on the subgroup averages or subgroup standard de-
viations have been proposed in the literature so far. For diffuse
We also include other procedures to obtain σ̂ . The first is a outliers, however, an individuals chart should detect outliers
variant of Rocke (1989). Rocke’s procedure first estimates σ by more quickly. We therefore propose a screening method based
the mean subgroup range on an individuals chart. The algorithm first calculates the resid-
uals by subtracting the subgroup median from each observation
1
k
in the corresponding subgroup. This ensures that the variabil-
R̄ = Ri , (4) ity is measured within and not between subgroups. Next, an
k i=1
individuals chart of the residuals is constructed. The location
where Ri is the range of the ith subgroup. An unbiased es- of the chart (µ) is estimated by the mean of the subgroup me-
timator of σ under normality is R̄/d2 (n), where d2 (n) is the dians, which is zero because the subgroup medians have been
expected range of a random N (0, 1) subgroup of size n. Val- subtracted from the observations, and σ is estimated by MD.
ues of d2 (n) can be found in Duncan (1974, table M). Any For simplicity, the factors for the individuals chart are 3 and –3
subgroup that exceeds the Phase I control limits is deleted and (see Table 2). The residuals that fall outside the control limits
R̄ is recomputed from the remaining subgroups. Our approach is are excluded from the dataset. Then the procedure is repeated:
similar but continues until all subgroup ranges fall between the the median values of the adjusted subgroups are determined, the

= Un R̄/d2 (n) and
Phase I control limits. These are set at UCL residuals are calculated, and the control limits of the individuals

= Ln R̄/d2 (n). We derive the factors Un and Ln from the
LCL chart are computed. The residuals that now exceed the limits
0.99865 and 0.00135 quantiles of the distribution of R̄/d2 (n). are removed. This continues until all residuals fall within the
Table 2 shows the factors for n = 4, 5, 9 as well as the con- control limits. Simulation revealed that the resulting estimates
stants added to obtain unbiased estimates from the screened of σ are slightly biased under normality. The constants neces-
data. The factors as well as the constants are obtained by simu- sary to obtain an unbiased estimate can be found in Table 2. The
i
lation. Note that the factors and the constants are the same for unbiased estimator is denoted by MD .
k = 20, 50, 100. The resulting estimator is denoted by R̄ s . The above procedure does not use the spread of the subgroups.
In addition, we evaluate an adaptive trimmer where the esti- Therefore, we finally propose an algorithm that combines the
mate of σ is obtained by the mean subgroup average deviation use of an individuals chart with subgroup screening. First, an
from the median instead of R̄. The mean subgroup average de- initial estimate of σ is obtained via MD (see (5)). This estimate is
viation from the median is given by then used to construct a standard deviation control chart so that
the subgroups can be screened. Adopting R as a charting statistic
1
k
will result in the exclusion of many subgroups, including many
MD = MDi , (5) uncontaminated observations, when diffuse disturbances are
k i=1
present. For this reason, we employ IQR for screening pur-
where MDi is the average absolute deviation from the median poses. The constants required to obtain an unbiased estimate of
Mi of subgroup i defined by σ based on IQR are 0.594 for n = 4, 0.990 for n = 5, and 1.144
for n = 9. The values chosen for the Phase I control limits are

n
presented in Table 2. The subgroup screening is continued until
MDi = |Xij − Mi |/n.
all IQRs fall within the limits. The resulting estimates of σ are
j =1

Table 2. Factors for Phase I control limits for k = 20, 50, 100

n=4 n=5 n=9

Chart Un Ln Constant Un Ln Constant Un Ln Constant

R̄ s 2.321 0.170 1 2.305 0.172 1 1.950 0.330 1

s
MD 2.321 0.170 0.998 2.305 0.172 1 1.950 0.330 1
i
MD 3 –3 0.990 3 –3 0.975 3 –3 0.986
i,s
MD 4.703 0.0018 1 3.225 0.035 1 2.485 0.142 1
3 –3 0.988 3 –3 0.975 3 –3 0.986

TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1

76 MARIT SCHOONHOVEN AND RONALD J.M.M. DOES

Table 3. Melt index measurements

Sample Observations R/d2 (4) S/c4 (4) MD/t2 (4) IQR/dIQR (4)

1 218 224 220 231 6.31 6.23 6.46 6.73

2 238 236 247 234 6.31 6.23 5.70 3.37
3 280 228 228 221 28.65 29.70 22.42 0
4 210 249 241 246 18.94 19.51 16.72 8.42
5 243 240 230 230 6.31 7.33 8.74 16.84
6 225 250 258 244 16.03 15.26 14.82 10.10
7 240 238 240 243 2.43 2.24 1.90 0
8 244 248 265 234 15.06 14.02 13.30 6.73
9 238 233 252 243 9.23 8.80 9.12 8.42
10 228 238 220 230 8.74 8.03 7.60 3.37
11 218 232 230 226 6.80 6.72 6.84 6.73
12 226 231 236 242 7.77 7.43 7.98 8.42
13 224 221 230 222 4.37 4.38 4.18 3.37
14 230 220 227 226 4.86 4.55 4.18 1.68
15 224 228 226 240 7.77 7.80 6.84 3.37
16 232 240 241 232 4.37 5.35 6.46 13.47
17 243 250 248 250 3.40 3.59 3.42 3.37
18 247 238 244 230 8.26 8.14 8.74 10.10
19 224 228 228 246 10.68 10.69 8.36 0
20 236 230 230 232 2.91 3.07 3.04 3.37

unbiased and are used to screen observations with an individuals 24.78 and –24.78. One residual in subgroup 3 and one residual
i
control chart (the procedure used to derive MD ). Simulation in subgroup 4 fall outside the control limits. The corresponding
revealed that the final estimates of σ are slightly biased. The observations are deleted from the dataset. The subgroup me-
constants necessary to obtain an unbiased estimate can be dians are determined from the remaining observations and the
i,s
found in Table 2. The unbiased estimator is denoted by MD . residuals are recalculated. The second estimate of σ is 6.82 and
the control limits are now 20.47 and –20.47. One residual in
2.2 Real Data Example subgroup 6 falls below the lower control limit and so one obser-
vation is removed. Again, the medians are determined from the
In this section, we demonstrate the estimation of σ in Phase I. remaining observations and the residuals are recomputed. The
Our dataset was supplied by Wadsworth, Stephens, and Godfrey third estimate of σ is 6.49 and the control chart has limits at
(2001, pp. 235–237). The operation concerns the melt index of 19.47 and –19.47. There are now no residuals that fall outside
a polyethylene compound. The data consist of 20 subgroups of the control limits. The resulting unbiased estimate is 6.55.
size 4 (Table 3).
The factors used for the n = 4, k = 20 case are presented in
Table 4. Residuals of melt index measurements
Table 2. Note that d2 (4) = 2.06, c4 (4) = 0.92, t2 (4) = 0.66, and
dIQR (4) = 0.59. The estimates of σ obtained by S̃ and D7 are Sample Residuals
determined in one iteration and are 10.14 and 6.59, respec-
s
tively. The values obtained by R̄ s and MD incorporate subgroup 1 −4.0 2.0 −2.0 9.0
screening. The initial value of R̄ s is 8.96 and the respective up- 2 1.0 −1.0 10.0 −3.0
per and lower control limits are 20.80 and 1.52. The unbiased 3 52.0 0.0 0.0 −7.0
estimate of the range (i.e., R̄/d2 (4)) of subgroup 3 falls above 4 −33.5 5.5 −2.5 2.5
the control limit and so this subgroup is deleted. The second 5 8.0 5.0 −5.0 −5.0
6 −22.0 3.0 11.0 −3.0
estimate of R̄ s equals 7.92 and the corresponding Phase I upper
7 0.0 −2.0 0.0 3.0
and lower control limits are 18.38 and 1.35. Now subgroup 4
8 −2.0 2.0 19.0 −12.0
does not meet the Phase I upper control limit and is removed. 9 −2.5 −7.5 11.5 2.5
The third estimate of R̄ s is 7.31 and the control limits are 16.97 10 −1.0 9.0 −9.0 1.0
and 1.24. There are no further subgroups whose R/d2 (4) ex- 11 −10.0 4.0 2.0 −2.0
ceeds the upper control limit. The resulting unbiased estimate 12 −7.5 −2.5 2.5 8.5
s
of σ is 7.31. The MD procedure works in a similar way. In this 13 1.0 −2.0 7.0 −1.0
case, subgroups 3 and 4 are again deleted. The final unbiased 14 3.5 −6.5 0.5 −0.5
estimate is 7.03. 15 −3.0 1.0 −1.0 13.0
i
For the MD chart, we use a procedure based on the individ- 16 −4.0 4.0 5.0 −4.0
uals control chart for the residuals. The residuals are calculated 17 −6.0 1.0 −1.0 1.0
18 6.0 −3.0 3.0 −11.0
by subtracting the subgroup median from each observation in
19 −4.0 0.0 0.0 18.0
the corresponding subgroup (see Table 4). The initial value of
20 5.0 −1.0 −1.0 1.0
σ is 8.26 and the control limits of the individuals chart are
TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1
 A ROBUST STANDARD DEVIATION CONTROL CHART 77

Table 5. Summary of estimates of σ and data deletions where σ̂ i is the value of the unbiased estimate in the ith simula-
tion run and N is the number of simulation runs. Comparisons are
Chart σ̂ Deleted subgroup Deleted observation made under normality and four types of disturbances (cf. Tatum
S̃ 10.14 1997), but with an error rate of 6% in each case. In general, we
D7 6.59 expect that a higher error rate would result in more pronounced
R̄ s 7.31 3; 4 differences between the estimators. The four disturbances are
s
MD 7.03 3; 4 captured in:
i
MD 6.55 3:1; 4:1; 6:1
i,s 1. A model for diffuse symmetric disturbances in which
MD 6.87 3; 7; 19 4:1; 6:1
each observation has a 94% probability of being drawn
from the N (0, 1) distribution and a 6% probability of
being drawn from the N (0, a) distribution, with a =
i,s
For the MD chart, the first part of the procedure screens the 1.5, 2.0, . . . , 5.5, 6.0.
subgroup IQR. The respective upper and lower control limits 2. A model for diffuse asymmetric variance disturbances in
of the IQR chart are 38.86 and 0.015. The IQR of subgroups which each observation is drawn from the N (0, 1) dis-
3, 7, and 19 are 0 and so these subgroups are deleted. It is tribution and has a 6% probability of having a multiple
not necessary to delete any further subgroups. Next, individual of a χ12 variable added to it, with the multiplier equal to
observations are screened. The estimate of σ is 7.81 and the 0.5, 1.0, . . . , 4.5, 5.0.
upper and lower control limits for the residuals are 23.45 and 3. A model for localized variance disturbances in which all
–23.45. An outlier in subgroup 4 is deleted. The next estimate observations in three (when k = 50) or six (when k = 100)
of σ is 7.18 with corresponding control limits 21.55 and –21.55. subgroups are drawn from the N (0, a) distribution, with
The outlier in subgroup 6 is removed from the dataset. Now σ is a = 1.5, 2.0, . . . , 5.5, 6.0.
set at 6.79 with corresponding control limits 20.37 and –20.37. 4. A model for diffuse mean disturbances in which each
No further deletions are required. The unbiased estimate for the observation has a 94% probability of being drawn
i,s
MD chart is 6.87. from the N (0, 1) distribution and a 6% probability of
The final estimates for σ as well as the data deletions are being drawn from the N (b, 1) distribution, with b =
presented in Table 5. The estimate based on S̃ is higher than the 0.5, 1.0, . . . , 9.0, 9.5.
other estimates. This is because S̃ is more sensitive to outliers
The MSE is obtained for n = 5, 9 and k = 50, 100. The num-
than the other estimators. Note, however, that the question of
ber of simulation runs N is equal to 50,000. Note that Tatum
which estimator has done the best job cannot be resolved from
(1997) used 10,000 simulation runs. Below we only present
such a limited dataset.
the results for k = 50 because the conclusions for k = 100 are
very similar. The figures comparing k = 50 and k = 100 are
2.3 Efficiency of the Proposed Estimators
available as Supplementary Material.
To evaluate Phase I performance, we now assess the MSE of Figure 1 shows the MSE values when diffuse symmetric vari-
the proposed Phase I estimators. The MSE is estimated as ance disturbances are present. The y-intercepts show that the
pooled standard deviation (S̃) has the lowest MSE when no dis-
turbances are present. However, when the size of the disturbance
1  i
N
MSE = (σ̂ − σ )2 , (6) (a) increases, the MSE increases quickly. The other estimators
N i=1 are more robust against outliers of this type. Those that use an

0.025 0.035
S̃ S̃
D7 D7
R̄s 0.03 R̄s
s s
0.02 M¯D M¯D
i i
M¯D M¯D
i,s i,s
M¯D 0.025 M¯D

0.015
0.02
MSE

MSE

0.015
0.01

0.01
0.005
0.005

0 0
1 2 3 4 5 6 1 2 3 4 5 6
a a

(a) (b)

Figure 1. MSE of estimators when symmetric diffuse variance disturbances are present: (a) n = 5, k = 50 and (b) n = 9, k = 50.
TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1
78 MARIT SCHOONHOVEN AND RONALD J.M.M. DOES

0.025 0.02
S̃ S̃
D7 0.018 D7
R̄s R̄s
s s
0.02 M¯D 0.016 M¯D
i i
M¯D M¯D
i,s i,s
M¯D 0.014 M¯D

0.015 0.012
MSE

MSE
0.01

0.01 0.008

0.006

0.005 0.004

0.002

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Multiplier Multiplier
(a) (b)

Figure 2. MSE of estimators when asymmetric diffuse variance disturbances are present: (a) n = 5, k = 50 and (b) n = 9, k = 50.

individuals control chart to identify individual outliers, that are, Figure 3 shows the results in situations with localized dis-
i i,s
MD and MD , coincide and perform best, followed by D7. The turbances. The estimators incorporating subgroup screening
s i,s
estimators based on only subgroup screening, namely R̄ s and (R̄ s and MD ) perform best. The estimator MD performs bet-
s i
MD , turn out to perform less well in this situation. The reason ter than D7 in this situation. Finally, MD does not perform as
is that they screen subgroup dispersion and ignore individual well in this case because it does not take into account informa-
s
outliers. Note that R̄ s falls far short of MD , because R̄ s uses tion on the subgroup spread.
R̄ (rather than MD) to estimate σ . As R̄ is more sensitive to out- The results for the fourth type of disturbance are shown in
liers, the Phase I limits are broader, making it more difficult to Figure 4. We can conclude that S̃ and R̄ s coincide for n = 9 and
s
detect outliers. This effect is particularly significant for n = 9, perform far worse than the other estimators. MD performs
because a larger subgroup is more likely to be infected with better but not as well as D7 and not as well as the estimators
an outlier. using an individuals chart to identify individual outliers. The
s
When asymmetric diffuse disturbances are present (Figure 2), reason is that MD is less capable of detecting such outliers.
i,s i
the results are comparable to the situation with diffuse symmet- The estimators MD and MD coincide and perform best in
i i,s
ric disturbances: MD and MD coincide and perform best, this situation.
s
followed by D7 and MD . Note that in this situation, S̃ and, for Out main conclusion from the above results is that the esti-
n = 9, R̄ perform badly. i,s
mator MD performs best overall.

0.008
S̃ S̃
0.016 D7 D7
R̄s 0.007 R̄s
s s
M¯D M¯D
i i
0.014 M¯D 0.006 M¯D
i,s i,s
M¯D M¯D
0.012
0.005
0.01
MSE

MSE

0.004
0.008
0.003
0.006
0.002
0.004

0.002 0.001

0 0
1 2 3 4 5 6 1 2 3 4 5 6
a a

(a) (b)

Figure 3. MSE of estimators when localized variance disturbances are present: (a) n = 5, k = 50 and (b) n = 9, k = 50.

TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1

 A ROBUST STANDARD DEVIATION CONTROL CHART 79

0.2 0.08
S̃ S̃
0.18 D7 D7
R̄s 0.07 R̄s
s s
0.16 M¯D M¯D
i i
M¯D 0.06 M¯D
i,s i,s
0.14 M¯D M¯D
0.05
0.12
MSE

MSE
0.1 0.04

0.08
0.03
0.06
0.02
0.04
0.01
0.02

0 0
0 2 4 6 8 10 0 2 4 6 8 10
b b

(a) (b)

Figure 4. MSE of estimators when diffuse mean disturbances are present: (a) n = 5, k = 50 and (b) n = 9, k = 50.

3. DERIVATION OF THE PHASE II CONTROL LIMITS the MSE with a, b and the multiplier equal to 4 to simulate the
contaminated cases (see Section 2.3).
Equation (1) gives control limits for the standard deviation The performance of the control charts is assessed in terms
control chart with σ estimated in Phase I. We estimate λσ in of the unconditional ARL and SDRL. We compute these run
Phase II by S/c4 (n) for all charts. One of the criteria used to length characteristics in an in-control situation and several out-
assess Phase II performance is the ARL. To allow comparison, of-control situations. We consider different shifts in the standard
Un and Ln are chosen such that the unconditional ARL equals deviation λσ , setting λ equal to 0.6, 1, 1.2, and 1.4. The per-
370 and, for each chart, the ARLs for the upper and lower formance characteristics are obtained by simulation. The next
control limits are similar. Un and Ln cannot be obtained easily section describes the simulation method, followed by the re-
in analytic form and are obtained from 50,000 simulation runs. sults for the control charts constructed in the uncontaminated
Table 6 presents Un and Ln for n = 5, 9 and k = 50, 100. situation and various contaminated situations.

4. CONTROL CHART PERFORMANCE 4.1 Simulation Procedure

We now evaluate the effect of the proposed estimators on the For each Phase I dataset of k subgroups of size n, we determine
σ̂ and the control limits UCL
and LCL.
Let σ̂i be an estimate
Phase II performance of the standard deviation control chart.
We consider the same Phase I estimators as those used to asses of λσ based on the ith subgroup Yij , j = 1, 2, . . . , n. Further,

or below LCL.
let Fi denote the event that σ̂i is above UCL
We
define P (Fi |σ̂ ) as the probability that subgroup i generates a
Table 6. Factors Un and Ln to determine Phase II
signal given σ̂ , that is,
control limits for n = 5 and n = 9
 

or σ̂i > UCL|
P (Fi |σ̂ ) = P σ̂i < LCL
σ̂ .

k = 50 k = 100 Given σ̂ , the distribution of the run length is geometric with

parameter P (Fi |σ̂ ). Consequently, the conditional ARL is
n σ̂ Un Ln Un Ln
given by
5 S̃ 2.230 0.163 2.236 0.169 1
D7 2.225 0.162 2.236 0.167 E(RL|σ̂ ) = .
P (Fi |σ̂ )
R̄ s 2.226 0.163 2.236 0.168
s
MD 2.226 0.162 2.237 0.167 When we take the expectation over the Xij ’s we get the uncon-
i
MD 2.217 0.160 2.333 0.166 ditional ARL
i,s
MD 2.217 0.160 2.333 0.166 1
ARL = E .
9 S̃ 1.832 0.343 1.835 0.347 P (Fi |σ̂ )
D7 1.830 0.341 1.834 0.346
R̄ s 1.831 0.342 1.835 0.347
This expectation is obtained by simulation: numerous datasets
MD
s
1.830 0.342 1.835 0.346
are generated from the normal distribution or contami-
i nated normal distribution, and for each dataset, E(RL|σ̂ ) is
MD 1.829 0.341 1.833 0.346
i,s computed. By averaging these values, we obtain the un-
MD 1.829 0.341 1.833 0.346
conditional value. The unconditional standard deviation is

TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1

80 MARIT SCHOONHOVEN AND RONALD J.M.M. DOES

Table 7. ARL and SDRL under normality for k = 50

ARL SDRL

n Chart λ = 0.6 λ = 1.0 λ = 1.2 λ = 1.4 λ = 0.6 λ = 1.0 λ = 1.2 λ = 1.4

5 S̃ 131 378 69.5 17.5 136 412 87.0 19.6

D7 135 373 69.6 17.4 141 414 90.0 20.0
R̄ s 132 369 68.9 17.3 137 407 88.4 19.7
s
MD 135 375 69.7 17.4 141 414 90.0 20.0
i
MD 143 371 70.0 17.3 151 421 94.6 20.6
i,s
MD 143 371 69.9 17.4 151 421 95.0 20.7
9 S̃ 28.4 371 43.6 9.02 29.3 392 51.6 9.30
D7 29.5 373 43.9 9.06 30.8 397 53.5 9.50
R̄ s 29.2 392 43.9 9.05 30.5 392 53.9 9.52
s
MD 29.1 369 43.9 9.02 30.3 393 53.4 9.45
i
MD 29.8 368 44.7 9.11 31.4 395 56.0 9.71
i,s
MD 29.9 366 44.4 9.05 31.6 394 56.4 9.65

determined by: dard errors for ARL. The relative standard error never
exceeds 0.76%.
 
SDRL = var(RL) = E(var(RL|σ̂ )) + var(E(RL|σ̂ ))

 2  2
1 1 1 4.2 Simulation Results
= 2E − E −E .
p(Fi |σ̂ ) p(Fi |σ̂ ) p(Fi |σ̂ )
The ARL and SDRL are obtained in the in-control situa-
tion (λ = 1) and in the out-of-control situation (λ
= 1). When
Enough replications of the above procedure were per- the process is in control, we want the ARL and SDRL to be
formed to obtain sufficiently small relative estimated stan- close to their intended values, namely 370. In the out-of-control

Table 8. ARL and SDRL when contaminations are present in Phase I for k = 50

ARL SDRL

n Chart λ = 0.6 λ = 1.0 λ = 1.2 λ = 1.4 λ = 0.6 λ = 1.0 λ = 1.2 λ = 1.4

N (0, 1) 5 S̃ 43.9 297 425 303 50.8 337 452 391

& D7 102 484 159 34.7 108 498 215 47.2
N (0, 4) R̄ s 92.2 464 206 50.1 99.9 479 286 87.2
s
(symm) MD 101 474 171 38.2 108 492 171 58.6
i
MD 123 443 114 25.3 131 479 164 34.2
i,s
MD 122 446 114 25.6 131 481 164 34.6
9 S̃ 5.47 114 317 286 5.92 148 354 351
D7 19.9 433 109 17.0 20.8 440 144 20.1
R̄ s 14.3 336 238 47.5 16.2 368 311 94.5
s
MD 18.9 410 128 19.6 20.3 421 178 26.3
i
MD 23.8 420 75.2 12.9 25.3 433 102 15.0
i,s
MD 23.7 421 75.9 13.0 25.2 434 103 15.0
N (0, 1) 5 S̃ 23.2 149 231 266 38.4 243 325 355
& D7 112 461 121 26.9 118 483 164 34.2
N (0, 4) R̄ s 108 450 133 29.9 115 472 191 43.6
s
(asymm) MD 115 449 119 26.6 121 475 168 35.7
i
MD 130 419 92.7 21.7 138 461 131 27.5
i,s
MD 130 422 95.0 21.9 138 463 135 27.9
9 S̃ 2.34 33.3 93.6 165 3.27 79.5 186 263
D7 22.7 435 80.0 13.5 23.7 443 104 15.2
R̄ s 18.8 399 140 22.6 20.6 415 207 39.7
s
MD 22.5 418 83.3 14.0 23.9 429 116 16.7
i
MD 26.0 407 61.1 11.2 27.5 425 80.5 12.6
i,s
MD 25.7 409 61.7 11.3 27.3 426 81.4 12.7

TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1

 A ROBUST STANDARD DEVIATION CONTROL CHART 81

Table 9. ARL and SDRL when contaminations are present in Phase I for k = 50

ARL SDRL

n Chart λ = 0.6 λ = 1.0 λ = 1.2 λ = 1.4 λ = 0.6 λ = 1.0 λ = 1.2 λ = 1.4

N (0, 1) 5 S̃ 42.8 293 436 308 48.3 330 459 390

& D7 118 442 103 23.5 124 469 139 28.7
N (0, 4) R̄ s 129 384 77.1 18.7 134 422 103 22.2
s
(loc) MD 131 391 78.5 19.0 137 431 106 22.9
i
MD 127 428 100 22.9 136 468 143 29.5
i,s
MD 135 404 85.9 20.4 144 451 122 25.7
9 S̃ 5.34 110 324 295 5.55 136 357 254
D7 24.3 427 67.9 12.1 25.4 438 86.9 13.3
R̄ s 28.8 372 46.0 9.31 30.2 396 57.8 9.93
s
MD 28.6 372 46.0 9.31 30.0 396 57.5 9.91
i
MD 23.9 420 74.4 12.8 25.5 433 100 14.7
i,s
MD 28.9 377 49.1 9.68 30.7 404 64.4 10.6
N (0, 1) 5 S̃ 40.2 280 470 335 42.8 300 480 395
& D7 80.3 471 286 72.8 86.7 483 359 117
N (4, 1) R̄ s 54.1 358 431 207 61.1 384 466 292
s
MD 65.0 409 394 144 73.0 431 450 220
i
MD 115 447 152 64.4 127 481 237 64.4
i,s
MD 115 449 152 63.5 127 483 234 63.5
9 S̃ 5.04 101 325 321 4.87 114 348 363
D7 14.4 358 230 35.1 15.3 380 291 51.4
R̄ s 5.59 117 341 292 5.84 144 367 353
s
MD 9.53 231 370 99.0 10.4 266 404 151
i
MD 22.8 409 93.6 15.2 25.0 425 142 20.9
i,s
MD 22.8 410 92.1 15.2 25.0 425 140 21.0

situation, we want to detect changes in the standard deviation the presence of any contamination, the chart outperforms the
as soon as possible, so the ARL should be as low as possible. alternatives.
The results are obtained for n = 5, 9 and k = 50, 100. How-
ever, as was done for the MSE comparison, we only present the
results for k = 50 because the conclusions for k = 100 are not 5. CONCLUDING REMARKS
all that different. The tables comparing k = 50 and k = 100 are
available as Supplementary Material. In this article, we consider several estimators of the stan-
Table 7 shows the ARL and SDRL for the situation when dard deviation in Phase I of the control charting process. We
the Phase I data are uncontaminated and normally distributed. have found that the performance of certain robust estimators
The ARL is very similar across charts and the SDRL is slightly is almost identical to the pooled subgroup standard deviation
i i,s
higher for the MD and MD charts. under normality, while the benefit of using such robust estima-
Tables 8 and 9 show that when there are disturbances in tors can be substantial when there are disturbances. Following
the Phase I data, the ARL values increase (decrease) consider- Rocke (1989, 1992), we have considered estimators that include
ably for λ > 1 (λ < 1) relative to the normal situation. Thus, a procedure for subgroup screening, but whereas Rocke used R̄,
when the Phase I data are contaminated, changes in the pro- we have used the average deviation from the median. This es-
cess standard deviation are less likely to be detected when timator performs better when there are localized disturbances
λ > 1, while there are more signals when λ < 1. With diffuse and is much more robust against diffuse disturbances. However,
disturbances (Table 8 and second half of Table 9), their im- when there are diffuse mean disturbances, the procedure loses
i i,s efficiency.
pact is smallest for the charts based on MD , MD , and D7.
To address this problem, we have proposed other algorithms
When there are localized disturbances (first half of Table 9),
s i,s based on a procedure that also screens for individual outliers.
the charts based on R̄ s , MD , and MD perform best, because The algorithms remove the variation between subgroups so that
these charts trim extreme subgroups. Note that in a number only the variation within subgroups is measured. We have shown
s
of cases, the S̃, R̄ s , and MD charts are ARL-biased: the in- that these algorithms are very effective when there are diffuse
control ARL is lower than the out-of-control ARL (cf. Jensen disturbances. When there might also be localized disturbances,
et al. 2006). the method can be combined with subgroup screening based on
i,s
Overall, the MD chart performs best. Under normality, this the IQR. The latter procedure reveals a performance very similar
chart almost matches the standard chart based on S̃, and in to the robust estimator for the standard deviation control chart

TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1

82 MARIT SCHOONHOVEN AND RONALD J.M.M. DOES

proposed by Tatum (1997). We think that this is a noteworthy Jensen, W. A., Jones-Farmer, L. A., Champ, C. W., and Woodall, W. H. (2006),
outcome since the procedure is simple and intuitive. Moreover, “Effects of Parameter Estimation on Control Chart Properties: A Literature
Review,” Journal of Quality Technology, 38, 349–364. [73,74,81]
it can be used to estimate σ in other practical applications. Mahmoud, M. A., Henderson, G. R., Epprecht, E. K., and Woodall, W. H. (2010),
“Estimating the Standard Deviation in Quality-Control Applications,” Jour-
nal of Quality Technology, 42, 348–357. [74]
SUPPLEMENTARY MATERIAL Quesenberry, C. P. (1993), “The Effect of Sample Size on Estimated Lim-
its for X and X Control Charts,” Journal of Quality Technology, 25,
Figures S1– S4 and Tables S1– S3. 237–247. [73]
Riaz, M., and Saghir, A. (2009), “A Mean Deviation-Based Approach to Monitor
Process Variability,” Journal of Statistical Computation and Simulation, 79,
ACKNOWLEDGMENTS 1173–1193. [75]
Rocke, D. M. (1989), “Robust Control Charts,” Technometrics, 31, 173–184.
The authors are very grateful to the Editor, the Associate [73,74,81]
Editor, and a referee for their helpful comments, which led to —— (1992), “X̄Q and RQ Charts: Robust Control Charts,” The Statistician, 41,
97–104. [73,81]
improvements in this article. Schoonhoven, M., Riaz, M., and Does, R. J. M. M. (2011), “Design and Anal-
ysis of Control Charts for Standard Deviation With Estimated Paramaters,”
[Received November 2010. Revised October 2011.] Journal of Quality Technology, 44, 307–333. [74,75]
Shewhart, W. A. (1931). Economic Control of Quality of Manufactured Product,
Princeton, NJ: Van Nostrand. [73]
REFERENCES Tatum, L. G. (1997), “Robust Estimation of the Process Standard Deviation for
Control Charts,” Technometrics, 39, 127–141. [73,74,77,81]
Chen, G. (1998), “The Run Length Distributions of the R, S and s 2 Con- Wadsworth, H. M., Stephens, K. S., and Godfrey, A. B. (2001), Modern
trol Charts When σ Is Estimated,” Canadian Journal of Statistics, 26, Methods for Quality Control and Improvement (2nd ed.), New York:
311–322. [73] Wiley. [76]
Davis, C. M., and Adams, B. M. (2005), “Robust Monitoring of Contaminated Woodall, W.H., and Montgomery, D.C. (1999), “Research Issues and Ideas
Data,” Journal of Quality Technology, 37, 163–174. [73] in Statistical Process Control,” Journal of Quality Technology, 31,
Duncan, A.J. (1974), Quality Control and Industrial Statistics (4th ed.), Home- 376–386. [73]
wood, IL: R.D. Irwin. [75] Yang, C., and Hillier, F. (1970), “Mean and Variance Control Chart Limits Based
Hillier, F. (1969), “X and R Chart Control Limits Based on a Small Number of on a Small Number of Subgroups,” Journal of Quality Technology, 2, 9–
Subgroups,” Journal of Quality Technology, 1, 17–26. [73] 16. [73]

TECHNOMETRICS, FEBRUARY 2012, VOL. 54, NO. 1