Stochastics and Quality Control 2019; 34(2): 95ś102

Research Article

Sai Sarada Vedururu*, M. Subbarayudu and K. V. S. Sarma

A New Method of Estimating the Process

Capability Index with Exponential Distribution
Using Interval Estimate of the Parameter
https://doi.org/10.1515/eqc-2019-0005
Received February 23, 2019; revised September 16, 2019; accepted September 16, 2019

Abstract: This paper deals with a new method of deriving the Process Capability Index (PCI) when the quality
characteristic X follows a positively skewed distribution. The focus of the paper is to derive a new estimate
of PCI by taking into account the 100(1 − α) % Conődence Intervals (CI) of the parameter (s) and arriving at
a new expression. The formula C s , proposed by Wright (1995) which contains a component for skewness, is
reexamined and a new estimate is constructed by utilizing the lower, middle and upper values of the CI of
the parameter. The weighted average of the three possible estimates of C s is proposed as the new estimate
by taking the weights inversely proportional to the squared deviation from the hypothetical value of C s . The
properties of the estimate are studied by simulation using one parameter exponential distribution.

Keywords: Exponential Distribution, Conődence Intervals, Process Capability Index

MSC 2010: 62P30

1 Introduction
Let F(x) be the probability distribution function of a process characteristic, e.g., task completion time with
mean μ and standard deviation σ. One of the measures of process performance is the Process Capability Index
(PCI) given by
USL − LSL
Cp = , (1.1)
6σ
where LSL, USL denote the lower and upper speciőcation limits. It is assumed that the process operates within
LSL and USL according to a normal distribution. A simple estimate of C p is obtained by replacing σ with the
sample standard deviation. A discussion on PCI measures can be found in [5].
Vännman [11] proposed a new expression for the PCI by taking into account both the eccentricity and
spread, calling it a super structure given by
d − u℘μ − M℘
C p (u, v) = , (1.2)
3√ σ2 ⋇ v(μ − T)2

where d = USL 2− LSL , M = USL 2⋇ LSL and (u, v) ≥ 0 are constants. Note that when u = v = 0, (1.2) reduces to (1.1).
The formula in (1.2) does not work well for non-normal distributions. Duncan [3] and Montgomery [5] have
discussed some data transformations like BoxśCox’s or Johnson’s method to bring in normality of the data.

*Corresponding author: Sai Sarada Vedururu, Department of Statistics, S. V. University, Tirupati-517 502, India,
e-mail: [email protected]
M. Subbarayudu, Department of Statistics, S. V. University, Tirupati-517 502, India, e-mail: [email protected]
K. V. S. Sarma, Biostatistician, Sri Venkateswara Institute of Medical Sciences, Tirupati-517 502, India,
e-mail: [email protected]

Brought to you by | Göteborg University - University of Gothenburg

Authenticated
Download Date | 12/4/19 5:06 AM
96 | S. S. Vedururu et al., A New Method of Estimating the Process Capability Index

Wright’s method [13]. This method makes use of skewness of the distribution to handle non-normality, using
the third central moment (μ3 ) in addition to μ and σ and given as
d − ℘μ − T℘
Cs = . (1.3)
3√ σ2 ⋇ (μ − T)2 ⋇ ℘ μσ3 ℘

This formula can be used for both skewed and symmetric distributions as well. Pearn and Chang [6] extended
the Wright’s method to some skewed distributions and found that the bias in the estimate of C s increases
when the skewness increases. A detailed review of various PCI measures can be found in [1].
This paper aims at proposing a new estimate of C s by utilizing the 100(1 − α) % Conődence Interval (CI)
of the process parameter (s). In essence, we wish to replace the point estimates with interval estimates and
arrive at a new formula for the PCI.

2 A Novel View of PCI Estimation

Let Θ = {θ1 , θ2 , . . . , θ k } be the set of k parameters of F(x), x = {x1 , x2 , . . . , x n } a random sample of size n from
the process and C = {C1 , C2 , . . . , C m } a set of pre-deőned constants unrelated to the process distribution.
Deőne g(C, Θ) to be a metric for the process capability (like PCI). For instance, the classical C p given in (1.1)
has k = 1 and m = 2 (LSL and ULS), while formula (1.2) has two parameters (k = 2) and őve constants (m = 5).
A moment estimator of g(C, Θ) is g(C, Θ), ̂ where Θ̂ is the set of point estimates θ̂ j of θ j for j = 1, 2, . . . , k.
Koch and Link [4] observed that the single-valued point estimator is very likely to be incorrect in some
practical situations and hence an interval estimator allowing a range of values is often useful. When the point
estimates of process center and spread turn out to be incorrect, the estimate of PCI itself becomes unreliable.
J. Sinsomboonthong [10] has addressed issues of constructing CIs using a method calling it TestSTAT method.
He observed that the coverage probability and the expected length of CI by this method were similar to those
obtained by the method of exact CIs.

Remark 2.1. In general, the interval estimator of a parameter θ is speciőed with two end points l x and u x
(both are functions of x) in such way that CI = (l x , u x ) will have a probability (1 − α) of capturing θ.

The quality of the CI is expressed in terms of

(i) coverage probability: P θ = P{θ ∈ (l x , u x )},
(ii) width of the interval: w = (u x − l x ).
The width is estimated as Eθ(W) over a range of sample intervals. In general, a good CI should have high
coverage probability and a small width. An interesting account of CIs, their creation and coverage probability
in the case of one parameter exponential distribution can be found in J. Sinsomboonthong [10].

Remark 2.2. The potential beneőt of the interval estimate in place of a point estimate is that the former takes
into account the dispersion of values within sample and indirectly addresses the problem of precision.

We can summarize the interval estimate of a metric into a single value by taking the weighted average of
selected values from the interval. One way is to consider the lower, middle and upper values of the CI as three
possible candidates for θ denoted by the triplet {l x , m x , u x }, where m x is the usual point estimate of θ. The
end points l x and u x will be equidistant around m x when the underlying distribution is symmetric but not
so when the distribution is skewed. If we use the triplet for each of the k parameters and attempt to evaluate
the metric g(C, Θ), we get 3k possible values. The Vännman’s superstructure given in (1.2) produces 32 = 9
possible outcomes, as it has k = 2. Similarly, the Wright’s formula has k = 3 and g(C, Θ) ends up with 33 = 27
possible outcomes. One way of evaluating the metric is to őnd all the 3k possible values and take their convex
combination with suitable weights. We call this the triplet method of estimation.
The weights can be determined several ways like using a őxed set of values (a priori weights) or deriv-
ing them from data (data driven). Weights can also be theoretical based on the quantiles of the sampling
distribution used to generate CI.

Brought to you by | Göteborg University - University of Gothenburg

Authenticated
Download Date | 12/4/19 5:06 AM
 S. S. Vedururu et al., A New Method of Estimating the Process Capability Index | 97

Vardhan and Sarma [12] have used the CI method in the context of deriving the area under the ROC
curve for binary classiőcation which needs a combination of nine different estimates. They have used a priori
weights with 0.5 for the middle value and the rest divided equally among the remaining eight values. It is a
case of triplet estimates with k = 2.
Rani, Sarada and Sarma [8] applied the triplet estimate method and derived a single sampling plan for
attribute type of inspection. They have used data driven weights instead of a priori weights. Simulation exper-
iments have shown that the triplet method has lower standard error than the classical method of using the
point estimates.
Sarada, Subbarayudu and Sarma [9] used triplet estimation method to estimate PCI of a normally
distributed process on the lines of Vännman’s formula given in (1.2). They have used weights inversely
proportional to the squared deviation from the target PCI.
In the following sections we deal with a non-normal process and apply the triplet estimation in the
Wright’s formula given in (1.3). A new estimate of C s is proposed assuming that X follows exponential dis-
tribution with parameter θ by using the triplet {l x , m x , u x } for θ. We will show by simulation that the new
formula has better properties than the PCI obtained with the point estimates of μ and σ.
In Section 3 we brieŕy review some basic results of exponential distribution. In Section 4 we develop
the new estimator and in Section 5 the performance of the new estimator will be discussed with simulated
experiments.

3 Interval Estimates of Exponential Parameter

Let X follow exponential distribution with density

{θe−θx , x ≥ 0,
f(x, θ) = {
{
0, otherwise.

For this distribution

1 1
μ = E[X] = and σ2 = V[X] = 2 .
θ θ
The third central moment is μ3 = θ23 and the maximum likelihood estimate of the parameter θ is θ̂ = 1x , where
x is the sample mean of n observations.
The 100(1 − α) % conődence intervals for the mean 1θ is given by
2n 1 2n
< < , (3.1)
̂θχ2 α ̂
θ θχ2α
1− ,2n
2 ,2n 2

where θ̂ is the point estimator of θ and θχ̂ 2 α

1− 2 ,2n
̂ 2α
and θχ are the upper and lower percentile points respec-
2 ,2n
tively on the Chi Square distribution.
Replacing θ with θ̂ the CI given in (3.1) becomes
2nx 1 2nx
< < 2 .
χ21− α ,2n θ χ α ,2n
2 2

For instance, let us take a random sample of size 10 from the exponential distribution with μ = 2.5 which
means θ = 0.4 and let the sample mean be x = 2.16 (θ̂ = 0.46). Taking α = 0.05, the 95 % CI will be
{1.264, 4.504} and this interval contains the true mean 2.5. Further the width of this interval is w = 3.240.
Hence the triplet estimate of μ is {1.264, 2.160, 4.504}. We note that the right gap (2.344) from the center
is wider than the left gap (0.896). Suppose we take of sample of 50 and let the sample remain as 2.16. Then
the triplet estimate of μ is {1.667, 2.160, 2.901} with width w = 1.24. This is a shorter interval than the one
obtained with n = 10. In this case, the right gap is 0.741 and the left gap is 0.493. As a rule we may consider
larger samples to obtain shorter CIs.
Consider the following illustration.

Brought to you by | Göteborg University - University of Gothenburg

Authenticated
Download Date | 12/4/19 5:06 AM
98 | S. S. Vedururu et al., A New Method of Estimating the Process Capability Index

Figure 1: Visualization of CI (n = 10). Figure 2: Visualization of CI (n = 50).

Example 3.1. Suppose the task time for őxing a mechanical assembly has exponential distribution with mean
of 2.5 minutes. Let us simulate 10,000 samples of size n = 10 from this distribution. We get 1000 CIs of the
form (l x , u x ) with conődence level 95 %. The coverage probability is estimated to be 0.951 (proportion of
samples that contained mean of 2.5). It means that only 4.1 % of cases have missed capturing the true θ.
A random sample of 50 out of 1000 cases of simulation is plotted to display the width of each CI as shown
in Figure 1. The vertical line shows the target mean and it is easy to see that u x is far away from m x when
compared to l x . Figure 2 shows a similar plot with n = 50. Long right tails are noticed when n = 10 but the
pattern is better when n = 50. We observe that though each CI can capture θ with 95 % conődence, pattern
of lower and upper limits around the central value conveys additional information that can be utilized to
improve the quality of the point estimate.

In the light of the discussion above, we proceed in the following section to develop a new estimate of the PCI
by utilizing the interval estimates of both μ and σ in the Wright’s formula given in (1.3). The moment μ3 will
be replaced with its point estimator 2̂ 3 .
θ

4 New Estimate for C s

Substituting the point estimate θ̂ for θ in (1.3) we obtain the estimate
󵄨 󵄨
d − 󵄨󵄨󵄨 1̂ − T 󵄨󵄨󵄨
̂
Cs = θ
,
2 󵄨 󵄨
3√( 1̂ 2 ) ⋇ ( 1̂ − T) ⋇ 󵄨󵄨󵄨 2̂ 2 󵄨󵄨󵄨
θ θ θ

where θ̂ = 1x . Using the CI given in equation (3.1), we get three possible estimates for the parameter θ which
are deőned below. The triplet of interval estimate {l x , m x , u x } leads to {θ̂ 1 , θ̂ 2 , θ̂ 3 }, where
1 2nx 1 1 2nx
= , = x, = .
̂θ1 χ2 α θ̂ 2 ̂θ3 χ2α
1− ,2n 2 2,2n

For each θ i for i = 1, 2, 3 the estimate is

󵄨 󵄨
d − 󵄨󵄨󵄨 1̂ − T 󵄨󵄨󵄨
C si = θi
, i = 1, 2, 3. (4.1)
3√ ⋇ ( 1̂ − T)
3 2
θ̂ 2i θi

We discuss two approaches of deriving weights viz., (i) based on squared deviation from a given target, and
(ii) drawing weights from the probability density function of Chi Square distribution at C si for i = 1, 2, 3. In
both cases the weights will be normed.

Brought to you by | Göteborg University - University of Gothenburg

Authenticated
Download Date | 12/4/19 5:06 AM
 S. S. Vedururu et al., A New Method of Estimating the Process Capability Index | 99

(a) Weights Based on Deviation from Target PCI. Suppose the hypothetical value of θ is given as θ0 (known).
Then the value of C S can be calculated which, gives the target PCI denoted by T0 as
󵄨 󵄨
d − 󵄨󵄨󵄨 θ10 − T 󵄨󵄨󵄨
T0 = .
3√( θ32 ) ⋇ ( θ10 − T)
2
0

We will combine the three estimates in (4.1) by taking weights inversely proportional to the squared deviation
from T0 . We call this adaptive weight method since the weights are adopted from the sample data itself. Deőne
d i = (C si − T0 )−2 for i = 1, 2, 3 and w i = ∑ddi i , so that ∑ wi = 1. It means the weights are inversely proportional
to the squared deviation from T0 . By using (4.1), the new estimate becomes

s = w1 C s1 ⋇ w2 C s2 ⋇ w3 C s3 .
CCI (4.2)
The value given by C s2 is the commonly used point estimate.

(b) Weights Based on Chi Square Distribution. When the target is not speciőed for the process parameter, it is
not possible to use (4.2). Since the conődence limits for θ are based on chi square distribution, we propose the
new weights as vi = f(χ2(3) , C si ) for i = 1, 2, 3, where f( ⋅ , ⋅ ) is a certain function. The normed weights can be
taken as b i = ∑vvi i , so that ∑ b i = 1. The degrees of freedom are taken as 3, since all the three are independent
values in the triplet. The new estimate is given by

s = b 1 C s1 ⋇ b 2 C s2 ⋇ b 3 C s3 .
CCI (4.3)
For instance, let the values of C si be {0.0376, 0.0669, 0.1365} from a sample of size 10. The corresponding
ordinates of χ2(3) at these three values will be {0.0759, 0.0998, 0.1377} so that the vector of weights becomes
{0.2423, 0.3184, 0.4393}. The new estimate works out to 0.0904. We call this density-based weights method.
In the following section the model is evaluated by simulated data.

5 Properties of New Estimate ś A Simulation Study

Simulated experiments are conducted to evaluate the performance of the new estimator CCI s with different
sample (subgroup) sizes using exponential distribution with θ = 0.40. The other parameters are LSL = 0,
USL = 15 and T = 7.5. The target PCI expected with these parameters works out to T0 = 0.126. A macro
written with SPSS syntax given is used for simulation and data is saved for further analysis. The mean, stan-
dard error and bootstrap standard error of CCI s are obtained from 10,000 simulations run with sample size
n = 10, 20, 30, 50, 100 for selected values of θ. A sample of 5 records with sample size n = 50 and θ = 0.4
has produced the results shown in Table 1.
From the complete experiment, it is found that C s2 has mean 0.1247 with standard error 0.00047 while
the new estimate CCIs has mean 0.1255 with standard error 0.00014. It means the new estimate is more con-
sistent than the classical point estimate C s . A comparison of simulated values is shown in Table 2. The mean
and 95 % Bootstrap CI of C s based on the classical and new estimates of the PCI for different sample sizes
and different values of θ are summarized in Table 2. From Table 2 it is clear that the new estimate has shorter
CI than the classical estimate.

Sample Triplet of Means C s at the Triplet Values Weights C sCI

lx mx ux C s1 C s2 C s3 W1 W2 W3

1 1.4269 1.8488 2.4909 0.0725 0.0949 0.1256 0.0001 0.0002 0.9997 0.1256
2 1.7860 2.3140 3.1176 0.0916 0.1177 0.1494 0.0493 0.8447 0.1060 0.1198
3 1.6493 2.1369 2.8791 0.0844 0.1093 0.1412 0.0680 0.4221 0.5099 0.1239
4 1.6665 2.1591 2.9090 0.0853 0.1104 0.1423 0.0713 0.4837 0.4450 0.1228
5 1.7123 2.2185 2.9890 0.0878 0.1132 0.1451 0.0716 0.6419 0.2865 0.1205

Table 1: Simulated means, triplets, weights and the new estimate (target T0 = 0.126).

Brought to you by | Göteborg University - University of Gothenburg

Authenticated
Download Date | 12/4/19 5:06 AM
 100 | S. S. Vedururu et al., A New Method of Estimating the Process Capability Index
Mean (μ) Parameter (θ) True PCI n Skewness Component C s2 (Classical Estimate) C sCI (New Estimate)
Mid-point SE Bootstrap CI Width Mid-point SE Bootstrap CI Width
Lower Upper Lower Upper

1.0 1.00 0.0409 10 2.0603 0.0482 0.00050 0.0472 0.0492 0.0020 0.0469 0.00016 0.0466 0.0472 0.0006
20 2.0945 0.0496 0.00037 0.0489 0.0503 0.0014 0.0483 0.00011 0.0481 0.0486 0.0004
30 2.0741 0.0496 0.00033 0.0490 0.0503 0.0013 0.0487 0.00009 0.0485 0.0488 0.0004
50 2.0449 0.0497 0.00023 0.0492 0.0501 0.0009 0.0492 0.00007 0.0491 0.0493 0.0003
100 2.0169 0.0495 0.00016 0.0493 0.0499 0.0006 0.0493 0.00005 0.0492 0.0494 0.0002
1.5 0.67 0.0409 10 4.8076 0.0750 0.00076 0.0735 0.0765 0.0030 0.0730 0.00023 0.0725 0.0734 0.0009
20 4.7115 0.0762 0.00054 0.0750 0.0772 0.0022 0.0747 0.00016 0.0744 0.0750 0.0006
30 4.6808 0.0766 0.00046 0.0757 0.0775 0.0018 0.0754 0.00014 0.0751 0.0757 0.0006
50 4.5975 0.0765 0.00036 0.0757 0.0772 0.0014 0.0756 0.00010 0.0754 0.0758 0.0004
100 4.5513 0.0765 0.00025 0.0760 0.0770 0.0010 0.0761 0.00007 0.0760 0.0763 0.0003
2.0 0.50 0.1026 10 9.0491 0.1013 0.00095 0.0995 0.1032 0.0037 0.0996 0.00027 0.0991 0.1001 0.0011
20 8.3292 0.1010 0.00067 0.0997 0.1023 0.0026 0.1010 0.00021 0.1006 0.1014 0.0008
Brought to you by | Göteborg University - University of Gothenburg

30 8.0924 0.1009 0.00055 0.0997 0.1019 0.0022 0.1012 0.00016 0.1009 0.1015 0.0006
50 8.0465 0.1015 0.00043 0.1006 0.1024 0.0018 0.1016 0.00014 0.1014 0.1019 0.0005
100 8.0278 0.1020 0.00032 0.1014 0.1027 0.0013 0.1022 0.00009 0.1021 0.1024 0.0004
2.5 0.40 0.1260 10 13.5516 0.1202 0.00097 0.1183 0.1222 0.0038 0.1237 0.00031 0.1231 0.1243 0.0012
20 13.4245 0.1249 0.00071 0.1234 0.1262 0.0028 0.1254 0.00022 0.1249 0.1258 0.0009
30 12.8886 0.1242 0.00061 0.1231 0.1254 0.0023 0.1251 0.00020 0.1247 0.1255 0.0008
50 12.6729 0.1247 0.00047 0.1238 0.1256 0.0018 0.1255 0.00014 0.1252 0.1257 0.0005
100 12.5841 0.1253 0.00033 0.1246 0.1259 0.0013 0.1257 0.00010 0.1255 0.1259 0.0004
3.0 0.33 0.1455 10 19.4157 0.1385 0.00091 0.1367 0.1403 0.0036 0.1444 0.00031 0.1438 0.1450 0.0013
20 18.8369 0.1417 0.00072 0.1403 0.1431 0.0028 0.1448 0.00021 0.1444 0.1452 0.0008
30 18.7003 0.1434 0.00057 0.1422 0.1444 0.0022 0.1454 0.00016 0.1451 0.1457 0.0006
Download Date | 12/4/19 5:06 AM

50 18.2012 0.1435 0.00046 0.1426 0.1445 0.0019 0.1453 0.00014 0.1451 0.1456 0.0005
100 18.1514 0.1446 0.00034 0.1439 0.1452 0.0013 0.1454 0.00010 0.1452 0.1456 0.0004

Table 2: Sensitivity of the classical and new estimates to changes in θ and n.

Authenticated
 S. S. Vedururu et al., A New Method of Estimating the Process Capability Index | 101

Figure 3: Violin plot of 10,000 simulations of classical and new PCI (n = 50, θ = 0.4).

Parameter Interval Mid-Point Standard Error 95 % CI

PCI-Classical 0.1255 0.0005 (0.1246, 0.1264)

PCI-Adaptive Weights 0.1256 0.0001 (0.1253, 0.1259)
PCI-Density Based Weights 0.1286 0.0004 (0.1278, 0.1295)

s from 1000 bootstrap simulations (target PCI = 0.1259).

Table 3: Comparative summary C CI

Figure 3 shows a plot of C S2 (Classical PCI) and CCIs (New PCI) with θ = 0.4 obtained from simulated sam-
ples of size 50. The violin plots in Figure 3 show that the PCI obtained by utilizing the CIs of the process param-
eters provides a more stable estimate than the PCI obtained at the point estimates of the process parameters
(classical method). We have also studied the new PCI from 10,000 simulations having n = 50 and θ = 0.4,
with density-based weights using (4.3) and the results are compared with the other methods as shown in
Table 3. It can be seen that not much difference exists in the standard error of the estimate between the
PCI-Classical and the PCI-Density based method. Therefore, if there is a known target, the adaptive weights
method is a better way of utilizing the information from interval estimates of a sample.

6 Conclusions
In this paper we explored PCI values that can be calculated at the three nominal values of the CI and com-
bined them as a weighted average where the weights are adopted from the data itself. We observed that the
new estimate of PCI obtained by exploiting the boundary values of the CIs of process parameters are more
consistent than those obtained by using the point estimates for the unknown moments. An issue of interest
is the method of determining the weights for mixing the different estimates. The method of adaptive weights
based on the deviation from a targeted PCI is a good way of estimation. We emphasize that satisfying the unbi-
asedness property is not always enough and the precision of the point estimate can be utilized to improve the
estimate of true parameter effectively.

References
[1] C. G. Borroni, M. Cazzaro and P. M. Chiodini, Measuring process capability under non-classical assumptions: A purposive
review of the relevant literature, Statist. Appl. 22 (2010), no. 3, 279ś306.
[2] J. A. Clements, Process capability calculations for non-normal distributions, Quality Progr. 22 (1989), 95ś100.

Brought to you by | Göteborg University - University of Gothenburg

Authenticated
Download Date | 12/4/19 5:06 AM
102 | S. S. Vedururu et al., A New Method of Estimating the Process Capability Index

[3] A. J. Duncan, Quality Control and Industrial Statistics, 3rd edition, D B Taraporevala Sons & Co Pvt Ltd., Bombay, 1999.
[4] G. S. Koch, Jr. and R. F. Link, Statistical Analysis of Geological Data, Dover Publications, New York, 2002.
[5] D. C. Montgomery, Introduction to Statistical Quality Control, 6th edition, John Wiley & Sons, New York, 2009.
[6] W. L. Pearn and C. S. Chang, The performance of process capability index C s on skewed distributions, Comm. Statist.
Simulation Comput. 26 (1997), 1361ś1377.
[7] W. L. Pearn and K. S. Chen, Capability indices for non-normal distributions with an application electrolytic capacitor
manufacturing, Microelect. Reliab. 37 (1997), 1853ś1858.
[8] K. Reddi Rani, V. Sai Sarada and K. V. S. Sarma, A new method of implementing single sampling plan by variables using
interval estimates of process mean and variance, Int. J. Sci. Eng. Res. 8 (2017), no. 7, 1702ś1705.
[9] V. Sai Sarada, M. Subbarayudu and K. V. S. Sarma, Estimation of process capability index using conődence intervals of
process parameters, Res. Rev. J. Stat. 7 (2018), no. 3, 49ś57.
[10] J. Sinsomboonthong, Conődence interval estimations of the parameter for one parameter exponential distribution,
IAENG Int. J. Appl. Math. 45 (2015), no. 4, 343ś353.
[11] K. Vännman, A uniőed approach to capability indices, Statist. Sinica 5 (1995), 805ś820.
[12] R. Vishnu Vardhan and K. V. S. Sarma, Estimation of the area under ROC curve using conődence interval of mean,
ANU J. Phys. Sci. 2 (2010), no. 1, 29ś39.
[13] P. A. Wright, A process capability index sensitive to skewness, J. Stat. Comput. Simul. 52 (1995), 195ś203.

Brought to you by | Göteborg University - University of Gothenburg

Authenticated
Download Date | 12/4/19 5:06 AM