Revista Colombiana de Estadística

Volumen 30 No. 2. pp. 301 a 316. Diciembre 2007

Comparison of Process Capability Indices under
Autocorrelated Data
Comparación de índices de capacidad de procesos con datos
autocorrelacionados
Rubén Darío Guevara1,a , José Alberto Vargas2,b
1 Programa

de Ciencias Básicas, Universidad de Ciencias Aplicadas y Ambientales,
Bogotá, Colombia

2 Departamento

de Estadística, Facultad de Ciencias, Universidad Nacional de
Colombia, Bogotá, Colombia

Resumen
The process capability indices provide a measure of how a process fits
within the specification limits. In calculating indices is usual to assume that
the process data are independent. However, in industrial applications data
are often autocorrelated. This paper deals with the indices Cp , Cpk , Cpm and
Cpmk when data are autocorrelated. Variances for their estimators are derived and coverage probabilities of some confidence intervals are calculated.
Palabras clave: Autocorrelation, Process analysis, Estimation, Process capability indices, SPC.
Abstract
Los índices de capacidad de un proceso suministran una información numérica acerca de cómo el proceso se ajusta a unos límites de especificación.
En el cálculo de estos índices se asume que las observaciones son independientes; sin embargo, en aplicaciones industriales frecuentemente los datos
están autocorrelacionados. Este artículo analiza los índices Cp , Cpk , Cpm y
Cpmk cuando los datos presentan autocorrelación, se encuentran las varianzas para sus estimadores cuando los procesos son gaussianos y se calculan
los porcentajes de cobertura para algunos intervalos de confianza.
Key words: autocorrelación, análisis de procesos, estimación, índice de capacidad de procesos, SPC.
a Profesor.
b Profesor

E-mail: r.guevara@udca.edu.co
titular. E-mail: javargasn@unal.edu.co

301

Additionally. A basic assumption in process capability analysis is that process data are independent and identically normally distributed. Though the literature on process capability indices is voluminous. Cpk . In Section 4. data exhibit some degree of autocorrelation. when it is present. widely used in industrial quality improvement programs. d = U SL − LSL and m = (U SL + LSL). Cpk . U SL − LSL Cp Cpm = p =p 2 2 6 σ + (µ − T ) 1 + ξ2 Revista Colombiana de Estadística 30 (2007) 301–316 . in several industrial processes. some conclusions are given. a = (U SL − LSL)/2. However. by measuring the coverage probability of confidence intervals constructed for some of the indices studied. (2003). in stationary gaussian processes. inferential procedures are performed. Subsequently. S 2 and S in these processes. Cpm and Cpmk . Cpm and Cpmk for stationary gaussian processes. In Section 6. Introduction Process capability analysis is an important tool. This is carried out through a simulation study with AR(1) processes. In the first part of this article we introduce the definition of the capability indices Cp . Cpk = m´ın  U SL − µ µ − LSL . Zhang (1998) studied the indices Cp and Cpk for autocorrelated data. Section 5. Process Capability Indices Among the several capability indices defined in the literature. the most extensively used in the industry are Cp . some results have been taken from Guevara (2005) in his master’s thesis. Cpm and Cpmk . as can be observed in Kotz & Johnson (2002) and in Spiring et al. we present: 1) some relevant terminology about stationary gaussian processes. 3σ 3σ      a − |µ − b| d − |2µ − m| = = 3σ 6σ where µ is the process mean.302 Rubén Darío Guevara & José Alberto Vargas 1. We extend Zhang’s study to the capability indices Cpm and Cpmk and compare the four indices Cp . Cpk . Finally. 2. we calculate the variances of the estimators of Cpm and Cpmk . b = (U SL + LSL)/2. defined as follows: Cp = U SL − LSL 6σ where U SL and LSL are the upper and lower specification limits respectively and σ is the process standard deviation. should not be ignored for calculating any of the four indices. there is not significant research when data are autocorrelated. shows why the autocorrelation structure of process data. and 2) the expected values and variances of X.

Xn } be a sample  Pn 2 Pn Xi i=1 (Xi −X ) 2 of size n from the process {Xt }. Xt ) γX (0) (see Brockwell & Davis 1996). r. E(Xt ) = m. X2 . ρi ) (3) V ar S 2 = (n − 1)2  1/2 1/2  (4) = σX f (n. ρi ) n  2 E S 2 = σX f (n. with index set Z. and γX (t + h. and S = n−1 be the sample mean and the sample variance respectively. s. {Xt } is a process gaussian if all of its joint distributions are multivariate normal. is said to be stationary if E|Xt |2 < ∞. Xs ) = E (Xr − EXr )(Xs − EXs ) . ∀t ∈ Z. ρi ) E(S) ≈ E S 2 Revista Colombiana de Estadística 30 (2007) 301–316 . A process is said to be stationary gaussian if it is stationary and gaussian simultaneously (see Brockwell & Davis 1996). Stationary Gaussian Processes If {Xt } is a process such that V ar(Xt ) < ∞ for each t ∈ W ⊂ R. s) = Cov(Xr . Let {Xt } be a stationary gaussian process. ∀r. ·) of {Xt } is:   γX (r. s + t). . t) is independent of t for each h ∈ Z. therefore µX(t) is independent of t. where R is the set of real numbers. Let X = i=1 n . . the autocovariance function (ACVF) of {Xt } is γX (h) = Cov(Xt+h . U SL − µ µ − LSL p . . t ∈ Z}. Zhang (1998) gives the expected values and variances of X. . t ∈ Z. S 2 and S:  E X = µx  σ2 (1) V ar X = X g(n. ρi ) (2) 4  2σX F (n.Comparison of Process Capability Indices under Autocorrelated Data where T is the target value and ξ = Cpmk = m´ın ( µ−T σ 303 . s) = γX (r + t. s ∈ W The time series {Xt . Xt ) and the autocorrelation function (ACF) is given by ρX (h) = γX (h) = Cor(Xt+h . then the autocovariance function γX (·. If {Xt } is a stationary process. p 2 2 3 σ + (µ − T ) 3 σ 2 + (µ − T )2 ) = a − |µ − b| Cpk p = p 2 3 σ 2 + (µ − T )2 1+ξ 3. ∀t ∈ Z and γX (r. Let {X1 .

ρi ) = n + 2 n n i=0 j=0 i=1 i=1 n−1 X and n−1 g(n. . Variances of Process Capability Indices Estimators Let {Xt } be a stationary gaussian process. 2 2  =  3 S2 + X − T 3 S2 + X − T bpk a − |X − b| C q = q 2 3 S2 + X − T 1 + ξb2 bp and Zhang (1998) found the following approximations for the variances of C b Cpk :  F (n. Cpk . ρ ) 2 i V ar S 2 F (n. Cpm y Cpmk are: bp = U SL − LSL C 6S       bpk = m´ın U SL − X . ρi ) = 1 − n−1 X 2 (n − i)ρi n(n − 1) i=1 #2 " n−1 n−1n−i X 2 XX 1 (n − i)ρi − (n − i − j)ρi ρj (n − i)ρ2i + 2 n + 2 F (n. . ρi ) 4E S where ρi = ρX (i). . ρi ) bp ≈ Cp2 V ar C 2(n − 1)2 f 3 (n. ρi ) 2(n − 1)2 f (n. f (n. and     X − LSL U SL − X q q = m´ın . ρi ) Revista Colombiana de Estadística 30 (2007) 301–316 . for i = 1. at lag i. . . is the autocorrelation of X. . . ρi ) = 1 + 2X (n − i)ρi n i=1 4.304 Rubén Darío Guevara & José Alberto Vargas and  4   2σX F (n. Let {X1 . n. Xn } be a sample of size n from {Xt }. ρi ) 2  = σX  =  (n−1) V ar(S) ≈ 2 2 4σX f (n. X − LSL = a − |X − b| = d − |2X − m| C 3S 3S 3S 6S bp C bpm = qU SL − LSL C 2 = q 6 S2 + X − T 1 + ξb2 where ξb = bpmk C X−T S . X2 . . The usual estimators of Cp .

ρi ) 2(n − 1)2 f 3 (n. and which equals the variance of C "  2 # 2  Cpk  F (n. pi ) 1 6ξ + + 2(n − 1)2 [f (n. ρi ) 2 + 2(n − 1)2 f 2 (n. pi ) 2(n − 1)2 f 2 (n. ρi ) bp . pi ) g(n. ρ ) f (n. pi ) 2(n − 1)2 f 2 (n. simultaneously the observations are independent and µ = T . pi ) + ξ 2 ] where ξ = µ−T σ . pi ) b V ar Cpmk ≈ + f (n. we have that " 2 #    Cp2 2 bpm ≈ Cp (n−1) = bp V ar C ≈ V ar C 4 2(n − 1) Revista Colombiana de Estadística 30 (2007) 301–316 .Comparison of Process Capability Indices under Autocorrelated Data and  bpk ≈ V ar C 305 " # 2 Cpk g(n. If µ = T then  bpm ≈ Cp2 V ar C F (n. which equals the variance of C When the observations are independent. ρi ) F (n. pi ) 9nCpk bpk . ρi 4g(n. pi ) + ξ 2 (  2 ) F (n.ρi ) (n−1)2 + 4[f (n. pi ) + ξ 2 ]2 9n Cpk 2[f (n. pi ) + = 2 f (n. pi ) 1 4σ 2 g(n. ρi ) 9nCpk i bpm and C bpmk .ρi )ξ 2 n ) + ξ 2 ]3      1 2 b × V ar Cpmk ≈ Cpk f (n. pi ) n a − |2µ − b| " # 2 Cpk F (n. we found the following approximations For the variances of C (see appendix A and B):   bpm ≈ C 2  V ar C p and 2F (n. pi ) g(n. the variances reduce to:      1 1 ξ2 2 b V ar Cpm ≈ Cpm + 1 + ξ 2 2(n − 1) n and   2 bpmk ≈ Cpmk V ar C (  2 ) 1 1 6ξ 1 + + 2(n − 1)[1 + ξ 2 ]2 9n Cpk 2[1 + ξ 2 ] If.

796 0.847 0. We observe in Table 1 that the higher the autocorrelation level the lower the capability index value. Xt = Xt−1 + et .776 0.598 *d = µ − T Through a simulation study we analyze the effect of the autocorrelation in the expected value of the sample mean and in the expected value of the standard error. 42.659 Cpm *d = 2 *d = 3 0.962 0.23 8.968 0.866 0. For Cp = 1.08 10. 20. Cpm and Cpmk are calculated. Similar results are obtained for V ar Cpmk . Different target values are considered: 40. the standard deviation and the capability indices Cp .619 *d = 5 0. A quality characteristic is normally distributed with mean 40 and standard deviation 7. 5.650 0. 200.919 0.306 Rubén Darío Guevara & José Alberto Vargas and   bpmk ≈ V ar C 2 Cpk ( 1 1 + 2 2(n − 1) 9nCpk ) bpk ≈ Var C  These last two results are equal to those found by Bissel (1990). this is. Tabla 1: Mean.000 0. bpm and V ar C n increases. φ = 0.661 *d = 1 0.737 0. Cpk .933 0. For fixed values of φ and n. 41.661 *d = 0 1. V ar C variability of V ar C  b increases.25 0.000 0. For each process. the mean.00 7. 100. ξ = 0. a simulation study was carried out for a first order stationary autorregressive process with parameter φ.75 Mean 40 40 40 40 STD 7.636 *d = 4 0.  0. To compare the variances of these estimators.866 0. standard deviation (STD). Cp and Cpm of a process not autocorrelated vs. 200. where {et } is a series of uncorrelated errors. 45 and 50. Revista Colombiana de Estadística 30 (2007) 301–316 . 50.50 0.968 0. bpm and V ar C V ar C 5. .58 Cp 1. Figure 1 shows great bpm bpm for n < 100. Fixing Cp .25.841 0.814 0. 41. 10.812 0.859 0. Now. n and ξ as |φ| increases. V ar C when ξ increases. We then compare two processes.  bpmk decreases.33. The specification limits are U SL = 61 and LSL = 19.75 and n = 10. We do not show the values of Cpk and Cpmk because Cp = Cpk and Cpm = Cpmk see Table 1. substituting Cpm for Cpmk and Cp for Cpk as can be seen in Figure 2. if φ and ξ are fixed values and   bpmk decreases. a process following an AR(1) model. |φ| No auto 0.50.990 0. σe2 ) and σe = 7. T = 40. . We generated 1000 samples from a no autocorrelated model and 1000 samples from an AR(1) model for each of the following cases: n = 15. .894 0. Autocorrelation Effects on Process Capability Indices We consider the example given in Shore (1997). 0.959 0. A process with independent observations and a process with observations following an AR(1) model. et ∼ N (0. the  target value is far away from the mean of the process. .868 0.

25.0008 + 0. 0. 0.00004 φ: ´ ` bpm V ar C + 0.00001 × bc × bc × + bc × + bc × + + + bc + bc bc + bc + bc bc × × 0.75 0.0001 × bc × bc × + bc × + bc × + + + bc + bc + bc bc + bc bc × × × 0.50 × 0.25 0.50 × 0.0002 + + bc bc × 0. 10 Figura 1: Variance of C and φ = 0.307 Comparison of Process Capability Indices under Autocorrelated Data ξ=0 0.5 + + 0.00000 0 20 bc 0.75.0007 + 0.0004 + bc 0.25 20 40 60 80 100 × × 120 × + bc × 140 + + + + + × × × × × bc bc 160 bc bc 180 bc 200 n ξ = 10 + 0.33.00003 + bc 0.00005 + 0. Revista Colombiana de Estadística 30 (2007) 301–316 .2 + bc 0.50 × 0.25 40 60 80 100 × × 120 × × 140 + bc × + + + + + × × × × × bc 160 bc bc 180 bc bc 200 n bpm in function of the sample size with Cp = 1.75 0.0005 0.75 φ: ´ ` bpm V ar C 0.0 0 20 + bc × bc × 40 + bc × + + × × bc 60 bc + c ×b 80 +bc × +bc × 100 +bc × +bc × 120 +×bc +×bc 140 +×bc +×bc 160 +×bc +×bc 180 +×bc 200 n ξ=5 0.1 × bc × 0.3 + 0.00002 × + + bc bc × 0. 5. ξ = 0.0003 × 0.0006 + 0.50.4 ´ ` bpm V ar C φ: bc 0.0000 0 bc 0.

2 + bc 0.308 Rubén Darío Guevara & José Alberto Vargas ξ=0 + 0.25 0.25.50.50 × 0.0002 × + bc × 0.0005 φ: + × 0.004 + + 0. 5.33.1 × bc × + bc × 0.001 φ: bc × + bc × bc × × 0.0 0 20 bc × 40 + bc × + + × + × × bc 60 bc bc 80 + bc × +bc × 100 +bc × +bc × 120 +bc × +bc × 140 +bc × +×bc 160 +×bc +×bc 180 +×bc 200 n ξ=5 ´ ` bpmk V ar C 0.0004 ´ ` bpmk V ar C bc 0. Revista Colombiana de Estadística 30 (2007) 301–316 .50 + bc × 60 + bc × + bc × 80 + + + + × + × × × + × × bc bc 100 bc bc 120 bc bc 140 + bc × +bc × 160 +bc × +bc × 180 +bc × 200 n ξ = 10 + 0.4 bc 0.75 φ: ´ ` bpmk V ar C 0.25 + bc × 0. 10 Figura 2: Variance of C and φ = 0. 0.003 + + bc 0.75 0. ξ = 0.0001 + bc × bc × bc × + bc × + bc × + + + bc + bc + bc bc bc × × 0.000 0 20 40 bc 0.25 0.50 + bc 0.002 × 0.75.3 + 0. 0.0003 0.0000 0 20 40 60 80 100 × × 120 × + bc × 140 + bc × + + + + + × × × × × bc 160 bc bc 180 bc bc 200 n bpmk in function of the sample size with Cpk = 1.75 0.5 + 0.0006 + 0.

0.50 −0.70 7.79 1.97 0.88 1.50 0.18 0. −0.96 1.88 Revista Colombiana de Estadística 30 (2007) 301–316 .38.54 0.90 39.99 0.77 Std No Auto 6.48 Auto 10.02 1.69 0. Tabla 2: Expected values and standard errors of the sample mean and sample standard deviation for processes no autocorrelated and processes following an AR(1) model.86 1.75 Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Average Std Error Mean No Auto Auto 39.76 1.97 1.69 2.73 40.18 0.02 39.50 0.51 8.08 39.41 40.98 0.25 15 0.02 40.98 1.55 40.70 6.50 −0.99 0.75.34 7.82 3.15 40.34 39.53 10.72 7. Let us remember that V ar(Xt ) = 1−φ 2 where σe is the white noise variance.50 6.08 0.75 −0.5 is 7.32 6.59 40.44 40.71 8.88.00 40.24 40.50 −0.94 1.75 −0.19 1.01 40.25 the estimated expected value of the standard error is 7.00 3. for φ = −0.09 8.84 1.03 0. 45 and φ = −0.96 0.02 40.52 40.75.53 1.309 Comparison of Process Capability Indices under Autocorrelated Data 42.41 39.99 40.76 8.99 1.13 in autocorrelated processes.50.95 and for φ = −0.50.15 40.96 0.98 40.97 0.56 7.72 6. 44.01 0.95 1.84 10.28 6.85 1.04 10.79 1. 0.00 0.75 −0.70 6.02 0.71 7.04 0.25 0.38 2. n φ −0.69 6.51 1.00 0.00 40.85 6.00 39. for n = 15 and φ = −0.49 1. 43.99 40.98 0. 6.95 0.51 6.50 0.99 1.70 0.33 6. For example.98 40.14 1.50 6.23 0.88 1.36 1.23 0.34 6.32 40.82 2.00 0.54 7.48 6.48 40.00 0.06 0.75 is 10.25.01 0.99 39.95 1.99 10.25 0.03 39.52 7.97 0.70 1. −0.32 6.75 −0. For independent observations the values are 6.00 1.25 50 0.85 and 6.97 0.25 0.58 10.75 Table 2 shows that the autocorrelation does not affect the expected value of the sample mean while a different situation occurs with the expected value of the σe2 2 standard error.36 7.80 39.01 0.25 100 0.26 6.71 0.57 0.02 40.13 1.25.98 0.82 7.03 7.06 40.93 40.05 1.97 0.95 39.75 −0. 0.99 0.68 0.66 40.33 39.03 0. Table 2 shows partial results of these simulations.50 6.97 0.00 0.

0. C fb C fb.668 respectively.84. AR(1) processes with normally distributed white noise are generated with φ = −0. 7.310 Rubén Darío Guevara & José Alberto Vargas respectively. 0. the coverage probabilities increase. the values of X. C bpm . for fb = n 1 + δb 1 + 2δb . For each sample. −0.968.0.81 when µ − T = 0. For each combination of n. Confidence Intervals for the Capability Indices under Stationary Gaussian Processes Let us assume that µ = 50. 50 y 100 the expected values of C are 0. . 0.α/2 fb.75 the coverage probabilities are very similar for different values of σ. For n = 15 and φ = 0. This interval is defined as follows: r r  . 2 and the expected values of C 3 respectively. The coverage probabilities for Cpm and Cpmk are not shown here because these were very low. σ bcpm and σ C bcp . 1. We also calculate through simulations the coverage probabilities of the confidence interval for Cpm . σ proportion of times that the true index is contained in the interval \index ± k(b capability σcapability index ) The simulations showed a coverage probability for Cp of about 95 % for k = 2 and 99 % for k = 3 .673 and 0. 5000 random samples from a normal distribution were generated. bias that decreases as n increases.50. while the true values are 0. the estimated expected value of the standard error increases slightly for autocorrelated data. −0. We observe that for φ = −0. Tables 4 and 5 show the coverage probabilities of the intervals for Cp and Cpk .23 respectively. As n increases.841 and 0. 0. Comparing the estimated expected values of the capability indices estimators shown in Table 3 with the theoretical values shown in Table 1. 100 are 7.75 and n = 15.25. U SL = 3 and LSL = −3. 0.75. σ bcpmk are obtained.25. 6. φ and σ.25 and n = 15. 0. In Table 5 it can be observed that coverage percentage for Cpk decreases when σ increases.83 and 0.75. proposed by Boyles (1991). For Cpk the coverage probability was of around 90 % for k = 2. 0.1−α/2 where χ2fb.661. bpm are 0.25 the estimated expected values for n = 15. For example. 50. For φ = 0. C bpmk . 50. 100 and standard deviation values σ = 0.50. Through other simulation study we analyze the performance of the capability indices estimators.05.938 while the true value is 0.937.50 and 0.866. 2. For example. We calculated the bcpk .α denotes the 100α % percentile of a chi-square distribution with fb de2   2  2 grees of freedom. In Tables 4 and 5 we observe that as n increases. bp .703. S.859. δb = X − T σ b and σ b2 = Revista Colombiana de Estadística 30 (2007) 301–316 .812. it can be observed that for autocorrelated processes the estimators are slightly biased.935 and 0. being more evident for large values of φ. 0.84. 0.  bpm χ2 bpm χ2 fb.5. bp are 0. 1. For sample sizes n = 25. for φ = −0.18 and 7. 100 the expected values of C bpk true value is 0. C bpk . while the 50. for φ = 0.

016 0.5 0.311 Comparison of Process Capability Indices under Autocorrelated Data S 2 (n − 1)/n.08 0.09 0.19 0.10 0.97 0.97 0.00 0.06 0.963 0.66 0.072 1.703 0.96 0.19 0.10 0.018 0.97 0.137 0.08 0.668 0.96 0.83 0.103 0.10 0.059 0.50 Av St −0.206 0.62 0.13 0.10 0.00 0.098 0.077 0.877 0.12 0.82 0.07 0.13 0.95 0.13 1.110 0.75 Av St −0.241 1.5 0.93 0.208 0.64 0.10 0.84 0.60 0.62 0.60 0.13 0.060 0.61 0.66 0.81 0.995 0.96 0.921 0.20 0.90 0.935 0.103 1.05 0.989 0.85 0.11 0.96 0.213 0.86 0.072 1.0 0.794 0.08 0.75 Av St −0.18 0.978 0.03 0.08 0.0 0.134 0.0 1.10 0.01 0.245 0.102 1.098 0.10 µ−T = 1 No Auto Auto 1.80 0.09 µ−T =3 No Auto Auto 0.975 0.967 0.83 0.17 0.75 Av St Av: Average St: Std Error No Auto 1.60 0.08 0.18 0.131 0.69 0.982 0.093 0.10 1.18 0.14 0.97 0.077 0.75 0.206 0.02 0.5 0.207 0.949 0.96 0.07 0.236 1.19 0.20 1.91 0.99 0.093 0.938 0.68 0.67 0.99 0.228 0.873 0.50 Av St 0.955 0.15 0.81 0.99 0.099 0.009 0.206 1.80 0.973 0.21 1.22 1.06 0.75 Av St −0.50 Av St 0.978 0.977 2.84 0.01 0.944 0.18 0.976 0.18 0.06 0.14 0.093 0.103 1.00 0.87 0.673 0.860 0.00 0.67 0.25 Av 15 St 0.90 0.0 0.937 0.12 0.09 µ−T =5 No Auto Auto 0.84 0.17 0.658 0.07 0.92 0.015 0.812 0.13 1.16 1.103 µ−T = 0 No Auto Auto 1.983 0.077 0.987 0.99 0.25 Av St 0.10 0.009 0.988 0.885 0.01 0.10 0.97 0.873 0.07 0.09 1.00 0.21 1.985 0.07 0.970 φ= 1.08 Tabla 4: Estimation average coverage rate of Cp for AR(1) processes using intervals bp ± 2σ b .10 0.10 0.100 0.76 0.08 0.200 1.99 0.02 0.5 0.88 0.007 0.07 0.10 0.12 0.952 0.95 0.077 Auto 0.99 0.077 0.98 0.18 0.93 0.987 2.966 0.10 0.19 0.019 0.82 0.25 1.81 0.17 0.10 0.92 0.92 0.84 0.19 0.18 0.50 Av St 0.97 0.07 0.218 0.10 0.14 1.920 0.204 0.92 0.20 1.957 0.21 1.072 1.690 0.05 0.983 0.102 1.96 0.86 0.00 0.07 0.92 0.05 0.00 0.10 0.104 0.08 0.00 0.82 0.132 0.08 0.109 0.09 0.59 0.88 0.86 0.14 0.75 Av St −0.01 0.18 0.89 0.06 0.14 1.85 0.92 0.960 0.004 0.83 0.06 0.05 0.80 0.985 0.10 0.92 0.21 1.05 0.20 0.96 0.19 0.198 1.17 0.961 0.02 0.985 0.19 0.13 1.09 0.97 0.19 0.07 0.90 0.949 0.59 0.06 0.991 0.200 1.102 1.08 0.09 1.14 0.009 0.70 0.92 0.135 0.84 0.97 0.99 0.01 0.202 1.99 0.987 0.75 0.230 0.07 0.982 0.12 0.103 No Auto 0.66 0.110 0.921 0.05 0.75 Av St −0.10 0.98 0.658 0.65 0.09 0.60 0.954 0.19 0.137 0.12 0.137 0.09 0.979 0.21 0.03 0.93 0.138 0.08 0.00 0.02 1.972 0.82 0.978 0.80 0.19 0.10 0.0 0.83 0.84 0.11 0.74 0.980 0.25 Av 100 St 0.94 0.963 0.99 0.09 0.74 0.110 0.22 1.07 0.19 0.237 0.19 0.88 0.00 0.966 2.08 0.22 1.057 0.19 0.85 0.973 0.072 1.975 φ = −0.10 0.973 0.83 0.75 1.526 0.99 0.10 0.987 0.95 0.15 0.982 Revista Colombiana de Estadística 30 (2007) 301–316 .64 0.82 0.00 0.90 0.916 0.15 0.109 0.07 0.00 0.82 0.5 0.07 0.986 0.16 0.14 0.10 1.967 0.08 0.75 1.109 0.50 Av St −0.95 0.232 0. C Cp n\σ 25 50 100 0.94 0.203 1.96 0.0 0.06 0.987 φ= 1.88 0.204 0.97 0.02 0.99 0.23 1.09 0.82 0.077 0.99 0.09 0.130 0.876 0.84 0.09 0.863 0.578 0.012 0.102 1.00 0.67 0.65 0.94 0.944 0.09 0.074 0.07 0.180 0. Cp n φ −0.14 0.17 0.82 0.988 0.97 0.08 0.00 0.84 0.58 0.5 0.11 0.63 0.50 Av St −0.013 0.09 0.07 0.10 0.666 0.67 0.08 0. this is the coverage probabilities obtained are low.988 0.071 1.96 0.80 0.15 0.14 0.15 0.14 0.25 Av 50 St 0.982 0.094 0.918 0.915 0.01 0.674 0. We did not find reliable results for Cpm .829 0. Tabla 3: Effect of the autocorrelation in the expected values and standard errors of the capability indices for processes following an AR(1) model.07 0.00 0.19 0.989 0.070 0.009 0.10 0.11 0.82 0.92 0.132 0.09 0.92 0.13 1.05 0.969 0.97 0.82 0.12 1.10 1.12 0.978 0.07 0.64 0.60 0.954 0.135 0.16 1.96 0.97 0.129 0.90 0.98 0.009 0.19 0.25 Av St 0.10 Cpm µ−T = 2 No Auto Auto 0.25 Av St 0.74 0.92 0.62 0.64 0.95 0.988 0.048 0.000 0.072 Cpk Auto 0.06 0.82 0.07 0.82 0.031 0.17 0.10 0.12 0.07 0.13 1.99 0.81 0.200 1.597 0.12 0.64 0.08 0.82 0.19 0.16 0.00 0.23 1.206 0.673 0.09 0.93 0.94 0.09 0.

2 75 76 76 76 76 77 77 78 78 78 78 78 78 78 78 78 78 79 79 79 79 2.966 0. we constructed confidence intervals of the form bpm − k1 σ bpm + k2 σ C bCpm .961 0.2 2.3 3.815 0.986 φ= 1.976 φ = −0. We believe that the low coverage probabilities are due to the bias of the estimators.9 3.940 0.975 0. σ = 1.0 74 75 75 75 75 76 76 77 77 77 77 77 77 77 77 77 77 78 78 78 78 2.5.7 3. which offer a coverage between 74 % and 85 % for the index Cpm for n = 50.8 2.0 0. C bCpmk  for Cpmk for different values of k1 and k2 .2 3.0 82 83 83 83 83 84 84 85 85 85 85 85 85 85 85 85 85 86 86 86 86 3.5 0. φ = 0.6 86 87 87 87 87 88 88 89 89 89 89 89 89 89 89 89 89 90 90 90 90 3.0 2.75 1.9 88 89 89 89 89 90 90 91 91 91 91 91 91 91 91 91 91 92 92 92 92 4.0 1.4 3.1 2.794 0.6 3.969 0.25 1.920 0.920 0.0 2. Tabla 6: Interval estimation average coverage rate of Cpm for AR(1) processes with φ = 0.4 86 87 87 87 87 88 88 89 89 89 89 89 89 89 89 89 89 90 90 90 90 3.1 74 75 75 75 75 76 76 77 77 77 77 77 77 77 77 77 77 78 78 78 78 2. this is.854 0.5 0.50.4 2.5 0.75 1.9 4.963 0. Table 6 presents some limited values of k1 and k2 .0 0.979 0.984 0.8 80 81 81 81 81 82 82 83 83 83 83 83 83 83 83 83 83 84 84 84 84 2.879 2.3 75 76 76 76 76 77 77 78 78 78 78 78 78 78 78 78 78 79 79 79 79 2.984 0. k1 2. σ = 1. C bCpm for Cpm and  bpmk − k1 σ bpmk + k2 σ C bCpmk .960 0.7 2.981 Finally.947 0.903 0.6 78 79 79 79 79 80 80 81 81 81 81 81 81 81 81 81 81 82 82 82 82 2.5 0.7 87 88 88 88 88 89 89 90 90 90 90 90 90 90 90 90 90 91 91 91 91 3. σ = 1.759 0.977 2. see Table 7.50.5 86 87 87 87 87 88 88 89 89 89 89 89 89 89 89 89 89 90 90 90 90 3.976 0.3 2.7 and k2 ≥ 3.949 2.5 and |µ − T | = 5 when 2 ≤ k1 ≤ 3 and 2 ≤ k2 ≤ 3.4 77 78 78 78 78 79 79 80 80 80 80 80 80 80 80 80 80 81 81 81 81 2. C Cpk n\σ 25 50 100 φ= 1. φ = 0.946 0.975 0.1 84 85 85 85 85 86 86 87 87 87 87 87 87 87 87 87 87 88 88 88 88 3.2 84 85 85 85 85 86 86 87 87 87 87 87 87 87 87 87 87 88 88 88 88 3.915 0.0 0.820 0.964 0.1 3. This coverage is greater or equal than 90 % for k1 ≥ 2.5 and |µ − T | = 5 when 2 ≤ k1 ≤ 3 and 2 ≤ k2 ≤ 3. the coverage probabilities obtained are low.0 0.954 0.957 0.8 88 89 89 89 89 90 90 91 91 91 91 91 91 91 91 91 91 92 92 92 92 3. We did not find reliable results for Cpmk .940 0.895 0.0 3.5 0. for example the coverage is between 33 % and 48 % for n = 50.5 3.5 0.0 0.967 0.7.954 0.6 2.312 Rubén Darío Guevara & José Alberto Vargas Tabla 5: Estimation average coverage rate of Cpk for AR(1) processes using intervals bpk ± 2σ b .7 80 81 81 81 81 82 82 83 83 83 83 83 83 83 83 83 83 84 84 84 84 2. n = 50.9 80 81 81 81 81 82 82 83 83 83 83 83 83 83 83 83 83 84 84 84 84 k2 3.5 2.8 3.5 78 79 79 79 79 80 80 81 81 81 81 81 81 81 81 81 81 82 82 82 82 2.50.3 85 86 86 86 86 87 87 88 88 88 88 88 88 88 88 88 88 89 89 89 89 3.0 89 90 90 90 90 91 91 92 92 92 92 92 92 92 92 92 92 93 93 93 93 Revista Colombiana de Estadística 30 (2007) 301–316 .

1 44 45 45 45 46 46 46 47 48 48 48 49 49 49 49 51 51 52 52 53 53 3.2 2.0 2.3 36 37 37 37 38 38 38 39 40 40 40 41 41 41 41 43 43 44 44 45 45 2.5 46 47 47 47 48 48 48 49 50 50 50 51 51 51 51 53 53 54 54 55 55 3.4 46 47 47 47 48 48 48 49 50 50 50 51 51 51 51 53 53 54 54 55 55 3.7 41 42 42 42 43 43 43 44 45 45 45 46 46 46 46 48 48 49 49 50 50 2.1 3. A. we show that the higher the autocorrelation level the lower the capability index value.2 44 45 45 45 46 46 46 47 48 48 48 49 49 49 49 51 51 52 52 53 53 3.0 3.50. that increases slightly for autocorrelated data when n increases.7 2.4 2.9 49 50 50 50 51 51 51 52 53 53 53 54 54 54 54 56 56 57 57 58 58 4. for µ = T and µ 6= T .1 2. Conclusions bpm and C bpmk for stationary We found approximations of the variances of C gaussian processes.8 3.9 42 43 43 43 44 44 44 45 46 46 46 47 47 47 47 49 49 50 50 51 51 k2 3. In particular. Boyles. The autocorrelation does not affect the expected value of the sample mean of the capability indices estimators but affect the estimated expected value of the standard error. we show expressions for the variance of these estimators when the observations are independent. Revista Colombiana de Estadística 30 (2007) 301–316 .4 37 38 38 38 39 39 39 40 41 41 41 42 42 42 42 44 44 45 45 46 46 2.313 Comparison of Process Capability Indices under Autocorrelated Data Tabla 7: Interval estimation average coverage rate of Cpmk for AR(1) processes with φ = 0.8 2. Applied Statistics 39.5 3.6 39 40 40 40 41 41 41 42 43 43 43 44 44 44 44 46 46 47 47 48 48 2. Journal of Quality Technology 23.6 46 47 47 47 48 48 48 49 50 50 50 51 51 51 51 53 53 54 54 55 55 3.8 41 42 42 42 43 43 43 44 45 45 45 46 46 46 46 48 48 49 49 50 50 2.0 33 34 34 34 35 35 35 36 37 37 37 38 38 38 38 40 40 41 41 42 42 2. Through a simulation study.3 45 46 46 46 47 47 47 48 49 49 49 50 50 50 50 52 52 53 53 54 54 3.6 2. We also observed that for autocorrelated processes the estimators are slightly biased. 17–26. Recibido: julio de 2007 Aceptado: octubre de 2007 Referencias Bissel.2 3. k1 2. bias that decreases as n increases.5.2 35 36 36 36 37 37 37 38 39 39 39 40 40 40 40 42 42 43 43 44 44 2. ‘How Reliable is your Capability Index?’.4 3. 331–340. (1991).0 44 45 45 45 46 46 46 47 48 48 48 49 49 49 49 51 51 52 52 53 53 3. R.8 46 47 47 47 48 48 48 49 50 50 50 51 51 51 51 53 53 54 54 55 55 3.5 38 39 39 39 40 40 40 41 42 42 42 43 43 43 43 45 45 46 46 47 47 2. ‘The Taguchi Capability Index’.0 49 50 50 50 51 51 51 52 53 53 53 54 54 54 54 56 56 57 57 58 58 7.1 33 34 34 34 35 35 35 36 37 37 37 38 38 38 38 40 40 41 41 42 42 2.6 3. σ = 1.3 3.9 4.9 3.3 2.0 2. n = 50. F.7 46 47 47 47 48 48 48 49 50 50 50 51 51 51 51 53 53 54 54 55 55 3. (1990). A.5 2.7 3.

& Yeung. In this appendix. (1997). Spiring.. McGrawHill. P. Comparación de los índices de capacidad cp . F. Mood. H. R. Journal of Quality Technology 34. Shore. (1998). 615–626. Cheng. Tesis de Maestría. 559–574. Facultad de Ciencias. Quality and Reliability Engineering International 19. Kotz.314 Rubén Darío Guevara & José Alberto Vargas Brockwell. ‘Estimating Process Capability Indexes for Autocorrelated Data’. Zhang. N. ‘Process Capability Analysis when Data are Autocorrelated’. Bogotá. Guevara. (2003). Using the estimator of Cpm : bpm = qU SL − LSL C 2 6 S2 + X − T  bpm . Departamento de Estadística. Leung. Estadística. & Davis. Introduction to the Theory of Statistics. 445– 460. we derive an approximation of the variance of C Let {Xt } be a stationary gaussian process. B. F. S. Quality Engineering 9(4). S. (2005). 2–19. cpm y cpmk con datos autocorrelacionados. New York. (2002). cpk . & Johnson. A. J. ‘A Bibliography of Process Capability Papers’. V ar C bpm is the variance of C  bpm = V ar C  U SL − LSL 6 2 V ar ( 2 S + X−T 2 − 12 ) Using the approximation 2 . D. ‘Process Capability Indices – A Review. L. N. Graybill & Boes (1974). A.. Universidad Nacional de Colombia. A. (1996). R. Apéndice A. Introduction to Time Series and Forecasting. 19922000’. Journal of Applied Statistics 25. Springer. New York. bpm .

∂ .

+ h(x. y).

∂x µx. µy 2    .

.

.

∂ ∂ ∂ .

.

.

y). Y ] h(x. V ar[Y ] + 2Cov[X.

h(x. y).

h(x. y).

µy   V ar h(X. Y ) ≈ V ar[X]  (see Mood et al. . 1974): Revista Colombiana de Estadística 30 (2007) 301–316 . µy ∂y µx. µy µx. ∂y ∂x µx.

X   h 2 i− 21 2 = V ar S + X − T  2 2 − 32 .315 Comparison of Process Capability Indices under Autocorrelated Data  h 2 i− 12 Let h S 2 . then  2 V ar h S . X = S 2 + X − T .

.

1 2 ≈ V ar(S ) − S + X − T .

2 2 E(S ).E(X) 2     2 − 32 .

.

1 + V ar X − S 2 + X − T 2 X −T .

2 E(S 2 ).E(X)   3.

    2 − 2 .

X − S 2 + X − T . 1 + 2Cov S 2 .

2 2 E(S ).E(X)   3 .

  −   1 2 2 .

2 X −T .

ρi ) + n g(n. ρi )[µ − 4[σ 2 f (n. X  ≈ ≈ σ2 2σ4 (n−1)2 F (n. ∗ − S2 + X − T 2 E(S 2 ). ρi ) + (µ − T )2 ]3 σ2  T ]2 2σ2 4 (n−1)2 F (n. (3) and (4)  2 V ar h S .E(X) 2 where Cov[S 2 . X       2 − 23 2 1 2 − E S + E X −T ≈ V ar S 2     2 i− 32    2  1h 2 − 2 E X −T E S + E X −T + V ar X 2 2     2 i−3 n   2 o 1h E S2 + E X − T V ar S 2 + 4V ar X E X − T 4    2 V ar S 2 + 4V ar X E X − T ≈ h   2 i3 4 E S2 + E X − T ≈ Using (1). (2). ρi ) + 4 n g(n. X] = 0 (see Zhang 1998). ρi ) + (µ − T )2 ]3 T ]2 Revista Colombiana de Estadística 30 (2007) 301–316 . Then  2 V ar h S . ρi )[µ − 4[σ 2 f (n.

ρi ) + (µ − T )2 ]3     2 2σ2 4 4 2 U SL − LSL  σ (n−1)2 F (n. ρi )[µ − T ] U SL − LSL bpm ≈   V ar C 6 4[σ 2 f (n. the variance of C    2  2σ2 4 2 σ 2 (n−1)  2 F (n. ρi ) σ  ≈ Cp2   h  i3 µ−T 2 4 f (n. ρi ) + σ Let ξ = µ−T σ . ρi ) + µ−T σ n  µ−T 2 o  4 2  (n−1)2 F (n. . bpmk .   bpm ≈ C 2  V ar C p 2F (n. ρi ) + n g(n. ρi 4g(n. ρi )[µ − T ]  ≈ 6σ 4[σ 2 f (n. ρi ) + n g(n.316 Rubén Darío Guevara & José Alberto Vargas bpm is Therefore. Writing the estimator of Cpmk as function of X and S 2 . ρi )[µ − T ] (n−1)  ≈ Cp2  4[σ 2 f (n. ρ ) + σ i 2 i (n−1) n σ   ≈ Cp2   i3 h  2 4σ 6 f (n. Let {Xt } In this appendix we derive an approximation of the variance of C be a stationary gaussian process. ρi ) + (µ − T )2 ]3    4 2σ2 2 4 σ 2 F (n.ρi ) (n−1)2 + 4[f (n. ρi ) + (µ − T )2 ]3  n  µ−T 2 o  4 2 6 g(n. ρ ) F (n.ρi )ξ 2 n ) + ξ 2 ]3   Apéndice B. ρi ) + n g(n. ρi ) + n g(n.

.

 a − .

2X − b.

 2 2   ∂h(S 2 . b Cpmk = h = h X. S 2 2 i 12 6 S2 + X − T we have. x) .

.

∂h(S 2 . x) .

.

+ V ar X .

.

S = a−|2x−b| − 2 S + (x − T )2 2 after some algebra.µX 3  1  2 − 2 ) where ∂h(x. pi ) g(n.µx µS 2 . we obtain ∂S 2 6    1 2 [ × V ar Cpmk ≈ Cpk f (n. pi ) + ξ 2 (  2 ) F (n. pi ) 1 6ξ + + 2(n − 1)2 [f (n. pi ) + ξ 2 ]2 9n Cpk 2[f (n. pi ) + ξ 2 ]     bpmk ≈ V ar S 2 V ar C Revista Colombiana de Estadística 30 (2007) 301–316 . ∂S 2 ∂x µS 2 .