Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Research ArticleData Transformation for Confidence IntervalImprovement An Application to the Estimation ofStress-Strength Model Reliability
Alessandro Barbiero
Department of Economics Management and Quantitative Methods Universita degli Studi di Milano Via Conservatorio7-20122 Milan Italy
Correspondence should be addressed to Alessandro Barbiero alessandrobarbierounimiit
Received 31 July 2014 Revised 19 September 2014 Accepted 29 September 2014 Published 23 October 2014
Academic Editor David Bulger
Copyright copy 2014 Alessandro BarbieroThis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
In many statistical applications it is often necessary to obtain an interval estimate for an unknown proportion or probability ormore generally for a parameter whose natural space is the unit interval The customary approximate two-sided confidence intervalfor such a parameter based on some version of the central limit theorem is known to be unsatisfactory when its true value is closeto zero or one or when the sample size is small A possible way to tackle this issue is the transformation of the data through a properfunction that is able to make the approximation to the normal distribution less coarse In this paper we study the application ofseveral of these transformations to the context of the estimation of the reliability parameter for stress-strengthmodels with a specialfocus on Poisson distribution From this work some practical hints emerge on which transformation may more efficiently improvestandard confidence intervals in which scenarios
1 Introduction
In many fields of applied statistics it is often necessary toobtain an interval estimate for an unknown proportion orprobability or more generally for a parameter whose naturalspace is the unit interval [1] If 119901 is the unknown parameterof a binomial distribution the customary approximate two-sided confidence interval (CI) for 119901 is known to be unsatis-factory when its true value is close to zero or one or when thesample size is small In fact estimation can cause difficultiesbecause the variance of the corresponding point estimatoris dependent on 119901 itself and because its distribution can beskewed A number of papers have been devoted to the devel-opment of more refined CIs for 119901 (see eg [1ndash4]) Here wewill consider the estimation of the probability 119877 = 119875(119883 lt 119884)where 119883 and 119884 are two independent rvrsquos If 119884 representsthe strength of a certain system and 119883 the stress on it 119877represents the probability that the strength overcomes thestress and then the system works (119877 is then referred to asthe ldquoreliabilityrdquo parameter) Such a statistical model is usually
called the ldquostress-strength modelrdquo and in the last decades hasattracted much interest from various fields [5 6] rangingfrom engineering to biostatistics In these works inferentialissues have been dealt with mainly in the parametric contextThe problem of constructing interval estimators for 119877 hasbeen considered when an exact analytical solution is notavailable approximations based on the delta method andasymptotic normality of point estimators are carried outsome of themmaking use of some data transformation of thepoint estimate of 119877
In this work we concentrate on the case of Poisson-dis-tributed stress and strength Approximate large-sample CIsfor the reliability 119877 have already been built and assessed fordifferent parameter and sample size configurations and havebeen proved to give satisfactory results unless 119877 is close to 1(or symmetrically 0) or the sample sizes are too small [7]With the aim of improving the performance of such CIsfour transformation functions (logit probit arcsine rootand complementary log-log) are selected and applied to themaximum likelihood estimate of 119877 and the resulting CIs are
Hindawi Publishing CorporationAdvances in Decision SciencesVolume 2014 Article ID 485629 10 pageshttpdxdoiorg1011552014485629
2 Advances in Decision Sciences
empirically compared in terms of coverage probability andexpected length
The paper is laid out as follows in Section 2 a brief dis-cussion on data transformations is presented with a specialfocus on those ordinarily used in connection with the esti-mation of the reliability parameter of a stress-strengthmodelSection 3 introduces the stress-strengthmodel with indepen-dent Poisson-distributed stress and strength recalling theformulas for reliability maximum likelihood estimator andstandard large-sampleCI and introducing refinements for thelatter based on data transformations Section 4 is devoted totheMonte Carlo simulation study which empirically assessesthe statistical performance of CIs An example of applicationon real data is provided in Section 5 and Section 6 concludesthe paper with some final remarks
2 Transformations and Application tothe Estimation of a Stress-Strength Model
In almost all fields of research one has to deal with datathat are not normal It is common practice to transform thenonnormal data at hand in order to exploit theoretical resultsthat strictly hold only for the normal distribution with theobjective of building plausible or more efficient estimatesCiting [8] ldquotransformations of statistical variables are usedin applied statistics mainly for two purposesrdquo The first one isldquovariance stabilizationrdquo ldquoA transformation is applied in orderto make it possible to use at any rate approximately thestandard techniques associated with continuous normal vari-ation for example the methods of analysis of variance Inparticular transformations are required which stabilize thevariance that is which make the variance of the trans-formed variable approximately independent of for examplethe binomial probability or the mean value of the Poissondistribution [sdot sdot sdot ] The constant-variance condition has ledto the introduction of the inverse sine inverse sinh andsquare-root transformations which are used nowadays inmany fields of applicationsrdquo In linear regression models theunequal variance of the error terms produces estimates thatare unbiased but are no longer best in the sense of having thesmallest variance [9]The second purpose is ldquonormalizationrdquoldquoA transformation is used in order to facilitate the computa-tion of tail sums of the distribution by the aid of the normalprobability integral [sdot sdot sdot ] A review of the literature shows thata considerable number of transformations of binomial nega-tive binomial Poisson and1205942 variables have been proposedrdquoIn regression models for example nonnormality of theresponse variable invalidates the standard tests of significancewith small samples since they are based on the normalityassumption [9] Reference [10] noted how ldquoapproximatelysymmetrizing transformation of a random variable may bea more effective method of normalizing it than stabilizing itsvariancerdquo Reference [11] although dated and themore recentreference [12] provide an exhaustive review of transformationused in statistical data analysis
If a random variable 119883 whose probability mass functionor density function depends on a parameter 120579 is transformedby a function 119884 = 119891(119883) which we suppose henceforth
to be strictly monotone the standard deviation of 119884 accord-ing to the deltamethod (see eg [13]) is given approximatelyby
120590119891(119909)(120579) asymp
120597119891
120597119909(120583119909(120579)) sdot 120590
119909(120579) (1)
where 120583119909(120579) and 120590
119909(120579) are the expected value and standard
deviation of119883Then the (approximate) standard deviation ofthe transformed random variable 119884 can be made equal to aconstant if the function119891 is chosen so to satisfy the followingrelationship
119891 (119909) = int
119909 119896
120590119909(120579)119889120583119909(120579) (2)
Thus for example if 119883 is distributed as a Poisson randomvariable with parameter 120582 being 120583
119909= 1205902
119909= 120582 we obtain
that the function 119891(119909) = int119909 119896120582minus12119889120582 = 119896lowastradic119909 is a variancestabilizing function with 119896 and 119896lowast proper positive cons-tants Formula (2) has been empirically shown to providereasonable stabilization in various applications as confirmedby its extensive use but other criteria can be employedbased on different ldquonotionsrdquo of variance stabilization [8 12]Otherwise modifications to the function derived by (2) canbe proposed for example in order to reduce or removethe bias [14] Reference [15] studied the root transformation119891(119909) = radic119909 + 119888 for a Poisson-distributed random variable119883 and demonstrated that for 119888 = 14 the root-transformedvariable radic119883 + 119888 has vanishing first-order bias and almostconstant variance
Whichever criterion is selected very often the propertransformation to be adopted is tied to the particular statis-tical distribution underlying the data However as much asoften the exact distribution of an estimator for a probabil-ityproportion is not easily derivable even if the distributionthat the sample data came from is known This sometimeshappens for example for the maximum likelihood estimator(MLE) of the reliability parameter 119877 of a stress-strengthmodel In this case we do not know ldquoa priorirdquo whichtransformation may fit the data (ie the distribution of theMLE) best
The focus of this paper is the estimation of a probabilitythus we will confine ourselves to transformation of propor-tions For this case among the most used transformationshere we recall the logit probit arcsine and complementarylog-log Table 1 reports the expression the image the firstderivative and the inverse of these four functions
The logit and probit transformations are widely used inthe homonym models [16] and more generally when dealingwith skewed proportion distributions They are similar sincethey are both antisymmetric functions around the point 119909 =05 that is 119891(119909 + 05) = minus119891(minus119909 + 05) The differencestands in the fact that the logit function takes absolute valueslarger than the probit as can be noted looking at Figure 1With regard to the estimation of the reliability parameter for
Advances in Decision Sciences 3
Table 1 Transformations
119891(119909) Codomain 1198911015840(119909) 119891
minus1(119909)
log[119909(1 minus 119909)] (minusinfin +infin) 1[119909(1 minus 119909)] e119909(1 + e119909)Φminus1(119909) (minusinfin +infin) 1120601(Φ
minus1(119909)) Φ(119909)
arcsin(radic119909) (0 +1205872) 1[2radic119909(1 minus 119909)] sin2 (119909)log[minus log(1 minus 119909)] (minusinfin +infin) 1[(1 minus 119909)| log(1 minus 119909)|] 1 minus eminuse
119909
00 02 04 06 08 10
minus4
minus2
0
2
4
x
Logit
Logit
Probit
Probit
Arcsin
Arcsin
Log-log
Log-log
Figure 1 Plot of transformation functions and dotted horizontallines of equation 119910 = 0 and 119909 = 05
the stress-strength model the logit transformation has beenby far the most used transformation for improving thestatistical performance of standard large-sample CIs amongothers it has been considered by [17ndash19] all these contri-butions concerning the Weibull distribution by [20] for theLindley distribution by [21] when stress and strength followa bivariate exponential distribution and by [22 23] in anonparametric context
Through the years the arcsine root transformation hasgained great favor and application among practitioners [24]perhaps more than its real merit As noted previously itshould be chosen for stabilizing binomial data but is oftenused to stabilize sample proportions as well (ie relativebinomial data) However it presents a drawback highlightedin [25] and related to the fact that its codomain is the limitedinterval (0 1205872) thus its normalizing effect may turn out tobe meagre as already pointed out in some studies where ithas been used for estimating the reliability of stress-strengthmodels [19 25 26]
The complementary log-log function which is sometimesused in binomial regression models is slightly different fromlogit and probit since it assumes negative values for 119909 lt 1 minuseminus1 (and positive values for 119909 gt 1 minus eminus1) and takes large posi-tive values only for values of 119909 very close to 1 for example ittakes values larger than 1 if and only if 119909 gt 1 minus eminuse = 0934
In the next section we will apply these transformations tothe estimation of a stress-strengthmodel with both stress andstrength following a Poisson distribution
3 Inference on the ReliabilityParameter for a Stress-Strength Model withIndependent Poisson Stress and Strength
Let119883 and119884be independent rvrsquosmodeling stress and strengthrespectively with 119883 sim Poisson (120582
119909) and 119884 sim Poisson (120582
119910)
Then the reliability 119877 = 119875(119883 lt 119884) of the stress-strengthmodel is given by (see [6 page 103])
119877 = 119875 (119883 lt 119884) =
+infin
sum
119894=0
120582119894
119909119890minus120582119909
119894[
[
1 minus
119894
sum
119895=0
120582119895
119910119890minus120582119910
119895]
]
= lim119896rarrinfin
119896
sum
119894=0
120582119894
119909119890minus120582119909
119894[
[
1 minus
119894
sum
119895=0
120582119895
119910119890minus120582119910
119895]
]
(3)
If two simple random samples of size 119899119909and 119899
119910from 119883 and
119884 respectively are available the reliability parameter can beestimated by the ML estimator obtained substituting in (3)the MLEs of the unknown parameters 120582
119909and 120582
119910
=
+infin
sum
119894=0
119909119894119890minus119909
119894[
[
1 minus
119894
sum
119895=0
119910119895119890minus119910
119895]
]
(4)
Deriving the expression of the variance of the MLE is notstraightforward however an approximate expression hasbeen easily derived through the delta method in [7] Basedupon this approximation a large-sample 1 minus 120572 CI for 119877 hasbeen built Such an interval estimator has the usual expression
(119877119871 119877119880) = ( minus 119911
1minus1205722radicV () + 119911
1minus1205722radicV ()) (5)
where V() is the sample estimate of the (asymptotic) varianceof (see [7] for details)
Although such an estimator has been proved to havea satisfactory behavior in terms of coverage for severalcombinations of sample sizes and values of the reliabilityparameter some decay of the performance is observed when119877 gets close to the extreme values 0 and 1 and when samplesizes are small In these cases the approximation to thenormal distribution is in fact very rough especially because ofthe skewness of the distribution of Thus transformationsof the values of the estimates can be considered in orderto ldquomake the data more normalrdquo and produce CIs based ontransformed data that have a coverage closer to the nominallevel
4 Advances in Decision Sciences
Table 2 CIs of the reliability parameter 119877 under variance-stabilizingnormalizing transformations
Transformation 120579 CI for 120579 (120579119871 120579119880) CI for 119877 (119877
119871 119877119880)
logit log[(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )]) (exp (120579
119871) (1 + exp (120579
119871)) exp (120579
119880) (1 + exp (120579
119880)))
Probit Φminus1() (120579 ∓ 119911
1minus1205722radicV()120601(Φminus1())) (Φ(120579
119871) Φ(120579
119880))
Arcsin arcsinradic (120579 ∓ 1199111minus1205722radicV()[2radic(1 minus )]) (sin2 (120579
119871) sin2 (120579
119880))
Loglog log[minus log(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )| log(1 minus )|]) (1 minus exp (minus exp (120579
119871)) 1 minus exp (minus exp (120579
119880)))
Recalling Table 1 and considering the logit transforma-tion of the reliability parameter 119877 120579 = log[119877(1 minus 119877)] theMLE of 120579 is 120579 = log[(1minus )] and an approximate (1minus120572) CIfor 120579 is
(120579119871 120579119880) = (120579 minus 119911
1minus1205722sdot
radicV ()
[ (1 minus )]
120579 + 1199111minus1205722
sdot
radicV ()
[ (1 minus )])
(6)
Then an approximate (1 minus 120572) CI for 119877 is
(119877119871 119877119880) = (
exp (120579119871)
[1 + exp (120579119871)]
exp (120579119880)
[1 + exp (120579119880)]) (7)
For the other transformations following the same stepsjust outlined approximate ldquotransformedrdquo CIs for 119877 can beobtained which are alternative to the standard naıve CI of(5) they are synthesized in Table 2
An alternative way to make inference on the reliabilityparameter 119877 for the stress-strength model with independentPoisson random variables can be summarized as follows (1)to transform (normalize) samples from 119883 and 119884 accordingto the a proper transformation for the Poisson distribution(2) to compute point and interval estimates of 119877 usingthe methods for a stress-strength model with independentnormal variables with known variances [6 page 112] LettingΞ = radic119883 + 14 and Υ = radic119884 + 14 then Ξ and Υ are inde-pendent and approximately distributed as normal randomvariables with variance 1205902
120585= 1205902
120592= 14 and expected value
120583120585= radic120582
119909and 120583
120592= radic120582119910 respectively Then instead of esti-
mating 119877 = 119875(119883 lt 119884) one can estimate 1198771015840 = 119875(Ξ lt Υ) a1 minus 120572 confidence interval for 1198771015840 is given by
(Φ[120578 minus1199111minus1205722
radic119872] Φ[120578 +
1199111minus1205722
radic119872]) (8)
with119872 = (1205902120585+1205902
120592)(1205902
120585119899119909+1205902
120592119899119910) and 120578 = (120592 minus 120585)120590 being
an estimator of 120578 = (120583120592minus 120583120585)120590 where 120585 and 120592 are the sample
means of Ξ and Υ respectively and 120590 = radic1205902120585+ 1205902120592= radic12
An alternative procedure tomake inference about the reli-ability parameter 119877 can be provided by parametric bootstrap
[27 pages 53ndash56] In the basic version it works as follows forthe estimation problem at hand
(1) Estimate the unknown parameters of the Poisson rvrsquos119883 and 119884 through their sample means 119909 and 119910
(2) Draw independently a bootstrap sample 119909lowast of size119899119909from a Poisson rv 119883lowast with parameter 119909 and a
bootstrap sample 119910lowast of size 119899119910from a Poisson rv 119884lowast
with parameter 119910
(3) Estimate the reliability parameter say lowast on samples119909lowast and 119910lowast using the very same expression in (4)
(4) Repeat steps 2 and 3 119861 times (119861 sufficiently large eg2000) thus obtaining the bootstrap distribution oflowast
(5) Estimate a (1 minus 120572) bootstrap percentile CI for 119877 fromlowast distribution taking the 1205722 and 1 minus 1205722 quantiles
(lowast
1205722 lowast
1minus1205722) (9)
4 Simulation Study
41 Scope and Design The simulation study aims at empir-ically comparing the performance of the interval estimatorspresented in the previous section namely the standard CI of(5) labeled ldquoANrdquo those of Table 2 (labeled ldquologitrdquo ldquoprobitrdquoldquoarcsinerdquo ldquologlogrdquo) and the CIs of (8) and (9) in terms ofcoverage rate (and also lower and upper uncoverage rates)and expected length In this Monte Carlo (MC) study thevalue of parameter 120582
119909of the Poisson distribution for stress
119883 is set equal to a ldquoreferencerdquo value 2 and the parameter120582119910of the Poisson distribution modeling strength is varied in
order to obtainmdashaccording to (3)mdashseveral different levels ofreliability119877 namely 01 02 03 04 05 06 07 08 09 and095
For each couple (120582119909 120582119910) a large number (119899119878 = 5000) of
samples of size 119899119909and of size 119899
119910are drawn from 119883 sim
Poisson (120582119909) and 119884 sim Poisson (120582
119910) independently Different
and unequal sample sizes are here considered (all the 16possible combinations between the values 119899
119909= 5 10 20 50
and 119899119910= 5 10 20 50) On each pair of samples the 95 CIs
for 119877 listed above are built and the MC coverage rate andthe length of the CIs are computed over the 119899119878MC samplesMoreover the lower and upper uncoverage or error rates arecomputed that is the proportion of CIs for which 119877 lt 119877
119871
and 119877 gt 119877119880 respectively
Advances in Decision Sciences 5
Note that for the smallest sample size that is when 119899119909= 5
(or 119899119910= 5) a practical problem arises as the sample values of
119883 (or 119884) may all be 0 in this case the variance estimate V()cannot be computed and a standard approximate CI cannotbe built We decided to discard these samples from the5000 MC samples planned for the simulation study andto compute the quantities of interest only on the ldquofeasiblerdquosamples Indeed the rate of such ldquononfeasiblerdquo samples is inany case very low under each scenario For the worst onescharacterized by 119877 = 01 and 119899
119910= 5 the theoretical rate of
ldquononfeasiblerdquo samples is about 5Since results for the coverage rates of some CIs are often
close to the nominal level 095 we performed a test in order tocheck if the actual coverage rate is significantly different from095 that is to state if such CIs are conservative or liberalThe null hypothesis is that the true rate 120578 of CIs that do notcover the real value 119877 is equal to 120572 = 005119867
0 120578 = 120572 = 005
whereas the alternative hypothesis is that the rate 120578 is differentfrom120572119867
1 120578 = 120572We employ the test suggested in [28 pages
518-519] which is based on the statistic
F =119899119877 + 1
119899119878 minus 119899119877sdot1 minus 120572
120572 (10)
where 119899119877 is the number of rejections of1198670in 119899119878 (5000 in our
case unless there are some nonfeasible samples) iterations ofaMC simulation plan Under119867
0F follows an119865 distribution
with ]1= 2(119899119878 minus 119899119877) and ]
2= 2(119899119877 + 1) degrees of freedom
The test rejects 1198670at level 120574 if either F le F
1205742]1]2
or F geF1minus1205742]
1]2
where F120574]1]2
denotes the 120574 percentile point ofan 119865 distribution with ]
1and ]2degrees of freedom The test
was performed on each scenario for each kind of CI at a level120574 = 1
42 Results and Discussion The simulation results are (par-tially) reported in Table 3 (120582
119910= 2 varying 119877 and the
common size 119899119909= 119899119910) The results about the CIs (8) and
(9) are not reported here as their performances are overallunsatisfactory Even if theoretically appealing the proce-dure leading to the CI in (8) practically fails some of thescenarios of the simulation plan were considered and theCI built following this alternative procedure never providessatisfactory results For the smallest sample size (119899 = 5) thecoverage proves to be larger than that provided by the ANCI but however smaller than the nominal one for the largersample size (119899 = 10 20 50) the coverage rate dramaticallydecreases (even to values as low as 60) Paradoxically thediscrete nature of the Poisson variable and thus the qualityof the normal approximation of step (1) affect results to amore relevant extent as the sample size increases As to thebootstrap procedure yielding the CI in (9) this solution iscomputationally cumbersome it becomes even more timeconsuming if bias-corrected accelerated CIs [27 pages 184ndash188] have to be calculated and it provides barely satisfactoryresults as proven by a preliminary simulation study notreported here that confirms the findings reported in [25] forparametric bootstrap inference on the reliability parameterin the bivariate normal case The rejection of 119867
0based on
the test statistic (10) described in the previous subsection
is indicated by a ldquolowastrdquo near the actual coverage rate of eachCI Figures 2 and 3 graphically display the values of lowerand upper uncoverage rates for the five interval estimatorsvarying 119877 in 05 06 07 08 09 095 for 119899
119909= 119899119910= 5 and
119899119909= 119899119910= 50 respectively These results comprise only the
equal sample size scenarios those for unequal sample sizescenarios do not add much value to the general discussionwe are going to outline
(i) First the improvement provided by the four trans-formations to the standard naıve CI of 119877 in terms ofcoverage is evidentThis improvement is considerablefor a small sample size and extreme values of 119877(say 095) when the coverage rate of the standardnaıve CI tends to decrease dramatically even below90 Note that for the sample sizes 5 10 and 20the actual coverage rate of the standard naıve CIis always significantly different (smaller) than thenominal level Under some scenarios namely for 119877 =03ndash07 the increase in coverage rate is accompaniedalso by a reduction in the CI average length On thecontrary for the other values of 119877 there is an increasein the average width of the modified CIs which ismuch more apparent for the logit transformation
(ii) Examining closely Table 3 one can note that the CIbased on the arcsine root transformation is exceptfor one scenario uniformly worse in terms of actualcoverage than the other three CIs based on logitprobit and complementary log-log transformationsFor small sample sizes (119899
119909= 5 10) the coverage rate
actually provided is significantly different (smaller)than the nominal one It can also be claimed that thisunsatisfactory performance is due to its incapabilityof symmetrizing the distribution as can be seen byglancing at the lower and upper uncoverage rates forvalues of 119877 getting close to 1 there is an apparentundercoverage effect on the left side (just a bit smallerthan that of the ldquooriginalrdquo standard approximate CI)
(iii) Logit probit and complementary log-log trans-formed CIs have overall a satisfactory performancein terms of closeness to the nominal confidence levelLogit and probit CIs exhibit a similar behavior as boththe lower and upper uncoverage rates they provide areclose to the nominal value (25) On the contrarycomplementary log-log function (as well as arcsineroot) often produces a larger undercoverage on oneside (here left) which is partially balanced by an over-coverage on the other side (here right) This is clearevidence of the higher symmetrizing capability ofthe logit and probit transformations However takinginto account the results of the hypothesis test based onthe statistic (10) the probit and complementary log-log functions are those that overall perform best thestatistical hypothesis that their actual coverage rate isequal to the nominal one is always accepted exceptfor one case (119877 = 01 and 119899 = 5) whereas thesame hypothesis is sometimes rejected (10 times outof 40) for the logit function (which tends to producesignificantly larger coverage for small sample sizes)
6 Advances in Decision Sciences
Table3Simulationresults
(120582119909=2)
119877119899119909119899119910
AN
Logit
Prob
itArcsin
Loglog
AN
Logit
Prob
itArcsin
Loglog
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
Width
01
55
8994lowast
034
972
9717lowast
266
017
9673lowast
207
120
9435
112
453
9620lowast
363
017
02836
03835
03491
03152
04204
1010
8849lowast
040
1111
9623lowast
251
126
9557
185
259
9252lowast
106
642
9565
323
112
02153
02501
02354
02214
02622
2020
9226lowast
036
738
9520
306
174
9494
232
274
9362lowast
144
494
9512
326
162
01620
01709
01653
01600
01750
5050
9334lowast
094
572
9544
232
224
9528
198
274
9450
156
394
9548
254
198
01020
01044
01029
01015
01054
02
55
8751lowast
166
1083
9601lowast
311
088
9511
253
237
9166lowast
166
668
9519
429
052
04360
04717
04553
044
2505123
1010
9068lowast
112
820
9530
268
202
9472
222
306
9314lowast
174
512
9486
364
150
03395
03429
03358
03312
03597
2020
9320lowast
136
544
9522
266
212
9522
218
260
9432
180
388
9490
322
188
02469
02473
02444
02428
02537
5050
9396lowast
138
466
9542
248
210
9516
220
264
9478
196
326
9520
276
204
01569
01572
01563
01559
01588
03
55
8911lowast
250
839
9642lowast
224
134
9472
224
304
9222lowast
224
554
9472
448
080
05440
05284
05245
05264
05664
1010
9110lowast
192
698
9540
248
212
9490
206
304
9370lowast
192
438
9428
412
160
04149
03991
03981
040
0204166
2020
9318lowast
182
500
9534
256
210
9504
236
260
9398lowast
232
370
9496
348
156
03012
02947
02945
02956
03018
5050
9438
182
380
9546
242
212
9532
220
248
9510
196
294
9540
280
180
01927
01909
01909
01912
01929
04
55
8898lowast
386
716
9652lowast
210
138
9448
232
320
9240lowast
290
470
9504
414
082
060
6805592
05627
05733
05891
1010
9214lowast
270
516
9586lowast
208
206
9478
228
294
9372lowast
246
382
9498
382
162
04572
04305
04335
04397
044
5520
209364lowast
260
376
9558
244
198
9522
244
234
9470
248
282
9514
332
154
03319
03210
03226
03254
03274
5050
9472
218
310
9548
220
232
9528
220
252
9506
220
274
9564
252
184
02128
02098
02103
02111
02116
05
55
8932lowast
504
564
9672lowast
156
172
9504
238
258
9278lowast
348
374
9490
434
076
06285
05691
05752
05889
05864
1010
9198lowast
382
420
9554
204
242
9494
236
270
9384lowast
278
338
9488
362
150
04709
044
06044
5004526
04501
2020
9350lowast
332
318
9536
230
234
9480
254
266
9464
262
274
9476
350
174
03413
03290
03312
03345
03332
5050
9462
246
292
9538
218
244
9516
232
252
9496
234
270
9542
272
186
02191
02157
02163
02173
02169
06
55
8944lowast
636
420
9652lowast
144
204
9470
250
280
9266lowast
404
330
9478
428
094
060
9505587
05628
05741
05601
1010
9178lowast
484
338
9552
178
270
9488
232
280
9392lowast
306
302
9492
348
160
04576
04305
04337
044
0104324
2020
9340lowast
402
258
9544
202
254
9512
234
254
9446
298
256
9500
334
166
03304
03197
03212
03240
03206
5050
9506
268
226
9554
198
248
9538
214
248
9524
228
248
9558
258
184
02120
02090
02095
02103
02093
07
55
8952lowast
754
294
9606lowast
148
246
9504
250
246
9256lowast
476
268
9454
454
092
05490
05266
05242
05279
05094
1010
9170lowast
582
248
9532
164
304
9502
222
276
9378lowast
366
256
9492
344
164
04162
03991
03987
040
1403919
2020
9328lowast
468
204
9562
190
248
9518
234
248
9444
326
230
9530
326
144
02984
02920
02918
02930
02889
5050
9474
324
202
9548
188
264
9540
216
244
9524
252
224
9538
264
198
01910
01893
01893
01896
01885
08
55
8894lowast
960
146
9600lowast
124
276
9492
232
276
9254lowast
578
168
9500
418
082
044
1904657
04526
044
3604289
1010
9166lowast
702
132
9518
156
326
9496
234
270
9370lowast
418
212
9522
336
142
03411
03408
03349
03315
03242
2020
9322lowast
564
114
9522
188
290
9494
252
254
9432
364
204
9500
334
166
02424
02425
02399
02386
02356
5050
9460
394
146
9534
194
272
9520
232
248
9504
286
210
9540
270
190
01543
01544
01537
01533
01525
09
55
8832lowast
1114
054
9528
114
358
9504
258
238
9196lowast
676
128
9470
438
092
02799
03557
03295
03043
03029
1010
9080lowast
850
070
9512
144
344
9476
252
272
9340lowast
498
162
9508
330
162
02175
02410
02294
02187
02181
2020
9248lowast
692
060
9524
174
302
9512
236
252
9422lowast
428
150
9520
320
160
01550
01621
01574
01532
01531
5050
9452
462
086
9510
198
292
9530
230
240
9506
322
172
9540
262
198
00976
00996
00983
00971
00971
095
55
8686lowast
1296
018
9516
112
372
9510
262
228
9152lowast
772
076
9502
412
086
01691
02593
02289
01973
02058
1010
8990lowast
984
026
9480
148
372
9498
250
252
9262lowast
628
110
9484
346
170
01277
01611
01489
01359
01398
2020
9230lowast
750
020
9522
174
304
9512
248
240
9402lowast
476
122
9492
332
176
00931
01021
00975
00925
00941
5050
9420lowast
532
048
9530
188
282
9524
240
236
9508
346
146
9532
278
190
00579
006
0200590
00577
00582
Legend
ldquocovrdquo
=actualcoverage
rateldquoLE
rdquo=lower
errorrateldquoU
Erdquo=up
pere
rror
rate
lowastTh
erejectio
nof
H0
Advances in Decision Sciences 7
R = 05
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 08
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 09
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 095
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 06
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 07
AN
Logit
Probit
Arcsin
Log-log
Figure 2 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 5
(iv) Differences in behavior among the four ldquotransfor-medrdquo CIs tend to become more apparent for smallsample sizes and for extreme values of 119877 (close to 0 or1) as can be seen looking at the bottom right graph ofFigure 2 Clearly the same differences tend to vanishas the sample sizes increase and119877 tends to 05 (see thetop left graph in Figure 3)
(v) Contrary to what one would predict the perfor-mances in terms of coverage rate of logit probit andcomplementary log-log-based CIs do not seem to benegatively affected by sample size in fact as under-lined before an (sometimes significant) increase ofthe coverage rate of the logit CIs is noticed when thesample size is small (119899
119909= 5) whereas the other two
CIs show an almost constant trendThe CI exploitingthe arcsine root transformation is the one takingmost advantage from the increase in sample sizeObviously there is an increase in the average length ofthe CIs moving to larger sample sizes
(vi) As one could expect there is a symmetry in the behav-ior of each CI when considering two complementaryvalues of119877 (ie values of119877 summing to 1) keeping thesample size fixed the values of the coverage rate andaverage width are similar As to the uncoverage ratesthey are nearly exchangeable for AN logit probitand arscin CIs that is a lower uncoverage error fora fixed value of 119877 is similar to the upper uncoverageerror for its complementary value (or at least theirrelativemagnitude is exchangeable)This feature does
not hold for log-log CIs which always present a left-side uncoverage error larger than that of the right-side These features are weakened for 119877 equal to 09(and 1 minus 119877 then equal to 01) probably because ofthe presence of ldquononfeasiblerdquo samples as discussed inSection 41 which distorts this symmetry condition
(vii) Finallymdashthe relative results are not reported here forthe sake of brevitymdashmoving to larger values (ielarger than 2) of 120582
119909 keeping 119877 constant seems to
bring benefit to the performance of all the CIs ascan be quantitatively noted by the reduced numberof rejections of the null hypothesis of equivalencebetween actual and theoretical coverage This may beexplained by the fact that for a large parameter 120582 thePoisson distribution tends to a normal distributiontherefore larger values of 120582
119909imply a better normal
approximation to Poisson and presumptively to theMLE
In Figure 4 for illustrative purposes the MC distributionof and the transformed data according to the four transfor-mations are displayed (119877 = 08 and 119899
119909= 119899119910= 5) We can
note at a glance that logit and probit produce distributionscloser to normality than arcsine root and log-log comple-mentary functions which do not seem completely able toldquocorrectrdquo the skewness of the original distribution Howeverimplementing the Shapiro-Wilk test of normality on the fourtransformed distributions leads to very low 119875-values practi-cally equal to 0 except for the probit function whose 119875-valueis 003132 thus proving that in this case all the transformeddata distributions are still far from normality
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
2 Advances in Decision Sciences
empirically compared in terms of coverage probability andexpected length
The paper is laid out as follows in Section 2 a brief dis-cussion on data transformations is presented with a specialfocus on those ordinarily used in connection with the esti-mation of the reliability parameter of a stress-strengthmodelSection 3 introduces the stress-strengthmodel with indepen-dent Poisson-distributed stress and strength recalling theformulas for reliability maximum likelihood estimator andstandard large-sampleCI and introducing refinements for thelatter based on data transformations Section 4 is devoted totheMonte Carlo simulation study which empirically assessesthe statistical performance of CIs An example of applicationon real data is provided in Section 5 and Section 6 concludesthe paper with some final remarks
2 Transformations and Application tothe Estimation of a Stress-Strength Model
In almost all fields of research one has to deal with datathat are not normal It is common practice to transform thenonnormal data at hand in order to exploit theoretical resultsthat strictly hold only for the normal distribution with theobjective of building plausible or more efficient estimatesCiting [8] ldquotransformations of statistical variables are usedin applied statistics mainly for two purposesrdquo The first one isldquovariance stabilizationrdquo ldquoA transformation is applied in orderto make it possible to use at any rate approximately thestandard techniques associated with continuous normal vari-ation for example the methods of analysis of variance Inparticular transformations are required which stabilize thevariance that is which make the variance of the trans-formed variable approximately independent of for examplethe binomial probability or the mean value of the Poissondistribution [sdot sdot sdot ] The constant-variance condition has ledto the introduction of the inverse sine inverse sinh andsquare-root transformations which are used nowadays inmany fields of applicationsrdquo In linear regression models theunequal variance of the error terms produces estimates thatare unbiased but are no longer best in the sense of having thesmallest variance [9]The second purpose is ldquonormalizationrdquoldquoA transformation is used in order to facilitate the computa-tion of tail sums of the distribution by the aid of the normalprobability integral [sdot sdot sdot ] A review of the literature shows thata considerable number of transformations of binomial nega-tive binomial Poisson and1205942 variables have been proposedrdquoIn regression models for example nonnormality of theresponse variable invalidates the standard tests of significancewith small samples since they are based on the normalityassumption [9] Reference [10] noted how ldquoapproximatelysymmetrizing transformation of a random variable may bea more effective method of normalizing it than stabilizing itsvariancerdquo Reference [11] although dated and themore recentreference [12] provide an exhaustive review of transformationused in statistical data analysis
If a random variable 119883 whose probability mass functionor density function depends on a parameter 120579 is transformedby a function 119884 = 119891(119883) which we suppose henceforth
to be strictly monotone the standard deviation of 119884 accord-ing to the deltamethod (see eg [13]) is given approximatelyby
120590119891(119909)(120579) asymp
120597119891
120597119909(120583119909(120579)) sdot 120590
119909(120579) (1)
where 120583119909(120579) and 120590
119909(120579) are the expected value and standard
deviation of119883Then the (approximate) standard deviation ofthe transformed random variable 119884 can be made equal to aconstant if the function119891 is chosen so to satisfy the followingrelationship
119891 (119909) = int
119909 119896
120590119909(120579)119889120583119909(120579) (2)
Thus for example if 119883 is distributed as a Poisson randomvariable with parameter 120582 being 120583
119909= 1205902
119909= 120582 we obtain
that the function 119891(119909) = int119909 119896120582minus12119889120582 = 119896lowastradic119909 is a variancestabilizing function with 119896 and 119896lowast proper positive cons-tants Formula (2) has been empirically shown to providereasonable stabilization in various applications as confirmedby its extensive use but other criteria can be employedbased on different ldquonotionsrdquo of variance stabilization [8 12]Otherwise modifications to the function derived by (2) canbe proposed for example in order to reduce or removethe bias [14] Reference [15] studied the root transformation119891(119909) = radic119909 + 119888 for a Poisson-distributed random variable119883 and demonstrated that for 119888 = 14 the root-transformedvariable radic119883 + 119888 has vanishing first-order bias and almostconstant variance
Whichever criterion is selected very often the propertransformation to be adopted is tied to the particular statis-tical distribution underlying the data However as much asoften the exact distribution of an estimator for a probabil-ityproportion is not easily derivable even if the distributionthat the sample data came from is known This sometimeshappens for example for the maximum likelihood estimator(MLE) of the reliability parameter 119877 of a stress-strengthmodel In this case we do not know ldquoa priorirdquo whichtransformation may fit the data (ie the distribution of theMLE) best
The focus of this paper is the estimation of a probabilitythus we will confine ourselves to transformation of propor-tions For this case among the most used transformationshere we recall the logit probit arcsine and complementarylog-log Table 1 reports the expression the image the firstderivative and the inverse of these four functions
The logit and probit transformations are widely used inthe homonym models [16] and more generally when dealingwith skewed proportion distributions They are similar sincethey are both antisymmetric functions around the point 119909 =05 that is 119891(119909 + 05) = minus119891(minus119909 + 05) The differencestands in the fact that the logit function takes absolute valueslarger than the probit as can be noted looking at Figure 1With regard to the estimation of the reliability parameter for
Advances in Decision Sciences 3
Table 1 Transformations
119891(119909) Codomain 1198911015840(119909) 119891
minus1(119909)
log[119909(1 minus 119909)] (minusinfin +infin) 1[119909(1 minus 119909)] e119909(1 + e119909)Φminus1(119909) (minusinfin +infin) 1120601(Φ
minus1(119909)) Φ(119909)
arcsin(radic119909) (0 +1205872) 1[2radic119909(1 minus 119909)] sin2 (119909)log[minus log(1 minus 119909)] (minusinfin +infin) 1[(1 minus 119909)| log(1 minus 119909)|] 1 minus eminuse
119909
00 02 04 06 08 10
minus4
minus2
0
2
4
x
Logit
Logit
Probit
Probit
Arcsin
Arcsin
Log-log
Log-log
Figure 1 Plot of transformation functions and dotted horizontallines of equation 119910 = 0 and 119909 = 05
the stress-strength model the logit transformation has beenby far the most used transformation for improving thestatistical performance of standard large-sample CIs amongothers it has been considered by [17ndash19] all these contri-butions concerning the Weibull distribution by [20] for theLindley distribution by [21] when stress and strength followa bivariate exponential distribution and by [22 23] in anonparametric context
Through the years the arcsine root transformation hasgained great favor and application among practitioners [24]perhaps more than its real merit As noted previously itshould be chosen for stabilizing binomial data but is oftenused to stabilize sample proportions as well (ie relativebinomial data) However it presents a drawback highlightedin [25] and related to the fact that its codomain is the limitedinterval (0 1205872) thus its normalizing effect may turn out tobe meagre as already pointed out in some studies where ithas been used for estimating the reliability of stress-strengthmodels [19 25 26]
The complementary log-log function which is sometimesused in binomial regression models is slightly different fromlogit and probit since it assumes negative values for 119909 lt 1 minuseminus1 (and positive values for 119909 gt 1 minus eminus1) and takes large posi-tive values only for values of 119909 very close to 1 for example ittakes values larger than 1 if and only if 119909 gt 1 minus eminuse = 0934
In the next section we will apply these transformations tothe estimation of a stress-strengthmodel with both stress andstrength following a Poisson distribution
3 Inference on the ReliabilityParameter for a Stress-Strength Model withIndependent Poisson Stress and Strength
Let119883 and119884be independent rvrsquosmodeling stress and strengthrespectively with 119883 sim Poisson (120582
119909) and 119884 sim Poisson (120582
119910)
Then the reliability 119877 = 119875(119883 lt 119884) of the stress-strengthmodel is given by (see [6 page 103])
119877 = 119875 (119883 lt 119884) =
+infin
sum
119894=0
120582119894
119909119890minus120582119909
119894[
[
1 minus
119894
sum
119895=0
120582119895
119910119890minus120582119910
119895]
]
= lim119896rarrinfin
119896
sum
119894=0
120582119894
119909119890minus120582119909
119894[
[
1 minus
119894
sum
119895=0
120582119895
119910119890minus120582119910
119895]
]
(3)
If two simple random samples of size 119899119909and 119899
119910from 119883 and
119884 respectively are available the reliability parameter can beestimated by the ML estimator obtained substituting in (3)the MLEs of the unknown parameters 120582
119909and 120582
119910
=
+infin
sum
119894=0
119909119894119890minus119909
119894[
[
1 minus
119894
sum
119895=0
119910119895119890minus119910
119895]
]
(4)
Deriving the expression of the variance of the MLE is notstraightforward however an approximate expression hasbeen easily derived through the delta method in [7] Basedupon this approximation a large-sample 1 minus 120572 CI for 119877 hasbeen built Such an interval estimator has the usual expression
(119877119871 119877119880) = ( minus 119911
1minus1205722radicV () + 119911
1minus1205722radicV ()) (5)
where V() is the sample estimate of the (asymptotic) varianceof (see [7] for details)
Although such an estimator has been proved to havea satisfactory behavior in terms of coverage for severalcombinations of sample sizes and values of the reliabilityparameter some decay of the performance is observed when119877 gets close to the extreme values 0 and 1 and when samplesizes are small In these cases the approximation to thenormal distribution is in fact very rough especially because ofthe skewness of the distribution of Thus transformationsof the values of the estimates can be considered in orderto ldquomake the data more normalrdquo and produce CIs based ontransformed data that have a coverage closer to the nominallevel
4 Advances in Decision Sciences
Table 2 CIs of the reliability parameter 119877 under variance-stabilizingnormalizing transformations
Transformation 120579 CI for 120579 (120579119871 120579119880) CI for 119877 (119877
119871 119877119880)
logit log[(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )]) (exp (120579
119871) (1 + exp (120579
119871)) exp (120579
119880) (1 + exp (120579
119880)))
Probit Φminus1() (120579 ∓ 119911
1minus1205722radicV()120601(Φminus1())) (Φ(120579
119871) Φ(120579
119880))
Arcsin arcsinradic (120579 ∓ 1199111minus1205722radicV()[2radic(1 minus )]) (sin2 (120579
119871) sin2 (120579
119880))
Loglog log[minus log(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )| log(1 minus )|]) (1 minus exp (minus exp (120579
119871)) 1 minus exp (minus exp (120579
119880)))
Recalling Table 1 and considering the logit transforma-tion of the reliability parameter 119877 120579 = log[119877(1 minus 119877)] theMLE of 120579 is 120579 = log[(1minus )] and an approximate (1minus120572) CIfor 120579 is
(120579119871 120579119880) = (120579 minus 119911
1minus1205722sdot
radicV ()
[ (1 minus )]
120579 + 1199111minus1205722
sdot
radicV ()
[ (1 minus )])
(6)
Then an approximate (1 minus 120572) CI for 119877 is
(119877119871 119877119880) = (
exp (120579119871)
[1 + exp (120579119871)]
exp (120579119880)
[1 + exp (120579119880)]) (7)
For the other transformations following the same stepsjust outlined approximate ldquotransformedrdquo CIs for 119877 can beobtained which are alternative to the standard naıve CI of(5) they are synthesized in Table 2
An alternative way to make inference on the reliabilityparameter 119877 for the stress-strength model with independentPoisson random variables can be summarized as follows (1)to transform (normalize) samples from 119883 and 119884 accordingto the a proper transformation for the Poisson distribution(2) to compute point and interval estimates of 119877 usingthe methods for a stress-strength model with independentnormal variables with known variances [6 page 112] LettingΞ = radic119883 + 14 and Υ = radic119884 + 14 then Ξ and Υ are inde-pendent and approximately distributed as normal randomvariables with variance 1205902
120585= 1205902
120592= 14 and expected value
120583120585= radic120582
119909and 120583
120592= radic120582119910 respectively Then instead of esti-
mating 119877 = 119875(119883 lt 119884) one can estimate 1198771015840 = 119875(Ξ lt Υ) a1 minus 120572 confidence interval for 1198771015840 is given by
(Φ[120578 minus1199111minus1205722
radic119872] Φ[120578 +
1199111minus1205722
radic119872]) (8)
with119872 = (1205902120585+1205902
120592)(1205902
120585119899119909+1205902
120592119899119910) and 120578 = (120592 minus 120585)120590 being
an estimator of 120578 = (120583120592minus 120583120585)120590 where 120585 and 120592 are the sample
means of Ξ and Υ respectively and 120590 = radic1205902120585+ 1205902120592= radic12
An alternative procedure tomake inference about the reli-ability parameter 119877 can be provided by parametric bootstrap
[27 pages 53ndash56] In the basic version it works as follows forthe estimation problem at hand
(1) Estimate the unknown parameters of the Poisson rvrsquos119883 and 119884 through their sample means 119909 and 119910
(2) Draw independently a bootstrap sample 119909lowast of size119899119909from a Poisson rv 119883lowast with parameter 119909 and a
bootstrap sample 119910lowast of size 119899119910from a Poisson rv 119884lowast
with parameter 119910
(3) Estimate the reliability parameter say lowast on samples119909lowast and 119910lowast using the very same expression in (4)
(4) Repeat steps 2 and 3 119861 times (119861 sufficiently large eg2000) thus obtaining the bootstrap distribution oflowast
(5) Estimate a (1 minus 120572) bootstrap percentile CI for 119877 fromlowast distribution taking the 1205722 and 1 minus 1205722 quantiles
(lowast
1205722 lowast
1minus1205722) (9)
4 Simulation Study
41 Scope and Design The simulation study aims at empir-ically comparing the performance of the interval estimatorspresented in the previous section namely the standard CI of(5) labeled ldquoANrdquo those of Table 2 (labeled ldquologitrdquo ldquoprobitrdquoldquoarcsinerdquo ldquologlogrdquo) and the CIs of (8) and (9) in terms ofcoverage rate (and also lower and upper uncoverage rates)and expected length In this Monte Carlo (MC) study thevalue of parameter 120582
119909of the Poisson distribution for stress
119883 is set equal to a ldquoreferencerdquo value 2 and the parameter120582119910of the Poisson distribution modeling strength is varied in
order to obtainmdashaccording to (3)mdashseveral different levels ofreliability119877 namely 01 02 03 04 05 06 07 08 09 and095
For each couple (120582119909 120582119910) a large number (119899119878 = 5000) of
samples of size 119899119909and of size 119899
119910are drawn from 119883 sim
Poisson (120582119909) and 119884 sim Poisson (120582
119910) independently Different
and unequal sample sizes are here considered (all the 16possible combinations between the values 119899
119909= 5 10 20 50
and 119899119910= 5 10 20 50) On each pair of samples the 95 CIs
for 119877 listed above are built and the MC coverage rate andthe length of the CIs are computed over the 119899119878MC samplesMoreover the lower and upper uncoverage or error rates arecomputed that is the proportion of CIs for which 119877 lt 119877
119871
and 119877 gt 119877119880 respectively
Advances in Decision Sciences 5
Note that for the smallest sample size that is when 119899119909= 5
(or 119899119910= 5) a practical problem arises as the sample values of
119883 (or 119884) may all be 0 in this case the variance estimate V()cannot be computed and a standard approximate CI cannotbe built We decided to discard these samples from the5000 MC samples planned for the simulation study andto compute the quantities of interest only on the ldquofeasiblerdquosamples Indeed the rate of such ldquononfeasiblerdquo samples is inany case very low under each scenario For the worst onescharacterized by 119877 = 01 and 119899
119910= 5 the theoretical rate of
ldquononfeasiblerdquo samples is about 5Since results for the coverage rates of some CIs are often
close to the nominal level 095 we performed a test in order tocheck if the actual coverage rate is significantly different from095 that is to state if such CIs are conservative or liberalThe null hypothesis is that the true rate 120578 of CIs that do notcover the real value 119877 is equal to 120572 = 005119867
0 120578 = 120572 = 005
whereas the alternative hypothesis is that the rate 120578 is differentfrom120572119867
1 120578 = 120572We employ the test suggested in [28 pages
518-519] which is based on the statistic
F =119899119877 + 1
119899119878 minus 119899119877sdot1 minus 120572
120572 (10)
where 119899119877 is the number of rejections of1198670in 119899119878 (5000 in our
case unless there are some nonfeasible samples) iterations ofaMC simulation plan Under119867
0F follows an119865 distribution
with ]1= 2(119899119878 minus 119899119877) and ]
2= 2(119899119877 + 1) degrees of freedom
The test rejects 1198670at level 120574 if either F le F
1205742]1]2
or F geF1minus1205742]
1]2
where F120574]1]2
denotes the 120574 percentile point ofan 119865 distribution with ]
1and ]2degrees of freedom The test
was performed on each scenario for each kind of CI at a level120574 = 1
42 Results and Discussion The simulation results are (par-tially) reported in Table 3 (120582
119910= 2 varying 119877 and the
common size 119899119909= 119899119910) The results about the CIs (8) and
(9) are not reported here as their performances are overallunsatisfactory Even if theoretically appealing the proce-dure leading to the CI in (8) practically fails some of thescenarios of the simulation plan were considered and theCI built following this alternative procedure never providessatisfactory results For the smallest sample size (119899 = 5) thecoverage proves to be larger than that provided by the ANCI but however smaller than the nominal one for the largersample size (119899 = 10 20 50) the coverage rate dramaticallydecreases (even to values as low as 60) Paradoxically thediscrete nature of the Poisson variable and thus the qualityof the normal approximation of step (1) affect results to amore relevant extent as the sample size increases As to thebootstrap procedure yielding the CI in (9) this solution iscomputationally cumbersome it becomes even more timeconsuming if bias-corrected accelerated CIs [27 pages 184ndash188] have to be calculated and it provides barely satisfactoryresults as proven by a preliminary simulation study notreported here that confirms the findings reported in [25] forparametric bootstrap inference on the reliability parameterin the bivariate normal case The rejection of 119867
0based on
the test statistic (10) described in the previous subsection
is indicated by a ldquolowastrdquo near the actual coverage rate of eachCI Figures 2 and 3 graphically display the values of lowerand upper uncoverage rates for the five interval estimatorsvarying 119877 in 05 06 07 08 09 095 for 119899
119909= 119899119910= 5 and
119899119909= 119899119910= 50 respectively These results comprise only the
equal sample size scenarios those for unequal sample sizescenarios do not add much value to the general discussionwe are going to outline
(i) First the improvement provided by the four trans-formations to the standard naıve CI of 119877 in terms ofcoverage is evidentThis improvement is considerablefor a small sample size and extreme values of 119877(say 095) when the coverage rate of the standardnaıve CI tends to decrease dramatically even below90 Note that for the sample sizes 5 10 and 20the actual coverage rate of the standard naıve CIis always significantly different (smaller) than thenominal level Under some scenarios namely for 119877 =03ndash07 the increase in coverage rate is accompaniedalso by a reduction in the CI average length On thecontrary for the other values of 119877 there is an increasein the average width of the modified CIs which ismuch more apparent for the logit transformation
(ii) Examining closely Table 3 one can note that the CIbased on the arcsine root transformation is exceptfor one scenario uniformly worse in terms of actualcoverage than the other three CIs based on logitprobit and complementary log-log transformationsFor small sample sizes (119899
119909= 5 10) the coverage rate
actually provided is significantly different (smaller)than the nominal one It can also be claimed that thisunsatisfactory performance is due to its incapabilityof symmetrizing the distribution as can be seen byglancing at the lower and upper uncoverage rates forvalues of 119877 getting close to 1 there is an apparentundercoverage effect on the left side (just a bit smallerthan that of the ldquooriginalrdquo standard approximate CI)
(iii) Logit probit and complementary log-log trans-formed CIs have overall a satisfactory performancein terms of closeness to the nominal confidence levelLogit and probit CIs exhibit a similar behavior as boththe lower and upper uncoverage rates they provide areclose to the nominal value (25) On the contrarycomplementary log-log function (as well as arcsineroot) often produces a larger undercoverage on oneside (here left) which is partially balanced by an over-coverage on the other side (here right) This is clearevidence of the higher symmetrizing capability ofthe logit and probit transformations However takinginto account the results of the hypothesis test based onthe statistic (10) the probit and complementary log-log functions are those that overall perform best thestatistical hypothesis that their actual coverage rate isequal to the nominal one is always accepted exceptfor one case (119877 = 01 and 119899 = 5) whereas thesame hypothesis is sometimes rejected (10 times outof 40) for the logit function (which tends to producesignificantly larger coverage for small sample sizes)
6 Advances in Decision Sciences
Table3Simulationresults
(120582119909=2)
119877119899119909119899119910
AN
Logit
Prob
itArcsin
Loglog
AN
Logit
Prob
itArcsin
Loglog
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
Width
01
55
8994lowast
034
972
9717lowast
266
017
9673lowast
207
120
9435
112
453
9620lowast
363
017
02836
03835
03491
03152
04204
1010
8849lowast
040
1111
9623lowast
251
126
9557
185
259
9252lowast
106
642
9565
323
112
02153
02501
02354
02214
02622
2020
9226lowast
036
738
9520
306
174
9494
232
274
9362lowast
144
494
9512
326
162
01620
01709
01653
01600
01750
5050
9334lowast
094
572
9544
232
224
9528
198
274
9450
156
394
9548
254
198
01020
01044
01029
01015
01054
02
55
8751lowast
166
1083
9601lowast
311
088
9511
253
237
9166lowast
166
668
9519
429
052
04360
04717
04553
044
2505123
1010
9068lowast
112
820
9530
268
202
9472
222
306
9314lowast
174
512
9486
364
150
03395
03429
03358
03312
03597
2020
9320lowast
136
544
9522
266
212
9522
218
260
9432
180
388
9490
322
188
02469
02473
02444
02428
02537
5050
9396lowast
138
466
9542
248
210
9516
220
264
9478
196
326
9520
276
204
01569
01572
01563
01559
01588
03
55
8911lowast
250
839
9642lowast
224
134
9472
224
304
9222lowast
224
554
9472
448
080
05440
05284
05245
05264
05664
1010
9110lowast
192
698
9540
248
212
9490
206
304
9370lowast
192
438
9428
412
160
04149
03991
03981
040
0204166
2020
9318lowast
182
500
9534
256
210
9504
236
260
9398lowast
232
370
9496
348
156
03012
02947
02945
02956
03018
5050
9438
182
380
9546
242
212
9532
220
248
9510
196
294
9540
280
180
01927
01909
01909
01912
01929
04
55
8898lowast
386
716
9652lowast
210
138
9448
232
320
9240lowast
290
470
9504
414
082
060
6805592
05627
05733
05891
1010
9214lowast
270
516
9586lowast
208
206
9478
228
294
9372lowast
246
382
9498
382
162
04572
04305
04335
04397
044
5520
209364lowast
260
376
9558
244
198
9522
244
234
9470
248
282
9514
332
154
03319
03210
03226
03254
03274
5050
9472
218
310
9548
220
232
9528
220
252
9506
220
274
9564
252
184
02128
02098
02103
02111
02116
05
55
8932lowast
504
564
9672lowast
156
172
9504
238
258
9278lowast
348
374
9490
434
076
06285
05691
05752
05889
05864
1010
9198lowast
382
420
9554
204
242
9494
236
270
9384lowast
278
338
9488
362
150
04709
044
06044
5004526
04501
2020
9350lowast
332
318
9536
230
234
9480
254
266
9464
262
274
9476
350
174
03413
03290
03312
03345
03332
5050
9462
246
292
9538
218
244
9516
232
252
9496
234
270
9542
272
186
02191
02157
02163
02173
02169
06
55
8944lowast
636
420
9652lowast
144
204
9470
250
280
9266lowast
404
330
9478
428
094
060
9505587
05628
05741
05601
1010
9178lowast
484
338
9552
178
270
9488
232
280
9392lowast
306
302
9492
348
160
04576
04305
04337
044
0104324
2020
9340lowast
402
258
9544
202
254
9512
234
254
9446
298
256
9500
334
166
03304
03197
03212
03240
03206
5050
9506
268
226
9554
198
248
9538
214
248
9524
228
248
9558
258
184
02120
02090
02095
02103
02093
07
55
8952lowast
754
294
9606lowast
148
246
9504
250
246
9256lowast
476
268
9454
454
092
05490
05266
05242
05279
05094
1010
9170lowast
582
248
9532
164
304
9502
222
276
9378lowast
366
256
9492
344
164
04162
03991
03987
040
1403919
2020
9328lowast
468
204
9562
190
248
9518
234
248
9444
326
230
9530
326
144
02984
02920
02918
02930
02889
5050
9474
324
202
9548
188
264
9540
216
244
9524
252
224
9538
264
198
01910
01893
01893
01896
01885
08
55
8894lowast
960
146
9600lowast
124
276
9492
232
276
9254lowast
578
168
9500
418
082
044
1904657
04526
044
3604289
1010
9166lowast
702
132
9518
156
326
9496
234
270
9370lowast
418
212
9522
336
142
03411
03408
03349
03315
03242
2020
9322lowast
564
114
9522
188
290
9494
252
254
9432
364
204
9500
334
166
02424
02425
02399
02386
02356
5050
9460
394
146
9534
194
272
9520
232
248
9504
286
210
9540
270
190
01543
01544
01537
01533
01525
09
55
8832lowast
1114
054
9528
114
358
9504
258
238
9196lowast
676
128
9470
438
092
02799
03557
03295
03043
03029
1010
9080lowast
850
070
9512
144
344
9476
252
272
9340lowast
498
162
9508
330
162
02175
02410
02294
02187
02181
2020
9248lowast
692
060
9524
174
302
9512
236
252
9422lowast
428
150
9520
320
160
01550
01621
01574
01532
01531
5050
9452
462
086
9510
198
292
9530
230
240
9506
322
172
9540
262
198
00976
00996
00983
00971
00971
095
55
8686lowast
1296
018
9516
112
372
9510
262
228
9152lowast
772
076
9502
412
086
01691
02593
02289
01973
02058
1010
8990lowast
984
026
9480
148
372
9498
250
252
9262lowast
628
110
9484
346
170
01277
01611
01489
01359
01398
2020
9230lowast
750
020
9522
174
304
9512
248
240
9402lowast
476
122
9492
332
176
00931
01021
00975
00925
00941
5050
9420lowast
532
048
9530
188
282
9524
240
236
9508
346
146
9532
278
190
00579
006
0200590
00577
00582
Legend
ldquocovrdquo
=actualcoverage
rateldquoLE
rdquo=lower
errorrateldquoU
Erdquo=up
pere
rror
rate
lowastTh
erejectio
nof
H0
Advances in Decision Sciences 7
R = 05
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 08
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 09
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 095
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 06
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 07
AN
Logit
Probit
Arcsin
Log-log
Figure 2 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 5
(iv) Differences in behavior among the four ldquotransfor-medrdquo CIs tend to become more apparent for smallsample sizes and for extreme values of 119877 (close to 0 or1) as can be seen looking at the bottom right graph ofFigure 2 Clearly the same differences tend to vanishas the sample sizes increase and119877 tends to 05 (see thetop left graph in Figure 3)
(v) Contrary to what one would predict the perfor-mances in terms of coverage rate of logit probit andcomplementary log-log-based CIs do not seem to benegatively affected by sample size in fact as under-lined before an (sometimes significant) increase ofthe coverage rate of the logit CIs is noticed when thesample size is small (119899
119909= 5) whereas the other two
CIs show an almost constant trendThe CI exploitingthe arcsine root transformation is the one takingmost advantage from the increase in sample sizeObviously there is an increase in the average length ofthe CIs moving to larger sample sizes
(vi) As one could expect there is a symmetry in the behav-ior of each CI when considering two complementaryvalues of119877 (ie values of119877 summing to 1) keeping thesample size fixed the values of the coverage rate andaverage width are similar As to the uncoverage ratesthey are nearly exchangeable for AN logit probitand arscin CIs that is a lower uncoverage error fora fixed value of 119877 is similar to the upper uncoverageerror for its complementary value (or at least theirrelativemagnitude is exchangeable)This feature does
not hold for log-log CIs which always present a left-side uncoverage error larger than that of the right-side These features are weakened for 119877 equal to 09(and 1 minus 119877 then equal to 01) probably because ofthe presence of ldquononfeasiblerdquo samples as discussed inSection 41 which distorts this symmetry condition
(vii) Finallymdashthe relative results are not reported here forthe sake of brevitymdashmoving to larger values (ielarger than 2) of 120582
119909 keeping 119877 constant seems to
bring benefit to the performance of all the CIs ascan be quantitatively noted by the reduced numberof rejections of the null hypothesis of equivalencebetween actual and theoretical coverage This may beexplained by the fact that for a large parameter 120582 thePoisson distribution tends to a normal distributiontherefore larger values of 120582
119909imply a better normal
approximation to Poisson and presumptively to theMLE
In Figure 4 for illustrative purposes the MC distributionof and the transformed data according to the four transfor-mations are displayed (119877 = 08 and 119899
119909= 119899119910= 5) We can
note at a glance that logit and probit produce distributionscloser to normality than arcsine root and log-log comple-mentary functions which do not seem completely able toldquocorrectrdquo the skewness of the original distribution Howeverimplementing the Shapiro-Wilk test of normality on the fourtransformed distributions leads to very low 119875-values practi-cally equal to 0 except for the probit function whose 119875-valueis 003132 thus proving that in this case all the transformeddata distributions are still far from normality
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Advances in Decision Sciences 3
Table 1 Transformations
119891(119909) Codomain 1198911015840(119909) 119891
minus1(119909)
log[119909(1 minus 119909)] (minusinfin +infin) 1[119909(1 minus 119909)] e119909(1 + e119909)Φminus1(119909) (minusinfin +infin) 1120601(Φ
minus1(119909)) Φ(119909)
arcsin(radic119909) (0 +1205872) 1[2radic119909(1 minus 119909)] sin2 (119909)log[minus log(1 minus 119909)] (minusinfin +infin) 1[(1 minus 119909)| log(1 minus 119909)|] 1 minus eminuse
119909
00 02 04 06 08 10
minus4
minus2
0
2
4
x
Logit
Logit
Probit
Probit
Arcsin
Arcsin
Log-log
Log-log
Figure 1 Plot of transformation functions and dotted horizontallines of equation 119910 = 0 and 119909 = 05
the stress-strength model the logit transformation has beenby far the most used transformation for improving thestatistical performance of standard large-sample CIs amongothers it has been considered by [17ndash19] all these contri-butions concerning the Weibull distribution by [20] for theLindley distribution by [21] when stress and strength followa bivariate exponential distribution and by [22 23] in anonparametric context
Through the years the arcsine root transformation hasgained great favor and application among practitioners [24]perhaps more than its real merit As noted previously itshould be chosen for stabilizing binomial data but is oftenused to stabilize sample proportions as well (ie relativebinomial data) However it presents a drawback highlightedin [25] and related to the fact that its codomain is the limitedinterval (0 1205872) thus its normalizing effect may turn out tobe meagre as already pointed out in some studies where ithas been used for estimating the reliability of stress-strengthmodels [19 25 26]
The complementary log-log function which is sometimesused in binomial regression models is slightly different fromlogit and probit since it assumes negative values for 119909 lt 1 minuseminus1 (and positive values for 119909 gt 1 minus eminus1) and takes large posi-tive values only for values of 119909 very close to 1 for example ittakes values larger than 1 if and only if 119909 gt 1 minus eminuse = 0934
In the next section we will apply these transformations tothe estimation of a stress-strengthmodel with both stress andstrength following a Poisson distribution
3 Inference on the ReliabilityParameter for a Stress-Strength Model withIndependent Poisson Stress and Strength
Let119883 and119884be independent rvrsquosmodeling stress and strengthrespectively with 119883 sim Poisson (120582
119909) and 119884 sim Poisson (120582
119910)
Then the reliability 119877 = 119875(119883 lt 119884) of the stress-strengthmodel is given by (see [6 page 103])
119877 = 119875 (119883 lt 119884) =
+infin
sum
119894=0
120582119894
119909119890minus120582119909
119894[
[
1 minus
119894
sum
119895=0
120582119895
119910119890minus120582119910
119895]
]
= lim119896rarrinfin
119896
sum
119894=0
120582119894
119909119890minus120582119909
119894[
[
1 minus
119894
sum
119895=0
120582119895
119910119890minus120582119910
119895]
]
(3)
If two simple random samples of size 119899119909and 119899
119910from 119883 and
119884 respectively are available the reliability parameter can beestimated by the ML estimator obtained substituting in (3)the MLEs of the unknown parameters 120582
119909and 120582
119910
=
+infin
sum
119894=0
119909119894119890minus119909
119894[
[
1 minus
119894
sum
119895=0
119910119895119890minus119910
119895]
]
(4)
Deriving the expression of the variance of the MLE is notstraightforward however an approximate expression hasbeen easily derived through the delta method in [7] Basedupon this approximation a large-sample 1 minus 120572 CI for 119877 hasbeen built Such an interval estimator has the usual expression
(119877119871 119877119880) = ( minus 119911
1minus1205722radicV () + 119911
1minus1205722radicV ()) (5)
where V() is the sample estimate of the (asymptotic) varianceof (see [7] for details)
Although such an estimator has been proved to havea satisfactory behavior in terms of coverage for severalcombinations of sample sizes and values of the reliabilityparameter some decay of the performance is observed when119877 gets close to the extreme values 0 and 1 and when samplesizes are small In these cases the approximation to thenormal distribution is in fact very rough especially because ofthe skewness of the distribution of Thus transformationsof the values of the estimates can be considered in orderto ldquomake the data more normalrdquo and produce CIs based ontransformed data that have a coverage closer to the nominallevel
4 Advances in Decision Sciences
Table 2 CIs of the reliability parameter 119877 under variance-stabilizingnormalizing transformations
Transformation 120579 CI for 120579 (120579119871 120579119880) CI for 119877 (119877
119871 119877119880)
logit log[(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )]) (exp (120579
119871) (1 + exp (120579
119871)) exp (120579
119880) (1 + exp (120579
119880)))
Probit Φminus1() (120579 ∓ 119911
1minus1205722radicV()120601(Φminus1())) (Φ(120579
119871) Φ(120579
119880))
Arcsin arcsinradic (120579 ∓ 1199111minus1205722radicV()[2radic(1 minus )]) (sin2 (120579
119871) sin2 (120579
119880))
Loglog log[minus log(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )| log(1 minus )|]) (1 minus exp (minus exp (120579
119871)) 1 minus exp (minus exp (120579
119880)))
Recalling Table 1 and considering the logit transforma-tion of the reliability parameter 119877 120579 = log[119877(1 minus 119877)] theMLE of 120579 is 120579 = log[(1minus )] and an approximate (1minus120572) CIfor 120579 is
(120579119871 120579119880) = (120579 minus 119911
1minus1205722sdot
radicV ()
[ (1 minus )]
120579 + 1199111minus1205722
sdot
radicV ()
[ (1 minus )])
(6)
Then an approximate (1 minus 120572) CI for 119877 is
(119877119871 119877119880) = (
exp (120579119871)
[1 + exp (120579119871)]
exp (120579119880)
[1 + exp (120579119880)]) (7)
For the other transformations following the same stepsjust outlined approximate ldquotransformedrdquo CIs for 119877 can beobtained which are alternative to the standard naıve CI of(5) they are synthesized in Table 2
An alternative way to make inference on the reliabilityparameter 119877 for the stress-strength model with independentPoisson random variables can be summarized as follows (1)to transform (normalize) samples from 119883 and 119884 accordingto the a proper transformation for the Poisson distribution(2) to compute point and interval estimates of 119877 usingthe methods for a stress-strength model with independentnormal variables with known variances [6 page 112] LettingΞ = radic119883 + 14 and Υ = radic119884 + 14 then Ξ and Υ are inde-pendent and approximately distributed as normal randomvariables with variance 1205902
120585= 1205902
120592= 14 and expected value
120583120585= radic120582
119909and 120583
120592= radic120582119910 respectively Then instead of esti-
mating 119877 = 119875(119883 lt 119884) one can estimate 1198771015840 = 119875(Ξ lt Υ) a1 minus 120572 confidence interval for 1198771015840 is given by
(Φ[120578 minus1199111minus1205722
radic119872] Φ[120578 +
1199111minus1205722
radic119872]) (8)
with119872 = (1205902120585+1205902
120592)(1205902
120585119899119909+1205902
120592119899119910) and 120578 = (120592 minus 120585)120590 being
an estimator of 120578 = (120583120592minus 120583120585)120590 where 120585 and 120592 are the sample
means of Ξ and Υ respectively and 120590 = radic1205902120585+ 1205902120592= radic12
An alternative procedure tomake inference about the reli-ability parameter 119877 can be provided by parametric bootstrap
[27 pages 53ndash56] In the basic version it works as follows forthe estimation problem at hand
(1) Estimate the unknown parameters of the Poisson rvrsquos119883 and 119884 through their sample means 119909 and 119910
(2) Draw independently a bootstrap sample 119909lowast of size119899119909from a Poisson rv 119883lowast with parameter 119909 and a
bootstrap sample 119910lowast of size 119899119910from a Poisson rv 119884lowast
with parameter 119910
(3) Estimate the reliability parameter say lowast on samples119909lowast and 119910lowast using the very same expression in (4)
(4) Repeat steps 2 and 3 119861 times (119861 sufficiently large eg2000) thus obtaining the bootstrap distribution oflowast
(5) Estimate a (1 minus 120572) bootstrap percentile CI for 119877 fromlowast distribution taking the 1205722 and 1 minus 1205722 quantiles
(lowast
1205722 lowast
1minus1205722) (9)
4 Simulation Study
41 Scope and Design The simulation study aims at empir-ically comparing the performance of the interval estimatorspresented in the previous section namely the standard CI of(5) labeled ldquoANrdquo those of Table 2 (labeled ldquologitrdquo ldquoprobitrdquoldquoarcsinerdquo ldquologlogrdquo) and the CIs of (8) and (9) in terms ofcoverage rate (and also lower and upper uncoverage rates)and expected length In this Monte Carlo (MC) study thevalue of parameter 120582
119909of the Poisson distribution for stress
119883 is set equal to a ldquoreferencerdquo value 2 and the parameter120582119910of the Poisson distribution modeling strength is varied in
order to obtainmdashaccording to (3)mdashseveral different levels ofreliability119877 namely 01 02 03 04 05 06 07 08 09 and095
For each couple (120582119909 120582119910) a large number (119899119878 = 5000) of
samples of size 119899119909and of size 119899
119910are drawn from 119883 sim
Poisson (120582119909) and 119884 sim Poisson (120582
119910) independently Different
and unequal sample sizes are here considered (all the 16possible combinations between the values 119899
119909= 5 10 20 50
and 119899119910= 5 10 20 50) On each pair of samples the 95 CIs
for 119877 listed above are built and the MC coverage rate andthe length of the CIs are computed over the 119899119878MC samplesMoreover the lower and upper uncoverage or error rates arecomputed that is the proportion of CIs for which 119877 lt 119877
119871
and 119877 gt 119877119880 respectively
Advances in Decision Sciences 5
Note that for the smallest sample size that is when 119899119909= 5
(or 119899119910= 5) a practical problem arises as the sample values of
119883 (or 119884) may all be 0 in this case the variance estimate V()cannot be computed and a standard approximate CI cannotbe built We decided to discard these samples from the5000 MC samples planned for the simulation study andto compute the quantities of interest only on the ldquofeasiblerdquosamples Indeed the rate of such ldquononfeasiblerdquo samples is inany case very low under each scenario For the worst onescharacterized by 119877 = 01 and 119899
119910= 5 the theoretical rate of
ldquononfeasiblerdquo samples is about 5Since results for the coverage rates of some CIs are often
close to the nominal level 095 we performed a test in order tocheck if the actual coverage rate is significantly different from095 that is to state if such CIs are conservative or liberalThe null hypothesis is that the true rate 120578 of CIs that do notcover the real value 119877 is equal to 120572 = 005119867
0 120578 = 120572 = 005
whereas the alternative hypothesis is that the rate 120578 is differentfrom120572119867
1 120578 = 120572We employ the test suggested in [28 pages
518-519] which is based on the statistic
F =119899119877 + 1
119899119878 minus 119899119877sdot1 minus 120572
120572 (10)
where 119899119877 is the number of rejections of1198670in 119899119878 (5000 in our
case unless there are some nonfeasible samples) iterations ofaMC simulation plan Under119867
0F follows an119865 distribution
with ]1= 2(119899119878 minus 119899119877) and ]
2= 2(119899119877 + 1) degrees of freedom
The test rejects 1198670at level 120574 if either F le F
1205742]1]2
or F geF1minus1205742]
1]2
where F120574]1]2
denotes the 120574 percentile point ofan 119865 distribution with ]
1and ]2degrees of freedom The test
was performed on each scenario for each kind of CI at a level120574 = 1
42 Results and Discussion The simulation results are (par-tially) reported in Table 3 (120582
119910= 2 varying 119877 and the
common size 119899119909= 119899119910) The results about the CIs (8) and
(9) are not reported here as their performances are overallunsatisfactory Even if theoretically appealing the proce-dure leading to the CI in (8) practically fails some of thescenarios of the simulation plan were considered and theCI built following this alternative procedure never providessatisfactory results For the smallest sample size (119899 = 5) thecoverage proves to be larger than that provided by the ANCI but however smaller than the nominal one for the largersample size (119899 = 10 20 50) the coverage rate dramaticallydecreases (even to values as low as 60) Paradoxically thediscrete nature of the Poisson variable and thus the qualityof the normal approximation of step (1) affect results to amore relevant extent as the sample size increases As to thebootstrap procedure yielding the CI in (9) this solution iscomputationally cumbersome it becomes even more timeconsuming if bias-corrected accelerated CIs [27 pages 184ndash188] have to be calculated and it provides barely satisfactoryresults as proven by a preliminary simulation study notreported here that confirms the findings reported in [25] forparametric bootstrap inference on the reliability parameterin the bivariate normal case The rejection of 119867
0based on
the test statistic (10) described in the previous subsection
is indicated by a ldquolowastrdquo near the actual coverage rate of eachCI Figures 2 and 3 graphically display the values of lowerand upper uncoverage rates for the five interval estimatorsvarying 119877 in 05 06 07 08 09 095 for 119899
119909= 119899119910= 5 and
119899119909= 119899119910= 50 respectively These results comprise only the
equal sample size scenarios those for unequal sample sizescenarios do not add much value to the general discussionwe are going to outline
(i) First the improvement provided by the four trans-formations to the standard naıve CI of 119877 in terms ofcoverage is evidentThis improvement is considerablefor a small sample size and extreme values of 119877(say 095) when the coverage rate of the standardnaıve CI tends to decrease dramatically even below90 Note that for the sample sizes 5 10 and 20the actual coverage rate of the standard naıve CIis always significantly different (smaller) than thenominal level Under some scenarios namely for 119877 =03ndash07 the increase in coverage rate is accompaniedalso by a reduction in the CI average length On thecontrary for the other values of 119877 there is an increasein the average width of the modified CIs which ismuch more apparent for the logit transformation
(ii) Examining closely Table 3 one can note that the CIbased on the arcsine root transformation is exceptfor one scenario uniformly worse in terms of actualcoverage than the other three CIs based on logitprobit and complementary log-log transformationsFor small sample sizes (119899
119909= 5 10) the coverage rate
actually provided is significantly different (smaller)than the nominal one It can also be claimed that thisunsatisfactory performance is due to its incapabilityof symmetrizing the distribution as can be seen byglancing at the lower and upper uncoverage rates forvalues of 119877 getting close to 1 there is an apparentundercoverage effect on the left side (just a bit smallerthan that of the ldquooriginalrdquo standard approximate CI)
(iii) Logit probit and complementary log-log trans-formed CIs have overall a satisfactory performancein terms of closeness to the nominal confidence levelLogit and probit CIs exhibit a similar behavior as boththe lower and upper uncoverage rates they provide areclose to the nominal value (25) On the contrarycomplementary log-log function (as well as arcsineroot) often produces a larger undercoverage on oneside (here left) which is partially balanced by an over-coverage on the other side (here right) This is clearevidence of the higher symmetrizing capability ofthe logit and probit transformations However takinginto account the results of the hypothesis test based onthe statistic (10) the probit and complementary log-log functions are those that overall perform best thestatistical hypothesis that their actual coverage rate isequal to the nominal one is always accepted exceptfor one case (119877 = 01 and 119899 = 5) whereas thesame hypothesis is sometimes rejected (10 times outof 40) for the logit function (which tends to producesignificantly larger coverage for small sample sizes)
6 Advances in Decision Sciences
Table3Simulationresults
(120582119909=2)
119877119899119909119899119910
AN
Logit
Prob
itArcsin
Loglog
AN
Logit
Prob
itArcsin
Loglog
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
Width
01
55
8994lowast
034
972
9717lowast
266
017
9673lowast
207
120
9435
112
453
9620lowast
363
017
02836
03835
03491
03152
04204
1010
8849lowast
040
1111
9623lowast
251
126
9557
185
259
9252lowast
106
642
9565
323
112
02153
02501
02354
02214
02622
2020
9226lowast
036
738
9520
306
174
9494
232
274
9362lowast
144
494
9512
326
162
01620
01709
01653
01600
01750
5050
9334lowast
094
572
9544
232
224
9528
198
274
9450
156
394
9548
254
198
01020
01044
01029
01015
01054
02
55
8751lowast
166
1083
9601lowast
311
088
9511
253
237
9166lowast
166
668
9519
429
052
04360
04717
04553
044
2505123
1010
9068lowast
112
820
9530
268
202
9472
222
306
9314lowast
174
512
9486
364
150
03395
03429
03358
03312
03597
2020
9320lowast
136
544
9522
266
212
9522
218
260
9432
180
388
9490
322
188
02469
02473
02444
02428
02537
5050
9396lowast
138
466
9542
248
210
9516
220
264
9478
196
326
9520
276
204
01569
01572
01563
01559
01588
03
55
8911lowast
250
839
9642lowast
224
134
9472
224
304
9222lowast
224
554
9472
448
080
05440
05284
05245
05264
05664
1010
9110lowast
192
698
9540
248
212
9490
206
304
9370lowast
192
438
9428
412
160
04149
03991
03981
040
0204166
2020
9318lowast
182
500
9534
256
210
9504
236
260
9398lowast
232
370
9496
348
156
03012
02947
02945
02956
03018
5050
9438
182
380
9546
242
212
9532
220
248
9510
196
294
9540
280
180
01927
01909
01909
01912
01929
04
55
8898lowast
386
716
9652lowast
210
138
9448
232
320
9240lowast
290
470
9504
414
082
060
6805592
05627
05733
05891
1010
9214lowast
270
516
9586lowast
208
206
9478
228
294
9372lowast
246
382
9498
382
162
04572
04305
04335
04397
044
5520
209364lowast
260
376
9558
244
198
9522
244
234
9470
248
282
9514
332
154
03319
03210
03226
03254
03274
5050
9472
218
310
9548
220
232
9528
220
252
9506
220
274
9564
252
184
02128
02098
02103
02111
02116
05
55
8932lowast
504
564
9672lowast
156
172
9504
238
258
9278lowast
348
374
9490
434
076
06285
05691
05752
05889
05864
1010
9198lowast
382
420
9554
204
242
9494
236
270
9384lowast
278
338
9488
362
150
04709
044
06044
5004526
04501
2020
9350lowast
332
318
9536
230
234
9480
254
266
9464
262
274
9476
350
174
03413
03290
03312
03345
03332
5050
9462
246
292
9538
218
244
9516
232
252
9496
234
270
9542
272
186
02191
02157
02163
02173
02169
06
55
8944lowast
636
420
9652lowast
144
204
9470
250
280
9266lowast
404
330
9478
428
094
060
9505587
05628
05741
05601
1010
9178lowast
484
338
9552
178
270
9488
232
280
9392lowast
306
302
9492
348
160
04576
04305
04337
044
0104324
2020
9340lowast
402
258
9544
202
254
9512
234
254
9446
298
256
9500
334
166
03304
03197
03212
03240
03206
5050
9506
268
226
9554
198
248
9538
214
248
9524
228
248
9558
258
184
02120
02090
02095
02103
02093
07
55
8952lowast
754
294
9606lowast
148
246
9504
250
246
9256lowast
476
268
9454
454
092
05490
05266
05242
05279
05094
1010
9170lowast
582
248
9532
164
304
9502
222
276
9378lowast
366
256
9492
344
164
04162
03991
03987
040
1403919
2020
9328lowast
468
204
9562
190
248
9518
234
248
9444
326
230
9530
326
144
02984
02920
02918
02930
02889
5050
9474
324
202
9548
188
264
9540
216
244
9524
252
224
9538
264
198
01910
01893
01893
01896
01885
08
55
8894lowast
960
146
9600lowast
124
276
9492
232
276
9254lowast
578
168
9500
418
082
044
1904657
04526
044
3604289
1010
9166lowast
702
132
9518
156
326
9496
234
270
9370lowast
418
212
9522
336
142
03411
03408
03349
03315
03242
2020
9322lowast
564
114
9522
188
290
9494
252
254
9432
364
204
9500
334
166
02424
02425
02399
02386
02356
5050
9460
394
146
9534
194
272
9520
232
248
9504
286
210
9540
270
190
01543
01544
01537
01533
01525
09
55
8832lowast
1114
054
9528
114
358
9504
258
238
9196lowast
676
128
9470
438
092
02799
03557
03295
03043
03029
1010
9080lowast
850
070
9512
144
344
9476
252
272
9340lowast
498
162
9508
330
162
02175
02410
02294
02187
02181
2020
9248lowast
692
060
9524
174
302
9512
236
252
9422lowast
428
150
9520
320
160
01550
01621
01574
01532
01531
5050
9452
462
086
9510
198
292
9530
230
240
9506
322
172
9540
262
198
00976
00996
00983
00971
00971
095
55
8686lowast
1296
018
9516
112
372
9510
262
228
9152lowast
772
076
9502
412
086
01691
02593
02289
01973
02058
1010
8990lowast
984
026
9480
148
372
9498
250
252
9262lowast
628
110
9484
346
170
01277
01611
01489
01359
01398
2020
9230lowast
750
020
9522
174
304
9512
248
240
9402lowast
476
122
9492
332
176
00931
01021
00975
00925
00941
5050
9420lowast
532
048
9530
188
282
9524
240
236
9508
346
146
9532
278
190
00579
006
0200590
00577
00582
Legend
ldquocovrdquo
=actualcoverage
rateldquoLE
rdquo=lower
errorrateldquoU
Erdquo=up
pere
rror
rate
lowastTh
erejectio
nof
H0
Advances in Decision Sciences 7
R = 05
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 08
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 09
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 095
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 06
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 07
AN
Logit
Probit
Arcsin
Log-log
Figure 2 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 5
(iv) Differences in behavior among the four ldquotransfor-medrdquo CIs tend to become more apparent for smallsample sizes and for extreme values of 119877 (close to 0 or1) as can be seen looking at the bottom right graph ofFigure 2 Clearly the same differences tend to vanishas the sample sizes increase and119877 tends to 05 (see thetop left graph in Figure 3)
(v) Contrary to what one would predict the perfor-mances in terms of coverage rate of logit probit andcomplementary log-log-based CIs do not seem to benegatively affected by sample size in fact as under-lined before an (sometimes significant) increase ofthe coverage rate of the logit CIs is noticed when thesample size is small (119899
119909= 5) whereas the other two
CIs show an almost constant trendThe CI exploitingthe arcsine root transformation is the one takingmost advantage from the increase in sample sizeObviously there is an increase in the average length ofthe CIs moving to larger sample sizes
(vi) As one could expect there is a symmetry in the behav-ior of each CI when considering two complementaryvalues of119877 (ie values of119877 summing to 1) keeping thesample size fixed the values of the coverage rate andaverage width are similar As to the uncoverage ratesthey are nearly exchangeable for AN logit probitand arscin CIs that is a lower uncoverage error fora fixed value of 119877 is similar to the upper uncoverageerror for its complementary value (or at least theirrelativemagnitude is exchangeable)This feature does
not hold for log-log CIs which always present a left-side uncoverage error larger than that of the right-side These features are weakened for 119877 equal to 09(and 1 minus 119877 then equal to 01) probably because ofthe presence of ldquononfeasiblerdquo samples as discussed inSection 41 which distorts this symmetry condition
(vii) Finallymdashthe relative results are not reported here forthe sake of brevitymdashmoving to larger values (ielarger than 2) of 120582
119909 keeping 119877 constant seems to
bring benefit to the performance of all the CIs ascan be quantitatively noted by the reduced numberof rejections of the null hypothesis of equivalencebetween actual and theoretical coverage This may beexplained by the fact that for a large parameter 120582 thePoisson distribution tends to a normal distributiontherefore larger values of 120582
119909imply a better normal
approximation to Poisson and presumptively to theMLE
In Figure 4 for illustrative purposes the MC distributionof and the transformed data according to the four transfor-mations are displayed (119877 = 08 and 119899
119909= 119899119910= 5) We can
note at a glance that logit and probit produce distributionscloser to normality than arcsine root and log-log comple-mentary functions which do not seem completely able toldquocorrectrdquo the skewness of the original distribution Howeverimplementing the Shapiro-Wilk test of normality on the fourtransformed distributions leads to very low 119875-values practi-cally equal to 0 except for the probit function whose 119875-valueis 003132 thus proving that in this case all the transformeddata distributions are still far from normality
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
4 Advances in Decision Sciences
Table 2 CIs of the reliability parameter 119877 under variance-stabilizingnormalizing transformations
Transformation 120579 CI for 120579 (120579119871 120579119880) CI for 119877 (119877
119871 119877119880)
logit log[(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )]) (exp (120579
119871) (1 + exp (120579
119871)) exp (120579
119880) (1 + exp (120579
119880)))
Probit Φminus1() (120579 ∓ 119911
1minus1205722radicV()120601(Φminus1())) (Φ(120579
119871) Φ(120579
119880))
Arcsin arcsinradic (120579 ∓ 1199111minus1205722radicV()[2radic(1 minus )]) (sin2 (120579
119871) sin2 (120579
119880))
Loglog log[minus log(1 minus )] (120579 ∓ 1199111minus1205722radicV()[(1 minus )| log(1 minus )|]) (1 minus exp (minus exp (120579
119871)) 1 minus exp (minus exp (120579
119880)))
Recalling Table 1 and considering the logit transforma-tion of the reliability parameter 119877 120579 = log[119877(1 minus 119877)] theMLE of 120579 is 120579 = log[(1minus )] and an approximate (1minus120572) CIfor 120579 is
(120579119871 120579119880) = (120579 minus 119911
1minus1205722sdot
radicV ()
[ (1 minus )]
120579 + 1199111minus1205722
sdot
radicV ()
[ (1 minus )])
(6)
Then an approximate (1 minus 120572) CI for 119877 is
(119877119871 119877119880) = (
exp (120579119871)
[1 + exp (120579119871)]
exp (120579119880)
[1 + exp (120579119880)]) (7)
For the other transformations following the same stepsjust outlined approximate ldquotransformedrdquo CIs for 119877 can beobtained which are alternative to the standard naıve CI of(5) they are synthesized in Table 2
An alternative way to make inference on the reliabilityparameter 119877 for the stress-strength model with independentPoisson random variables can be summarized as follows (1)to transform (normalize) samples from 119883 and 119884 accordingto the a proper transformation for the Poisson distribution(2) to compute point and interval estimates of 119877 usingthe methods for a stress-strength model with independentnormal variables with known variances [6 page 112] LettingΞ = radic119883 + 14 and Υ = radic119884 + 14 then Ξ and Υ are inde-pendent and approximately distributed as normal randomvariables with variance 1205902
120585= 1205902
120592= 14 and expected value
120583120585= radic120582
119909and 120583
120592= radic120582119910 respectively Then instead of esti-
mating 119877 = 119875(119883 lt 119884) one can estimate 1198771015840 = 119875(Ξ lt Υ) a1 minus 120572 confidence interval for 1198771015840 is given by
(Φ[120578 minus1199111minus1205722
radic119872] Φ[120578 +
1199111minus1205722
radic119872]) (8)
with119872 = (1205902120585+1205902
120592)(1205902
120585119899119909+1205902
120592119899119910) and 120578 = (120592 minus 120585)120590 being
an estimator of 120578 = (120583120592minus 120583120585)120590 where 120585 and 120592 are the sample
means of Ξ and Υ respectively and 120590 = radic1205902120585+ 1205902120592= radic12
An alternative procedure tomake inference about the reli-ability parameter 119877 can be provided by parametric bootstrap
[27 pages 53ndash56] In the basic version it works as follows forthe estimation problem at hand
(1) Estimate the unknown parameters of the Poisson rvrsquos119883 and 119884 through their sample means 119909 and 119910
(2) Draw independently a bootstrap sample 119909lowast of size119899119909from a Poisson rv 119883lowast with parameter 119909 and a
bootstrap sample 119910lowast of size 119899119910from a Poisson rv 119884lowast
with parameter 119910
(3) Estimate the reliability parameter say lowast on samples119909lowast and 119910lowast using the very same expression in (4)
(4) Repeat steps 2 and 3 119861 times (119861 sufficiently large eg2000) thus obtaining the bootstrap distribution oflowast
(5) Estimate a (1 minus 120572) bootstrap percentile CI for 119877 fromlowast distribution taking the 1205722 and 1 minus 1205722 quantiles
(lowast
1205722 lowast
1minus1205722) (9)
4 Simulation Study
41 Scope and Design The simulation study aims at empir-ically comparing the performance of the interval estimatorspresented in the previous section namely the standard CI of(5) labeled ldquoANrdquo those of Table 2 (labeled ldquologitrdquo ldquoprobitrdquoldquoarcsinerdquo ldquologlogrdquo) and the CIs of (8) and (9) in terms ofcoverage rate (and also lower and upper uncoverage rates)and expected length In this Monte Carlo (MC) study thevalue of parameter 120582
119909of the Poisson distribution for stress
119883 is set equal to a ldquoreferencerdquo value 2 and the parameter120582119910of the Poisson distribution modeling strength is varied in
order to obtainmdashaccording to (3)mdashseveral different levels ofreliability119877 namely 01 02 03 04 05 06 07 08 09 and095
For each couple (120582119909 120582119910) a large number (119899119878 = 5000) of
samples of size 119899119909and of size 119899
119910are drawn from 119883 sim
Poisson (120582119909) and 119884 sim Poisson (120582
119910) independently Different
and unequal sample sizes are here considered (all the 16possible combinations between the values 119899
119909= 5 10 20 50
and 119899119910= 5 10 20 50) On each pair of samples the 95 CIs
for 119877 listed above are built and the MC coverage rate andthe length of the CIs are computed over the 119899119878MC samplesMoreover the lower and upper uncoverage or error rates arecomputed that is the proportion of CIs for which 119877 lt 119877
119871
and 119877 gt 119877119880 respectively
Advances in Decision Sciences 5
Note that for the smallest sample size that is when 119899119909= 5
(or 119899119910= 5) a practical problem arises as the sample values of
119883 (or 119884) may all be 0 in this case the variance estimate V()cannot be computed and a standard approximate CI cannotbe built We decided to discard these samples from the5000 MC samples planned for the simulation study andto compute the quantities of interest only on the ldquofeasiblerdquosamples Indeed the rate of such ldquononfeasiblerdquo samples is inany case very low under each scenario For the worst onescharacterized by 119877 = 01 and 119899
119910= 5 the theoretical rate of
ldquononfeasiblerdquo samples is about 5Since results for the coverage rates of some CIs are often
close to the nominal level 095 we performed a test in order tocheck if the actual coverage rate is significantly different from095 that is to state if such CIs are conservative or liberalThe null hypothesis is that the true rate 120578 of CIs that do notcover the real value 119877 is equal to 120572 = 005119867
0 120578 = 120572 = 005
whereas the alternative hypothesis is that the rate 120578 is differentfrom120572119867
1 120578 = 120572We employ the test suggested in [28 pages
518-519] which is based on the statistic
F =119899119877 + 1
119899119878 minus 119899119877sdot1 minus 120572
120572 (10)
where 119899119877 is the number of rejections of1198670in 119899119878 (5000 in our
case unless there are some nonfeasible samples) iterations ofaMC simulation plan Under119867
0F follows an119865 distribution
with ]1= 2(119899119878 minus 119899119877) and ]
2= 2(119899119877 + 1) degrees of freedom
The test rejects 1198670at level 120574 if either F le F
1205742]1]2
or F geF1minus1205742]
1]2
where F120574]1]2
denotes the 120574 percentile point ofan 119865 distribution with ]
1and ]2degrees of freedom The test
was performed on each scenario for each kind of CI at a level120574 = 1
42 Results and Discussion The simulation results are (par-tially) reported in Table 3 (120582
119910= 2 varying 119877 and the
common size 119899119909= 119899119910) The results about the CIs (8) and
(9) are not reported here as their performances are overallunsatisfactory Even if theoretically appealing the proce-dure leading to the CI in (8) practically fails some of thescenarios of the simulation plan were considered and theCI built following this alternative procedure never providessatisfactory results For the smallest sample size (119899 = 5) thecoverage proves to be larger than that provided by the ANCI but however smaller than the nominal one for the largersample size (119899 = 10 20 50) the coverage rate dramaticallydecreases (even to values as low as 60) Paradoxically thediscrete nature of the Poisson variable and thus the qualityof the normal approximation of step (1) affect results to amore relevant extent as the sample size increases As to thebootstrap procedure yielding the CI in (9) this solution iscomputationally cumbersome it becomes even more timeconsuming if bias-corrected accelerated CIs [27 pages 184ndash188] have to be calculated and it provides barely satisfactoryresults as proven by a preliminary simulation study notreported here that confirms the findings reported in [25] forparametric bootstrap inference on the reliability parameterin the bivariate normal case The rejection of 119867
0based on
the test statistic (10) described in the previous subsection
is indicated by a ldquolowastrdquo near the actual coverage rate of eachCI Figures 2 and 3 graphically display the values of lowerand upper uncoverage rates for the five interval estimatorsvarying 119877 in 05 06 07 08 09 095 for 119899
119909= 119899119910= 5 and
119899119909= 119899119910= 50 respectively These results comprise only the
equal sample size scenarios those for unequal sample sizescenarios do not add much value to the general discussionwe are going to outline
(i) First the improvement provided by the four trans-formations to the standard naıve CI of 119877 in terms ofcoverage is evidentThis improvement is considerablefor a small sample size and extreme values of 119877(say 095) when the coverage rate of the standardnaıve CI tends to decrease dramatically even below90 Note that for the sample sizes 5 10 and 20the actual coverage rate of the standard naıve CIis always significantly different (smaller) than thenominal level Under some scenarios namely for 119877 =03ndash07 the increase in coverage rate is accompaniedalso by a reduction in the CI average length On thecontrary for the other values of 119877 there is an increasein the average width of the modified CIs which ismuch more apparent for the logit transformation
(ii) Examining closely Table 3 one can note that the CIbased on the arcsine root transformation is exceptfor one scenario uniformly worse in terms of actualcoverage than the other three CIs based on logitprobit and complementary log-log transformationsFor small sample sizes (119899
119909= 5 10) the coverage rate
actually provided is significantly different (smaller)than the nominal one It can also be claimed that thisunsatisfactory performance is due to its incapabilityof symmetrizing the distribution as can be seen byglancing at the lower and upper uncoverage rates forvalues of 119877 getting close to 1 there is an apparentundercoverage effect on the left side (just a bit smallerthan that of the ldquooriginalrdquo standard approximate CI)
(iii) Logit probit and complementary log-log trans-formed CIs have overall a satisfactory performancein terms of closeness to the nominal confidence levelLogit and probit CIs exhibit a similar behavior as boththe lower and upper uncoverage rates they provide areclose to the nominal value (25) On the contrarycomplementary log-log function (as well as arcsineroot) often produces a larger undercoverage on oneside (here left) which is partially balanced by an over-coverage on the other side (here right) This is clearevidence of the higher symmetrizing capability ofthe logit and probit transformations However takinginto account the results of the hypothesis test based onthe statistic (10) the probit and complementary log-log functions are those that overall perform best thestatistical hypothesis that their actual coverage rate isequal to the nominal one is always accepted exceptfor one case (119877 = 01 and 119899 = 5) whereas thesame hypothesis is sometimes rejected (10 times outof 40) for the logit function (which tends to producesignificantly larger coverage for small sample sizes)
6 Advances in Decision Sciences
Table3Simulationresults
(120582119909=2)
119877119899119909119899119910
AN
Logit
Prob
itArcsin
Loglog
AN
Logit
Prob
itArcsin
Loglog
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
Width
01
55
8994lowast
034
972
9717lowast
266
017
9673lowast
207
120
9435
112
453
9620lowast
363
017
02836
03835
03491
03152
04204
1010
8849lowast
040
1111
9623lowast
251
126
9557
185
259
9252lowast
106
642
9565
323
112
02153
02501
02354
02214
02622
2020
9226lowast
036
738
9520
306
174
9494
232
274
9362lowast
144
494
9512
326
162
01620
01709
01653
01600
01750
5050
9334lowast
094
572
9544
232
224
9528
198
274
9450
156
394
9548
254
198
01020
01044
01029
01015
01054
02
55
8751lowast
166
1083
9601lowast
311
088
9511
253
237
9166lowast
166
668
9519
429
052
04360
04717
04553
044
2505123
1010
9068lowast
112
820
9530
268
202
9472
222
306
9314lowast
174
512
9486
364
150
03395
03429
03358
03312
03597
2020
9320lowast
136
544
9522
266
212
9522
218
260
9432
180
388
9490
322
188
02469
02473
02444
02428
02537
5050
9396lowast
138
466
9542
248
210
9516
220
264
9478
196
326
9520
276
204
01569
01572
01563
01559
01588
03
55
8911lowast
250
839
9642lowast
224
134
9472
224
304
9222lowast
224
554
9472
448
080
05440
05284
05245
05264
05664
1010
9110lowast
192
698
9540
248
212
9490
206
304
9370lowast
192
438
9428
412
160
04149
03991
03981
040
0204166
2020
9318lowast
182
500
9534
256
210
9504
236
260
9398lowast
232
370
9496
348
156
03012
02947
02945
02956
03018
5050
9438
182
380
9546
242
212
9532
220
248
9510
196
294
9540
280
180
01927
01909
01909
01912
01929
04
55
8898lowast
386
716
9652lowast
210
138
9448
232
320
9240lowast
290
470
9504
414
082
060
6805592
05627
05733
05891
1010
9214lowast
270
516
9586lowast
208
206
9478
228
294
9372lowast
246
382
9498
382
162
04572
04305
04335
04397
044
5520
209364lowast
260
376
9558
244
198
9522
244
234
9470
248
282
9514
332
154
03319
03210
03226
03254
03274
5050
9472
218
310
9548
220
232
9528
220
252
9506
220
274
9564
252
184
02128
02098
02103
02111
02116
05
55
8932lowast
504
564
9672lowast
156
172
9504
238
258
9278lowast
348
374
9490
434
076
06285
05691
05752
05889
05864
1010
9198lowast
382
420
9554
204
242
9494
236
270
9384lowast
278
338
9488
362
150
04709
044
06044
5004526
04501
2020
9350lowast
332
318
9536
230
234
9480
254
266
9464
262
274
9476
350
174
03413
03290
03312
03345
03332
5050
9462
246
292
9538
218
244
9516
232
252
9496
234
270
9542
272
186
02191
02157
02163
02173
02169
06
55
8944lowast
636
420
9652lowast
144
204
9470
250
280
9266lowast
404
330
9478
428
094
060
9505587
05628
05741
05601
1010
9178lowast
484
338
9552
178
270
9488
232
280
9392lowast
306
302
9492
348
160
04576
04305
04337
044
0104324
2020
9340lowast
402
258
9544
202
254
9512
234
254
9446
298
256
9500
334
166
03304
03197
03212
03240
03206
5050
9506
268
226
9554
198
248
9538
214
248
9524
228
248
9558
258
184
02120
02090
02095
02103
02093
07
55
8952lowast
754
294
9606lowast
148
246
9504
250
246
9256lowast
476
268
9454
454
092
05490
05266
05242
05279
05094
1010
9170lowast
582
248
9532
164
304
9502
222
276
9378lowast
366
256
9492
344
164
04162
03991
03987
040
1403919
2020
9328lowast
468
204
9562
190
248
9518
234
248
9444
326
230
9530
326
144
02984
02920
02918
02930
02889
5050
9474
324
202
9548
188
264
9540
216
244
9524
252
224
9538
264
198
01910
01893
01893
01896
01885
08
55
8894lowast
960
146
9600lowast
124
276
9492
232
276
9254lowast
578
168
9500
418
082
044
1904657
04526
044
3604289
1010
9166lowast
702
132
9518
156
326
9496
234
270
9370lowast
418
212
9522
336
142
03411
03408
03349
03315
03242
2020
9322lowast
564
114
9522
188
290
9494
252
254
9432
364
204
9500
334
166
02424
02425
02399
02386
02356
5050
9460
394
146
9534
194
272
9520
232
248
9504
286
210
9540
270
190
01543
01544
01537
01533
01525
09
55
8832lowast
1114
054
9528
114
358
9504
258
238
9196lowast
676
128
9470
438
092
02799
03557
03295
03043
03029
1010
9080lowast
850
070
9512
144
344
9476
252
272
9340lowast
498
162
9508
330
162
02175
02410
02294
02187
02181
2020
9248lowast
692
060
9524
174
302
9512
236
252
9422lowast
428
150
9520
320
160
01550
01621
01574
01532
01531
5050
9452
462
086
9510
198
292
9530
230
240
9506
322
172
9540
262
198
00976
00996
00983
00971
00971
095
55
8686lowast
1296
018
9516
112
372
9510
262
228
9152lowast
772
076
9502
412
086
01691
02593
02289
01973
02058
1010
8990lowast
984
026
9480
148
372
9498
250
252
9262lowast
628
110
9484
346
170
01277
01611
01489
01359
01398
2020
9230lowast
750
020
9522
174
304
9512
248
240
9402lowast
476
122
9492
332
176
00931
01021
00975
00925
00941
5050
9420lowast
532
048
9530
188
282
9524
240
236
9508
346
146
9532
278
190
00579
006
0200590
00577
00582
Legend
ldquocovrdquo
=actualcoverage
rateldquoLE
rdquo=lower
errorrateldquoU
Erdquo=up
pere
rror
rate
lowastTh
erejectio
nof
H0
Advances in Decision Sciences 7
R = 05
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 08
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 09
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 095
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 06
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 07
AN
Logit
Probit
Arcsin
Log-log
Figure 2 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 5
(iv) Differences in behavior among the four ldquotransfor-medrdquo CIs tend to become more apparent for smallsample sizes and for extreme values of 119877 (close to 0 or1) as can be seen looking at the bottom right graph ofFigure 2 Clearly the same differences tend to vanishas the sample sizes increase and119877 tends to 05 (see thetop left graph in Figure 3)
(v) Contrary to what one would predict the perfor-mances in terms of coverage rate of logit probit andcomplementary log-log-based CIs do not seem to benegatively affected by sample size in fact as under-lined before an (sometimes significant) increase ofthe coverage rate of the logit CIs is noticed when thesample size is small (119899
119909= 5) whereas the other two
CIs show an almost constant trendThe CI exploitingthe arcsine root transformation is the one takingmost advantage from the increase in sample sizeObviously there is an increase in the average length ofthe CIs moving to larger sample sizes
(vi) As one could expect there is a symmetry in the behav-ior of each CI when considering two complementaryvalues of119877 (ie values of119877 summing to 1) keeping thesample size fixed the values of the coverage rate andaverage width are similar As to the uncoverage ratesthey are nearly exchangeable for AN logit probitand arscin CIs that is a lower uncoverage error fora fixed value of 119877 is similar to the upper uncoverageerror for its complementary value (or at least theirrelativemagnitude is exchangeable)This feature does
not hold for log-log CIs which always present a left-side uncoverage error larger than that of the right-side These features are weakened for 119877 equal to 09(and 1 minus 119877 then equal to 01) probably because ofthe presence of ldquononfeasiblerdquo samples as discussed inSection 41 which distorts this symmetry condition
(vii) Finallymdashthe relative results are not reported here forthe sake of brevitymdashmoving to larger values (ielarger than 2) of 120582
119909 keeping 119877 constant seems to
bring benefit to the performance of all the CIs ascan be quantitatively noted by the reduced numberof rejections of the null hypothesis of equivalencebetween actual and theoretical coverage This may beexplained by the fact that for a large parameter 120582 thePoisson distribution tends to a normal distributiontherefore larger values of 120582
119909imply a better normal
approximation to Poisson and presumptively to theMLE
In Figure 4 for illustrative purposes the MC distributionof and the transformed data according to the four transfor-mations are displayed (119877 = 08 and 119899
119909= 119899119910= 5) We can
note at a glance that logit and probit produce distributionscloser to normality than arcsine root and log-log comple-mentary functions which do not seem completely able toldquocorrectrdquo the skewness of the original distribution Howeverimplementing the Shapiro-Wilk test of normality on the fourtransformed distributions leads to very low 119875-values practi-cally equal to 0 except for the probit function whose 119875-valueis 003132 thus proving that in this case all the transformeddata distributions are still far from normality
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Advances in Decision Sciences 5
Note that for the smallest sample size that is when 119899119909= 5
(or 119899119910= 5) a practical problem arises as the sample values of
119883 (or 119884) may all be 0 in this case the variance estimate V()cannot be computed and a standard approximate CI cannotbe built We decided to discard these samples from the5000 MC samples planned for the simulation study andto compute the quantities of interest only on the ldquofeasiblerdquosamples Indeed the rate of such ldquononfeasiblerdquo samples is inany case very low under each scenario For the worst onescharacterized by 119877 = 01 and 119899
119910= 5 the theoretical rate of
ldquononfeasiblerdquo samples is about 5Since results for the coverage rates of some CIs are often
close to the nominal level 095 we performed a test in order tocheck if the actual coverage rate is significantly different from095 that is to state if such CIs are conservative or liberalThe null hypothesis is that the true rate 120578 of CIs that do notcover the real value 119877 is equal to 120572 = 005119867
0 120578 = 120572 = 005
whereas the alternative hypothesis is that the rate 120578 is differentfrom120572119867
1 120578 = 120572We employ the test suggested in [28 pages
518-519] which is based on the statistic
F =119899119877 + 1
119899119878 minus 119899119877sdot1 minus 120572
120572 (10)
where 119899119877 is the number of rejections of1198670in 119899119878 (5000 in our
case unless there are some nonfeasible samples) iterations ofaMC simulation plan Under119867
0F follows an119865 distribution
with ]1= 2(119899119878 minus 119899119877) and ]
2= 2(119899119877 + 1) degrees of freedom
The test rejects 1198670at level 120574 if either F le F
1205742]1]2
or F geF1minus1205742]
1]2
where F120574]1]2
denotes the 120574 percentile point ofan 119865 distribution with ]
1and ]2degrees of freedom The test
was performed on each scenario for each kind of CI at a level120574 = 1
42 Results and Discussion The simulation results are (par-tially) reported in Table 3 (120582
119910= 2 varying 119877 and the
common size 119899119909= 119899119910) The results about the CIs (8) and
(9) are not reported here as their performances are overallunsatisfactory Even if theoretically appealing the proce-dure leading to the CI in (8) practically fails some of thescenarios of the simulation plan were considered and theCI built following this alternative procedure never providessatisfactory results For the smallest sample size (119899 = 5) thecoverage proves to be larger than that provided by the ANCI but however smaller than the nominal one for the largersample size (119899 = 10 20 50) the coverage rate dramaticallydecreases (even to values as low as 60) Paradoxically thediscrete nature of the Poisson variable and thus the qualityof the normal approximation of step (1) affect results to amore relevant extent as the sample size increases As to thebootstrap procedure yielding the CI in (9) this solution iscomputationally cumbersome it becomes even more timeconsuming if bias-corrected accelerated CIs [27 pages 184ndash188] have to be calculated and it provides barely satisfactoryresults as proven by a preliminary simulation study notreported here that confirms the findings reported in [25] forparametric bootstrap inference on the reliability parameterin the bivariate normal case The rejection of 119867
0based on
the test statistic (10) described in the previous subsection
is indicated by a ldquolowastrdquo near the actual coverage rate of eachCI Figures 2 and 3 graphically display the values of lowerand upper uncoverage rates for the five interval estimatorsvarying 119877 in 05 06 07 08 09 095 for 119899
119909= 119899119910= 5 and
119899119909= 119899119910= 50 respectively These results comprise only the
equal sample size scenarios those for unequal sample sizescenarios do not add much value to the general discussionwe are going to outline
(i) First the improvement provided by the four trans-formations to the standard naıve CI of 119877 in terms ofcoverage is evidentThis improvement is considerablefor a small sample size and extreme values of 119877(say 095) when the coverage rate of the standardnaıve CI tends to decrease dramatically even below90 Note that for the sample sizes 5 10 and 20the actual coverage rate of the standard naıve CIis always significantly different (smaller) than thenominal level Under some scenarios namely for 119877 =03ndash07 the increase in coverage rate is accompaniedalso by a reduction in the CI average length On thecontrary for the other values of 119877 there is an increasein the average width of the modified CIs which ismuch more apparent for the logit transformation
(ii) Examining closely Table 3 one can note that the CIbased on the arcsine root transformation is exceptfor one scenario uniformly worse in terms of actualcoverage than the other three CIs based on logitprobit and complementary log-log transformationsFor small sample sizes (119899
119909= 5 10) the coverage rate
actually provided is significantly different (smaller)than the nominal one It can also be claimed that thisunsatisfactory performance is due to its incapabilityof symmetrizing the distribution as can be seen byglancing at the lower and upper uncoverage rates forvalues of 119877 getting close to 1 there is an apparentundercoverage effect on the left side (just a bit smallerthan that of the ldquooriginalrdquo standard approximate CI)
(iii) Logit probit and complementary log-log trans-formed CIs have overall a satisfactory performancein terms of closeness to the nominal confidence levelLogit and probit CIs exhibit a similar behavior as boththe lower and upper uncoverage rates they provide areclose to the nominal value (25) On the contrarycomplementary log-log function (as well as arcsineroot) often produces a larger undercoverage on oneside (here left) which is partially balanced by an over-coverage on the other side (here right) This is clearevidence of the higher symmetrizing capability ofthe logit and probit transformations However takinginto account the results of the hypothesis test based onthe statistic (10) the probit and complementary log-log functions are those that overall perform best thestatistical hypothesis that their actual coverage rate isequal to the nominal one is always accepted exceptfor one case (119877 = 01 and 119899 = 5) whereas thesame hypothesis is sometimes rejected (10 times outof 40) for the logit function (which tends to producesignificantly larger coverage for small sample sizes)
6 Advances in Decision Sciences
Table3Simulationresults
(120582119909=2)
119877119899119909119899119910
AN
Logit
Prob
itArcsin
Loglog
AN
Logit
Prob
itArcsin
Loglog
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
Width
01
55
8994lowast
034
972
9717lowast
266
017
9673lowast
207
120
9435
112
453
9620lowast
363
017
02836
03835
03491
03152
04204
1010
8849lowast
040
1111
9623lowast
251
126
9557
185
259
9252lowast
106
642
9565
323
112
02153
02501
02354
02214
02622
2020
9226lowast
036
738
9520
306
174
9494
232
274
9362lowast
144
494
9512
326
162
01620
01709
01653
01600
01750
5050
9334lowast
094
572
9544
232
224
9528
198
274
9450
156
394
9548
254
198
01020
01044
01029
01015
01054
02
55
8751lowast
166
1083
9601lowast
311
088
9511
253
237
9166lowast
166
668
9519
429
052
04360
04717
04553
044
2505123
1010
9068lowast
112
820
9530
268
202
9472
222
306
9314lowast
174
512
9486
364
150
03395
03429
03358
03312
03597
2020
9320lowast
136
544
9522
266
212
9522
218
260
9432
180
388
9490
322
188
02469
02473
02444
02428
02537
5050
9396lowast
138
466
9542
248
210
9516
220
264
9478
196
326
9520
276
204
01569
01572
01563
01559
01588
03
55
8911lowast
250
839
9642lowast
224
134
9472
224
304
9222lowast
224
554
9472
448
080
05440
05284
05245
05264
05664
1010
9110lowast
192
698
9540
248
212
9490
206
304
9370lowast
192
438
9428
412
160
04149
03991
03981
040
0204166
2020
9318lowast
182
500
9534
256
210
9504
236
260
9398lowast
232
370
9496
348
156
03012
02947
02945
02956
03018
5050
9438
182
380
9546
242
212
9532
220
248
9510
196
294
9540
280
180
01927
01909
01909
01912
01929
04
55
8898lowast
386
716
9652lowast
210
138
9448
232
320
9240lowast
290
470
9504
414
082
060
6805592
05627
05733
05891
1010
9214lowast
270
516
9586lowast
208
206
9478
228
294
9372lowast
246
382
9498
382
162
04572
04305
04335
04397
044
5520
209364lowast
260
376
9558
244
198
9522
244
234
9470
248
282
9514
332
154
03319
03210
03226
03254
03274
5050
9472
218
310
9548
220
232
9528
220
252
9506
220
274
9564
252
184
02128
02098
02103
02111
02116
05
55
8932lowast
504
564
9672lowast
156
172
9504
238
258
9278lowast
348
374
9490
434
076
06285
05691
05752
05889
05864
1010
9198lowast
382
420
9554
204
242
9494
236
270
9384lowast
278
338
9488
362
150
04709
044
06044
5004526
04501
2020
9350lowast
332
318
9536
230
234
9480
254
266
9464
262
274
9476
350
174
03413
03290
03312
03345
03332
5050
9462
246
292
9538
218
244
9516
232
252
9496
234
270
9542
272
186
02191
02157
02163
02173
02169
06
55
8944lowast
636
420
9652lowast
144
204
9470
250
280
9266lowast
404
330
9478
428
094
060
9505587
05628
05741
05601
1010
9178lowast
484
338
9552
178
270
9488
232
280
9392lowast
306
302
9492
348
160
04576
04305
04337
044
0104324
2020
9340lowast
402
258
9544
202
254
9512
234
254
9446
298
256
9500
334
166
03304
03197
03212
03240
03206
5050
9506
268
226
9554
198
248
9538
214
248
9524
228
248
9558
258
184
02120
02090
02095
02103
02093
07
55
8952lowast
754
294
9606lowast
148
246
9504
250
246
9256lowast
476
268
9454
454
092
05490
05266
05242
05279
05094
1010
9170lowast
582
248
9532
164
304
9502
222
276
9378lowast
366
256
9492
344
164
04162
03991
03987
040
1403919
2020
9328lowast
468
204
9562
190
248
9518
234
248
9444
326
230
9530
326
144
02984
02920
02918
02930
02889
5050
9474
324
202
9548
188
264
9540
216
244
9524
252
224
9538
264
198
01910
01893
01893
01896
01885
08
55
8894lowast
960
146
9600lowast
124
276
9492
232
276
9254lowast
578
168
9500
418
082
044
1904657
04526
044
3604289
1010
9166lowast
702
132
9518
156
326
9496
234
270
9370lowast
418
212
9522
336
142
03411
03408
03349
03315
03242
2020
9322lowast
564
114
9522
188
290
9494
252
254
9432
364
204
9500
334
166
02424
02425
02399
02386
02356
5050
9460
394
146
9534
194
272
9520
232
248
9504
286
210
9540
270
190
01543
01544
01537
01533
01525
09
55
8832lowast
1114
054
9528
114
358
9504
258
238
9196lowast
676
128
9470
438
092
02799
03557
03295
03043
03029
1010
9080lowast
850
070
9512
144
344
9476
252
272
9340lowast
498
162
9508
330
162
02175
02410
02294
02187
02181
2020
9248lowast
692
060
9524
174
302
9512
236
252
9422lowast
428
150
9520
320
160
01550
01621
01574
01532
01531
5050
9452
462
086
9510
198
292
9530
230
240
9506
322
172
9540
262
198
00976
00996
00983
00971
00971
095
55
8686lowast
1296
018
9516
112
372
9510
262
228
9152lowast
772
076
9502
412
086
01691
02593
02289
01973
02058
1010
8990lowast
984
026
9480
148
372
9498
250
252
9262lowast
628
110
9484
346
170
01277
01611
01489
01359
01398
2020
9230lowast
750
020
9522
174
304
9512
248
240
9402lowast
476
122
9492
332
176
00931
01021
00975
00925
00941
5050
9420lowast
532
048
9530
188
282
9524
240
236
9508
346
146
9532
278
190
00579
006
0200590
00577
00582
Legend
ldquocovrdquo
=actualcoverage
rateldquoLE
rdquo=lower
errorrateldquoU
Erdquo=up
pere
rror
rate
lowastTh
erejectio
nof
H0
Advances in Decision Sciences 7
R = 05
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 08
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 09
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 095
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 06
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 07
AN
Logit
Probit
Arcsin
Log-log
Figure 2 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 5
(iv) Differences in behavior among the four ldquotransfor-medrdquo CIs tend to become more apparent for smallsample sizes and for extreme values of 119877 (close to 0 or1) as can be seen looking at the bottom right graph ofFigure 2 Clearly the same differences tend to vanishas the sample sizes increase and119877 tends to 05 (see thetop left graph in Figure 3)
(v) Contrary to what one would predict the perfor-mances in terms of coverage rate of logit probit andcomplementary log-log-based CIs do not seem to benegatively affected by sample size in fact as under-lined before an (sometimes significant) increase ofthe coverage rate of the logit CIs is noticed when thesample size is small (119899
119909= 5) whereas the other two
CIs show an almost constant trendThe CI exploitingthe arcsine root transformation is the one takingmost advantage from the increase in sample sizeObviously there is an increase in the average length ofthe CIs moving to larger sample sizes
(vi) As one could expect there is a symmetry in the behav-ior of each CI when considering two complementaryvalues of119877 (ie values of119877 summing to 1) keeping thesample size fixed the values of the coverage rate andaverage width are similar As to the uncoverage ratesthey are nearly exchangeable for AN logit probitand arscin CIs that is a lower uncoverage error fora fixed value of 119877 is similar to the upper uncoverageerror for its complementary value (or at least theirrelativemagnitude is exchangeable)This feature does
not hold for log-log CIs which always present a left-side uncoverage error larger than that of the right-side These features are weakened for 119877 equal to 09(and 1 minus 119877 then equal to 01) probably because ofthe presence of ldquononfeasiblerdquo samples as discussed inSection 41 which distorts this symmetry condition
(vii) Finallymdashthe relative results are not reported here forthe sake of brevitymdashmoving to larger values (ielarger than 2) of 120582
119909 keeping 119877 constant seems to
bring benefit to the performance of all the CIs ascan be quantitatively noted by the reduced numberof rejections of the null hypothesis of equivalencebetween actual and theoretical coverage This may beexplained by the fact that for a large parameter 120582 thePoisson distribution tends to a normal distributiontherefore larger values of 120582
119909imply a better normal
approximation to Poisson and presumptively to theMLE
In Figure 4 for illustrative purposes the MC distributionof and the transformed data according to the four transfor-mations are displayed (119877 = 08 and 119899
119909= 119899119910= 5) We can
note at a glance that logit and probit produce distributionscloser to normality than arcsine root and log-log comple-mentary functions which do not seem completely able toldquocorrectrdquo the skewness of the original distribution Howeverimplementing the Shapiro-Wilk test of normality on the fourtransformed distributions leads to very low 119875-values practi-cally equal to 0 except for the probit function whose 119875-valueis 003132 thus proving that in this case all the transformeddata distributions are still far from normality
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
6 Advances in Decision Sciences
Table3Simulationresults
(120582119909=2)
119877119899119909119899119910
AN
Logit
Prob
itArcsin
Loglog
AN
Logit
Prob
itArcsin
Loglog
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
cov
LEUE
Width
01
55
8994lowast
034
972
9717lowast
266
017
9673lowast
207
120
9435
112
453
9620lowast
363
017
02836
03835
03491
03152
04204
1010
8849lowast
040
1111
9623lowast
251
126
9557
185
259
9252lowast
106
642
9565
323
112
02153
02501
02354
02214
02622
2020
9226lowast
036
738
9520
306
174
9494
232
274
9362lowast
144
494
9512
326
162
01620
01709
01653
01600
01750
5050
9334lowast
094
572
9544
232
224
9528
198
274
9450
156
394
9548
254
198
01020
01044
01029
01015
01054
02
55
8751lowast
166
1083
9601lowast
311
088
9511
253
237
9166lowast
166
668
9519
429
052
04360
04717
04553
044
2505123
1010
9068lowast
112
820
9530
268
202
9472
222
306
9314lowast
174
512
9486
364
150
03395
03429
03358
03312
03597
2020
9320lowast
136
544
9522
266
212
9522
218
260
9432
180
388
9490
322
188
02469
02473
02444
02428
02537
5050
9396lowast
138
466
9542
248
210
9516
220
264
9478
196
326
9520
276
204
01569
01572
01563
01559
01588
03
55
8911lowast
250
839
9642lowast
224
134
9472
224
304
9222lowast
224
554
9472
448
080
05440
05284
05245
05264
05664
1010
9110lowast
192
698
9540
248
212
9490
206
304
9370lowast
192
438
9428
412
160
04149
03991
03981
040
0204166
2020
9318lowast
182
500
9534
256
210
9504
236
260
9398lowast
232
370
9496
348
156
03012
02947
02945
02956
03018
5050
9438
182
380
9546
242
212
9532
220
248
9510
196
294
9540
280
180
01927
01909
01909
01912
01929
04
55
8898lowast
386
716
9652lowast
210
138
9448
232
320
9240lowast
290
470
9504
414
082
060
6805592
05627
05733
05891
1010
9214lowast
270
516
9586lowast
208
206
9478
228
294
9372lowast
246
382
9498
382
162
04572
04305
04335
04397
044
5520
209364lowast
260
376
9558
244
198
9522
244
234
9470
248
282
9514
332
154
03319
03210
03226
03254
03274
5050
9472
218
310
9548
220
232
9528
220
252
9506
220
274
9564
252
184
02128
02098
02103
02111
02116
05
55
8932lowast
504
564
9672lowast
156
172
9504
238
258
9278lowast
348
374
9490
434
076
06285
05691
05752
05889
05864
1010
9198lowast
382
420
9554
204
242
9494
236
270
9384lowast
278
338
9488
362
150
04709
044
06044
5004526
04501
2020
9350lowast
332
318
9536
230
234
9480
254
266
9464
262
274
9476
350
174
03413
03290
03312
03345
03332
5050
9462
246
292
9538
218
244
9516
232
252
9496
234
270
9542
272
186
02191
02157
02163
02173
02169
06
55
8944lowast
636
420
9652lowast
144
204
9470
250
280
9266lowast
404
330
9478
428
094
060
9505587
05628
05741
05601
1010
9178lowast
484
338
9552
178
270
9488
232
280
9392lowast
306
302
9492
348
160
04576
04305
04337
044
0104324
2020
9340lowast
402
258
9544
202
254
9512
234
254
9446
298
256
9500
334
166
03304
03197
03212
03240
03206
5050
9506
268
226
9554
198
248
9538
214
248
9524
228
248
9558
258
184
02120
02090
02095
02103
02093
07
55
8952lowast
754
294
9606lowast
148
246
9504
250
246
9256lowast
476
268
9454
454
092
05490
05266
05242
05279
05094
1010
9170lowast
582
248
9532
164
304
9502
222
276
9378lowast
366
256
9492
344
164
04162
03991
03987
040
1403919
2020
9328lowast
468
204
9562
190
248
9518
234
248
9444
326
230
9530
326
144
02984
02920
02918
02930
02889
5050
9474
324
202
9548
188
264
9540
216
244
9524
252
224
9538
264
198
01910
01893
01893
01896
01885
08
55
8894lowast
960
146
9600lowast
124
276
9492
232
276
9254lowast
578
168
9500
418
082
044
1904657
04526
044
3604289
1010
9166lowast
702
132
9518
156
326
9496
234
270
9370lowast
418
212
9522
336
142
03411
03408
03349
03315
03242
2020
9322lowast
564
114
9522
188
290
9494
252
254
9432
364
204
9500
334
166
02424
02425
02399
02386
02356
5050
9460
394
146
9534
194
272
9520
232
248
9504
286
210
9540
270
190
01543
01544
01537
01533
01525
09
55
8832lowast
1114
054
9528
114
358
9504
258
238
9196lowast
676
128
9470
438
092
02799
03557
03295
03043
03029
1010
9080lowast
850
070
9512
144
344
9476
252
272
9340lowast
498
162
9508
330
162
02175
02410
02294
02187
02181
2020
9248lowast
692
060
9524
174
302
9512
236
252
9422lowast
428
150
9520
320
160
01550
01621
01574
01532
01531
5050
9452
462
086
9510
198
292
9530
230
240
9506
322
172
9540
262
198
00976
00996
00983
00971
00971
095
55
8686lowast
1296
018
9516
112
372
9510
262
228
9152lowast
772
076
9502
412
086
01691
02593
02289
01973
02058
1010
8990lowast
984
026
9480
148
372
9498
250
252
9262lowast
628
110
9484
346
170
01277
01611
01489
01359
01398
2020
9230lowast
750
020
9522
174
304
9512
248
240
9402lowast
476
122
9492
332
176
00931
01021
00975
00925
00941
5050
9420lowast
532
048
9530
188
282
9524
240
236
9508
346
146
9532
278
190
00579
006
0200590
00577
00582
Legend
ldquocovrdquo
=actualcoverage
rateldquoLE
rdquo=lower
errorrateldquoU
Erdquo=up
pere
rror
rate
lowastTh
erejectio
nof
H0
Advances in Decision Sciences 7
R = 05
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 08
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 09
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 095
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 06
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 07
AN
Logit
Probit
Arcsin
Log-log
Figure 2 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 5
(iv) Differences in behavior among the four ldquotransfor-medrdquo CIs tend to become more apparent for smallsample sizes and for extreme values of 119877 (close to 0 or1) as can be seen looking at the bottom right graph ofFigure 2 Clearly the same differences tend to vanishas the sample sizes increase and119877 tends to 05 (see thetop left graph in Figure 3)
(v) Contrary to what one would predict the perfor-mances in terms of coverage rate of logit probit andcomplementary log-log-based CIs do not seem to benegatively affected by sample size in fact as under-lined before an (sometimes significant) increase ofthe coverage rate of the logit CIs is noticed when thesample size is small (119899
119909= 5) whereas the other two
CIs show an almost constant trendThe CI exploitingthe arcsine root transformation is the one takingmost advantage from the increase in sample sizeObviously there is an increase in the average length ofthe CIs moving to larger sample sizes
(vi) As one could expect there is a symmetry in the behav-ior of each CI when considering two complementaryvalues of119877 (ie values of119877 summing to 1) keeping thesample size fixed the values of the coverage rate andaverage width are similar As to the uncoverage ratesthey are nearly exchangeable for AN logit probitand arscin CIs that is a lower uncoverage error fora fixed value of 119877 is similar to the upper uncoverageerror for its complementary value (or at least theirrelativemagnitude is exchangeable)This feature does
not hold for log-log CIs which always present a left-side uncoverage error larger than that of the right-side These features are weakened for 119877 equal to 09(and 1 minus 119877 then equal to 01) probably because ofthe presence of ldquononfeasiblerdquo samples as discussed inSection 41 which distorts this symmetry condition
(vii) Finallymdashthe relative results are not reported here forthe sake of brevitymdashmoving to larger values (ielarger than 2) of 120582
119909 keeping 119877 constant seems to
bring benefit to the performance of all the CIs ascan be quantitatively noted by the reduced numberof rejections of the null hypothesis of equivalencebetween actual and theoretical coverage This may beexplained by the fact that for a large parameter 120582 thePoisson distribution tends to a normal distributiontherefore larger values of 120582
119909imply a better normal
approximation to Poisson and presumptively to theMLE
In Figure 4 for illustrative purposes the MC distributionof and the transformed data according to the four transfor-mations are displayed (119877 = 08 and 119899
119909= 119899119910= 5) We can
note at a glance that logit and probit produce distributionscloser to normality than arcsine root and log-log comple-mentary functions which do not seem completely able toldquocorrectrdquo the skewness of the original distribution Howeverimplementing the Shapiro-Wilk test of normality on the fourtransformed distributions leads to very low 119875-values practi-cally equal to 0 except for the probit function whose 119875-valueis 003132 thus proving that in this case all the transformeddata distributions are still far from normality
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Advances in Decision Sciences 7
R = 05
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 08
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 09
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 095
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 06
AN
Logit
Probit
Arcsin
Log-log
13 11 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6
R = 07
AN
Logit
Probit
Arcsin
Log-log
Figure 2 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 5
(iv) Differences in behavior among the four ldquotransfor-medrdquo CIs tend to become more apparent for smallsample sizes and for extreme values of 119877 (close to 0 or1) as can be seen looking at the bottom right graph ofFigure 2 Clearly the same differences tend to vanishas the sample sizes increase and119877 tends to 05 (see thetop left graph in Figure 3)
(v) Contrary to what one would predict the perfor-mances in terms of coverage rate of logit probit andcomplementary log-log-based CIs do not seem to benegatively affected by sample size in fact as under-lined before an (sometimes significant) increase ofthe coverage rate of the logit CIs is noticed when thesample size is small (119899
119909= 5) whereas the other two
CIs show an almost constant trendThe CI exploitingthe arcsine root transformation is the one takingmost advantage from the increase in sample sizeObviously there is an increase in the average length ofthe CIs moving to larger sample sizes
(vi) As one could expect there is a symmetry in the behav-ior of each CI when considering two complementaryvalues of119877 (ie values of119877 summing to 1) keeping thesample size fixed the values of the coverage rate andaverage width are similar As to the uncoverage ratesthey are nearly exchangeable for AN logit probitand arscin CIs that is a lower uncoverage error fora fixed value of 119877 is similar to the upper uncoverageerror for its complementary value (or at least theirrelativemagnitude is exchangeable)This feature does
not hold for log-log CIs which always present a left-side uncoverage error larger than that of the right-side These features are weakened for 119877 equal to 09(and 1 minus 119877 then equal to 01) probably because ofthe presence of ldquononfeasiblerdquo samples as discussed inSection 41 which distorts this symmetry condition
(vii) Finallymdashthe relative results are not reported here forthe sake of brevitymdashmoving to larger values (ielarger than 2) of 120582
119909 keeping 119877 constant seems to
bring benefit to the performance of all the CIs ascan be quantitatively noted by the reduced numberof rejections of the null hypothesis of equivalencebetween actual and theoretical coverage This may beexplained by the fact that for a large parameter 120582 thePoisson distribution tends to a normal distributiontherefore larger values of 120582
119909imply a better normal
approximation to Poisson and presumptively to theMLE
In Figure 4 for illustrative purposes the MC distributionof and the transformed data according to the four transfor-mations are displayed (119877 = 08 and 119899
119909= 119899119910= 5) We can
note at a glance that logit and probit produce distributionscloser to normality than arcsine root and log-log comple-mentary functions which do not seem completely able toldquocorrectrdquo the skewness of the original distribution Howeverimplementing the Shapiro-Wilk test of normality on the fourtransformed distributions leads to very low 119875-values practi-cally equal to 0 except for the probit function whose 119875-valueis 003132 thus proving that in this case all the transformeddata distributions are still far from normality
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
8 Advances in Decision Sciences
R = 05
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 08
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 09
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 06
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
R = 07
AN
Logit
Probit
Arcsin
Log-log
6 5 4 3 2 1 0 1 2 3
AN
Logit
Probit
Arcsin
Log-logR = 095
Figure 3 Lower and upper uncoverage rates 120582119909= 2 119877 ge 05 119899
119909= 119899119910= 50
06 08 10 12 14minus05 05 10 15 20
Probit(R)minus1 0 1 2 3 4
Logit(R)02 04 06 08 10
R log-log(R)minus15 minus05 00 05 10 15
Arcsin(R)
Figure 4 Histogram of the MC empirical distribution of and 120579 according to the transformations of Table 2 (120582119909= 2 119877 = 08 119899
119909= 119899119910= 5)
Superimposed the density function of the normal distribution with mean and standard deviation equal to the mean and standard deviationof the data Note that neither the 119910-axis nor 119909-axis shares the same scale in the different plots
5 An Application
The application we illustrate here is based on the datadescribed in [29] and already used in [7] to whichwe redirectfor full details On these data we build the four 95CIs basedon the transformations of Table 2 along with the standardone The results (lower and upper bounds and length of theCIs) are reported in Table 4 Although the five CIs are notmuch different from each other nevertheless we can notethat all four transformation-based intervals have both lowerand upper bounds smaller than the corresponding boundfor the standard naıve interval (ie they are left shifted withrespect to it) the logit transformation yields an interval abit wider than the standard one whereas the other threetransformations produce a slight decrease in its length
Table 4 Results for the application data
Type 119877119871
119877119880
WidthAN 07452 09232 01780Logit 07256 09054 01799Probit 07302 09080 01777Arcsine 07365 09128 01763Log-log 07363 09113 01750
6 Conclusions
This work provided an empirical analysis of the conve-nience of data transformations for estimating the reliability
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Advances in Decision Sciences 9
parameter in stress-strength models We focused on themodel involving two independent Poisson random variablessince this is a case where the distribution of the MLE is noteasily derivable and exact CIs cannot be built thus trans-formation of the estimates is a viable tool to refine thestandard naıve interval estimator derived by the central limittheorem Four transformationswere considered (logit probitarcsine root and complementary log-log) and empiricallyassessed under a number of scenarios in terms of the coverageand average length of the CI they produced Results are infavor of logit probit and complementary log-log functionsmdashalthough to varying degreesmdashwhich ensure a coverage rateclose to the nominal coverage even with small samplesizes On the contrary the arcsine root function althoughimproving the performance of the standard CI often keepsits coverage rates under the nominal confidence level Thesefindings were to some extent predictable since logit andprobit transformations are very popular link functions ingeneralized regression models but are a bit surprising withregard to the complementary log-log function whose use islimited
Future research will ascertain if such results hold truefor other distributions for the stress-strength model and willeventually inspect alternative data transformations
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
Theauthorwould like to thank the editor and two anonymousreferees whose suggestions and comments improved thequality of the original version of the paper
References
[1] J L Fleiss B Levin andM C Paik Statistical Methods for Ratesand Proportions John Wiley amp Sons Hoboken NJ USA 3rdedition 2003
[2] A Agresti and B A Coull ldquoApproximate is better than ldquoexactrdquofor interval estimation of binomial proportionsrdquoThe AmericanStatistician vol 52 no 2 pp 119ndash126 1998
[3] L D Brown T T Cai and A Dasgupta ldquoConfidence intervalsfor a binomial proportion and asymptotic expansionsrdquo TheAnnals of Statistics vol 30 no 1 pp 160ndash201 2002
[4] A M Pires and C Amado ldquoInterval estimators for a binomialproportion comparison of twenty methodsrdquo REVSTAT Statis-tical Journal vol 6 no 2 pp 165ndash197 2008
[5] B V Gnedenko and I A Ushakov Probabilistic Reliability Engi-neering John Wiley amp Sons New York NY USA 1995
[6] S Kotz Y Lumelskii and M PenskyThe Stress-strength Modeland Its Generalizations Theory and Applications World Scien-tific New York NY USA 2003
[7] A Barbiero ldquoInference for reliability of Poisson datardquo Journal ofQuality and Reliability Engineering vol 2013 Article ID 5305308 pages 2013
[8] G Blom ldquoTransformations of the binomial negative binomialpoisson and X2 distributionsrdquo Biometrika vol 41 pp 302ndash3161954
[9] D C Montgomery Design and Analysis of Experiment JohnWiley amp Sons New York NY USA 6th edition 2005
[10] Y P Chaubey and G S Mudholkar ldquoOn the symmetrizingtransformations of random variablesrdquo Concordia Universityunpublished preprint Mathematics and Statistics 1983
[11] M H Hoyle ldquoTransformations an introduction and a bibliog-raphyrdquo International Statistical Review vol 41 no 2 pp 203ndash223 1973
[12] A Foi ldquoOptimization of variance-stabilizing transformationsrdquosubmitted
[13] W Greene Econometric Analysis Prentice Hall New York NYUSA 7th edition 2012
[14] A Dasgupta Asymptotic Theory of Statistics and ProbabilitySpringer New York NY USA 2008
[15] L Brown T Cai R Zhang L Zhao and H Zhou ldquoThe root-unroot algorithm for density estimation as implemented viawavelet block thresholdingrdquo Probability Theory and RelatedFields vol 146 no 3-4 pp 401ndash433 2010
[16] J H Aldrich and F D Nelson Linear Probability Logit andProbit Models Sage Beverly Hills Calif USA 1984
[17] A Baklizi ldquoInference on Pr (119883 lt 119884) in the two-parameterWei-bull model based on recordsrdquo ISRN Probability and Statisticsvol 2012 Article ID 263612 11 pages 2012
[18] K Krishnamoorthy and Y Lin ldquoConfidence limits for stress-strength reliability involving Weibull modelsrdquo Journal of Statis-tical Planning and Inference vol 140 no 7 pp 1754ndash1764 2010
[19] S P Mukherjee and S S Maiti ldquoStress-strength reliability inthe Weibull caserdquo in Frontiers in Reliability vol 4 pp 231ndash248World Scientific Singapore 1998
[20] D K Al-Mutairi M E Ghitany and D Kundu ldquoInferences onstress-strength reliability from Lindley distributionsrdquo Commu-nications in Statistics Theory and Methods vol 42 no 8 pp1443ndash1463 2013
[21] M M Shoukri M A Chaudhary and A Al-Halees ldquoEstimat-ing P (119884 lt 119883) when 119883 and 119884 are paired exponential variablesrdquoJournal of Statistical Computation and Simulation vol 75 no 1pp 25ndash38 2005
[22] A Baklizi and O Eidous ldquoNonparametric estimation of(119883 lt 119884) using kernel methodsrdquo Metron vol 64 no 1 pp 47ndash60 2006
[23] G Qin and L Hotilovac ldquoComparison of non-parametricconfidence intervals for the area under the ROC curve of a con-tinuous-scale diagnostic testrdquo Statistical Methods in MedicalResearch vol 17 no 2 pp 207ndash221 2008
[24] W H Ahrens ldquoUse of the arcsine and square root transforma-tions for subjectively determined percentage datardquo Weed Sci-ence vol 38 no 4-5 pp 452ndash458 1990
[25] A Barbiero ldquoInterval estimators for reliability the bivariatenormal caserdquo Journal of Applied Statistics vol 39 no 3 pp 501ndash512 2012
[26] A Barbiero ldquoConfidence intervals for reliability of stress-strength models in the normal caserdquo Communications in Statis-ticsmdashSimulation and Computation vol 40 no 6 pp 907ndash9252011
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
10 Advances in Decision Sciences
[27] B Efron and R Tibshirani An Introduction to the BootstrapChapman and Hall London UK 1993
[28] I R Cardoso de Oliveira and D F Ferreira ldquoMultivariateextension of chi-squared univariate normality testrdquo Journal ofStatistical Computation and Simulation vol 80 no 5-6 pp 513ndash526 2010
[29] P CWang ldquoComparisons on the analysis of Poisson datardquoQua-lity and Reliability Engineering International vol 15 pp 379ndash383 1999
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of