Quality Engineering

ISSN: 0898-2112 (Print) 1532-4222 (Online) Journal homepage: [Link]

Nonparametric (distribution-free) control charts:

An updated overview and some results

S. Chakraborti & M. A. Graham

To cite this article: S. Chakraborti & M. A. Graham (2019): Nonparametric (distribution-

free) control charts: An updated overview and some results, Quality Engineering, DOI:
10.1080/08982112.2018.1549330

To link to this article: [Link]

Published online: 03 May 2019.

Submit your article to this journal

View Crossmark data

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

[Link]
QUALITY ENGINEERING
[Link]

Nonparametric (distribution-free) control charts: An updated overview and

some results
S. Chakrabortia and M. A. Grahamb
a
Department of Information Systems, Statistics, and Management Science, University of Alabama, Tuscaloosa, Alabama; bDepartment
of Science, Mathematics and Technology Education, University of Pretoria, Pretoria, South Africa

ABSTRACT KEYWORDS
Control charts that are based on assumption(s) of a specific form for the underlying process CUSUM chart; EWMA chart;
distribution are referred to as parametric control charts. There are many applications where median; multivariate; Phase I;
there is insufficient information to justify such assumption(s) and, consequently, control Phase II; precedence and
exceedance statistics; rank;
charting techniques with a minimal set of distributional assumption requirements are in robust; runlength; Shewhart
high demand. To this end, nonparametric or distribution-free control charts have been pro- chart; sign; univariate
posed in recent years. The charts have stable in-control properties, are robust against out-
liers and can be surprisingly efficient in comparison with their parametric counterparts.
Chakraborti and some of his colleagues provided review papers on nonparametric control
charts in 2001, 2007 and 2011, respectively. These papers have been received with consider-
able interest and attention by the community. However, the literature on nonparametric
statistical process/quality control/monitoring has grown exponentially and because of this
rapid growth, an update is deemed necessary. In this article, we bring these reviews forward
to 2017, discussing some of the latest developments in the area. Moreover, unlike the past
reviews, which did not include the multivariate charts, here we review both univariate and
multivariate nonparametric control charts. We end with some concluding remarks.

Introduction der Laan, and Bakir (2001) provided a thorough

account of univariate nonparametric control charting
Modern statistical process control and monitoring
literature up to the end of the year 2000. Chakraborti
methods include nonparametric (or distribution-free)
control (NSPC) charts which provide a more robust and Graham (2007) updated that review covering
alternative to standard parametric (e.g., normal the- much of the literature up until the end of 2007.
ory) charts when the form of the underlying process Following that, Chakraborti, Human, and Graham
distribution is unknown. A key advantage of nonpara- (2011) gave an updated overview of the nonparamet-
metric control charts is that one does not need to ric control charting literature from 2007 up until the
make very specific model (shape) assumptions such as end of the year 2010. The interested reader can also
normality and their in-control (IC) run-length distri- see the short book by Bakir (2011) for some newer
bution remains the same for all continuous process proposals. More recently, Qiu (2014) devoted two
distributions (of all shapes) under minimal assump- chapters to NSPC and reviewed some of the available
tions (such as continuity and symmetry). This prop- methods. However, the field of research in NSPC con-
erty ensures the full knowledge and the stability of the tinues to grow at a rapid pace and many new contri-
IC properties of the chart, which are crucial for chart butions have been added in the last few years. Our list
implementation and application in practice. This is of references will show that since 2010, at least 50
not the case for parametric charts whose IC run- papers have been published in the area, which is truly
length properties are exact only under the assumed remarkable. Thus, it seems to be a good point in time
distribution and departures from the assumed distri- to take account of what has happened in the recent
bution can affect the chart performance, often in a past. Motivated by this, our goal in this review is to
significantly negative way (such as too many false bring the reviews of Chakraborti et al. (2001, 2007
alarms). Recognizing the potential, Chakraborti, Van and 2011) forward to the main body of the published

CONTACT S. Chakraborti schakrab@[Link] Department of Information Systems, Statistics, and Management Science, University of Alabama,
Tuscaloosa, AL, USA.
ß 2019 Taylor & Francis Group, LLC
2 S. CHAKRABORTI AND M. A. GRAHAM

literature on nonparametric control charts to date. charts, called self-starting charts, that do not require
Note that for better coverage and more completeness, such a strict demarcation, where process monitoring
we also review a few of the articles that were not cov- can begin as soon as some amount of “trial” (training)
ered in Chakraborti et al. (2011). In addition, we data become available, but we mainly focus on Phase
cover the multivariate nonparametric control charting I and Phase II charts in this review.
literature that was not covered in any of the earlier Nonparametric charts are useful in both of these
reviews. This should be of interest to researchers as phases with perhaps more utility in Phase I where no
well as practitioners. knowledge about the distribution seems to be typically
available. However, note that although one might sug-
gest that in Phase II we have more knowledge about
A review of the literature
the process from the Phase I analysis which we can
For a review of the literature we follow the same for- leverage from, such as by fitting (estimating) a para-
mat of presentation as in Chakraborti et al. (2001) for metric distribution and using it as the model, we
consistency, so the charts are classified according to argue that in fact one often does not have the assur-
the three main categories, namely, Shewhart, cumula- ance of “knowing” a distribution even after a detailed
tive sum (CUSUM) and exponentially weighted mov- Phase I analysis, since conceivably processes can
ing average (EWMA). Within each category, we change during monitoring. Moreover, if one considers
discuss some of the key contributions in each of the a fitted model in Phase II, the uncertainty associated
two important phases (stages) namely Phase I and with the estimated model (estimated parameters and
Phase II as far as monitoring applications are con- the form) from Phase I will need to be accounted for.
cerned. Note that as mentioned earlier, nonparametric Thus NSPC is perhaps a safer bet and a reasonable
charts are also known as distribution-free charts as way to approach process monitoring in practice.
they can be applied without a specific (shape) para- The keys to nonparametric control charts are the
metric model assumption (e.g., normal) about the nonparametric statistical methods such as the tests
underlying distribution. and confidence intervals. Nonparametric methods are
typically based on order statistics, ranks and various
functions of them. The charting statistic is often
Univariate nonparametric process monitoring
selected adapting a distribution-free test statistic,
We begin with the univariate nonparametric charts. meaning that its IC distribution does not depend on
Consider the case of variables data where a continu- the underlying distribution. So these charts are IC
ous random variable, the quality characteristic of robust by definition. One can then construct control
interest, is monitored with respect to its distribution charts based on these statistics and their statistical
or distributional parameters, such as the mean, properties using a classical control charting paradigm
median or spread. There are generally two phases in such as the Shewhart, the CUSUM and the EWMA
this monitoring process. One, where the value(s) of (and their enhancements). In this sense, control charts
the parameter(s) of interest is (are) specified in the IC can be viewed as graphics which are functions of suit-
case, known as the known parameter case (denoted able charting statistics, where the charting statistics
Case K) and two, where values are unknown or depend are model dependent in the parametric case
unspecified, and thus need to be estimated before pro- and are distribution-free in the nonparametric case.
cess monitoring can start. The latter is known as the For example, in the normal theory case, we typically
unknown parameter case (denoted Case U). In Case use the sample means and the standard deviations as
K, process monitoring can start as soon as data the charting statistics. In the nonparametric case, we
become available, but in Case U, since one needs to use distribution-free statistics, based on, for example,
estimate the unknown parameter(s), a preliminary or the signs and the ranks, as will be seen next.
retrospective or Phase I analysis is performed so that
the process can be brought under control and from
Univariate nonparametric process monitoring:
the resulting IC process data, called the reference
Phase I charts
data, process parameters can be estimated, control
chart limits can be calculated and then prospective The importance of a proper and effective Phase I ana-
monitoring of incoming, new or test data can start. lysis has been recognized in the literature (see
This phase of the monitoring process is called Phase Chakraborti, Human, and Graham 2009; Capizzi and
II monitoring. Note that there are some control Masarotto 2013; Jones-Farmer et al. 2014). Typically
 QUALITY ENGINEERING 3

Shewhart-type charts have been recommended in which is the multivariate generalization of the univari-
Phase I since these charts are simple and can detect ate Phase I mean-rank chart. Graham, Human, and
larger shifts quickly. A Phase I control charting prob- Chakraborti (2010) proposed a nonparametric Phase I
lem bears similarities with what is known as the test median chart that was also covered in the review of
of homogeneity or the k-sample problem in the statis- Chakraborti et al. (2011) and is thus not repeated
tical hypothesis testing literature. Among the available here. While these charts are useful, one potential limi-
Phase I charts a nonparametric Phase I Shewhart-type tation is that they are not directly usable with individ-
chart, called the mean-rank chart, was proposed by ual data and there are constraints on the subgroup
Jones-Farmer, Jordan, and Champ (2009). This chart size and the number of subgroups. Along these lines,
is based on the well-known Kruskal-Wallis nonpara- Capizzi and Masarotto (2013) proposed a distribu-
metric test (see Gibbons and Chakraborti 2010). We tion-free strategy for detecting shifts in process loca-
start off by combining the observations from the m tion and/or scale in Phase I, called the recursive
Phase I samples in a single pooled sample of size segmentation and permutation (RS/P) procedure. This
N ¼ m  n; ordering the observations from the lowest methodology is based on a time-ordered segmentation
to the highest. Then ranks (Rij where i ¼ 1; 2; :::; m of the data, with significance determined using a per-
and j ¼ 1; 2; :::; n) are assigned to each observation of mutation approach. The authors provide an R package
this pooled sample. The average rank for the ith to implement the RS/P procedure in practice. Based
P
sample is given by R i ¼ n  1 nj¼1 Rij . Noting that the on the simulation results, the authors claim that the
expected value and the variance of the ranks, when RS/P methodology leads to an effective Phase I distri-
the process is IC, are given by Nþ1 2 and
bution-free procedure, both in terms of maintaining a
ðN  nÞðNþ1Þ FAP and out-of-control (OOC) performance. There is
12n ;
respectively (see Gibbons and Chakraborti
2010), the charting statistics are given by Zi ¼ a need for more work on univariate nonparametric
  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Phase I charts as Phase I is a very important part of
R i  Nþ1
2 = ðN  12nnÞðNþ1Þ
; i ¼ 1; 2; 3; ::: Jones-Farmer the overall monitoring regime. In terms of computa-
et al. (2009) gave two choices for the control limits, tional burden, nonparametric charts in general, and
namely, the simulated and the approximate normal nonparametric Phase I charts in particular, are on the
theory based control limits. For the latter choice, by same level as their normal theory counterparts. Next
the central limit theorem (for large n), marginally, the we consider Phase II charts.
standardized mean-rank Zi approximately follows a
N(0,1) distribution when the process is IC: However,
Univariate nonparametric process monitoring:
the charting statistics Z1 ; Z2 ; :::; Zm are dependent
Phase II charts
random variables and this dependence needs to be
properly accounted for. It can be shown that asymptot- Shewhart-type control charts
ically, the joint distribution of Z1 ; Z2 ; :::; Zm can be Shewhart-type control charts for monitoring location
approximated by a (singular) multivariate normal dis- in Case K. Control charts based on the sign and
tribution with means equal to zero, standard deviations signed-rank statistics. As in the case with normally
equal to one and a common pairwise correlation distributed data, different types of distribution-free
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qij ¼ 1=ðm  1Þ. As m increases this correlation charts have been proposed and studied. To this end,
tends to zero, so the lower control limit (LCL) and the note that the sign (SN) and the signed-rank (SR) tests
upper control limit (UCL) may be approximated using are among the simplest and yet versatile one-sample
the quantiles of the univariate N(0,1) distribution. In distribution-free tests (see, for example, Gibbons and
Phase I, the typical metric of IC performance is the Chakraborti, 2010). Adapting these to the SPC setting,
false alarm probability, FAP; which is the probability of Amin, Reynolds, and Bakir (1995) and Bakir (2004)
at least one false alarm. The control limits are chosen proposed the Shewhart-type charts based on the SN
such that the attained FAP does not exceed a desired and the SR statistics, called the Shewhart-SN and
nominal FAP; typically taken to be 0.01, 0.05 or 0.10. Shewhart-SR charts, respectively. However, these
Note that the Jones-Farmer et al. (2009) paper was papers and several of their extensions and generaliza-
covered in the review of Chakraborti et al. (2011) but tions were reviewed in the overview paper of
we provide the background details here since in a Chakraborti et al. (2011) and will not be discussed
later section on multivariate control charts, we discuss here. The key idea is to use distribution-free test sta-
the paper by Bell, Jones-Farmer, and Billor (2014) that tistics used in a hypothesis testing context, adapting
proposed a Phase I multivariate mean-rank chart them suitably in the process monitoring framework, for
4 S. CHAKRABORTI AND M. A. GRAHAM

example, constructing Shewhart, CUSUM and EWMA required for nonparametric control charts, since no spe-
charts and their generalizations based on these statistics. cific model assumptions are made for them.
Note that in Case K we work with what are known as
one-sample nonparametric test statistics such as the SN Shewhart-type control charts for monitoring spread
and the SR. In Case U, the monitoring regime consists in Case U. While a lot of work has been done on moni-
of an initial or retrospective phase, called Phase I, that toring the location, only a few papers are available on
yields a reference sample and a Phase II that yields test monitoring the spread, or the dispersion, or the variabil-
samples that need to be monitored, comparing each to ity of a process. Interested readers can consult Gibbons
the reference sample, in a sequential manner, and thus and Chakraborti (2010) for a discussion on these three
we work with what are known as two-sample nonpara- concepts which are not always the same and are not well
metric tests adapting them to construct Phase II control understood. Das (2008) proposed two charts for moni-
charts. This is the basic general strategy. We discuss a toring variability, when the location parameter is under
number of these charts below. control, based on two nonparametric tests for equality
of dispersion by Mood (1954) and Siegel and Tukey
Shewhart-type control charts for location with runs- (1960). Since the Mood tests, based on the statistic, say
type signaling rules in Case U. While the Shewhart- M, and the Siegel-Tukey test, based on the statistic, say
type chart is simple to use and is powerful in detecting R, are well-known (see for example, Gibbons and
larger, sudden shifts, in the process parameter, it lacks Chakraborti 2010) nonparametric tests for scale, details
sensitivity in detecting smaller shifts. Thus there have will not be given here. These are used to test the equality
been proposals to improve the effectiveness of the of dispersions of two populations assuming that their
Shewhart-type charts for detecting smaller shifts by add- medians are equal but unknown. Das (2008) used the
ing runs-type signaling rules. The original set of such standardized M and R statistics as the charting statistics
for each chart, respectively, with control limits LCL/
rules goes back to the Western Electric Company (1956)
UCL ¼ 6 3 and CL ¼ 0 for each chart. Das (2008) con-
and in the more recent literature there are both standard
cluded that, for all shifts under consideration, the chart
and improved runs-rules schemes. The former are typ-
based on Mood’s test performed best. Again, more work
ically of the form w-of-(w þ v) with w > 1 and v  0 and
is necessary in this important area. There are several
the latter is a combination of the classical 1-of-1 runs-
other nonparametric scale tests in the literature which
rule and the w-of-(w þ v) runs-rules. The improved
can be adapted to control charting. One key question
scheme has the advantage of improving the perform-
here is the assumption about the equality of locations.
ance of the charts in detecting larger shifts while main-
Typically, while monitoring the mean in Case U and in
taining its performance in detecting small to moderate
the normal theory setting the variances are assumed to
shifts. While there is a considerable literature on the
be equal and IC, but in the distribution-free case, one
normal theory charts, on the nonparametric side, requires the opposite. The full impact of this paradigm
Malela-Majika, Chakraborti, and Graham (2016) have shift (see Jones-Farmer and Champ, 2010) needs to be
implemented both schemes into the Shewhart chart investigated and better understood. The effect of the ref-
based on the Mann-Whitney statistic proposed by erence sample size on the estimation of the common
Chakraborti and Van de Wiel (2008). The authors com- median and its impact on the performance of the chart
pared their chart to the Shewhart-type precedence chart would be particularly interesting.
with and without signaling rules, see Chakraborti, Van While a number of control charts for monitoring
der Laan, and Van de Wiel (2004) and Chakraborti, the location and some for monitoring the scale have
Eryilmaz, and Human (2009), respectively, and with its been considered, there are situations where one
parametric counterpart, i.e., the parametric Shewhart-X monitors both the location and the scale parameters
chart with and without signaling rules and found that simultaneously. In fact, this is what is typically recom-
their proposed chart outperformed its competitors in mended while monitoring the mean of a normally dis-
detecting shifts under distributions of various shapes. tributed process, since the standard deviation appears
Given that the runs-rules can enhance the performance in the control limits and therefore needs to be con-
of Shewhart charts in detecting small shifts, while main- trolled (monitored) first. This has been called the joint
taining the basic simplicity, we envision more work in monitoring problem. The reader is referred to
this area, particularly with regard to the computation McCracken and Chakraborti (2013) for a detailed
and the effect of the size of the reference sample on review of the recent advances in joint monitoring.
chart performance. Generally speaking, more data are Joint process monitoring schemes in the normal
 QUALITY ENGINEERING 5

distribution setting use two separate charts, one for statistic which are necessary for setting up the control
the mean, and one for the standard deviation (or the chart. When the process is IC, F ¼ G, which means
variance) or a single chart using a function of these that h ¼ 0 and d ¼ 1. In this case
statistics. For example, a Shewhart X chart is often 1
used, along with a Shewhart R (or an S) chart, as a EðT1 jICÞ ¼ l1 ¼ nðN þ 1Þ (3)
2
scheme. Diko, Chakraborti, and Graham (2016)
showed that there are issues with this monitoring and
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
scheme with regard to the false alarm rate and inter- 1
pretation, and thus some researchers have recom- STDEV ðT1 jICÞ ¼ r1 ¼ mnðN þ 1Þ (4)
12
mended using a single chart for monitoring both
location and scale simultaneously. Although the respectively. Similarly for the AB statistic
8
majority of the overview of McCracken and >
> nN
< if N is even
Chakraborti (2013) is on parametric (normal theory) 4
EðT2 jICÞ ¼ l2 ¼ (5)
charts, two nonparametric papers were cited, namely, >
> nðN 2  1Þ
: if N is odd
Zou and Tsung (2010) and Mukherjee and 4N
Chakraborti (2012). We review the Mukherjee and
and
Chakraborti’s (2012) paper next. 8
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
> 1 N2  4
>
< mn if N is even
Shewhart-type control charts for joint monitoring of 48 N 1
STDEV ðT2 jICÞ ¼ r2 ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
> 1 ðN þ 1ÞðN 2 þ 3Þ
location and scale in Case U. Control charts based >
: mn if N is odd
48 N2
on the Lepage statistic. In Case U, recall that process
monitoring consists of both a Phase I and a Phase II (6)
and the key idea is to consider two-sample distribu- respectively. Again, one is referred to Gibbons and
tion-free test statistics and adapt them for use in pro- Chakraborti (2010) for details. For the jth test sample,
cess monitoring. In order to monitor both the the WRS and AB statistics T1j and T2j and the corre-
location and the scale simultaneously, Mukherjee and sponding standardized statistics, S1j and S2j , where
T l T l
Chakraborti (2012) proposed a Shewhart-type chart S1j ¼ 1jr1 1 and S2j ¼ 2jr2 2 , respectively. The charting
based on the Lepage (1971) statistic (denoted statistic for the Shewhart-LP chart for monitoring the
Shewhart-LP). The corresponding Lepage (1971) test jth Phase II (test) sample, is then given by
is a distribution-free test for testing the equality of the
location and the scales of two continuous distributions LPj ¼ S21j þ S22j (7)
based on two independent random samples. Consider where an OOC signal is given when LPj  a.
two continuous distributions (the U and the V distri- Note that since the LPj statistic can only be posi-
butions) with continuous cdf’s FðxÞ and tive, there is only an UCL for this chart. If the chart
GðyÞ ¼ Fðdy þ h), respectively, with d > 0 represent- signals, the question is whether the process location,
ing the unknown scale parameter,  1 < h < 1 or the process spread, or both have gone OOC. This
the unknown location parameter and F an unknown is called post signal diagnostics and is an important
and continuous cdf. The Lepage statistic is a combin- practical issue for all joint monitoring schemes, both
ation of the Wilcoxon rank-sum (WRS) test for loca- parametric and nonparametric. By the way, in the
tion and the Ansari-Bradley (AB) test for scale. The multivariate case, this is also an important issue, since
WRS test statistic, say T1 , is defined as if the chart signals a change, one needs to know
X
N which variable(s) is(are) responsible for the change.
T1 ¼ kZk (1) Mukherjee and Chakraborti (2012) proposed a follow-
k¼1
up procedure for this purpose where two more design
and the AB test statistic, say T2 , is defined as parameters are needed, say a1 and a2 , respectively. If
X
N
1 the chart signals at the jth sample and only S21j exceeds
T2 ¼ k  ðN þ 1ÞZk : (2) a1 , a shift in location is indicated; whereas if only S22j
k¼1
2
exceeds a2 , a shift in scale is indicated. Finally, if both
For more details on the WRS and the AB tests the S21j and S22j exceed a1 and a2 , respectively, a shift in
interested reader is referred to Gibbons and both location and scale is indicated. Performance of
Chakraborti (2010). Here we simply give the IC the Shewhart-LP chart has been studied and the
expected value and standard deviation for each reader is referred to their paper for more details.
6 S. CHAKRABORTI AND M. A. GRAHAM

Clearly, there is a need to develop more nonparamet- Control charts based on the exceedance statistic.
ric charts for monitoring process location and process Mukherjee et al. (2013) proposed a CUSUM chart
scale simultaneously. A systematic and thorough based on the exceedance statistic (denoted CUSUM-
examination of the post signal diagnostic scheme, for EX). An upper one-sided CUSUM control chart based
joint monitoring schemes, would also be worthwhile. on the EX statistic is defined as
There is also the need for easier access to computa-  
Cjþ ¼ max 0; Cj  1 þ ðUj;r  lU Þ  k for j ¼ 1; 2; [Link]
tional aides and better understanding the effect of the
size of the reference sample on chart performance. (8)
 
The Shewhart-type charts are simple and effective with the starting value C0þ ¼ 0, lU ¼ E Uj;r jXðrÞ ¼
for detecting larger shifts. Next we consider CUSUM- npr and which signals at the first j for which Cjþ  H.
type univariate nonparametric charts, those, like their The values of the design parameters, k and H, are
parametric counterparts, that are effective in detecting found such that a nominal IC average run-length
smaller shifts. (ICARL) value is obtained. Typically, k is specified
first and then, following this, H is found using a
CUSUM-type control charts search algorithm in order to obtain some nominal
CUSUM charts are useful and sometimes more natur- ICARL. The steps for choosing the two design param-
ally fitting in the process control environment in view eters, k and H, are the same as for other CUSUM-
of the sequential nature of data collection. We review type charts and will not be discussed each time. More
some nonparametric CUSUM-type charts next. work on the precedence and exceedance charts is
needed, particularly to understand their performance
CUSUM-type control charts for monitoring location in Phase II and how much Phase I data are needed to
in Case K. Control charts based on the sign and guarantee nominal IC and decent OOC performance.
signed-rank statistics. Using the SN and SR statistics, This is true for most of the Phase II (Case U) non-
Amin et al. (1995) and Bakir and Reynolds (1979) pro- parametric charts currently available in the literature.
posed CUSUM-type charts based on the SN and the SR Next we review a nonparametric Phase II CUSUM-
statistics, respectively. These charts are denoted type chart useful for joint monitoring adapting the
CUSUM-SN and CUSUM-SR, respectively, however, Lepage (1971) statistic.
these papers were covered in the overview paper of
Chakraborti et al. (2011) and will not be discussed here. CUSUM-type control charts for joint monitoring of
location and scale in Case U. Chowdhury,
CUSUM-type control charts for monitoring location Mukherjee, and Chakraborti (2015) considered a dis-
in Case U. As noted earlier, in Case U, the key idea is tribution-free CUSUM-type chart based on the LP
to consider two-sample distribution-free test statistics statistic. Being a CUSUM chart, this is expected to be
and adapt them for use in process monitoring. more sensitive than the Shewhart-LP chart (discussed
Chakraborti et al. (2004) considered a class of non- earlier) for small, sustained (upward or downward)
parametric Phase II Shewhart-type charts based on the shifts in the location. Distribution-free CUSUM charts
so-called precedence statistics, called the precedence based on the LP statistic are called CUSUM-LP charts
charts, however, this important paper was covered in from this point forward.
the overview paper of Chakraborti et al. (2011). Thus, An upper one-sided CUSUM chart based on the LP
it is logical to consider CUSUM-type charts based on statistic is defined as
h i
precedence statistics and this is considered in Cjþ ¼ max 0; Cjþ 1 þ ðLPj  2Þ  k for j ¼ 1; 2; [Link]
Mukherjee, Graham, and Chakraborti (2013). It turns
out that it is more convenient to work with the so- (9)
called exceedance statistics, which are closely related and the upper control limit, H, is found using a
to the precedence statistics. The number of Y observa- search algorithm in order to obtain some nominal
tions in the jth Phase II sample that precede XðrÞ ; the ICARL. Chowdhury et al. (2015) considered values of
rth ordered observation in the reference sample, is k ¼ 0, 3 and 6, respectively. For the post signal diag-
called the precedence statistic, whereas the number of nostics, Chowdhury et al. (2015) proposed a follow-up
Y observations in the jth Phase II sample that exceeds that makes use of the hypothesis testing connection
XðrÞ ; denoted Uj;r ; is called the exceedance statistic. and the p-values of the corresponding WSR and the
 
The probability pr ¼ P Y>XðrÞ j XðrÞ is called an AB statistics, denoted p1 and p2 , respectively. These p-
exceedance probability. values are with the reference value k  0, the starting
 QUALITY ENGINEERING 7

value C0þ ¼ 0 and which signals at the first j for which Zi ¼ kSNi þ ð1  kÞZi  1 Z0 ¼ 0 (11)
Cjþ  H. It should be noted that we subtract 2 in the
where 0 < k  1 is the smoothing constant to be
CUSUM since EðLPjICÞ ¼ 2. The values of the design
specified later. The exact control limits and the cen-
parameters, k and H, are found such that a nominal
terline (CL) of the EWMA-SN chart are given by
ICARL value is obtained. Typically, k is specified first rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
and then, calculated on the basis of the two samples; kn  2i

UCL=LCL ¼ 6L 1  ð1  kÞ and CL ¼ 0
one with m Phase I observations and the other with 2k
the n observations from the jth test sample (if a signal (12)
was given at sample number j). If p1 is very low but
where L > 0 is the distance of the control limits from
not p2 , this indicates a shift in location only.
the centerline.
Alternatively, if p1 is very high but not p2 , this indi-
The steady-state control limits (typically used when
cates a shift in scale only. Finally, if both p1 and p2 are
the EWMA chart has been running for several time
very low, this indicates a shift in location and scale.
periods) are based on the asymptotic standard devi-
The question about which Phase I sample order stat-
ation of the control statistic (Lucas and Saccucci,
istic to use in forming either the precedence (or the
1990) and are given by
exceedance) statistics is an interesting one. Following rffiffiffiffiffiffiffiffiffiffiffiffiffi
the work of Mukherjee et al. (2013) which was based on kn
UCL=LCL ¼ 6L and CL ¼ 0: (13)
the median, Graham, Mukherjee, and Chakraborti 2k
(2017) investigated this choice in the design of the
If any of the charting statistics Zi plots on or out-
CUSUM-EX control chart and gave some recommenda-
side either of the two control limits, the process is
tions from a practical point of view.
declared OOC and a search for assignable causes is
started. Otherwise, the process is considered IC and
EWMA-type control charts
charting continues. The two design parameters, k and
Like the CUSUM charts, the EWMA charts also take
L, are selected so that a nominal ICARL value is
advantage of the sequentially accumulating nature of
attained. Typically, k is chosen first with small values
the data arising in a typical SPC environment and are
of k being recommended for small shifts (k ¼ 0.05),
known to be efficient in detecting smaller shifts, how-
larger values of k being recommended for moderate
ever they are easier to implement. Following the
shifts (k ¼ 0.10) and a large value of k being recom-
Shewhart and CUSUM-type charts described earlier,
mended for large shifts (k ¼ 0.20). Once k is selected,
some nonparametric EWMA-type charts have been
then L is found using a search algorithm so that a
considered both in Case K and Case U. They are
nominal ICARL value is attained. The steps for choos-
described next.
ing the two design parameters, k and L, are the same
as for other EWMA-type charts and will not be dis-
EWMA-type control charts for monitoring location
cussed each time. Graham, Chakraborti, and Human
in Case K. Control charts based on the sign and
(2011b) proposed a nonparametric EWMA-type chart,
signed-rank statistics. As we have noted earlier, in
based on the SR statistics, denoted EWMA-SR chart.
Case K, one can simply use the SN and the SR statis-
The SR charts require symmetry of the underlying
tics in the monitoring framework. Following the
distribution but have performance comparable with
Shewhart and CUSUM-type charts, Graham,
normal theory EWMA charts based on the mean. The
Chakraborti, and Human (2011a) considered an
reader is referred to the paper for more details.
EWMA chart based on the SN statistic (denoted
Next we discuss some EWMA-type charts in
EWMA-SN). For this chart let Xi denote the ith indi-
Case U.
vidual measurement from an unknown continuous
distribution with cdf F with median h, that is to be
EWMA-type control charts for monitoring location
monitored. Define the SN statistic
in Case U. Control charts based on the exceedance
Xn
  statistic. In Case U, with a reference sample, one can
SNi ¼ sign Xij  h0 for i ¼ 1; 2; 3; ::: (10)
use the precedence/exceedance statistics, in a way
j¼1
similar to what was done for the CUSUM-EX chart.
where signðAÞ ¼  1; 0; 1 if A < 0, ¼ 0, > 0 and Hence, the charting statistic of the EWMA-EX chart
h0 is the known or the specified or the target value h. is obtained by sequentially accumulating the exceed-
The charting statistic of the EWMA-SN chart is ance statistics U1;r , U2;r , U3;r , … and the charting
defined as statistic is defined as
8 S. CHAKRABORTI AND M. A. GRAHAM

 
Zj ¼ kUj;r þ ð1  kÞZj  1 ; Z0 ¼ E Uj  k;r jXðrÞ ¼ npr UCL. The value of a is obtained by making use of simu-
(14) lation in order to find a desirable ARL value. The
EWMA-LP accumulates the statistics LP1 ; LP2 ; LP3 ; :::
where 0 < k  1 is the smoothing constant to be with the EWMA charting statistic defined as
specified later. Graham, Mukherjee, and Chakraborti
(2012) showed that the unconditional IC mean and Zj ¼ k LPj þ ð1  kÞZj  1 for j ¼ 1; 2; 3; ::: (18)
standard deviation of Zj are given by where the weighting constant 0 < k  1 with starting
   
j
E Zj jIC ¼ nð1  aÞ 1  ð1  kÞ and (15) value Z0 ¼ 2.
  The process is considered to be IC while all the
STDEV Zj jIC ¼ charting statistics Zj , j ¼ 1; 2; 3; ::: fall below the UCL,
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

 
nað1  aÞ  
j 2 kðm þ 1Þ  2j
 however as soon as a charting statistic falls on or
n 1  ð1  kÞ þ 1  ð1  kÞ ; above the UCL the process is declared OOC and typ-
mþ2 2k
ically a search for assignable causes would be started.
respectively, where a ¼ r=ðm þ 1Þ. The exact control This work is currently under review.
limits and the CL of the two-sided EWMA-EX chart Another contribution to the nonparametric EWMA
are given by literature is by Zou and Tsung (2010) who integrated
      the powerful goodness-of-fit test with the EWMA
UCL=LCL ¼ E Zj jIC 6L  STDEV Zj jIC and CL ¼ E Zj jIC
scheme, however, details are omitted.
(16) As noted earlier, in a global sense, more work on
where L > 0 is the distance of the control limits from the precedence and exceedance charts is needed, par-
the centerline. The steady-state control limits and CL ticularly to understand their performance in Phase II
are given by and how much Phase I data are needed to guarantee
UCL=LCL ¼ nð1  aÞ6L nominal IC and decent OOC performance. This is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

 ffi true for most of the Phase II (Case U) nonparametric
nað1  aÞ kðm þ 1Þ charts currently available in the literature.
nþ and CL ¼ nð1  aÞ:
mþ2 2k We review a few other univariate nonparametric
charts in the next section.
(17)
If any Zi plots on or outside either of the two con- Other univariate nonparametric charts
trol limits, the process is declared to be OOC and a GWMA-type control charts in Case K. A generaliza-
search for assignable causes is started. Otherwise, the tion of the EWMA charting scheme to monitor the
process is considered IC and charting continues. mean, under normality, was considered by Sheu and Lin
Following the work of Graham et al. (2012) which (2003). This is referred to as the generally weighted
was based on the median of the Phase I (reference) moving average (GWMA) charting scheme. They
sample, Graham et al. (2017) investigated the choice showed that the GWMA chart performs better in
of the Phase I (reference) sample order statistic used detecting small shifts in the process mean. In the non-
in the design of the EWMA-EX chart and gave some parametric setting, Lu (2015) considered a GWMA
recommendations for selecting this order statistic chart for the process median based on the SN statistic,
from a practical point of view. called the GWMA-SN chart. Let X be some quality
Next we review an EWMA-type chart for joint characteristic of a process with target value T. Define
monitoring in the unknown parameter case based on the deviation Y ¼ X  T, then p ¼ PðY > 0Þ is the
the Lepage (1971) statistic. process proportion with p ¼ 0:5 indicating the process
is IC and p 6¼ 0:5 indicating the process is OOC. For a
EWMA-type control chart for joint monitoring of sample of size n at time t define Yit ¼ Xit  T and Iit ¼
location and scale in Case U. Control charts based 1; 0 if Yit > 0, otherwise, for i ¼ 1;
P2; :::; n. Thus, the SN
on the Lepage statistic. Chowdhury et al. (2015) stated statistic used by Lu (2015), Nt ¼ ni¼1 Iit . It follows that
that it’s challenging to design the EWMA chart based Nt follows a BIN(n,0.5) distribution when the process is
on the Lepage statistic (denoted EWMA-LP chart), since IC. The charting statistic for the GWMA-SN chart is
“construction of exponentially weighted moving average defined as
chart based on the Lepage statistic will be challenging as  
Xt
there is no simple explicit form of conditional variance Gt ¼ q ðj  1Þa
 qja a
Nt  jþ1 þ qt T for t ¼ 1; 2; [Link]
j¼1
of Lepage statistic that can be obtained”. In order to cir-
cumvent this problem, we use a constant value a for the (19)
 QUALITY ENGINEERING 9

where 0  q < 1 and 0 < a  1 are two chart beyond the control limits until the current time point,
parameters. Note that the qðj  1Þ  qj
a a
are the which is called the CRL. If the CRL value is suffi-
weights from the most recent to the oldest observa- ciently small, an OOC signal is generated.
tion. It is seen that the GWMA-SN chart reduces to To the best of our knowledge, only three nonpara-
the EWMA-SN chart for a ¼ 1 and q ¼ 1  k; where metric synthetic control charts have been proposed in
0 < k  1; is the smoothing parameter of the the literature so far. Khilare and Shirke (2010) and
EWMA chart. Also, as is the case for the EWMA Pawar and Shirke (2010) proposed nonparametric
chart, the GWMA reduces to a Shewhart-type chart synthetic control charts based on the SN and the SR
for q ¼ 0 and a ¼ 1. Moreover, it can be shown that statistic, respectively, for monitoring location. Khilare
the IC expected value and the variance of the charting and Shirke (2012) proposed a nonparametric synthetic
statistic are n2 and Qt n4, respectively, where chart to monitor the process variation using the SN
P
Qt ¼ tj¼1 ðPðM ¼ jÞÞ2 . Accordingly, the exact statistic. First we give some definitions needed for our
Shewhart-type control limits for the GWMA-SN control discussions. When referring to the zero-state mode, it
chart are given by is assumed that there is a nonconforming sample at
LCL=UCL ¼ EðGt jICÞ6L  STDEV ðGt jICÞ time zero, whereas when referring to the steady-state
rffiffiffiffiffiffiffiffiffi mode, one assumes that the process starts and stays
n n n (20)
¼ 6L Qt with CL ¼ EðGt jICÞ ¼ : IC for a long time before a process shift occurs at
2 4 2
some ‘random time’.
For t ! 1, the asymptotic variance of the charting The synthetic SN chart integrates the operation of the
statistic Gt is given by limt!1 VARðGt Þ ¼ Qr2 , where nonparametric Shewhart-SN chart and the CRL chart.
Q ¼ limt!1 Qt . Hence the steady-state control limits Khilare and Shirke (2010) compared the (zero-state) ARL
and the CL are given by performance of the synthetic SN chart against the ordin-
rffiffiffiffiffiffiffi
n n n ary Shewhart-SN and X charts and observed that the
LCL=UCL ¼ 6L Q and CL ¼ : (21) synthetic SN chart performs best overall and that the
2 4 2
improvement in the ARL is more significant for small to
Lu (2015) found that the GWMA-SN chart is more moderate shifts. Note that the authors acknowledged that
efficient than the EWMA-SN chart in detecting smaller
this good performance might be due to the fact that they
shifts where larger values of q and a are recommended.
assumed a zero-state mode and that under a steady-state
For larger shifts, the GWMA-SN chart is not optimal,
this might not have been the case. Following their 2010
but can compete with the EWMA-SN chart. Clearly, a
paper, in 2012 the same authors proposed a nonparamet-
GWMA-SR chart can be considered based on the SR
ric synthetic chart to monitor the process variation using
statistic and assuming symmetry, and in fact, this has
the SN-based test assuming both the zero- and steady-
been done by Chakraborty et al. (2016). Their GWMA-
state modes.
SR chart performs better, as expected, for symmetric
The ARL performance of the synthetic SN chart is
distributions and smaller shifts.
compared with the Shewhart S2 and Shewhart-SN
As we noted earlier, attempts have been made to
charts and it is observed that the synthetic SN chart
enhance the smaller shift detection ability of the
Shewhart charts, under normality, for example, by performs best overall. However, superiority of the syn-
including supplementary runs-rules. Another attempt thetic SN chart is limited to the zero-state mode,
in this direction includes a new class charts, called the since, when in the steady-state mode, the ARL per-
synthetic charts (Wu and Spedding 2000). A lot of formance of the synthetic chart is poor compared to
work has been done on these charts, we briefly review the zero-state ARL. This occurs because the effect of
the nonparametric analogs. the head start feature has disappeared. Next, we dis-
cuss the synthetic SR chart proposed by Pawar and
Synthetic control charts in Case K. The synthetic Shirke (2010). This chart integrates the operation of
control chart integrates a Shewhart chart and what is the nonparametric Shewhart-SR chart and the CRL
called a conforming run-length (CRL) chart. Unlike chart. As perhaps expected, Pawar and Shirke (2010)
the Shewhart chart, a synthetic chart does not signal compared the (zero-state) ARL performance of the
when a single charting statistic plots on or beyond the synthetic SR chart against the Shewhart-SR chart (i.e.,
control limits. Instead, when that happens, the practi- the 1-of-1 chart), the 2-of-2 runs-type signaling rules
tioner determines how many samples have been taken SR chart and the Shewhart X chart. The authors
since the last time a charting statistic plotted on or observed that the synthetic SR chart outperforms all
10 S. CHAKRABORTI AND M. A. GRAHAM

the other charts for all upwards shifts in the pro- Change-point model based methods in Case U. As
cess medians. eluded in the above discussion, another popular for-
Note that although the synthetic charts have mulation of the monitoring problem in the literature
attracted a lot of interest in the literature, and more is the change-point formulation. Hawkins, Qiu, and
work can be done, some researchers have advised Kang (2003) studied the change-point model under
against their use. See for example, Knoth (2016) for a the assumption of normality. Hawkins and Deng
discussion. Again, to reiterate, more work on these (2010) developed a nonparametric analog, based on
charts is needed, particularly to understand their per- the Mann-Whitney statistic, however, this paper was
formance in Phase II and how much Phase I data are covered in the overview paper of Chakraborti et al.
needed to guarantee nominal IC and decent OOC (2011) and will not be discussed here. Liu, Zhang, and
performance. Zi (2015) proposed a dual nonparametric CUSUM
(DNC) chart based on a sequential rank for monitor-
Adaptive control charts. Another class of control ing the location of a process. Let xi denote the ith
charts studied in the literature is called adaptive independent observation from an unknown continu-
charts, which allows the user to vary the charting con- ous distribution, F. Let s denote the unknown
ditions depending on the recent outcomes observed in change-point and let l0 and l1 (6¼ l0 ) denote the IC
the data from the process. For example, if a control and OOC location parameters, respectively. Let Rn
P
chart shows that the process is IC and stable for a denote the nth sequential rank, Rn ¼ nj¼1 I fxn  xj g
long time, it might seem reasonable to extend the of an observation among n observations. As described
sampling interval. On the other hand, if the chart earlier, the standardized statistic is Rn ¼ ðRn 
indicates that the process might be heading toward EðRn jICÞÞ=SDðRn jICÞ where EðRn jICÞ ¼ ðn þ 1Þ=2 and
being OOC, one might want to collect data more fre- pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
SDðRn jICÞ ¼ ðn þ 1Þðn  1Þ=12, respectively. Then
quently, thus shortening the sampling time interval.
a CUSUM charting statistic can be based on the Rn
Psarakis (2015) provided an overview of the adaptive þ 
given by Sþ n ¼ max 0; Sn  1 þ Rn  k

and Sn ¼
control charting literature. In that paper the definition 
of an adaptive control chart is given as “A control min 0; Sn 1 þ Rn þ k ; with the starting values
 þ
chart is considered adaptive or dynamic if at least one Sn ¼ Sþn ¼ 0. Let Sn ¼ max Sn ; Sn and which sig-
of the parameters h, n, or k are allowed to change in nals at the first j for which Sn  H. The authors com-
real time depending on the actual values from the pare their chart to the well-known nonparametric
previous sample statistics”, where h denotes the sam- change-point chart by Hawkins and Deng (2010) and
pling interval, n is the sample size and k denotes the found that it has almost the same performance for a
constant which determines the width of the control wide range of shifts. The latter paper was covered in
limit. One type of adaptive chart has been proposed the review by Chakraborti et al. (2011) and will not
for the variable sampling interval (VSI) setting. be repeated here.
Reynolds et al. (1988) were the first to suggest an
X-chart with a VSI scheme. Advantages of using VSI Monitoring distributions. In many industries today
charts, or adaptive charts in general, include minimiz- there is an overwhelmingly large amount of data col-
ing the time for shift detection, as well as the number lected almost continuously. Such data sets provide a
of samples that are needed to detect a shift. In some rich environment with a large number of variables
instances this can greatly reduce the cost. that could be monitored to the advantage of the
Liu et al. (2013) considered an adaptive nonpara- manufacturer. Such process streams are referred to
metric EWMA chart, called the ANE chart, for moni- as being ‘high dimensional’ and SPC methods, for
toring the location of a process. The chart is based on working with multiple data streams, are useful. Ross,
the sequential ranks (see their paper for details). Liu Tasoulis, and Adams (2011) provided a framework
et al. (2015) considered a generalization of the ANE for detecting changes in data streams when the dis-
chart under the VSI setting. They found that adding tributional form is unknown. This is basically a non-
the VSI modification improves the performance of the parametric change-point model for detecting change
chart. Coelho, Graham, and Chakraborti (2017) pro- in the distribution and their methods make use of
posed a VSI SR chart and showed that it performs rank-based nonparametric tests in a streaming/moni-
similar or better than the VSI SN chart that was pro- toring context. Ross et al. (2011) started by extending
posed by Amin and Widmaier (1999). the change-point model framework so that changes
 QUALITY ENGINEERING 11

in the scale parameter can be found. Following this, Given these different classes of charts, using different
they considered the simultaneous monitoring of loca- charting statistics, one obvious question in practice is
tion and scale parameters. They continue by intro- which chart should be used and when. A discussion of
ducing the idea of stream discretization which saves this important question and the recommendations fol-
time by computing the ranks in a computationally low the guidelines in the parametric case, namely, if
effective way. Ross and Adams (2012) considered smaller shifts are of interest a CUSUM or an EWMA-
two Phase II nonparametric control charts for detect- type chart should be used, whereas for larger shifts a
ing arbitrary distribution changes. These charts are Shewhart-type chart is recommended. The choice
based on some well-known nonparametric goodness- between a parametric and a nonparametric chart
of-fit statistics, such as the Cramer-von-Mises (CvM) depends on what is known about the process and what
and the Kolmogorov-Smirnov (KS) statistic (see assumptions can be made and justified.
Gibbons and Chakraborti 2010). Suppose there is a
change-point s in a set of observations fX1 ; :::; Xt g Multivariate nonparametric process monitoring
then H0 : Xi  F0 ; i ¼ 1; :::; t and Ha : X1 ; :::; Xs 
F0 ; Xsþ1 ; :::; Xt  F1 where F0 and F1 are two con- In recent years the interest in multivariate process moni-
tinuous cdf’s. Immediately following any observation toring has increased significantly because in today’s data
Xk at time k; a change-point can be tested by partition- rich environment more quality features and variables are
ing the observations into two samples S1 ¼ fX1 ; :::; Xk g available and monitored that are possibly interdepend-
ent. Consequently, the interest in multivariate nonpara-
and S2 ¼ fXkþ1 ; :::; Xt g with sample sizes n1 ¼ k and
metric control charts has gone up tremendously because
n2 ¼ t  k, respectively. Then an appropriate two-
it is far more difficult, if not impossible, to justify a para-
sample hypothesis test can be performed. Both the
metric multivariate distribution assumption (such as the
CvM and the KS tests rely on the comparisons of the
multivariate normal) for the underlying distribution.
empirical distribution functions of the two
P Suppose that p quality characteristics are of interest and
samples defined as F ^ S1 ðxÞ ¼ 1 k I ðXi  xÞ and
Pt k i¼1 let X 1 ; X 2 ; ::: be random vectors, each of dimension
^
F S2 ðxÞ ¼ 1
I ðXi  xÞ, respectively. The KS p  1, which denotes the observations over time. It is
tk i¼kþ1
test is defined as Dk;t ¼ supx F^ S1 ðxÞ  F^ S2 ðxÞ and assumed that these vectors are independent with mean
Ð1 vectors l1 ; l2 ; ::: respectively and a common variance-
the CvM test is defined as Wk;t ¼  1 F^ S1 ðxÞ 
covariance matrix RX . Taking a cue from hypothesis
F^ S2 ðxÞ2 dF^ t ðxÞ where F^ t ðxÞ is the empirical cdf of the
testing, Hotelling’s v2 chart is probably the most popular
pooled sample. For the KS and the CvM test, the null
multivariate control chart for the mean, however, it
hypothesis that no change occurs is rejected at time requires the assumption of multivariate normality to
ðKSÞ ðCvMÞ
point k; if Dk;t > hk;t or if Wk;t > hk;t , respectively, maintain its nominal performance. Here we consider
ðÞ some nonparametric multivariate charts.
where hk;t represents some threshold. Now, for the
Note that although Chakraborti et al. (2001, 2007
CvM change-point model Ross and Adams (2012) used
  and 2011) presented extensive overviews on the non-
W l
the charting statistic Wt ¼ maxk k;trk;t k;t ; 1 < k < parametric control charting literature, they did not con-
t where lk;t and rk;t denotes the mean and the stand- sider multivariate charts. That was primarily because
ard deviation of Wk;t . A signal is given when Wt the literature was not quite mature at that point in time.
ðCvMÞ Some of the earlier papers in this literature include
exceeds ht . For the KS change-point model, the
Hayter and Tsui (1994) who proposed a Shewhart-type
authors proposed monitoring pk;t , the p-value associ-
multivariate nonparametric control charting scheme for
ated with Dk;t qt ¼ maxk ðqk;t Þ. Letting qk;t ¼ 1  pk;t ; a
the mean vector based on the M statistic; the maximum
signal is given when qt ¼ maxk ðqk;t Þ exceeds some deviation of the observations from their sample means.
threshold. Based on the performance results, they con- Kapatou and Reynolds (1994, 1998) proposed EWMA-
clude that the CvM chart outperforms the KS chart. type multivariate nonparametric control charting
Monitoring the underlying distribution is an interesting schemes for the SN and SR statistics, respectively.
problem which should include a post signal diagnostic Though, it should be noted that they are not exactly
analysis, in order to indicate the type of change that nonparametric since some elements of the covariance
may have taken place. More work needs to be done in matrix were estimated. Liu (1995) proposed a Shewhart-
this area. type multivariate nonparametric control charting scheme
12 S. CHAKRABORTI AND M. A. GRAHAM

based on a concept of data depth. Qiu and Hawkins outperforms the Hotelling T 2 chart when the underly-
(2003) proposed a CUSUM-type multivariate nonpara- ing process distribution is non-normal.
metric control charting scheme based on the so-called Li, Dai, and Wang (2014) proposed a Phase I
antiranks of the measurements. Hamurkarouglu, Mert, multivariate nonparametric change-point control
and Sayken (2004) proposed a nonparametric control charting scheme based on the concept of data depth
charting scheme based on the Mahalanobis depth. Qiu (CPDP), for individual data, which can be used to
(2008) proposed a CUSUM-type multivariate nonpara- monitor the mean and/or the variance, i.e., it moni-
metric control charting scheme based on log-linear mod- tors the mean vector, the covariance matrix or both.
eling. Zou and Tsung (2011) proposed an EWMA-type Suppose we have n independent observations from a p
multivariate nonparametric control charting scheme dimensional multivariate distribution, xi  FðpÞ ðui ; Ri Þ,
combined with the multivariate SN test; they integrated i ¼ 1; :::n: When the process is IC ui ¼ l and Ri ¼ R,
the spatial-SN test of Randles (2000) with an EWMA 8i; where the common l and R are unknown.
control charting scheme. Motivated by the use of the spa- Assume that a step shift has occurred in the mean
tial-SN test by Zou and Tsung (2011), Zi et al. (2013) pro- and/or the variance after s observations, then the
posed an EWMA-type multivariate nonparametric remaining n  s observations have mean and vari-
control charting scheme for monitoring location parame- ance l ð6¼ lÞ and R ð6¼ RÞ, respectively. The authors
ters. They adapted the directional spatial-SN test to proposed using a generalized Mann-Whitney statistic
P
online sequential monitoring by incorporating the Qðn1 Þ ¼ nj¼n1 þ1 Rn1 ðjÞ where
EWMA scheme. Boone and Chakraborti (2012) consid- n o
ered two simple Shewhart-type multivariate nonparamet- Rn1 ðjÞ ¼ xi jDFn1 ðxi Þ < DFn1 ðxj Þ; i ¼ 1; :::; n1
n o
ric control charting schemes based on the SN and the SR þ 0:5 xi jDFn1 ðxi Þ ¼ DFn1 ðxj Þ; i ¼ 1; :::; n1
statistics, respectively.
In this article, not all these papers will be covered where DFn1 ðxi Þ denotes the data depth of xi according to
or discussed in detail due to lack of space. The reader the empirical distribution of x1 ; :::; xn1 . Although the
is referred to Chapter 9 of Qiu (2014) where a good probability of ties is zero theoretically for continuous var-
discussion about multivariate nonparametric control iables, ties do occur in practice and the authors argued
charts can be found. As noted earlier in the section of that the problem of ties can be addressed by allocating a
univariate control charts, the Phase I analysis is a very probability of a 0.5 for observations that have the same
important part of the process monitoring regime and data depth. The standardized likelihood-ratio statistic
control charts are used in this phase regularly. In the    
is defined as SQðn1 Þ ¼ E Qðn1 ÞjIC  Qðn1 Þ =
multivariate setting, since more variables are involved,    
SD Qðn1 ÞjIC where E Qðn1 ÞjIC ¼ n1 ðn  n1 Þ=2 and
a Phase I analysis is just as, if not more, important.   pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
SD Qðn1 ÞjIC ¼ n1 ðn  n1 Þðn þ 1Þ=12, respectively.
We discuss some recent papers, but more work needs  
to be done in this area. Note that the numerator is E Qðn1 ÞjIC  Qðn1 Þ and
 
not the typical Qðn1 Þ  E Qðn1 ÞjIC , since the former
is positive with high probability. The charting statistics,
Multivariate nonparametric process monitoring:
SQðiÞ, are plotted against i ð1  i < n) and a signal is
Phase I charts
given when max1i < n SQðiÞ plots above some UCL. The
Bell et al. (2014) proposed a Phase I multivariate UCL is chosen such that a desired nominal FAP is
mean-rank (MMR) chart which is the multivariate attained. The authors compared their chart to the LRT
analog of Jones-Farmer et al. (2009)’s Phase I univari- chart (Sullivan and Woodall 2000) and found that it has
ate mean-rank chart. The chart is distribution-free similar performance under the normal distribution and
over the class of continuous multivariate elliptical dis- performs better when the underlying process distribution
tributions, such as the multivariate normal and the is non-normal.
multivariate t distribution. It should be noted that Cheng and Shiau (2015) proposed a Phase I chart
most multivariate charts only have an UCL and the based on the spatial SN statistic for monitoring the
process is declared OOC when the charting statistic, location parameter vector of a multivariate process.
which is usually a quadratic form, plots above the They compared their chart to four competing charts,
UCL and, typically, a search for assignable causes is the well-known Hotelling T 2 chart, the TSD 2
chart pro-
started for the post signal diagnostics. In Phase I the 2
posed by Sullivan and Woodall (1996) and the TMVE
UCL is chosen such that a desired FAP is attained. 2
and TMCD charts proposed by Jensen, Birch, and
The authors conclude that the MMR chart Woodall (2007), and found that their chart is the only
 QUALITY ENGINEERING 13

chart that is robust to the normality assumption and first kind given concordance (discordance) by hj ðsj Þ
that it is more powerful for the majority of OOC con- so that the null hypothesis can be expressed as
ditions with the exception being that their chart per-
H0 : hj ¼ sj ¼ 0:5; j ¼ 1; 2; :::; n:
forms worse than the Hotelling T 2 chart for large
shifts in the location vector. The charting statistic is given by T ¼
One practical issue with both Bell et al. (2014)’s ð4=CÞðC1  C=2Þ2 þ ð4=ðn  CÞÞðD1  ðn  CÞ=2Þ2
and Cheng and Shiau (2015)’s charts may be that they where Ci ðDi Þ denotes the number of concordances
require subgrouped data. To overcome this, Capizzi (discordances) of the ith kind, i ¼ 1; 2, and C ¼
and Masarotto (2017) proposed a distribution-free C1 þ C2 and D ¼ D1 þ D2 with C þ D ¼ n. The pro-
multivariate Phase I chart that works both for sub- cess is declared OOC (H0 is rejected) when T > UCL
grouped data and individual measurements, as the lat- with UCL ¼ v22;a and typically a search for assignable
ter is fairly typical in today’s data rich environments. causes would be started. Das (2009) concludes that
Their chart is based on multivariate SRs that integrate the chart outperforms the Hotelling T 2 chart when
spatial signs and ranks of the Mahalanobis depths the underlying process distribution is non-normal.
which is also an advantage over Bell et al. (2014)’s Similarly, Ghute and Shirke (2012) proposed a
chart who limited their focus to Mahalanobis depths multivariate nonparametric control charting scheme
and Cheng and Shiau (2015)’s chart who limited their based on the bivariate SR test of Bennett (1964). Let
focus to spatial signs. Capizzi and Masarotto (2017) Xi ¼ ðX1i ; X2i Þ, i ¼ 1; 2; :::; n be a subgroup sample
conclude that their proposed chart, which uses a per- from some symmetric continuous bivariate distribu-
mutation approach to calculate charting constants and tion with l and R being the location vector and
thus avoids a distributional assumption, shows wider covariance matrix, respectively. The chart is for the
applicability and has satisfactory performance for parameter known case, so without any loss of general-
many OOC conditions. They also provide an R pack-

 0 1 q
age for their proposed chart. ity assume l0 ¼ 0 0 and R0 ¼ , with
q 1
Next we consider some Phase II nonparametric
 1 < q < 1 and l0 and R0 are the known IC val-
control charts for multivariate data.
ues of the process mean and correlation matrix, which
means that the correlation coefficient q must also be
Multivariate nonparametric process monitoring: assumed known. The interest is in monitoring the
Phase II charts location, so we are interested in detecting shifts in l0 :
Shewhart-type control charts for location in Case K At each inspection point in time, a SR statistic is cal-
Das (2009) proposed a multivariate nonparametric culated for each variate in Xi ¼ ðX1i ; X2i Þ making use
control charting scheme based on the bivariate SN of n measurements in a sample. Recall that for the jth
test (the reader is referred to Puri and Sen (1976) for variable, the SR statistic is
more details on the bivariate SN test). Let X
n
Xj ¼ ðX1j ; X2j Þ0 , j ¼ 1; 2; :::; n be n independent sto- Tj ¼ CðXji ÞRðXji Þ for j ¼ 1; 2 (22)
chastic vectors with cdf’s F1 ðxÞ, … , Fn ðxÞ. We want to i¼1

test whether F1 , … , Fn have n specified pairs of mar- where CðAÞ ¼ 1; 0 if A > 0, < 0, RðXji Þ is the rank
ginal medians. Das (2009) chose the origins so that of Xji among Xj1 ; Xj2 ; ::::; Xjn and T1 and T2 are the
the assumption can be made that the pair of hypo- two SR statistics corresponding to two variables in a
thetical medians for each Xj , j ¼ 1; 2; :::; n, is  
 0 sample of size n. Let E Tj jl ¼ l0 ; IC ¼ vj for
0 ¼ 0 0 . The null hypothesis is given by
j ¼ 1; 2. Then the two SR statistics (T1 and T2 ) are
H0 : Fj ð0; 1Þ ¼ Fj ð  1; 0Þ ¼ 0:5; j ¼ 1; 2; :::; n: combined into a quadratic form, which is the charting
For each Xj define the events ðX1j  0; X2j  0Þ statistic
and ðX1j  0; X2j  0Þ as concordance events of the 0
W ¼ ðT  vÞ b^  1 ðT  vÞ (23)
first and second kind, respectively, and define the
events ðX1j  0; X2j  0Þ and ðX1j  0; X2j  0Þ as where T ¼ ðT1 ; T2 Þ0 is a 2  1 column vector of
discordance events of the first and second kind, coordinate-wise univariate SR statistics, v ¼ ðv1 ; v2 Þ0 is

respectively. Let 0 < cj < 1 denote the probability of ^ b11 b12

a 2  1 column vector and b ¼ is the
concordance of ðX1j ; X2j Þ. Finally, denote the condi- b21 b22
tional probability of concordance (discordance) of the variance-covariance matrix of T.
14 S. CHAKRABORTI AND M. A. GRAHAM

As in the univariate case, for an IC process, the ^ x ðX i  /0 Þ (and V will

cluster along the direction of A
nðnþ1Þ
means are v1 ¼ v2 ¼ 4 ; the variances are b11 ¼ point approximately toward that direction as well).

b22 ¼ nðnþ1Þð2nþ1Þ and the covariances are b12 ¼ b21 ¼ The test statistic (Randles (2000)) is given by Q ¼
Pn 24 0
npV V and H0 is rejected for large values of Q. Li
sgnðX1i ÞsgnðX2i ÞRðX1i ÞRðX2i Þ
i¼1
4 (see Dietz (1982)). The pro- (2015) uses a similar reasoning to test whether
cess is declared OOC when W > UCL with UCL ¼ fXs ; :::; Xt g has a similar location as fY1 ; :::; Ym g.
v22;a where v22;a is the upper 100a percentage point of Let Yi ¼ A ^ y ðYi  /^ Þ, i ¼ 1; :::; m, and let
y
the Chi-square distribution with 2 degrees of freedom  
X i ¼ A^ y ðX i  / ^ Þ, i ¼ s; :::; t where / ^ ;A
^ y is the
and a is the Type I error probability. y y

Clearly, both CUSUM and EWMA-type charts solution to

based on the bivariate SN and SR test statistics can be 0  1
X
m A ^
^ y Yi  /
considered. To this end, Li (2015) noted that the 1 B y C
@  2 A ¼ 0
majority of such charts need a pre-specified tuning m i¼1 ^ ^
Ay Y i  / y
parameter (k in the case of a CUSUM chart and k in 0    01
the case of the EWMA chart), which requires a know- ^ ^ ^ 0^
1 X BA y Y i  / y Y i  / y A y C 1
m
ledge about the OOC distribution for implementation @   2 A ¼ Ip
but, typically, the OOC distribution is unknown. This m i¼1 ^
^ y Yi  / p
A y
is in fact also true for their univariate counterparts,
both parametric and nonparametric. Li (2015) pro- where A ^ y is an upper triangular matrix of dimension
posed a nonparametric multivariate chart, that does p  p with positive diagonal elements and a 1 in the
not depend on a tuning parameter, using a hypothesis upper-left element. Similar to Randles (2000), the
testing approach, based on Randles (2000)’s multivari- above transformation makes the direction vectors of
ate SN test. Dovoedo and Chakraborti (2017) noted Y i centered at 0 and approximately uniformly distrib-
that Randle’s test is distribution-free over the class of uted on the p-dimensional unit sphere. If fX s ; :::; X t g
elliptically symmetric distributions and asymptotically and fY 1 ; :::; Y m g have the same distribution then the
distribution-free over the (larger) class of directionally above transformation will also center the direction
symmetric class of distributions. The reader is referred
vectors of X s ;:::;X t g at 0 and make them approxi-
to Randles (2000) for details about these classes of dis- mately uniformly distributed on the p-dimensional
tributions. The directionally symmetric family con- unit sphere. If fX s ; :::; X t g has a different location, say
tains distributions such as the multivariate normal
/x , the direction vectors of X s ;:::;Xt g will cluster
and the multivariate t distributions, and certain
along the direction of A ^ Þ. Define Vi ¼
^ y ð/x  /
skewed distributions. First we give a brief overview of y
Randles (2000)’s multivariate SN test. Assume that Xi =Xi then the test
statistic is given by
0 P
fX 1 ; X 2 ; :::; X n g is a random sample from a p-dimen- pðt  s þ where V s;t ¼ ti¼s Vi =ðt  s þ 1Þ.
1ÞV s;t V s;t
sional continuous population with location /. In The charting statistic is given by
order to test H0 : / ¼ /0 against Ha : / 6¼ /0 let 0
X i ¼ A ^ x ðX i  /0 Þ, i ¼ 1; 2; :::; n where A ^ x is an upper St ¼ max pðt  s þ 1ÞV s;t V s;t
1st
triangular matrix of dimension p  p with positive
diagonal elements and a 1 in the upper-left element and the process is declared OOC when St > UCL and
and satisfies typically post signal diagnostics, i.e., a search for assign-
! able causes would be started. Again, the UCL is found
1X n ^ x ðX i  /0 ÞðX i  /0 Þ0 A
A ^0 1 for a specified nominal ICARL. Li (2015) concludes that
x
¼ Ip
n i¼1 kA ^ x ðX i  /0 Þk 2
p the chart outperforms its nonparametric competitors.
Tables are provided for easy implementation. Note
where k  k and Ip denote the Euclidean norm and the
again that the charting statistic is based on a suitable
identity matrix of dimension p  p, respectively. Note multivariate two-sample distribution-free test statistic
that A ^ x ðX i  /0 Þ=A
^ x ðX i  /0 Þ gives the direction vec-
and there is room for further research along this line.

tor of X i . Under H0 the above transformation centers Dovoedo and Chakraborti (2017) conducted a
X i at 0 and makes the direction vectors, Vi ’s, approxi- study on the performance of multivariate nonparamet-
mately uniformly distributed on the p-dimensional ric Phase II control charts in terms of robustness,
Pn
unit sphere. Under H0 the average V ¼ i¼1 Vi =n under estimated parameters. Such a study is important
should be close to zero and under Ha the Vi ’s should due to the inherent complexity of multivariate tests,
 QUALITY ENGINEERING 15

particularly multivariate nonparametric tests. The into great detail about tests for shape parameters based
authors stated that “While a number of recent studies on a spatial-SN covariance matrix. For brevity we omit
have examined the IC robustness question related to the the details here and simply give the charting statistic of
size of the reference sample for both univariate and their chart. The EWMA-type charting statistic is given by
multivariate normal theory (parametric) charts, in this  
xi ¼ ð1  kÞxi  1 þ kvi v0i x0 ¼ EIC vi v0i (24)
article we study the effect of parameter estimation on
the performance of the multivariate nonparametric SN where 0 < k  1 is the smoothing constant and vi is
EWMA (MSEWMA) chart” (of Zou and Tsung 2011). the unit vector, i.e., the standardized xi observations.
The MSEWMA chart is appealing for several reasons: Thus, vi ¼ U ðA0 ðxi  h0 ÞÞ where h0 is the specified
(i) It is nonparametric, thus much more IC robust than multivariate center vector (affine-equivariant median)
the MEWMA chart; (ii) It is affine-invariant and has a and A0 is the associated transformation matrix. An
strictly distribution-free property over a large class of OOC signal is given when
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
population distributions or models (distributions with 2  k  
elliptical directions). That is, the IC run-length distribu- Qi ¼  tr p  xi  Ip Þ2 > L: (25)
k
tion can attain or is always close to the nominal one,
when the same control limits designed for the multivari- The two design parameters, k and L, are selected so that
ate normal distributions are used; (iii) when the process a nominal ICARL value is attained and this procedure is
distribution is one from the elliptical direction class, the discussed in the EWMA-SN section in more detail. The
ICARL can be computed quickly via a one-dimensional authors conclude that their proposed chart is very efficient,
Markov chain model; (iv) it is fast to implement with a especially for small to moderate shifts. Further work
similar computational effort to the MEWMA chart should consider extending this chart to the parameter
because only a multivariate median and the associated unknown case, which is a common practical situation.
transformation matrix need to be specified (estimated)
from the historical data before monitoring; and (v) it is Other multivariate nonparametric charts
also very efficient in detecting process shifts, especially Charts based on multiple testing
small or moderate shifts when the process distribution Another interesting idea is to use the multiple testing
is heavy tailed or skewed. The ICARL robustness and (comparisons) approach. Park and Jun (2015) proposed a
the OOC shift detection performance are both exam- multivariate nonparametric EWMA chart for monitoring
ined. Based on the results, it is seen that the required the location based on a series of the most recent T 2 statis-
amount of the Phase I data can be very (perhaps imprac- tics. The typical multivariate EWMA chart is given by
tically) high if one wants to use the control limits given Z i ¼ LX i þ ðI  LÞZ i  1 where X i is the data vector
by Zou and Tsung (2011) for the known parameter case with Z 0 ¼ 0 (a zero vector of order p) and L ¼
and maintain a nominal ICARL, which can limit the  
diag l1 ; :::; lp representing a diagonal matrix of p
implementation of these useful charts in practice. To
smoothing constants. Define Ti2 ¼ Z 0i RZi 1 Z i where RZi
remedy this situation, using simulations, the authors
obtain the “corrected for estimation” control limits that denotes the variance-covariance matrix of Z i . If there is
achieve a desired nominal ICARL value when parame- no a-prior information about the smoothing constants,
ters are estimated for a given set of Phase I data. This then one idea is to set them all equal to l so that the vari-
ance-covariance matrix simplifies to
should be very interesting and useful to practitioners. n   o
2i
The OOC performance of the MSEWMA chart with the RZi ¼ l 1  ð1  lÞ =2  l R, which, as i
correct control limits is also studied. The use of the cor- increases, reduces to RZi ¼ fl=2  lgR. The authors
rected control limits with specific amounts of available proposed multiple testing of r hypotheses at time i,
reference sample is recommended. Otherwise, the per- namely:
formance the MSEWMA chart may be seriously affected ðÞ ðÞ
under parameter estimation. H0;ii : li ¼ l0 vs H1;ii : li 6¼ l0
ðÞ ðÞ
H0;ii  1 : li  1 ¼ l0 vs H1;ii  1 : li  1 6¼ l0
EWMA-type control charts in Case K
..
Li et al. (2013) proposed a nonparametric EWMA-type .
multivariate control charting scheme for monitoring ðÞ ðÞ
H0;ii  rþ1 : li  rþ1 ¼ l0 vs H1;ii  rþ1 : li  rþ1 6¼ l0
shape parameters. Their methodology adapts a spatial-
SN covariance matrix to online sequential monitoring by The process is declared OOC if any of the hypothe-
incorporating the EWMA procedure. The authors go ses are rejected. The BH-procedure (see Benjamini
16 S. CHAKRABORTI AND M. A. GRAHAM

and Hochberg 1995) is used to control the FDR of the observations fx1 ; :::; xn g and set Rðxi Þ ¼ 2rðiÞ  n  1
abovementioned multiple testing as close to target in order to center the ranks. Then Rðxi Þ ¼
level q as possible. Then steps for the BH-EWMA Pn
j¼1 signðxi  xj Þ and the well-known Wilcoxon-
scheme, at time i, using an r-span and a target level
Mann-Whitney test statistic, for the difference in loca-
(q) are as follows. First one needs to calculate the sta-
ðiÞ ðiÞ tion between two samples fx1 ; :::; xk g and fxkþ1 ; :::; xn g,
tistics Ti2 ; :::; Ti2 rþ1 corresponding to H0;i ; :::; H0;i  rþ1 P
and the corresponding p-values (pk ). The authors state is given by uk ¼ ki¼1 Rðxi Þ. Next, in order to outline a
“A nonparametric density estimation based on the multivariate generalization of this test statistic, the sign
Parzen windows is adopted to approximate the distri- function is exchanged by a kernel function,
bution of the T-square statistics, from which the p- hðx; yÞ ¼  hðy; xÞ, so that hðx; xÞ ¼ 0 and hðx; yÞ rep-
values are calculated.” The readers are referred to the resents a measure of difference between x and y. Choi
paper for more details. Sort the p-values in ascending and Marden (1997) suggested making use of a direc-
order and let HðkÞ be the null hypothesis correspond- tional rank test by means of the kernel function
ing to pðkÞ ; which is the kth ordered p-value. At the hðx; yÞ ¼ ðx  yÞ=jjx  yjj. Then define Rn ðxi Þ ¼
outset take ^k ¼ 0 and then find ^k, the maximum value Pn
j¼1 hðxi ; xj Þ so that Rn ðxi Þ denotes the directional
of k such that pðkÞ  qk =r. For ^k non-zero the hypoth- rank of xi for i ¼ 1; 2; :::; n. Next, the authors intro-
eses are rejected and the process is declared to be duced notation for defining within-group rank vectors
OOC. The authors perform performance comparisons as follows. For each possible change-point,
via simulation in terms of the OOC and IC average P
k 2 f1; 2; :::; n  1g, let Rn;k ðxi Þ ¼ nj¼kþ1 hðxi ; xj Þ and
run-length according to various non-centrality param-
ðkÞ
Pk
r n ¼ k i¼1 Rn ðxi Þ. The authors then continue with a
1
eters. They conclude by stating that the proposed BH-
EWMA chart outperforms the existing MEWMA detailed discussion on the pooled and unpooled sample
chart. Again, one needs to perform post signal diag- covariance matrices; here we simply mention that the
nostic analysis and this would need further research. directional rank statistic for testing for differences in
Note also that in the above formulation, l0 is assumed location vector between fx1 ; :::; xk g and fxkþ1 ; :::; xn g is
ðkÞ 0  1 ðkÞ
known or given as in Case K. Clearly, a question to rk;n ¼ r n R ^ ^ 1 ¼
given by k;n r n with R
be addressed is the important Case U, that is when l0   k;n
ðn  kÞ=nk R ^ and where R ^ denotes the unpooled
is to be estimated. The application of p-values (moni- n n

toring based on p-values) may be an interesting area covariance matrix. The scheme proposed by Holland
within the SPC literature. and Hawkins (2014) entails maximizing rk;n across pos-
sible change-point values, rmax;c;n ¼
Change-point model based methods. Holland and maxc < k < n  c ðrk;n Þ. Note that c denotes the number of
Hawkins (2014) proposed a Phase II nonparametric observations at the beginning and at the end of
multivariate control chart for monitoring location. Their sequence that is not considered; see Holland and
chart is based on an approximately distribution-free Hawkins (2014) for a detailed motivation of why the so-
multivariate generalization of the Wilcoxon-Mann- called quarantine constant, c, was instigated. The pro-
Whitney test. They extended the work of Hawkins and cess is declared OOC when rmax;c;n > ha;p;c;n and typic-
Deng (2010), who developed a nonparametric univariate ally a search for assignable causes would be started. The
change-point analog based on the Mann-Whitney statis- authors concluded that their proposed scheme outper-
tic, to the multivariate setting. The authors caution that formed its parametric counterpart for small to moderate
although the underlying test statistic is approximately magnitudes of shift.
distribution-free, their procedure may not be suitable for
some multivariate distributions with unusual dependence Control charts using bootstrap. Determining the con-
structures between vector components. trol limits of multivariate nonparametric charts is time
A change-point model for the univariate case was consuming and complicated. Phaladiganon et al. (2011)
given earlier. Now, for the multivariate setting, given used a bootstrap approach, as opposed to kernel density
a sample of p  1 random vectors we have estimation (KDE) in order to find the control limits for
 the multivariate T 2 charts when the assumption of nor-
FðlÞ for i ¼ 1; 2; :::; s
xi  mality does not hold. The basis of bootstrapping is dis-
Fðl þ dÞ for i ¼ s þ 1; s þ [Link]
cussed in the univariate section of this article. Here we
where d represents an arbitrary sustained shift in loca- simply give an outline of the bootstrapping technique in
tion. For p ¼ 1 let rðiÞ denote the rank of xi among the
2ðiÞ 2ðiÞ 2ðiÞ
this case. Let T1 ; T2 ; :::; Tn be a set of n T 2 values
 QUALITY ENGINEERING 17

from the ith bootstrap sample, i ¼ 1; :::; B (with B typic- analysis (KLDA), respectively, in order to make the
ally being a large number), which are drawn randomly charting statistics continuous. Although these proposed
from the initial T 2 statistics (with replacement). For charts outperformed the traditional RTC chart, they do
each bootstrap sample the 100ð1  aÞth percentile value is not have the advantage of being able to be applied to
calculated with a being a value specified by the practi- various data types such as categorical data and missing
tioner beforehand. The control limit is determined by data. In order to overcome this problem, Jang, Park,
taking an average of the B percentile values, say and Baek (2017) proposed improved RTC charts mak-
T 2 100ð1  aÞ . The process is declared to be OOC if a chart- ing use of random forests with weighted voting. These
ing statistic plots beyond T 2 100ð1  aÞ . The authors found improved charts outperformed the traditional RTC
that the bootstrap technique outperforms the traditional chart and was shown to be more effective than the RTC
T 2 chart in all cases and performed similarly to the charts based on SVMS and KLDA.
KDE-based T 2 chart. Davila, Runger, and Tuv (2014) also made use of a
supervised learning approach and showed a practical
Miscellaneous applications. As the realm of monitor- application in public health surveillance which does
ing applications have grown and have become more not involve low dimensional data (which is usually
pervasive in many spheres of life, data collection has the case with application in public health surveillance)
become easier and quicker and computational resour- where 25 decision trees were used to monitor counts
ces have grown faster, several researchers have devel- of a disease. Chen, Zi, and Zou (2016) also showed a
oped distribution-free control charts for monitoring practical application of a semi-conductor manufactur-
big data and for monitoring higher dimensional pro- ing process which involved 591 variables and cau-
cess data. We briefly mention a few of these here. tioned against the assumption that all 591 variables
Zhang, Tsung, and Zou (2015) considered a method are normally distributed.
making use of both past IC and OOC data, where the Kang and Kim (2013) proposed a control chart that
charting statistic and the control limits were computed makes use of a k-means clustering analysis which sys-
making use of support vector machines (SVMs) meth- tematically divides the data into clusters by minimiz-
odology. One issue with this approach is that the real- ing within-group variation and by maximizing
time (current) Phase II data is not included in the calcu- between-group variation. The charting statistic for this
P
lation of their control limits and the classifier is only chart, sðxÞ ¼ mink ðx  lk Þ0 k 1 ðx  lk Þ where x is
trained once, based on a single data set, at the beginning the new observation that has not been classified into a
P
of the monitoring process. Note that earlier, Deng, cluster yet and lk and k are the mean vector and
Runger, and Tuv (2012) introduced a supervised learn- the covariance matrix of the kth cluster, respectively.
ing approach which is based on constantly updating the The number of clusters, k, should be pre-selected by
classifier with real-time data. They referred to this the practitioner, although Kang and Kim (2013)
approach as the RTC since it is based on real-time con- showed that making use of different values for k does
trasts between the IC reference data and the real-time not significantly change the result. A bootstrap
(current) data. RTC assigns one class label to the IC ref- method was used in order to find the control limit.
erence data and another class label to some available When the joint distribution of multiple quality
real-time data which converts the monitoring problem characteristics is unknown, Liang, Xiang, and Pu
into a dynamic classification problem. The advantage in (2016) have extended on the idea of incorporating the
doing this is that, in the previous literature, a classifier is least absolute shrinkage and selection operator
trained once based on a single data set at the beginning (LASSO) into the EWMA monitoring scheme
of the monitoring process but in the RTC approach a (referred to as the LEWMA chart), which was pro-
classifier is retrained with each new observation and the posed by Zou and Qiu (2009), by developing a robust
statistic (such as the classification error rate) that is counterpart of it that has an affine-invariance prop-
monitored is generated using the generalized likeli- erty and is nonparametric under the elliptical distribu-
hood-ratio principle. Although the RTC approach is tion family in that the control limit can be obtained
useful and can be applied to a variety of monitoring by making use of a standard multivariate normal dis-
problems, the typical problems with making use of a tribution. Since a sign statistic is used, the chart is
discrete charting statistic arises, i.e., that it might be less referred to as the SLEWMA chart. This chart, used
efficient in fault detection. He, Jiang, and Deng (2016) for Phase II monitoring, is compared to the LEWMA
and Wei et al. (2016) proposed distance-based control chart and the multivariate direction sign EWMA chart
charts based on SVMs and kernel linear discriminant (denoted MDSE), which was proposed by Zi et al.
18 S. CHAKRABORTI AND M. A. GRAHAM

(2013), and not only is the computation of the control for Papers” for this special issue, “Traditional control
limits for the SLEWMA much simpler than the other charts require the assumption that the process distri-
charts, but there are many cases where the proposed bution follows a parametric form (e.g., normal). In
SLEWMA chart outperforms its competitors. Future practice, however, this assumption may not hold, in
research could include the modification of Liang et al. other words, the process may not follow the pre-speci-
(2016)’s chart so that it can be used for Phase I data. fied parametric distribution. In the literature, it has
In general, multivariate control charting, particu- been well demonstrated that results from the trad-
larly multivariate nonparametric control charting, itional control charts using the pre-specified distribu-
remains an area open for more contributions. There is tion in their design may not be reliable because their
need for more theoretical insights and practical rec- actual false alarm rates could be substantially larger or
ommendations. For example, more studies are needed smaller than the nominal false alarm rate. A direct
for many proposed charts, along the lines of Dovoedo consequence of this could be that much labor and
and Chakraborti (2017), to understand the effects of many resources are wasted, or that many defective
parameter estimation, the amount of required data for products are manufactured without notice. Therefore,
nominal performance and perhaps considering how to in cases when no parametric form of the process dis-
adapt or modify the charts for a given amount of tribution is available or when no parametric form is
data. Post signal diagnostics is also an important area validated properly beforehand, control charts without
of further research. There is also the need for case requiring the specification of a parametric form for
studies and detailed explanations of what seems like a the process response distribution, or simply nonpara-
complicated methodology. To all of these ends, above metric (distribution-free) statistical process control
all, software development is urgently needed. (NSPC) charts, should be considered.”
We feel that nonparametric statistics in general and
nonparametric control charts in particular have much
Summary and recommendations
to offer to the quality practitioner. Thus the updated
NSPC charts provide a robust alternative for statistical overview of the available literature will be of interest
process monitoring in practice when the form of the and value to a broad spectrum of readers. These
underlying distribution is unknown. The goal of this charts can be beneficial when parametric model
article is to bring the review of nonparametric control assumptions cannot be objectively justified. They have
charting techniques, of Chakraborti et al. (2001, 2007 stable IC properties, are robust against outliers and
and 2011), forward to 2017, covering some of the their efficiency can be quite high. Further work in
major advances and contributions. In doing so, we NSPC research and practice are encouraged and will
found that while a lot of progress has been made in be welcome. As we have noted in various places in
the field of NSPC research, NSPC has not been fully the paper, there are a lot of open problems. Two
embraced in SPC practice. To this end, Woodall and major concerns are the effects of parameter estimation
Montgomery (2014) stated “Despite their advantages and examination of the reference sample size require-
in reducing the distributional assumptions required to ments. There has been very few work in this direction.
design control charts with specified IC performance, it It is generally agreed that nonparametric charts
does not seem that nonparametric methods are gain- require more data but a systematic study will be valu-
ing a foothold with practitioners. This could partially able. From the practical side, nonparametric charts do
be due to a lack of available statistical software for not have a heavy computation burden and require a
implementing the methods, a lack of familiarity, and a lot of extra computation compared to their parametric
lack of textbook coverage. Nevertheless, this research counterparts, but there is not much software available
area remains active.” To address these concerns, at the moment to apply all the control charts that are
Chakraborti, Qiu, and Mukherjee (2015), in the intro- available. This would be an important area for future
duction to a special issue on Nonparametric Statistical contributions.
Process Control Charts, that appeared in a recent
issue of Quality and Engineering Reliability
Available software
Engineering International, wrote, “Motivated by such
observations, our aim has been to bring NSPC charts Some software are now available in the area of non-
to the mainstream SPC arena, since we strongly feel parametric SPC. A vast majority of them is written in
that it has much to offer to both the SPC practitioner R. Most of these are focused on researchers and their
and the researcher. To reiterate, as stated in the “Call own work, without much standardization, including
 QUALITY ENGINEERING 19

notation and terminology. This limits their availability Acknowledgments

and accessibility to some degree. As a consequence, it
The authors thank the editor and two anonymous reviewers
is not easy for the interested users to apply nonpara- for their comments and suggestions that have improved
metric charts in routine SPC. In any case, we list clarity and presentation.
some of these R packages below which can be down-
loaded from the CRAN archives. More work is neces-
sary in this arena to make the full power of NSPC
About the authors
available to the users. S. Chakraborti holds a PhD in Statistics from the State
University of New York at Buffalo. He is Professor of
Statistics and Robert C. and Rosa P. Morrow Faculty
Excellence Fellow in the Department of Information Systems,
Statistics and Management Science at the Culverhouse
College of Commerce and Business Administration at the
Name Brief description University of Alabama, Tuscaloosa. He is a Fellow of the
dfphase1 Phase I control charts (with emphasis on American Statistical Association, an Elected member of the
distribution-free methods)
Includes statistical methods for retrospect-
International Statistical Institute and a Fulbright Senior
ively detecting changes in location and/or Scholar to South Africa. His specialty areas are
dispersion of univariate and multivariate Nonparametric and Robust Statistical Inference with applica-
variables. Data values are assumed to be tions in areas such as Statistical Process Control, Survival/
independent, can be individual (one
observation at each instant of time) or
Reliability Analysis, Econometrics, Statistical Computing, and
subgrouped (more than one observation Extreme Values. He has over one hundred publications in a
at each instant of time). Control limits are variety of outlets, including national and international peer-
computed, often using a permutation review journals and is a co-author of the book Nonparametric
approach, so that a prescribed false alarm Statistical Inference, fifth edition, published by Marcel
probability is guaranteed without making
any parametric assumptions on the stable Dekker. He has been a visiting professor at a great number of
(in-control) distribution. universities abroad, in India, Brazil, France, South Africa and
[Link] Methods for nonparametric change-point Turkey and has won a number of teaching and research excel-
detection lence awards. He has been serving as an Associate Editor of
another alternative for Phase I analysis -
only univariate, only individual Communications in Statistics for over fifteen years. He is a
observations member of the American Statistical Association and the
Cpm change-point methods - parametric and International Statistical Institute. His e-mail address is
nonparametric - both Phase I and Phase II schakrab@[Link].
NPMVCP nonparametric multivariate change-point
detection in Phase II
M. A. Graham holds a PhD in Mathematical Statistics
Spcadjust implements a bootstrap based method to from the University of Pretoria, South Africa, and holds a
adjust monitoring schemes to take esti- Y1 rating from the National Research Foundation (NRF).
mation error into account. Although the She has published in several accredited peer-reviewed jour-
spcadjust is not for nonparametric charts nals and has presented her research at national and inter-
specifically, they say that the package
covers the most common setups of national conferences. She has supervised over twenty
Shewhart, CUSUM and EWMA charts and Master's and Ph.D. students and her research interests are
that it is easy to add further charts in Statistical Process Control and Nonparametric Statistics.
SPC Model Nonpar updates (Xt  l  D=2Þ=r , no distribu-
Center Scale tional assumptions
SPCModelNonpar user defined updates, no distributional
assumptions References
Qcr Allows to generate Shewhart-type charts and
to obtain numerical results of interest for Amin, R. W., and O. Widmaier. 1999. Sign control charts
a process quality control (involving with variable sampling intervals. Communications in
continuous, attribute or count data). Statistics: Theory and Methods 28 (8):1961–85. doi:
This package provides basic functionality 10.1080/03610929908832398.
for univariable and multivariable
quality control analysis, including: xbar, Amin, R. W., M. R. Reynolds, Jr., and S. T. Bakir. 1995.
xbar-one, S, R, ewna, cusum, Nonparametric quality control charts based on the sign
mewna, mcusum and T2 charts. statistic. Communications in Statistics: Theory and Methods
Additionally have nonparametric 24 (6):1597–623. doi:10.1080/03610929508831574.
control charts multivariate
Mnspc Statistical process control for multivariate Bakir, S. T. 2004. A distribution-free shewhart quality con-
data that is not necessarily Gaussian dis- trol chart based on signed-ranks. Quality Engineering 16
tributed. A function is provided to cat- (4):613–23. doi:10.1081/QEN-120038022.
egorize components of multivariate Bakir, S. T. 2011. Distribution-free (nonparametric) quality
response vectors. Tools for setting up a
CUSUM procedure for the transformed
control charts part I.
data are included. The CUSUM scheme Bakir, S. T., and M. R. Reynolds. Jr. 1979. A nonparametric
can also be applied to the case when procedure for process control based on within-group
some (or all) of the multivariate response ranking. Technometrics 21 (2):175–83. doi:10.1080/
components are binary-categorical.
00401706.1979.10489747.
20 S. CHAKRABORTI AND M. A. GRAHAM

Bell, R. C., L. A. Jones-Farmer, and N. Billor. 2014. A distribu- Chen, N., X. Zi, and C. Zou. 2016. A distribution-free
tion-free multivariate phase I location control chart for sub- multivariate control chart. Technometrics 58 (4):448–59.
grouped data from elliptical distributions. Technometrics 56 doi:10.1080/00401706.2015.1049750.
(4):528–38. doi:10.1080/00401706.2013.879264. Cheng, C.-R., and J.-J. H. Shiau. 2015. A distribution-free
Benjamini, Y., and Y. Hochberg. 1995. Controlling the false multivariate control chart for phase I applications.
discovery rate: A practical and powerful approach to Quality and Reliability Engineering International 31 (1):
multiple testing. Journal of the Royal Statistical Society, 97–111. doi:10.1002/qre.1751.
Series B (Methodological ) 57 (1):289–300. Choi, K., and J. Marden. 1997. An approach to multivariate
Bennett, B. M. 1964. Non-parametric test for randomness in rank tests in multivariate analysis of variance. Journal of
a sequence of multinomial trials. Biometrics 20 (1): the American Statistical Association 92 (440):1581–90.
182–90. doi:10.2307/2527626. doi:10.1080/01621459.1997.10473680.
Boone, J. M., and S. Chakraborti. 2012. Two simple she- Chowdhury, S., A. Mukherjee, and S. Chakraborti. 2015.
whart-type multivariate nonparametric control charts. Distribution-free phase II CUSUM control chart for joint
Applied Stochastic Models in Business and Industry 28 (2): monitoring of location and scale. Quality and Reliability
130–40. doi:10.1002/asmb.900. Engineering International 31 (1):135–51. doi:10.1002/
Capizzi, G., and G. Masarotto. 2013. Phase I distribution- qre.1677.
free analysis of univariate data. Journal of Quality Coelho, M. L. I., M. A. Graham, and S. Chakraborti. 2017.
Technology 45 (3):273–84. doi:10.1080/ Nonparametric signed-rank control charts with variable
00224065.2013.11917938. sampling intervals. Quality and Reliability Engineering
Capizzi, G., and G. Masarotto. 2017. Phase I distribution- International 33 (8):2181–92. doi:10.1002/qre.2177.
free analysis of multivariate data. Technometrics 57 (4): Davila, D., S. Runger, G. Tuv. and E. 2014. Public health
484–95. doi:10.1080/00401706.2016.1272494. surveillance with ensemble-based supervised learning. IIE
Chakraborti, S., and M. A. Graham. 2007. Nonparametric Transactions 46 (8):770–89. doi:10.1080/
control charts. In Encyclopedia of statistics in quality and 0740817X.2014.894806.
reliability, Vol. 1, 415–429. New York: John Wiley. Das, N. 2008. Nonparametric control chart for controlling
Chakraborti, S., and M. A. Van de Wiel. 2008. A nonpara-
variability based on rank test. Economic Quality Control
metric control chart based on the Mann-Whitney statis-
23 (2):227–42.
tic. IMS Collections. Beyond Parametrics in
Das, N. 2009. A new multivariate nonparametric control
Interdisciplinary Research: Festschrift in Honor of
chart based on the sign test. Quality Technology &
Professor Pranab K. Sen 1:156–72.
Quantitative Management 6 (2):155–69. doi:10.1080/
Chakraborti, S., S. Eryilmaz, and S. W. Human. 2009. A
16843703.2009.11673191.
phase II nonparametric control chart based on prece-
Deng, H., G. Runger, and E. Tuv. 2012. System monitoring
dence statistics with runs-type signaling rules.
with Real-Time contrasts. Journal of Quality Technology
Computational Statistics and Data Analysis 53 (4):
44 (1):9–27. doi:10.1080/00224065.2012.11917878.
1054–65. doi:10.1016/[Link].2008.09.025.
Dietz, E. J. 1982. Bivariate nonparametric tests for the one-
Chakraborti, S., S. W. Human, and M. A. Graham. 2009.
Phase I statistical process control charts: An overview sample location problem. Journal of the American
and some results. Quality Engineering 21 (1):52–62. doi: Statistical Association 77 (377):163–9. doi:10.1080/
10.1080/08982110802445561. 01621459.1982.10477781.
Chakraborti, S., S. W. Human, and M. A. Graham. 2011. Diko, M. D., S. Chakraborti, and M. A. Graham. 2016.
Nonparametric (distribution-free) quality control charts. Monitoring the process mean when standards are
In Handbook of methods and applications of statistics: unknown: A classic problem revisited. Quality and
Engineering, Quality control, and physical sciences, ed. N. Reliability Engineering International 32 (2):609–22. doi:
Balakrishnan, 298–329. New York: John Wiley & Sons. 10.1002/qre.1776.
Chakraborti, S., P. Qiu, and A. Mukherjee. 2015. Dovoedo, Y. H., and S. Chakraborti. 2017. Effects of param-
Introduction to the special issue: Nonparametric statis- eter estimation on the multivariate distribution-free phase
tical process control charts. Quality and Reliability II sign EWMA chart. Quality and Reliability Engineering
Engineering International 31 (1):1–2. doi:10.1002/ International 33 (2):431–49. doi:10.1002/qre.2019.
qre.1759. Ghute, V. B., and D. T. Shirke. 2012. A nonparametric
Chakraborti, S., P. Van der Laan, and S. T. Bakir. 2001. signed-rank control chart for bivariate process location.
Nonparametric control charts: An overview and some Quality Technology & Quantitative Management 9 (4):
results. Journal of Quality Technology 33 (3):304–15. doi: 317–28. doi:10.1080/16843703.2012.11673296.
10.1080/00224065.2001.11980081. Gibbons, J. D., and S. Chakraborti. 2010. Nonparametric
Chakraborti, S., P. Van Der Laan, and M. A. Van de Wiel. statistical inference. 5th ed. Boca Raton, FL: Taylor and
2004. A class of distribution-free control charts. Journal Francis.
of the Royal Statistical Society: Series C (Applied Statistics) Graham, M. A., S. Chakraborti, and S. W. Human. 2011a.
53 (3):443–62. doi:10.1111/j.1467-9876.2004.0d489.x. A nonparametric EWMA sign chart for location based
Chakraborty, N., S. Chakraborti, S. W. Human, and N. on individual measurements. Quality Engineering 23 (3):
Balakrishnan. 2016. A generally weighted moving average 227–41. doi:10.1080/08982112.2011.575745.
signed-rank control chart. Quality and Reliability Graham, M. A., S. Chakraborti, and S. W. Human. 2011b.
Engineering International 32 (8):2835–45. doi:10.1002/ A nonparametric exponentially weighted moving average
qre.1968. signed-rank chart for monitoring location. Computational
 QUALITY ENGINEERING 21

Statistics and Data Analysis 55 (8):2490–503. doi:10.1016/ processes. International Journal of Production Research 51
[Link].2011.02.013. (18):5644–5657. doi:10.1080/00207543.2013.793427.
Graham, M. A., S. W. Human, and S. Chakraborti. 2010. A Kapatou, A., and M. R. Reynolds. Jr. 1994. Multivariate
phase I nonparametric shewhart-type control chart based nonparametric control charts using small sample sizes.
on the median. Journal of Applied Statistics 37 (11): Proceedings of the Section on Quality and Productivity,
1795–813. doi:10.1080/02664760903164913. American Statistical Association 241–6.
Graham, M. A., A. Mukherjee, and S. Chakraborti. 2012. Kapatou, A., and M. R. Reynolds. Jr. 1998. Multivariate
Distribution-free exponentially weighted moving average nonparametric control charts for the case of known R.
control charts for monitoring unknown location. Proceedings of the Section on Quality and Productivity,
Computational Statistics and Data Analysis 56 (8): American Statistical Association 77:82.
2539–61. doi:10.1016/[Link].2012.02.010. Khilare, S. K., and D. T. Shirke. 2010. A nonparametric syn-
Graham, M. A., A. Mukherjee, and S. Chakraborti. 2017. thetic control chart using sign statistic. Communications
Design and implementation issues for a class of distribu- in Statistics: Theory and Methods 39 (18):3282–93. doi:
tion-free phase II EWMA exceedance control charts. 10.1080/03610920903249576.
International Journal of Production Research 55 (8): Khilare, S. K., and D. T. Shirke. 2012. Nonparametric syn-
2397–430. doi:10.1080/00207543.2016.1249428. thetic control charts for process variation. Quality and
Hamurkarouglu, C., M. Mert, and Y. Sayken. 2004. Reliability Engineering International 28 (2):193–202. doi:
Nonparametric control charts based on mahalanobis 10.1002/qre.1233.
depth. Hacettepe Journal of Mathematics and Statistics 33: Knoth, S. 2016. The case against the use of synthetic control
57–67. charts. Journal of Quality Technology 48 (2):178–95. doi:
Hawkins, D. M., and Q. Deng. 2010. A nonparametric 10.1080/00224065.2016.11918158.
change-point control chart. Journal of Quality Technology Lepage, Y. 1971. A combination of wilcoxon’s and Ansari-
42 (2):165–73. doi:10.1080/00224065.2010.11917814. Bradley’s statistics. Biometrika 58 (1):213–7. doi:10.1093/
Hawkins, D. M., P. Qiu, and C. W. Kang. 2003. The biomet/58.1.213.
changepoint model for statistical process control. Journal Li, J. 2015. Nonparametric multivariate statistical process
of Quality Technology 35 (4):355–66. doi:10.1080/ control charts: a hypothesis testing-based approach.
00224065.2003.11980233. Journal of Nonparametric Statistics 27 (3):384–400. doi:
Hayter, A. J., and K.-L. Tsui. 1994. Identification and quan- 10.1080/10485252.2015.1062889.
tification in multivariate quality control problems. Li, Z., Y. Dai, and Z. Wang. 2014. Multivariate change point
Journal of Quality Technology 26(3):197–208. doi:10.1080/ control chart based on data depth for phase I analysis.
00224065.1994.11979526. Communications in Statistics – Simulation and Computation
He, S., W. Jiang, and H. Deng. 2016. A distance-based con- 43 (6):1490–507. doi:10.1080/03610918.2012.735319.
trol chart for monitoring multivariate processes using Li, Z., C. Zou, Z. Wang, and L. Huwang. 2013. A multivari-
support vector machines. Annals of Operations Research ate sign chart for monitoring process shape parameters.
1–17. Journal of Quality Technology 45 (2):149–65. doi:10.1080/
Holland, M. D., and D. M. Hawkins. 2014. A control chart 00224065.2013.11917923.
based on a nonparametric multivariate change-point Liang, W., D. Xiang, and X. Pu. 2016. A robust multivariate
model. Journal of Quality Technology 46 (1):63–77. doi: EWMA control chart for detecting sparse mean shifts.
10.1080/00224065.2014.11917954. Journal of Quality Technology 48 (3):265–83. doi:10.1080/
Jensen, W. A., J. B. Birch, and W. H. Woodall. 2007. High 00224065.2016.11918166.
breakdown estimation methods for phase I multivariate Liu, L.,. X. Zi, J. Zhang, and Z. Wang. 2013. A sequential rank-
control charts. Quality and Reliability Engineering based nonparametric adaptive EWMA control chart.
International 23 (5):615–29. doi:10.1002/qre.837. Communications in Statistics – Simulation and Computation
Jang, S., S. H. Park, and J. G. Baek. 2017. Real-time con- 42 (4):841–59. doi:10.1080/03610918.2012.655829.
trasts control chart using random forests with weighted Liu, L.,. B. Chen, J. Zhang, and X. Zi. 2015. Adaptive phase
voting. Expert Systems with Applications 71:358–69. doi: II nonparametric EWMA control chart with variable
10.1016/[Link].2016.12.002. sampling interval. Quality and Reliability Engineering
Jones-Farmer, L. A., V. Jordan, and C. W. Champ. 2009. International 31 (1):15–26. doi:10.1002/qre.1742.
Distribution-free phase I control charts for subgroup Liu, L.,. J. Zhang, and X. Zi. 2015. Dual nonparametric
location. Journal of Quality Technology 41 (3):304–16. CUSUM control chart based on ranks. Communications
doi:10.1080/00224065.2009.11917784. in Statistics - Simulation and Computation 44 (3):756–72.
Jones-Farmer, L. A., and C. W. Champ. 2010. A distribu- doi:10.1080/03610918.2013.784985.
tion-free phase I control chart for subgroup scale. Journal Liu, R. 1995. Control charts for multivariate processes.
of Quality Technology 42 (4):373–87. doi:10.1080/ Journal of the American Statistical Association 90 (432):
00224065.2010.11917834. 1380–7. doi:10.1080/01621459.1995.10476643.
Jones-Farmer, L. A., W. H. Woodall, S. Steiner, and C. W. Lu, S.-L. 2015. An extended nonparametric exponentially
Champ. 2014. An overview of phase I analysis for process weighted moving average sign control chart. Quality and
improvement and monitoring. Journal of Quality Technology Reliability Engineering International 31 (1):3–13. doi:
46 (3):265–80. doi:10.1080/00224065.2014.11917969. 10.1002/qre.1673.
Kang, J. H., and S. B. Kim. 2013. A clustering algorithm-based Lucas, J. M., and M. S. Saccucci. 1990. Exponentially
control chart for inhomogeneously distributed TFT-LCD weighted moving average control schemes: Properties and
22 S. CHAKRABORTI AND M. A. GRAHAM

enhancements. Technometrics 32 (1):1–12. doi:10.1080/ Ross, G. J., and N. M. Adams. 2012. Two nonparametric
00401706.1990.10484583. control charts for detecting arbitrary distribution changes.
Malela-Majika, J. C., S. Chakraborti, and M. A. Graham. Journal of Quality Technology 44 (2):102–16. doi:10.1080/
2016. Distribution-free phase II Mann-Whitney control 00224065.2012.11917887.
charts with runs-rules. The International Journal of Ross, G. J., D. K. Tasoulis, and N. M. Adams. 2011.
Advanced Manufacturing Technology 86 (1–4):723–35. Nonparametric monitoring of data streams for changes
doi:10.1007/s00170-015-8083-1. in location and scale. Technometrics 53 (4):379–89. doi:
McCracken, A. K., and S. Chakraborti. 2013. Control charts 10.1198/TECH.2011.10069.
for joint monitoring of mean and variance: an overview. Siegel, S., and J. W. Tukey. 1960. Nonparametric sum of
Quality Technology & Quantitative Management 10 (1): ranks procedure for relative spread in unpaired samples.
17–36. doi:10.1080/16843703.2013.11673306. Journal of the American Statistical Association 55 (291):
Mood, A. M. 1954. On the asymptotic efficiency of certain 429–45. doi:10.2307/2281906.
non-parametric two sample test. The Annals of Sheu, S.-H., and T.-C. Lin. 2003. The generally weighted
Mathematical Statistics 25 (3):514–22. doi:10.1214/aoms/ moving average control chart for detecting small shifts in
1177728719. the process mean. Quality Engineering 16 (2):209–31. doi:
Mukherjee, A., and S. Chakraborti. 2012. A distribution-free 10.1081/QEN-120024009.
control chart for the joint monitoring of location and Sullivan, J. H., and W. H. Woodall. 1996. A comparison of
scale. Quality and Reliability Engineering International 28 multivariate control charts for individual observations.
(3):335–52. doi:10.1002/qre.1249. Journal of Quality Technology 28 (4):398–408. doi:
Mukherjee, A., M. A. Graham, and S. Chakraborti. 2013. 10.1080/00224065.1996.11979698.
Distribution-free exceedance CUSUM control charts for Sullivan, J. H., and W. H. Woodall. 2000. Change-point
location. Communications in Statistics – Simulation and detection of mean vector or covariance matrix shifts
Computation 42 (5):1153–87. doi:10.1080/ using multivariate individual observations. IIE
03610918.2012.661638. Transactions – Quality and Reliability Engineering 32 (6):
Park, J., and C. H. Jun. 2015. A new multivariate EWMA 537–49. doi:10.1080/07408170008963929.
control chart via multiple testing. Journal of Process Wei, Q., W. Huang, W. Jiang, and W. Zhao. 2016. Real-
Control 26:51–5. doi:10.1016/[Link].2015.01.007. time process monitoring using kernel distances.
Pawar, V. Y., and D. T. Shirke. 2010. A nonparametric she- International Journal of Production Research 54 (21):
whart-type synthetic control chart. Communications in 6563–78. doi:10.1080/00207543.2016.1173257.
Statistics: Simulation and Computation 39 (8):1493–505. Western Electronic Company 1956. Statistical quality con-
doi:10.1080/03610918.2010.503014. trol handbook. Indianapolis, IN: Western Electric
Phaladiganon, P., S. B. Kim, V. C. Chen, J. G. Baek, and S. K. Corporation.
Park. 2011. Bootstrap-based T2 multivariate control charts. Woodall, W. H., and D. C. Montgomery. 2014. Some cur-
Communications in Statistics – Simulation and Computation rent directions in the theory and application of statistical
40 (5):645–62. doi:10.1080/03610918.2010.549989. process monitoring. Journal of Quality Technology 46 (1):
Psarakis, S. 2015. Adaptive control charts: Recent develop- 78–94. doi:10.1080/00224065.2014.11917955.
ments and extensions. Quality and Reliability Engineering Wu, Z., and T. A. Spedding. 2000. A synthetic control chart
International 31 (7):1265–80. doi:10.1002/qre.1850. for detecting small shifts in the process mean. Journal of
Puri, M. L., and P. K. Sen. 1976. Nonparametric methods in Quality Technology 32 (1):32–8. doi:10.1080/
multivariate analysis. New York, NY: John Wiley & Sons 00224065.2000.11979969.
Inc. Zhang, C., F. Tsung, and C. Zou. 2015. A general frame-
Qiu, P. 2014. Introduction to statistical process control. work for monitoring complex processes with both in-
Chapman and Hall. control and out-of-control information. Computers &
Qiu, P., and D. M. Hawkins. 2003. A nonparametric multi- Industrial Engineering 85:157–68. doi:10.1016/
variate cumulative sum procedure for detecting shifts in [Link].2015.03.007.
all directions. Journal of the Royal Statistical Society: Zi, X., C. Zou, Q. Zhou, and J. Wang. 2013. A directional
Series D (the Statistician) 52 (2):151–64. doi:10.1111/ multivariate sign EWMA control chart. Quality
1467-9884.00348. Technology & Quantitative Management 10 (1):115–32.
Qiu, P. 2008. Distribution-free multivariate process control doi:10.1080/16843703.2013.11673311.
based on log-linear modeling. IIE Transactions 40 (7): Zou, C., and P. Qiu. 2009. Multivariate statistical process control
664–77. doi:10.1080/07408170701744843. using LASSO. Journal of the American Statistical Association
Randles, R. H. 2000. A simpler, affine-invariant, multivariate, 104(488):1586–96. doi:10.1198/jasa.2009.tm08128.
distribution-free test. Journal of the American Statistical Zou, C., and F. Tsung. 2010. Likelihood ratio-based distribution-
Association 95 (452):1263–8. doi:10.2307/2669766. free EWMA control charts. Journal of Quality Technology 42
Reynolds, M. R., Jr., R. W. Amin, J. C. Arnold, and J. A. (2):174–96. doi:10.1080/00224065.2010.11917815.
Nachlas. 1988. Charts with variable sampling intervals. Zou, C., and F. Tsung. 2011. A multivariate sign EWMA
Technometrics 30 (2):181–92. doi:10.1080/00401706.1988. control chart. Technometrics 53 (1):84–97. doi:10.1198/
10488366. TECH.2010.09095.