2013 DissT Reparameterized Birnbaum Saunders Distr.

Universidade Federal de Pernambuco
Centro de Ciências Exatas e da Natureza

Departamento de Estatística
Pós-graduação em Estatística
Reparameterized Birnbaum-Saunders distribution: Estimation,

Modelling and Applications
Manoel Ferreira dos Santos Neto
Tese de Doutorado
Recife
02 de dezembro de 2013
Universidade Federal de Pernambuco
Centro de Ciências Exatas e da Natureza
Departamento de Estatística
Manoel Ferreira dos Santos Neto
Reparameterized Birnbaum-Saunders distribution: Estimation,

Modelling and Applications
Trabalho apresentado ao Programa de Pós-graduação em Es-

tatística do Departamento de Estatística da Universidade Federal
de Pernambuco como requisito parcial para obtenção do grau de
Doutor em Estatística.
Orientador: Prof. Dr. Francisco José de Azevêdo Cysneiros

Co-Orientador: Prof. Dr. Víctor Eliseo Leiva Sánchez
Recife
02 de dezembro de 2013
.
ii
To my parents.
iii
iii
Acknowledgements
I would like to thank my wife (Alice), parents (Regina e Jacinto), and sisters (Carol, Camila and
Cintia), for their unceasing support throughout my life.
There are so many persons contributing to this thesis. Words cannot express my gratitude to my
advisors, Francisco José de Azevêdo Cysneiros, Víctor Leiva and Michelli Barros, for their encourage-
ment, patience, and critical reviews of this manuscript. Without guidance and help of them, I could never
accomplish this difficult task.
I am also grateful to Liana (Santa Rita High School), first person to believe in my potential. I would
also like to thank the teachers Jadson (Santa Rita High School), Marquinhos (Santa Rita High School),
Viviane (Santa Rita High School), Vera (Santa Rita High School), Giovana (Santa Rita High School),
Maurício (UFC), Welliandre (UFC), Juvêncio (UFC) and Ailton (UFC), for their support and incentive.
My UFPE classmates provided me endless support both emotionally and professionally. I want
to thank Helton, Marcelo, William, Mariana, Jadson, Pollyana, Francisca, Josimar, Jeremias, Lutem-
berg, Priscila, Luz, Rodrigo, Fábio, Abraão, Tarciana, Hemílio, Tatiene, Izabel, Juliana, Manoel, Frank,
Gustavo and so on, for their friendship and professional collaboration.
My friends from Ceará, Recife and Campina Grande.
Without financial support, I would struggle in my living. I am thankful to CNPq and CAPES. I
also thank the Department of Mathematics and Statistics at Federal University of Campina Grande for
providing support.
iv
iv
Resumo
Neste trabalho, obtemos resultados baseados em uma nova parametrização da distribuição Birnba-
um-Saunders. Primeiro, propomos estimadores para os correspondentes parâmetros usando os méto-
dos de máxima verossimilhança, de momentos e de momentos generalizados, e obtemos, suas res-
pectivas distribuições assintóticas. Por meio de um estudo de simulação de Monte Carlo, verificamos o
desempenho dos estimadores propostos. Segundo, estendemos o modelo de regressão proposto por
Leiva et al. (2014) considerando uma estrutura de regressão para o parâmetro de precisão. Discutimos
quatro testes para verificar se a precisão deve ser constante, apresentamos quatro tipos de resíduos e
conduzimos um estudo de influência local para este modelo com precisão variável. Novamente, usando
simulações de Monte Carlo, avaliamos o poder dos testes e o desempenho dos resíduos apresenta-
dos e avaliamos suas distribuições empíricas. Terceiro, e finalmente, propomos um modelo de mistura
discreta-contínua que é construído usando uma massa de probabilidade no zero e uma componente
contínua baseada na mencionada distribuição Birnbaum-Saunders reparametrizada. Obtemos esti-
madores para os parâmetros desse modelo de mistura e através do método de Monte Carlo, avaliamos
o comportamento desses estimadores considerando diferentes proporções de zeros na amostra. Para
todas essas novas propostas, usamos conjuntos de dados reais que permitiram explorar a correspon-
dente metodologia, destacando seu potencial e sua flexibilidade. Códigos computacionais no programa
estatístico R foram desenvolvidos, permitindo que essas metodologias possam ser aplicadas em diver-
sas áreas da ciência.
Palavras-chave: Distribuição Birnbaum-Saunders reparametrizada, métodos de estimação, méto-

dos de Monte Carlo, modelos com precisão variável, modelos de mistura.
v
v
Abstract
In this work, we derive methodologies based on a new parameterization of the Birnbaum-Saunders

distribution. First, we propose estimators for the corresponding parameters using the maximum like-
lihood and generalized method of moments. By means of a Monte Carlo simulation study, we detect
the performance of the proposed estimators. Second, we develop a Birnbaum-Saunders regression
model with varying precision, which generalizes the proposal of Leiva et al. (2014). We discuss four
tests, present four types of residuals and conduct a study of local influence for this varying precision
model. Again using Monte Carlo simulation, we evaluate the performance of presented residuals and
determinate their distributions empirically. Third, and finally, we propose a discrete-continuous mixture
model, which is constructed by using a probability mass at zero and a continuous component based
on the mentioned reparameterized Birnbaum-Saunders distribution. We estimate the parameters of this
mixture model by using the maximum likelihood method. Once again using Monte Carlo simulation, we
assess the behavior of these estimators considering different proportions of zeros in the sample. For the
three new proposals, we use different real data sets which allow the corresponding methodologies to be
explored and their potentially and flexibility to be highlighted. Computational codes in the R statistical
software have been developed allowing these methodologies to be implemented and available to the
interested practitioners.
Keywords: Reparameterized Birnbaum-Saunders distribution, estimation methods, Monte Carlo

methods, varying precision models, mixture models.
vi
LIST OF FIGURES
2.1 PDF plots of the RBS model for different values of µ and δ . p. 12
2.2 Plots of the log-likelihood function and its respective contours, for the RBS. p. 13
2.3 Boxplots (a) and histogram with the estimated density of the RBS (b). p. 20
2.4 Boxplots (a) and histogram with the estimated density of the RBS (b). p. 21
3.1 Quantiles for δ = 4.5-(a), δ = 7.4(b) and δ = 20.1-(c) when n = 50. p. 36
3.2 Normal probability plots for the residuals r p,i , rs,i , rd,i and rq,i in case I. p. 41
3.3 Normal probability plots for the residuals r p,i , rs,i , rd,i and rq,i in case II. p. 42
3.4 QQ plot with envelope for rd,i , plot index values versus rd,i . p. 45
3.5 Index plots of Ci (·) for θ , β and α . p. 47
4.1 Plots of the ZIRBS PDF for the indicated values of µ (a), δ (b) and p (c). p. 51
! (a), δ! (b) and p! (c)
4.2 Empirical distributions of µ p. 60
! (a), δ! (b) and p! (c)
4.3 Empirical distributions of µ p. 61
4.4 Histograms with the estimated ZIRBS PDF p. 62
vii
LIST OF TABLES
1.1 The main packages used in this thesis. p. 8

2.1 Values of Kv (δ /2) for the indicated values of v. p. 11
2.7 Household expenditures for food in the US. p. 19
2.8 Descriptive statistics for the household expenditures for food (US$) p. 20
2.9 Point estimates and CIs for the indicated parameter and method using S1. p. 20
2.10 Lifetimes of coupons p. 21
2.11 Point estimates and CIs for the indicated parameter and method using S2. p. 22
2.2 Empirical mean of the estimator of the indicated parameter p. 23
2.3 Empirical bias of the estimator of the indicated parameter p. 24
2.4 Empirical SE of the estimator of the indicated parameter p. 25
√
2.5 Empirical MSE of the estimator of the indicated parameter p. 26
2.6 CP of 95% CIs for the indicated parameter p. 27
2 2
3.1 Results of dµ /dη , dδ /dτ ,d µ /dη 2 and d δ /dτ 2 . p. 29
3.2 Rejection rates of, ζR , ζS , ζW and ζG . p. 36
3.3 Powers of the ζR , ζS , ζW and ζG tests. p. 37
3.4 Descriptive summary of the indicated residual in case I. p. 40
3.5 Descriptive summary of the indicated residual in case II. p. 40
3.6 Estimates, SE, t-value and p-value for the coefficients of the model. p. 45
4.1 Empirical RB of the indicated parameter estimator p. 55
√
4.2 Empirical MSE of the indicated parameter estimator p. 56
4.3 Empirical CP of 90% CIs based on MC simulation p. 57
4.4 Empirical CP of 95% CIs based on MC simulations p. 58
4.5 Empirical CP of 99% CIs based on MC simulations p. 59
4.6 Descriptive measures for the indicated demand data set. p. 62
4.7 Statistics to test the null hypothesis of independence p. 63
4.8 Point and interval estimates of the parameters of the ZIRBS p. 63
4.9 Costs involved in generating a purchase order (OCk ). p. 67
4.10 Annual costs involved in the storage of a product (SCk ). p. 67
viii
viii
CONTENTS
LIST OF FIGURES p. vi
LIST OF TABLES p. vii
CONTENTS p. viii
1 Preliminary p. 1
1 Resumo p. 1
2 Introduction and bibliographical review p. 1
3 Background p. 3
3.1 Original parameterization for the BS distribution p. 3
3.2 A new parameterization for the BS distribution p. 4
3.3 Regression models based on BS distributions p. 4
4 Motivation of the thesis p. 5
5 Objectives of the thesis p. 6
6 Highlights p. 6
7 Products of the thesis p. 7
8 Organization of the thesis p. 7
9 Computing platform for implementation of results p. 7
2 Estimation in the RBS distribution p. 9

1 Resumo p. 9
2 Introduction p. 9
3 Moments of the RBS distribution p. 10
4 Estimation p. 12
4.1 Maximum likelihood estimation p. 12
4.2 Moment estimation p. 13
4.3 Modified moment estimation p. 15
4.4 Generalized Method of Moments estimation p. 17
5 Simulation p. 18
ix
ix
6 Applications p. 19
6.1 Data set I (S1): Griffiths et al. (1993) p. 19
6.2 Data set II (S2): Birnbaum and Saunders (1969b) p. 21
7 Conclusion p. 22
3 Reparameterized Birnbaum-Saunders regression models with varying precision p. 28

1 Resumo p. 28
2 Introduction p. 28
3 The model with varying precision p. 29
4 Diagnostics for varying precision p. 33
4.1 Simulation p. 34
5 Diagnostic techniques p. 36
5.1 Residuals p. 36
5.1.1 Simulation p. 38
5.2 Generalized leverage and local influence p. 39
5.3 Total local influence p. 40
5.4 Calculation of the curvatures p. 41
Scheme 1: Case-weight perturbation p. 41
Scheme 2: Response perturbation p. 42
Scheme 3: Predictor perturbation (x⊤
i ) p. 43
Scheme 4: ⊤
Predictor perturbation (zi ) p. 43
Scheme 5: Predictors perturbation (x⊤ ⊤
i , zi ) p. 43
6 Application p. 44
6.1 Diagnostic analysis p. 45
7 Conclusion p. 46
4 Zero-inflated reparameterized Birnbaum-Saunders distribution p. 48

1 Resumo p. 48
2 Introduction and Motivation p. 48
3 The demand statistical model p. 50
4 Estimation of ZIRBS parameters p. 52
5 Simulation study p. 53
6 Application to inventory management p. 60
6.1 Description of the problem p. 60
6.2 Statistical analysis p. 61
6.2.1 Computational implementation p. 61
6.2.2 Data sets p. 62
6.2.3 Exploratory data analysis and model selection p. 62
6.2.4 Estimation of the demand rate p. 63
6.3 Inventory management analysis p. 63
6.3.1 Inventory management model for a stable product p. 64
6.3.2 Inventory management model for a perishable product p. 64
6.4 Financial analysis p. 65
6.4.1 Calculation of prices and incomes p. 65
x
1
6.4.2 Calculation of contribution margins p. 66

6.5 Results of the financial analysis p. 68
7 Conclusion p. 68
5 Conclusions and Future Research p. 69

1 Resumo p. 69
2 Conclusions p. 69
3 Future Research p. 70
REFERENCES p. 71
xi
CHAPTER 1
Preliminary
1 Resumo
Neste capítulo, apresentamos uma revisão bibliográfica sobre o tema da tese. Adicionalmente,
oferecemos um conhecimento teórico que proporciona ao leitor um melhor entendimento de nossa
proposta, que inclui aspectos relacionados a uma nova parametrização da distribuição Birnbaum-
Saunders (veja Santos-Neto et al., 2012). Aspectos em modelos de regressão, sob essa nova parametri-
zação, propostos por Leiva et al. (2014) são discutidos. Além disso, motivações, objetivos e produtos
desta tese são apresentados. Por fim, descrevemos a organização da tese e a plataforma computa-
cional por meio da qual seus resultados foram obtidos.
2 Introduction and bibliographical review

Birnbaum and Saunders (1969a) proposed a statistical model for fatigue life of structures under
cyclic stress; see Freudenthal and Shinozuka (1979) for an earlier work on this model, but with a differ-
ent parameterization. Based on a setting that shows that the failures originate from the development of
a dominant crack produced by cumulative stress, they derived the Birnbaum-Saunders (BS) distribution,
that we call original BS. This distribution is unimodal and positively skewed, it models data that take
values greater than zero, and it has two parameters, which correspond to the shape and scale of the
distribution. For more details about the BS distribution, see Birnbaum and Saunders (1969a). The wide
interest on the BS distribution is due to its physical theoretical arguments, its attractive properties and its
relationship with the normal distribution. Specifically, we assume that the amount of cumulative damage,
that allows the BS distribution to be generated, follows a normal distribution. Over the past and recent
decades, theoretical and practical aspects of the BS model have been largely studied. For more details
on the BS distribution, see Johnson et al. (1995, pp. 651-663). Although the BS distribution has its
genesis from material fatigue, it has been used for applications in, agriculture, air contamination, busi-
nesses, economics, engineering, environment, finance, food, forest and textile industries, human mortal-
Chapter
___________1___
-_Preliminary
_____________________________________________________________________________________________________________________2
ity, informatics, insurance, medicine, microbiology, nutrition, pharmacology, psychology, quality control,
queue theory, toxicology, water quality, and wind energy; see Balakrishnan et al. (2007, 2009a,b, 2011),
Leiva et al. (2007, 2008a,c, 2009, 2010, 2011a,b, 2012, 2014), Podlaski (2008), Barros et al. (2008),
Kotz et al. (2010), Bhatti (2010), Ahmed et al. (2010), Vilca et al. (2010, 2011), Sanhueza et al. (2011),
Villegas et al. (2011), Azevedo et al. (2012), Paula et al. (2012), Ferreira et al. (2012), Marchant et al.
(2013a,b), and Santos-Neto et al. (2013a,b,c).
Several types of estimators for the original Birnbaum-Saunders distribution have been proposed by
a number of authors. Birnbaum and Saunders (1969b) found their maximum likelihood (ML) estimators.
In Bhattacharyya and Fries (1982) is mentioned that the lack of an exponential family structure for the BS
model complicates the statistical inference of its parameters. Engelhardt et al. (1981), Achcar (1993),
Chang and Tang (1994), and Dupuis and Mills (1998) proposed other types of estimators of the original
parameters. However, in all these cases, it is not possible to find explicit expressions for these estima-
tors, so that numerical procedures must be used. Ng et al. (2003) introduced a modified moment (MM)
method for estimating the BS model parameters, which provides easy analytical expressions to compute
them. From and Li (2006) presented and summarized several estimation methods for the BS distribu-
tion. Results on improved statistical inference for this distribution are attributed to Lemonte et al. (2007)
and Cysneiros et al. (2008). Thus, different estimation aspects related to the original BS distribution
have been considered by a number of authors.
Statistical modeling based on Birnbaum-Saunders distributions has considerably attracted the at-
tention of several researchers. Rieck and Nedelman (1991) were the pioneers in this line proposing BS
log-linear regression models. Galea et al. (2004) and Xie and Wei (2007) developed several diagnostic
tools for this model, whereas Leiva et al. (2007) formulated BS log-linear regression models and their
diagnostics based on censored data. Lemonte and Cordeiro (2009) proposed BS non-linear regres-
sion models that generalize the proposal of Rieck and Nedelman (1991), whereas Lemonte and Patriota
(2011) and Vanegas et al. (2012) performed diagnostic procedures for these non-linear models. In all
of these regression models, the original response must be transformed to a logarithmic scale, which
could provoke a reduction of the power of the study and difficulties of interpretation; see Huang and Qu
(2006). In addition, although in this scale one is modeling the mean, in the natural scale one is modeling
the median. This issue could make sense when lifetime data are modeled, but if we are modeling data
in other areas, such as business, for example, it makes more sense to model the mean.
Although, as mentioned, Freudenthal and Shinozuka (1979) proposed a parameterization of the
BS distribution, we recognize as its first parameterization that introduced by Birnbaum and Saunders
(1969a) based on the physics of materials. In our bibliographical review, we note that not much at-
tention has been paid to other parameterizations of the BS distribution, except the works proposed by
Volodin and Dzhungurova (2000), Ahmed et al. (2008) and Lio et al. (2010).
Santos-Neto et al. (2012) proposed and studied different parameterizations for the BS distribution.
One of these reparameterizations indexes the BS distribution by their mean and precision parameters,
which we call reparameterized BS (RBS) distribution. By using the RBS distribution, one can model
directly the mean with no transformations, such as Ferrari and Cribari-Neto (2004) did following a similar
idea to that of generalized linear models (GLM), but based on the BS distribution, which does not belong
to the exponential family. Leiva et al. (2014) proposed a RBS regression model following the line of GLM.
In this context, the mean response is related to the linear predictor through a link function. This linear
predictor encompasses unknown regressors and parameters. Differently from the existing BS regression
Chapter
___________1___
-_Preliminary
_____________________________________________________________________________________________________________________3
models at present, the approach proposed by Leiva et al. (2014) allows the data to be modeled in their
natural scale.
Despite being common to model dispersion, there exists few works in the literature where the pre-
cision is modeled. In this context, Smithson and Verkuilen (2006) and Ferrari et al. (2011) considered
a beta regression specification in which precision is not constant, but it is a function of explanatory
variables (predictor) and unknown parameters. In these papers some examples in which the modelling
of the precision parameter proved to be important are presented. The problem of modeling variances
has been largely discussed in the statistical literature. Cook and Weisberg (1983), Atkinson (1985) and
Tsai et al. (1998) presented some graphical methods to detect heteroscedasticity in models under nor-
mal errors. Taylor and Verbyla (2004) studied the dependence on predictors of the location and scale
parameters of the t distribution. Lin et al. (2009) considered tests for heteroscedasticity in t linear regres-
sion models. Wu et al. (2012) proposed a unified penalized likelihood method which can simultaneously
select significant variables on the mean and dispersion of the lognormal distribution. In the context
of generalized linear models, Smyth (1989) defined sub-models to describe the mean and dispersion,
whereas Smyth and Verbyla (1999) proposed an extension of generalized linear models that allows
both the mean and variance to be modeled. Finally, Cysneiros et al. (2007) and Cao, et al. (2010) con-
sidered heteroscedastic linear models with symmetric errors and discussed the diagnostic aspects of
these models.
Despite the wide use of the BS distribution, this is well-defined only for positive values, making
that data sets that contain values equal to zero, for example, daily demand data of a product, cannot
be modeled using this distribution. Thus, when zero-values are present, we should adapt the BS dis-
tribution for allowing zero-values to be captured with a positive probability. A way for obtain such a
model, that is, a model suitable for describing data observed on the interval [0, ∞), is to use a mixture
distribution of two components. One of them being a BS distribution (continuous component), and the
other one being a degenerate distribution (discrete component) at zero. It is well-known that the mix-
ture models are powerful and popular tools for generating flexible distributions with good properties; see
McLachlan and Peel (2000), for mixture distributions in general, and Balakrishnan et al. (2009a, 2011)
and Kotz et al. (2010), for some recent works on mixture distributions based on the BS distribution.
Some zero inflated models can be seen in Aitchison (1955), Feuerverger (1979), Yoo (2004), Cook et al.
(2004), Stasinopoulos and Rigby (2007), Ospina and Ferrari (2010) and Lesaffre et al. (2012).
3 Background
3.1 Original parameterization for the BS distribution
The original BS distribution is indexed in terms of shape (α ) and scale (β ) parameters. Thus, if
a random variable (RV) T follows the BS distribution with parameters α > 0 and β > 0, the notation
Y ∼ BS(α , β ) is used and the corresponding density function (PDF) is given by
" # $%
1 1 y β [y + β ]
f (y; α , β ) = √ exp − 2 + −2 & , y > 0, α > 0, β > 0. (1.1)
2π 2α β y 2 α β y3
Chapter
___________1___
-_Preliminary
_____________________________________________________________________________________________________________________4
3.2 A new parameterization for the BS distribution

&
The RBS distribution is reparameterized with respect to the original parameterization by α = 2/δ
and β = δ µ /[δ + 1], such that δ = 2/α 2 and µ = β [1 + α 2 /2], where δ > 0 and µ > 0 are shape and
mean parameters, respectively (see Santos-Neto et al., 2012). Motivations and justifications for this
new parameterization are detailed in the next section. Thus, if Y ∼ RBS(µ , δ ), then its PDF of is given by
√ # $ " # $%
exp (δ /2) δ + 1 δµ δ {δ + 1}y δµ
f (y; µ , δ ) = √ y+ exp − + , y > 0. (1.2)
4 π µ y3/2 δ +1 4 δµ {δ + 1}y
From (1.1) one can note that RBS and standard normal RVs are related by
⎡ )* ⎤2
+2
δµ ⎣ Z Z
Y = √ + √ + 1⎦ and (1.3)
δ +1 2δ 2δ
. /) ) 0
δ {δ + 1}Y µδ
Z = − .
2 µδ {δ + 1}Y
Thus, from (1.3), the cumulative distribution function (CDF) and the quantile functin (QF) of Y ∼ RBS(µ , δ )
are, respectively, given by
1. / ) ) 02
δ {δ + 1} y µδ
F(y; µ , δ ) = Φ − , y > 0, and (1.4)
2 µδ {δ + 1} y
⎡ )* ⎤2
+2
δ µ z(q)
⎣√ + z(q)
y(q; µ , δ ) = F −1 (q) = √ + 1⎦ , 0 < q < 1, (1.5)
δ +1 2δ 2δ
where z(q) is the qth quantile of the standard normal distribution and F −1 is the inverse CDF of Y . The
hazard rate function of Y is defined by
3 4 56
√ # $ exp − δ {δ +1} y
+ δµ
exp (δ /2) δ + 1 δµ 4 δµ {δ +1} y
h(y; µ , δ ) = & y+ " 7 47 7 5% , y > 0.
4 πµy 3 δ + 1 δ {δ +1} y µδ
Φ − 2 − µδ {δ +1} y
3.3 Regression models based on BS distributions

Rieck and Nedelman (1991) defined that if Y ∼ BS(α , β ), then Z = log(Y ) follows a log -BS distribution
with shape parameter α and location parameter γ = log(β ) ∈ R, denoted by Z ∼ log-BS(α , γ ). In this
regression model, the original response must be transformed to a logarithmic scale. Thus, although in
this scale one is modeling the mean, say γ = log(β ), in the natural scale one is modeling β = exp(γ ),
which in the BS case corresponds to the median.
Leiva et al. (2014) proposed a new approach to the BS regression model. Formally we have
Y1 , . . . ,Yn are independent random variables, where each Yi ∼ RBS(µi , δ ), for i = 1, . . ., n, and y = [y1 , . . . , yn ]⊤
the corresponding observations. Then, we define a statistical model based on (1.2) by the systematic
component
g(µi ) = ηi = x ⊤
i β, i = 1, 2, . . . , n, (1.6)
where β = [β1 , . . . , β p ]⊤ , for p < n, is a vector of unknown parameters to be estimated, and x i ⊤ =

[1, xi2 , . . . , xip ] represents the values of p regressors, such that µi = g−1 (xx⊤ −1
i β ), with g (·) being the inverse
function of g(·). In the model given in (1.6), the link function g: R → R+ is strictly monotone, positive, and
Chapter
___________1___
-_Preliminary
_____________________________________________________________________________________________________________________5
√
at least twice differentiable; for example, g(µ ) = log(µ ) or g(µ ) = µ.
Formally, we have that Var[Yi ] is a function of µi and, consequently, of the regressors x i . Then,
because we are modeling the mean based on a particular structure, we are also modeling the variance.
Therefore, situations in which the variance is not constant can be analyzed by using this model.
The log-likelihood function of the model given in (1.6) for θ = [β ⊤ , δ ]⊤ is ℓ(θ ; y) = ∑ni=1 ℓi (µi , δ ; yi ),
where " %
δ log(16π ) 1 {δ + 1} µi y3i {δ + 1} yi µi δ 2
ℓ i ( µi , δ ; yi ) = − − log − − . (1.7)
2 2 2 { δ yi + yi + δ µi } 2 4 µi 4{δ + 1} yi
The score functions for β j , with j = 1, . . . , p, and δ are, respectively, given by
n # $
∂ ℓ(θ ) 1 δ {δ + 1}yi δ2 1
Uβ j (θ ) = =∑ − + + 2
− ′
xi j
∂βj i=1 2 µ i { δ yi + yi + δ µ i } 4 µi 4{δ + 1}yi g (µi )
8 9: ;
ai
n
= ∑ [y∗i − µi∗] ai xi j
i=1
and
n # $ n
∂ ℓ(θ ) δ {yi + µi } yi δ {δ + 2}µi
Uδ (θ ) = =∑ + − − 2
= ∑ [y⋆i − µi⋆],
∂δ i=1 2{δ + 1} {δ yi + yi + δ µi } 4µi 4{δ + 1} yi i=1
where g′ (·) is the derivative of g(·). In matrix form, we have
Uβ (θ ) = X ⊤ A [yy∗ − µ ∗ ] and Uδ (θ ) = 1 ⊤ [yy⋆ − µ ⋆ ], (1.8)
δ [δ +1]yi δ2 1 [yi + µi ] yi δ [δ +2] µi

where y∗i = [δ yi +yi +δ µi ] + 4 µi 2
− 4[δ +1]yi , µi∗ = 2 µi , y⋆i = [δ yi +yi +δ µi ] − 4 µi − 4[δ +1]2 yi
, µi⋆ = − 2[δδ+1] , 1 =
[1, . . . , 1]⊤ is a vector of ones with n elements, X = (xx1 , . . . , x n ), with x i given in (1.6), for i = 1, . . . , n,
[yy∗ − µ ∗ ] = [{y∗1 − µ1∗ }, . . . , {y∗n − µn∗ }]⊤ , [yy⋆ − µ ⋆ ] = [{y⋆1 − µ1⋆ }, . . . , {y⋆n − µn⋆ }]⊤ and A = ai δinj = diag(aa) is n × n
matrice, with a = [a1 , . . . , an ]⊤ with δinj being the Kronecker Delta that is a function of the 3 arguments i, j
and n. If i and j are equals then the function takes the value 1, otherwise is equal to zero. The value n
is the order of the square matrix. Thus, the score vector is U(θ ) = [Uβ⊤ (θ ),Uδ (θ )]⊤ .
4 Motivation of the thesis

This work has the following motivations:
• For the original parameterization of the BS distribution is not possible to obtain estimators by
the moment method, because the solution of the equations of moments does not present unique
solution. However, with the RBS distribution is possible to obtain such estimators.
• In regression models based on the original BS distribution, it is necessary to transform the data to
a logarithmic scale. Using the RBS distribution, regression models are based on the natural scale
of the data, facilitating the interpretation of results because the modeling is made in terms of the
mean.
• Most of the BS regression models describe the behavior of the median (original BS) or mean
(RBS), but the precision is considered as constant. Consequently, few has been studied about the
precision of the model. Absence of modeling for the precision of the data can lead to misinterpre-
tations in the study. Thus, a BS regression model with varying precision parameter is needed.
• Data sets containing values equal to zero are frequent in diverse fields of application in statistics.
Chapter
___________1___
-_Preliminary
_____________________________________________________________________________________________________________________6
However, as mentioned, the BS distribution only allows data that take values greater than zero to
be modeled. Thus, a discrete-continuous mixture model that considers a probability mass at zero
and a continuous component based on the RBS distribution is needed.
5 Objectives of the thesis

This work has the following objectives:
• To propose estimators for the parameters of RBS distribution and to study their finite sample per-
formance.
• To formulate RBS regression models varying precision, including estimation, residuals and diag-
nostics, and study their the statistical behavior of the residuals.
• To develop a zero inflated RBS model, including estimation, and study their the statistical perfor-
mance of the corresponding estimators.
6 Highlights
This thesis has the following highlights:
• The BS distribution is an asymmetric probabilistic model that is receiving considerable attention

due to its good properties.
• Santos-Neto et al. (2012) proposed different parameterizations for the BS distribution. One of
these reparameterizations indexes the BS distribution by their mean and precision parameters,
which we have denoted by RBS distribution.
• Some mathematical properties and estimates of the parameters that index the RBS distribution,
using the ML, moments, MM methods and generalized method of moments (GMM), are derived,
as well as their asymptotic distributions. A simulation study evaluates the performance of these
estimators under different scenarios.
• A RBS regression model with varying precision is proposed. Tests, residuals and local influence
for this model are derived. A simulation study assesses the behavior of the tests and residuals
derived for this model.
• A methodology for optimized logistics and inventories based on a zero-inflated RBS (ZIRBS) model
is recommended. Some mathematical properties and estimates of the parameters of the ZIRBS
model are studied. A simulation study is conducted for evaluating the empirical behavior of these
estimators. The methodology includes also some financial aspects.
• All new proposals in this thesis have been implemented in to the R statistical software and applied
to real data sets, showing their convenience, flexibility and potentiality. In particular, the methodol-
ogy for optimized logistics and inventories allowed the contribution margins of a Chilean company
to be significantly improved.
Chapter
___________1___
-_Preliminary
_____________________________________________________________________________________________________________________7
7 Products of the thesis

This thesis allowed the following products to be obtained:
1. Santos-Neto, M., Cysneiros, F.J.A., Leiva, V., Barros, M. (2013) A reparameterized Birnbaum-
Saunders distribution and its moments, estimation and applications. Revstat - Statistical Journal
(in press).
2. Santos-Neto, M., Cysneiros, F.J.A., Leiva, V., Barros, M. (2013) A zero-inflated Birnbaum-Saunders
distribution useful for modeling demand applied to inventory managements (submitted).
3. Santos-Neto, M., Cysneiros, F.J.A., Leiva, V., Barros, M. (2013) Birnbaum-Saunders regression
model varying precision (submitted).
4. Leiva, V. Santos-Neto, M., Cysneiros, F.J.A., Barros, M. (2013) On a new method for Birnbaum-
Saunders modelling. 7th Workshop on Statistics, Mathematics and Computation. Tomar, Portugal
(oral presentation).
5. Santos-Neto, M., Cysneiros, F.J.A., Leiva, V., Barros, M. (2013) Distribuição Birn-baum-Saunders:
um novo enfoque. 58a RBRAs. Campina Grande, Brazil (oral presentation).
8 Organization of the thesis

This work contains four chapters taking into account this introduction chapter. In Chapter 2, we
conduct a study of point estimation based on the method of moments for RBS distribution and present
the asymptotic distributions of the estimators. In Chapter 3, we generalize the model proposed by
Leiva et al. (2014) considering a regression structure for the corresponding precision parameter. We
present also hypothesis tests, residuals and new diagnostic tools for this new model. In Chapter 4, we
propose a zero inflated RBS model and estimate its parameters. We conduct simulation studies to verify
the behavior of the estimators for different proportions of zeros in the sample. Finally, we present a
general conclusion this work.
9 Computing platform for implementation of results

In this thesis, we implement the obtained results in the R statistical software, whereas the math-
ematical formulas and text are written in the LATEX software. For simulation and graphical represen-
tation of results, we use the R software. The latest version of this software is freely available at
http://www.R-project.org. The main packages used in this thesis are present in Table 1.1. The
LATEX software was the chosen typesetting environment, and BibTeX as the references managing tool.
Chapter
___________1___
-_Preliminary
_____________________________________________________________________________________________________________________8
Table 1.1: The main packages used in this thesis.
Package Authors Used Functions

gbs Barros, M, Leiva, V. and Paula, G.A. descriptiveSummary()
gmm Pierre Chausse gmm(), SpecTest()
robustbase Peter Rousseeuw and Christophe Croux adjbox()
moments Lukasz Komsta and Frederick Novomestky kurtosis(), skewness()
nortest Juergen Gross ad.test(), cvm.test()
gamlss Mikis Stasinopoulos and Bob Rigby gamlss(), vcov()
base R Core Team summary(), eigen(), solve(), crossprod(),
diag(), matrix(), t(), c(), rnorm(),
qnorm(), apply(), hist(), plot(), qqplot()
runif(), qchisq(), segments(),
abline(), text(), legend()
pracma Hans W Borchers hessian()
9
CHAPTER 2
Estimation in the RBS distribution
1 Resumo
Neste capítulo, apresentamos alguns novos resultados da distribuição RBS, incluindo a sua função
característica e seus momentos. Estimação e inferência para a distribuição RBS, baseada nos métodos
de máxima verossimilhança, momentos, momentos modificados e momentos generalizados, são discu-
tidas. Também, discutimos o desempenho dos estimadores propostos, avaliados através de simulações
de Monte Carlo. Além disso, propomos um gerador de números pseudo-aleatórios e discutimos aspec-
tos computacionais com base em pacotes do software R. Por fim, concluímos este capítulo ilustrando
os resultados por meio da análise de dois conjuntos de dados reais.
2 Introduction
One of the main motivations to study the RBS distribution is its easier way to obtain statistical
properties which are not possible through the classical parametrization. Such a reparameterization is
useful because, first, moment estimates for the original parameterization of the BS distribution do not
have a closed-form, but this is possible with Santos-Neto et al. (2012)’s reparameterization and, second,
it allows a response variable to be modeled in its original scale (see Leiva et al., 2014), which is not
possible with the parameterizations proposed until now.
The objectives of this chapter are (i) to provide some results on moments of a reparameterized
version of the BS distribution and a generator of random numbers; (ii) to propose estimators for this
reparameterization; (iii) to study the performance of these estimators; and (iv) to apply the results to real
data. The studied estimators are based on ML, moment, MM methods and (GMM).
This chapter is organized as follows. In Section 3, we present some new results of the RBS distri-
bution. In Section 4, we develop estimation and inference for this parameterization based on the ML,
moment, MM methods and GMM. In Section 5, we evaluate the performance of the proposed estimators
through Monte Carlo (MC) simulations. In this section, a generator of random numbers and computa-
tional aspects based on packages of the R software are discussed. In Section 6, we illustrate the results
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
10
obtained in this work by means of the analysis for two real data sets. In the final section, we provide the
conclusions of this chapter.
3 Moments of the RBS distribution

Here, we provide some results on the characteristic function (CF) and moments of the RBS dis-
tribution. Results on moments for the original parameterization of the BS distribution can be seen in
Leiva et al. (2008a) and Balakrishnan et al. (2009a).
Another way to characterize a distribution is by using its CF. In the literature on the BS distribution,
practically, the CF is not studied. From the PDF given in (1.2), we obtain the CF of Y ∼ RBS(µ , δ ) in the
following theorem.
Theorem 1. Let Y ∼ RBS(µ , δ ). Then, the CF ϕ : R → C of Y is

</ √ 0
1 δ +1
ϕ (t) = E[exp(it Y )] = 1+ &
2 1 + δ − 4 µ it
1 √ & 2=
δ { δ + 1 − 1 + δ − 4 µ it}
× exp √ , t ∈ R,
2 δ +1
√
where i = −1 is the imaginary unit.
Proof. The result can be established by using algebraic and integration methods.
Corollary 1. Let Y ∼ RBS(µ , δ ) with CF ϕ as given in Theorem 1. Then, the rth derivative of ϕ with
respect to t, evaluated at the point t = 0, is
dr ϕY (t) >> >

> 1 ? 3 6
ϕ (0)(r) = > = ir
E[Y r
exp(it Y )] > = √ 3 ir r 2
µ δ exp δ
2
dt r t=0 t=0 2 π [δ + 1] 2
4 1 3@ 1
3 6 3 3
3 65A
× {δ r− 2 + δ r− 2 {δ + 1} 2 −r Kr+ 1 δ2 + δ r− 2 {δ + 1} 2 −r Kr− 1 δ2 ,
2 2
where Kv is the modified Bessel function of second type.
Table 2.1 displays the values of the function Kv (see Abramowitz and Stegun, 1972) for some values
of v, which are useful for calculating the non-central moments of the BS distribution.
By means of Theorem 1 and Corollary 1, it is possible to obtain the non-central moments of Y ∼
RBS(µ , δ ), in special, the four fist moments are:
[δ 2 + 4 δ + 6]
E[Y ] = µ, E[Y 2 ] = µ 2 ,
[δ + 1]2
[δ 3 + 9 δ 2 + 36 δ + 60]
E[Y 3 ] = µ3 , and
[δ + 1]3
[δ 4 + 16 δ 3 + 120 δ 2 + 460 δ + 840]
E[Y 4 ] = µ4 . (2.1)
[δ + 1]4
The rth central moment of Y ∼ RBS(µ , δ ), which we denote by µr , can be obtained using the following
formula: " %
r
r
µr = E[Y − µ ]r = ∑ j (−1)r− j E[Y j ] µ r− j , r = 2, 3, . . . . (2.2)
j=0
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
11
Table 2.1: Values of Kv (δ /2) for the indicated values of v.
v Value of the function

√
1 π exp(− 12 δ )
√
2 δ
3 3 6# 2
$
δ
K1 2 1 +
2 2 δ
3 6 # $
5 δ 6 12
K1 2 1+ + 2
2 2 δ δ
3 6 # $
7 δ 12 60 120
K1 2 1+ + 2 + 3
2 2 δ δ δ
3 6 # $
9 20 180 840 1680
K 1 δ2 1+ + 2 + 3 + 4
2 2 δ δ δ δ
From (2.1) and (2.2), we have that the variance of Y is Var[Y ] = µ 2 [2 δ + 5]/[δ + 1]2 , which allows the
parameter δ to be interpreted as a precision parameter because, for µ fixed, the variance of Y decreases
when δ increases. In addition, we can rewrite this variance as Var[Y ] = V (µ )/φ , where V (µ ) = µ 2 and
φ = [δ + 1]2 /[2δ + 5], with V (µ ) acting as a “variance function”, as in a GLM.
Another interesting result is that the RBS distribution preserves the reciprocation property of the
original BS distribution, that is, 1/Y is in the same family of distributions of Y . Thus, if Y ∼ RBS(µ , δ ),
then 1/Y ∼ RBS([δ + 1]2 / µ δ 2 , δ ) and, consequently,
[δ + 1]2 [2δ + 5][δ + 1]2

E[1/Y ] = and Var[1/Y ] = .
µ δ2 µ2 δ 4
Random numbers from the RBS distribution can be obtained by using a generator described in the
following algorithm:
Algorithm 1 Generator of BS random numbers

1: Generate a random number z from a RV Z ∼ N (0, 1), by using an appropriate method;
2: Set values for µ and δ of Y ∼ RBS(µ , δ ); and
3: Compute a random number y from Y ∼ RBS(µ , δ ), by using the expression given in (1.3).
Figures 2.1(a)-2.1(b) show shapes for the PDF of Y ∼ RBS(µ , δ ) considering different values of µ ,
when δ is fixed, and different values of δ , when µ is fixed. From Figure 2.1(a), note that the µ controls
the scale of the PDF, so that it is a scale parameter and also the mean of the distribution. This aspect
can be formally verified because bY ∼ RBS(b µ , δ ), with b > 0. From Figure 2.1(b), observe that the
parameter δ controls the shape of the PDF, making it more platykurtic as δ increases. Figure 2.1(c)
shows a plot of δ versus Var[Y ], for µ = 1.0. This figure allows the effect exerted by δ on the variance
of the distribution to be detected. Notice that this variance decreases as δ increases, and it converges
to 5.0, when δ goes to zero. Then, by means of this graphical analysis, we note that δ is a precision
parameter.
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
12
5
2.5
2.5
µ = 1.0 δ = 2.0
µ = 1.5
4
δ = 5.0
µ = 2.0 δ = 10.0
2.0
2.0
µ = 2.5 δ = 25.0
µ = 3.0 δ = 50.0
3
µ = 3.5 δ = 100.0
f (y; µ , δ )
f (y; µ , δ )
1.5
1.5
Var[Y ]
2
1.0
1.0
0.5
0.5
1
0.0
0.0
1 2 3 4 5 0.0 0.5 1.0 1.5 2.0 0 2 4 6 8 10
y y δ
(a) (b) (c)
Figure 2.1: PDF plots of the RBS model for different values of µ with δ = 100.0 (a), for different values of δ with
µ = 1.0 (b) and of δ versus Var[Y ] (c).
4 Estimation
In this section, we derive estimation and inference for the parameters, in the sequel denoted by
θ = [µ , δ ]⊤ , of the RBS distribution based on the ML, moment, MM methods and GMM.
4.1 Maximum likelihood estimation
Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample of size n from Y ∼ RBS(µ , δ ), with PDF as given in (1.2).
Then, the log-likelihood function for θ is given by
n
ℓ(θ ) = ∑ ℓi (θ ), (2.3)
i=1
where ℓi (θ ) is the logarithm of the PDF given in (1.2) replacing y by yi . Figure 4.2 displays plots of
the log-likelihood function and its respective contours, considering, as illustration, a sample from Y ∼
RBS(µ = 1.5, δ = 10). In this figure, note that the shape of the log-likelihood function is well behaved
and, through its contours, it is easy to see the region where the values that maximize the function ℓ(θ )
given in (2.3) are located.
In the case of the RBS distribution, the score vector for θ is given by U(θ ) = [Uµ (θ ),Uδ (θ )]⊤ , where
n # $
∂ ℓ(θ ) δ {δ + 1}yi δ2 1
Uµ (θ ) = =∑ + − − and
∂µ i=1 {δ yi + yi + δ µ } 4µ 2 4{δ + 1}yi 2µ
n # $
∂ ℓ(θ ) {yi + µ } yi δ {δ + 2}µ δ
Uδ (θ ) = =∑ − − + .
∂δ i=1 {δ yi + yi + δ µ } 4µ 4{δ + 1}2yi 2{δ + 1}
Such as in the case of the original BS parameterization, for the reparameterized version, it is not possible
to find closed-form estimators for its parameters. Then, we must use an iterative numerical method to
optimize the function ℓ(θ ) given in (2.3). For example, a Newton-Raphson type algorithm can be used in
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
13
25
-100
20
-120
ℓ(θ ) 15 -140
-160
10
-180
µ 5
δ
-200
1.0 1.2 1.4 1.6 1.8 2.0
(a) (b)
Figure 2.2: Plots of the log-likelihood function (a) and its respective contours (b), for the RBS(µ = 1.5, δ = 10)
distribution.
this case. The corresponding expected Fisher information matrix, denoted by i (θ ) = [iθ j θk ], has elements
# $ # $
∂ 2 ℓ(θ ) δ δ2
iµ µ = −E =n + I(θ ) ,
∂ µ2 2µ 2 {δ + 1}2
# 2 $ # $
∂ ℓ(θ ) 1 δµ
iδ µ = −E =n + I(θ ) and (2.4)
∂ µ∂ δ 2µ {δ + 1} {δ + 1}3
# 2 $ # 2 $
∂ ℓ(θ ) {δ + 3δ + 1} µ2
iδ δ = −E = n + I(θ ) ,
∂δ2 2δ 2 {δ + 1}2 {δ + 1}4
where iδ µ = iµδ and

B ∞# $−2
δµ
I(θ ) = y+ f (y; θ ) dy.
0 {δ + 1}
Under regularity conditions (see Cox and Hinkley, 1974), we have that the corresponding asymptotic
! , δ!] = i (θ )−1 , where i (θ ) is the expected Fisher information matrix
variance-covariance matrix is Cov[µ
with elements given by the expressions formulated in (2.4). In addition, in general, as is well-known, ML
! , δ!]⊤ is asymptoti-
estimators have an asymptotic bivariate normal joint distribution. Thus, in our case, [µ
cally distributed as 1/ 0 2
µ −1
N2 , i (θ ) .
δ
4.2 Moment estimation
Consider the following definition.
Definition 1 (Moment conditions (Mátyás, 1999)). Let Y = [Y1 , . . ., Yn ]⊤ be a random sample of size n
from any distribution. We want to estimate a p × 1 unknown parameter vector θ , with true value θ 0 . Let
f (Y j , θ ) be a continuous q × 1 vector function of θ , and assume that E[ f (Y j , θ )] exists and is finite for all
j and θ . Then, the moment conditions are that E[ f (Y j , θ 0 )] = 0 .
We want to estimate the parameter vector θ by using the moment conditions given in Definition 1.
First, we consider the case when p = q, that is, when θ is exactly identified by the moment conditions.
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
14
Thus, these conditions represent a set of p equations, with p unknown parameters. Solving these
equations, we find the value of θ that satisfies the mentioned moment conditions, that should be the
true value θ 0 . However, it is not possible to observe E[ f (Y j , θ )], but only f (y j , θ ). In this way, a natural
procedure is to define the sample moments of f (Y j , θ ), given by
n
1
fn (θ ) =
n ∑ f (Y j , θ ).
j=1
If the sample moments are estimators of the population moments with good properties, we then hope
that the estimator θC that holds the sample moment conditions fn (θ ) = 0 is a good estimator of the true
value θ 0 , which holds the population moment conditions E[ f (Y j , θ )] = 0 . Hence, θC is a moment estimator
of θ .
Theorem 2. Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample of size n from Y ∼ RBS(µ , δ ). Then, the moment
estimators of µ and δ are, respectively,
&
Ȳ 2 − S2 + Ȳ 4 + 3 Ȳ 2 S2
C = Ȳ
µ and δC = 2
,
S
where Ȳ = n−1 ∑nj=1 Y j and S2 = n−1 ∑nj=1 [Y j − Ȳ ]2 .
Proof. Recall from (2.1) and (2.2) that E[Y ] = µ and E[Y − µ ]2 = µ 2 [2δ + 5]/[δ + 1]2 . Recall θ = [µ , δ ]⊤
and define the vector of functions
# $⊤
2 µ 2 [2δ + 5]
f (Y j , θ ) = Y j − µ , [Y j − µ ] − .
[δ + 1]2
Then, the moment conditions are E[g(Y j , θ 0 )] = 0 . We have that fn (θC ) = 0 implies
1 n
1 n C 2 [2δC + 5]
µ
∑ Y j − µC = 0 and ∑ [Y j − µC ]2 − = 0,
n j=1 n j=1 [δC + 1]2
which, after some algebraic manipulations, results to be

√
−κC2 + 1 + 3κC2 + 1
C = Ȳ and
µ δC = , (2.5)
κC2
√ √
where κC = S2 /Ȳ is the sample coefficient of variation, and 0 < κC < 5. Therefore, we have that (4.2)
can be rewritten as &
Ȳ 2 − S2 + Ȳ 4 + 3 Ȳ 2 S2
C = Ȳ
µ and δC = .
S2
Theorem 3. Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample of size n from Y ∼ RBS(µ , δ ). Then, µ C and δC have
C , δC]⊤ approximately follows the distribution
an asymptotic bivariate normal joint distribution, that is, [µ
⎛ ⎡ ⎤⎞
/ 0 µ 2 {2 δ +5} µ {2 δ 2 +8 δ −3}
µ 1⎢ {δ +1}2
− {δ +1}{δ +4}
⎜ ⎥⎟
N2 ⎝ , ⎣ ⎦⎠ .
n δ 2 +8 δ −3} {2 δ 4 +28 δ 3 +122 δ 2 +126 δ +57}
δ − µ{{2
δ +1}{δ +4} {δ +4}2
Proof. let Y = [Y1 , . . . ,Yn ]⊤ be independent identically distributed (IID) RVs according to Y ∼ RBS(µ , δ )
C = f1 (Ȳ , S2 ) and δC = f2 (Ȳ , S2 ) be the moment estimators
and E[Y 4 ] given in (2.1) be finite. In addition, let µ
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
15
of the parameters µ and δ , respectively. Given by that the random vector

/ 0
√ Ȳ − E[Y ]
n
S2 − E[Y − µ ]2
converges in distribution to 1/ 0 2
0
N2 , Σ ,
0
where / 0
ν µ3
Σ= ,
µ3 µ4 − ν 2
with
µ 2 [2 δ + 5] 4[3δ + 11]µ 3
ν = Var[Y ] = , µ3 = and
[δ + 1]2 [δ + 1]3
8 µ 4 [δ 2 + 20δ + 76]
µ4 − ν 2 = .
[δ + 1]4
We want to determine the asymptotic joint distribution of the estimators
C = f1 (Ȳ , S2 ) and δC = f2 (Ȳ , S2 ).

µ
These estimators can be expressed as

&
x2 − y + x4 + 3 x2 y
f1 (x, y) = x and f2 (x, y) = .
y
Given that f1 (·) and f2 (·) are two continuous functions and by using the delta method (see Rao, 1965),
we obtain that the random vector / 0
√ C−µ
µ
n
δC − δ
converges in distribution to 1/ 0 2
0
N2 , Σ ,
0
where ⎡ ⎤
µ 2 {2 δ + 5} µ {2 δ 2 + 8 δ − 3}
⎢ − ⎥
⎢ {δ + 1}2 {δ + 1}{δ + 4} ⎥
Σ=⎢
⎢
⎥.
⎥
⎣ µ {2 δ 2 + 8 δ − 3} 4 3 2
{2 δ + 28 δ + 122 δ + 126 δ + 57} ⎦
−
{δ + 1}{δ + 4} {δ + 4}2
4.3 Modified moment estimation
Ng et al. (2003) used the fact that the BS distribution satisfies the reciprocation property to propose
MM estimates for the parameters of this distribution. The MM estimation method is a variation of the
moment estimation method, substituting the expression that equates the second population and sample
moments by equating the expected value of 1/Y with ∑nj=1 1/Y j . Because the reparameterized BS dis-
tribution preserves the reciprocation property, once again, the MM estimates of its parameters µ and δ
can be easily obtained.
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
16
Theorem 4. Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample of size n from Y ∼ RBS(µ , δ ). Then, the MM
estimators of µ and δ are, respectively,
⎡) ⎤−1
Ȳ
µ̆ = Ȳ and δ̆ = ⎣ − 1⎦ ,
Ȳh
where Ȳh = [{1/n} ∑nj=1 {1/Y j }]−1 .
Proof. Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample of size n from Y ∼ RBS(µ , δ ). Then, E[Y ] = µ and
E [1/Y ] = [δ + 1]2/[µδ 2 ]. Thus,
# $⊤
1 {δ + 1}2
f (Y j , ) = Y j − µ ,
θ − .
Yj µ δ2
Recall the moment conditions are E[ f (Y j , θ 0 )] = 0 . We have that fn (θ̆θ ) = 0 implies

n n
1 1 1 [δ̆ + 1]2
n ∑ Y j − µ̆ = 0 and
n ∑ Yj − µ̆ δ̆ 2
= 0. (2.6)
j=1 j=1
Hence, solving (4.3), we obtain the MM estimators

⎡) ⎤−1
Ȳ
µ̆ = Ȳ and δ̆ = ⎣ − 1⎦ ,
Ȳh
where Ȳh = [{1/n} ∑nj=1 {1/Y j }]−1 . In addition, we have that δ̆ is well-defined for Ȳh ̸= Ȳ , when Ȳh < Ȳ .
Theorem 5. Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample of size n from Y ∼ RBS(µ , δ ). Then, µ̆ and δ̆ have
an asymptotic bivariate normal joint distribution, that is, [µ̆ , δ̆ ]⊤ approximately follows the distribution
⎛ ⎡ ⎤⎞
µ 2 {2 δ + 5} 2µ δ
/ 0 −
⎜ µ ⎢
1 ⎢ {δ + 1}
2 {δ + 1} ⎥⎟
⎜ ⎥⎟
N2 ⎜ , ⎢ ⎥⎟ .
⎝ δ n⎣ 2µ δ ⎦⎠
− 2δ2
{δ + 1}
Proof. let Y = [Y1 , . . . ,Yn ]⊤ be IID RVs according to Y ∼ RBS(µ , δ ) and E[Y j4 ] < ∞. In addition, consider
the statistics / 0−1
n n
1 1 1
Ȳ =
n ∑ Yj and Ȳh =
n ∑ Yj .
j=1 j=1
The vector [Ȳ , Ȳh−1 ]⊤ follows a bivariate normal distribution, which implies that
/ 0 1/ 0 2
√ Ȳ − E[Y ] D 0
n → N2 , Σ ,
Ȳh−1 − E[Y −1 ] 0
D
where → denotes convergence in distribution, and
/ 0
Var[Y ] Cov[Y, Y −1 ]
Σ= ,
Cov[Y, Y −1 ] Var[Y −1 ]
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
17
with
µ 2 [2 δ + 5] −1 [δ + 1]2
Var[Y ] = , Cov[Y, Y ] = 1 − and
[δ + 1]2 δ2
[2 δ + 5] [δ + 1]2
Var[Y −1 ] = .
µ2 δ 4
However, our interest is to find the asymptotic joint distribution of
µ̆ = f1 (Ȳ , Ȳh−1 ) and δ̆ = f2 (Ȳ , Ȳh−1 ).
For these estimators, consider

√
f1 (x, y) = x, f2 (x, y) = [ x y − 1]−1 ,
two continuous functions and the delta method. Then,

/ 0 1/ 0 2
√ µ̆ − µ D 0
n → N2 , Σ ,
δ̆ − δ 0
where ⎡ ⎤
µ 2 {2 δ + 5} 2µ δ
⎢ {δ + 1}2 −
⎢ {δ + 1} ⎥
⎥
Σ=⎢ ⎥.
⎣ 2µ δ ⎦
− 2δ 2
{δ + 1}
4.4 Generalized Method of Moments estimation
The GMM is an estimation technique that generalizes of the usual moment estimation method. The
GMM provides estimators that are in general consistent, but in general not efficient. For more details
about this method, see, e.g., Mátyás (1999). Differently from the ML method that requires the complete
specification of the model and its probability distribution. The GMM does not require this sort of full
knowledge. It only demands the specification of a set of moment conditions which the model should
satisfy.
Definition 2 (GMM Estimator (Mátyás, 1999)). Let Y = [Y1 , . . ., Yn ]⊤ be a random sample of size n from
any distribution. We want to estimate an unknown parameter p × 1 vector θ , with true value θ 0 . Let
E[ f (Y j , θ 0 )] = 0 be a set of q moment conditions, and fn (Y j , θ ) the corresponding sample moments.
Define the criterion function
Qn (θ ) = fn (Y j , θ )⊤ A−1
n f n (Y j , θ ),
where A n is a O p (1) stochastic positive definite matrix. Then, the GMM estimator of θ is
θ = argmin Qn (θ ).
θ̌
θ
As mentioned, in general, the GMM estimation provides consistent estimators, but θ must be the
only solution of E[ f (Y j , θ )] and an element of a compact space. Some limit assumptions on high order
moments of f (Y j , θ ) also are needed. However, there are no restrictions on the model that generates
the data, except for the case of dependent data.
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
18
Theorem 6. Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample of size n from Y ∼ RBS(µ , δ ). Then, µ̌ and δ̌ have
an asymptotic bivariate normal joint distribution, that is, [µ̌ , δ̌ ]⊤ approximately follows the distribution
1/ 0 2
µ 1
N2 , V ,
δ n
asymptotically,
# $⊤ # $
∂ f (Y j , θ ) ∂ f (Y j , θ )
V =E A −1
n E .
∂θ ∂θ
Proof. Under some regularity conditions (see Mátyás, 1999, Section 1.3.2), as n goes to infinity, the
√ D
GMM estimator converges to a bivariate normal distribution and so the random vector n[θ̌θ − θ ] →
N2 (00,V
V ), where
# $⊤ # $
∂ f (Y j , θ ) ∂ f (Y j , θ )
V =E A−1
n E .
∂θ ∂θ
Considering q > p, we can perform the J -test (see Hansen, 1982) to verify whether the moment
conditions hold. In this case, the null hypothesis H0 : E[ f (Y j , θ 0 )] = 0 can be tested by using the statistic
−1
n fn (Y j , θ̌θ )⊤ Ǎ 2
An fn (Y j , θ̌θ ), which approximately follows the χq−p distribution; see Mátyás (1999).
5 Simulation
In this section, we conduct a study based on MC simulations for evaluating the performance of the
ML, moment, MM and GMM estimators for the RBS distribution. MC replications are based on Algorithm
1. For each replication generated by this algorithm, we calculate ML, moment, MM and GMM estimates.
The algorithm and estimation methods are implemented in the R software (www.R-project.org) by us-
ing the gamlss and gmm packages; see Stasinopoulos and Rigby (2007) and Chaussé (2010), respec-
tively. For details about generation of numbers from the BS distribution, see Leiva et al. (2008b) and
Barros et al. (2009). Then, the mean, bias, standard error (SE) and square root of the mean squared
√
error ( MSE) of these estimators are empirically computed. We obtain point estimates, confidence
intervals (CIs) and their coverage probabilities (CPs) of 95% level, based on the asymptotic results
associated with each estimator given in Section 4. The ML estimates are obtained from the gamlss
package and the GMM estimates from the gmm package. The CIs based on the GMM estimates are
obtained by using the R function confint(), where the main argument is an object of the gmm class.
The scenario of this simulation study considers 10000 MC replications in each case, sample sizes
n ∈ {30, 50, 75, 100, 200} and values for δ ∈ {0.5, 2.0, 8.0, 32.0, 200} (according to different levels of skew-
ness) and µ = 1.0. The obtained results are presented in Tables 2.2, 2.3, 2.4 and 2.5. To perform the
GMM estimation of the parameters of the RBS(µ , δ ) distribution, we consider the moment conditions
⎡ ⎤
µ − Yj
⎢ 2 ⎥
⎢ µ {2δ + 5} 2 ⎥
⎢ − {Y j − µ } ⎥
E [ f (Y j , θ )] = E ⎢ {δ + 1}2 ⎥ = 0,
⎢ 2
⎥
⎣ {δ + 1} 1 ⎦
2
−
µδ Yj
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
19
where the gradient function of fn (Y j , θ ) is given by

⎡ ⎤
1 0
⎢ ⎥
∂ fn (Y j , θ ) ⎢
⎢ 2 µ {2 δ + 5} 2 µ 2 {δ + 4} ⎥
⎥
G= = E⎢ − {2 µ − 2 Ȳ } − ⎥.
∂θ ⎢ {δ + 1}2 {δ + 1}3 ⎥
⎣ {δ + 1}2 2 {δ + 1} ⎦
− −
{ µ δ }2 µ δ3
From Tables 2.2 to 2.5, note that the ML, moment and MM estimators of the parameter µ present
√
similar statistical properties in relation to the empirical bias and MSE. However, the GMM estimator
presents similar properties to the other estimators only when the sample size is large. In the case of the
parameter δ , the ML and MM estimators present similar properties for the different sample sizes and
true values assumed for this parameter. Table 2.3 shows that, in general, the GMM underestimates the
√
true value of µ . From Tables 2.4 and 2.5, note that the values of the empirical SE and MSE increase
as δ increases, for all the considered methods, in the case of the parameter δ . Nevertheless, in the
√
case of the parameter µ , we have a reverse behavior, that is, the values of the empirical SE and MSE
decrease as δ increases, for all the considered methods. In addition, the GMM estimator presents
the worse behavior in terms of statistical properties, but, as the sample size increases, the estimators
obtained by this method turn to be more competitive, with respect to the other estimators considered.
Table 2.6 provides empirical CPs of 95% CIs for the parameters of the RBS(µ , δ ) distribution. Note
that the CIs based on the GMM estimates have CPs smaller than those from the other methods. How-
ever, as the sample size increases, the distance between CPs for the fixed confidence levels decreases.
Also, when the true value of δ increases, the distance between the nominal value (0.95) and the empiri-
cal CP decreases. Thus, such as in the study based on point estimation, for interval estimation, ML and
MM estimators present similar statistical properties and better than the other estimators considered.
6 Applications
In this section, we provide a practical illustration of the proposed estimation methods based on the
two real data sets, with small and large sample sizes and from two fields: economics and engineering.
6.1 Data set I (S1): Griffiths et al. (1993)
The first data set is presented in Griffiths et al. (1993) and corresponds to household expenditures
for food in the United States (US) expressed in (US$). These data are provided in Table 2.7.
Table 2.7: Household expenditures for food (in dollars) in the US provided by Griffiths et al. (1993).
15.998 16.652 21.741 7.431 10.481 13.548 23.256 17.976 14.161 8.825
14.184 19.604 13.728 21.141 17.446 9.629 14.005 9.160 18.831 7.641
13.882 9.670 21.604 10.866 28.980 10.882 18.561 11.629 18.067 14.539
19.192 25.918 28.833 15.869 14.910 9.550 23.066 14.751
Table 2.8 presents a descriptive summary of the household expenditures for food that include, mean
(Ȳ ), median (Ỹ ), standard deviation (SD), coefficient of variation (CV), coefficient skewness (CS), coeffi-
cient kurtosis (CK), minimum (Y(1) ) and maximum (Y(n) ) values. Note that the distribution of the response
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
20
is slightly positively skewed.
Table 2.8: Descriptive statistics for the household expenditures for food (US$)
Y(1) Ȳ Ỹ SD CV CS CK Y(n)
7.431 15.953 14.831 5.624 35.256 0.525 −0.444 28.980
0.08
25
0.06
20
Density
0.04
food
15
0.02
10
0.00
Adjusted Boxplot Original Boxplot 5 10 15 20 25 30

food
(a) (b)
Figure 2.3: Boxplots (a) and histogram with the estimated density of the RBS (b) for the household expenditures
for food.
Figure 2.3 presents the boxplots and histogram with the estimated density of the RBS, with µ =
15.95 and δ = 15.57, for the household expenditures for food. In Figure 2.3(a) note that the adjusted
and original boxplots exhibit the same behavior. This behavior makes sense since the data have little
asymmetry. Already in Figure 2.3(b) is possible to note that the data has good adjustments to the
distribution RBS and we confirm this behavior using measures of goodness-of-fit.
Chen and Balakrishnan (1995) proposed a general approximate goodness-of-fit test for the hypoth-
esis H0 : Y1 , . . . ,Yn with Yi following F(y; θ ), where the form of F is known but the p-vector θ is unknown.
This method is based on the Cramér-von Mises (CM) and Anderson-Darling (AD) statistics and, in gen-
eral, the smaller the values of these statistics, the better the fit. In this application, such methodology is
applied to provide goodness-of-fit tests of the RBS distribution to data set in study. The values (p-values)
of the CM and AD statistics obtained for this data set were 0.271 (0.656) and 0.043 (0.608), respectively.
Thus confirming that the RBS distribution present an proper adjust to data.
Point estimates for the proposed methods and 90% and 95% CIs for the parameters µ and δ are
displayed in Table 2.9.
Table 2.9: Point estimates and CIs for the indicated parameter and method using S1.
µ δ
Method Estimate CI(90%) CI(95%) Estimate CI(90%) CI(95%)
ML 15.95 [14.41;17.50] [14.11;17.79] 15.57 [ 9.70;21.45] [ 8.57;22.57]

Moment 15.95 [14.47;17.43] [14.19;17.72] 16.91 [ 9.51;24.31] [ 8.10;25.72]
MM 15.95 [14.41;17.50] [14.11;17.79] 15.57 [ 9.70;21.45] [ 8.57;22.57]
GMM 15.30 [14.31;16.30] [14.12;16.49] 15.94 [10.96;20.92] [10.00;21.87]
For S1 data, we obtain p-value = 0.430 when the J -test is applied, which once again supports that
the null hypothesis is not rejected for significance levels of 1%, 5% and 10%. Therefore, we do not have
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
21
evidence to conclude that the moment conditions are incorrect or that the RBS(µ , δ ) distribution does
not fit the data well in this case.
6.2 Data set II (S2): Birnbaum and Saunders (1969b)

The second data set is a classical one used in the literature on the topic. These data were intro-
duced by Birnbaum and Saunders (1969b) and correspond to lifetimes of 6061-T6 aluminum coupons
expressed in cycles (×10−3) at a maximum stress level of 3.1 psi (×104), until the failure to occur. These
coupons were cut parallel to the direction of rolling and oscillating at 18 cycles per seconds. The data
are displayed in Table 2.10.
Useful descriptive information for S2 data are n = 101, Ȳ = 133.73, Ȳh = 129.930 and SD = 22.360.
The values (p-value) of the CM and AD statistics obtained for this data set were 0.502 (0.202) and
0.086 (0.169), respectively. Thus confirming that the data present an proper adjust to RBS distribution.
In Figure 2.4(a) note that the adjusted and original boxplots exhibit the same behavior. This behavior
makes sense since the data have little asymmetry. Already in Figure 2.4(b) is possible to note that really
the data has good adjustments to the distribution RBS .
Table 2.10: Lifetimes (in cycles ×10−3 ) of coupons provided by Birnbaum and Saunders (1969b).
70 90 96 97 99 100 103 104 104 105 107 108 108 108 109
109 112 112 113 114 114 114 116 119 120 120 120 121 121 123
124 124 124 124 124 128 128 129 129 130 130 130 131 131 131
131 131 132 132 132 133 134 134 134 134 134 136 136 137 138
138 138 139 139 141 141 142 142 142 142 142 142 144 144 145
146 148 148 149 151 151 152 155 156 157 157 157 157 158 159
162 163 163 164 166 166 168 170 174 196 212
200
0.015
180
160
0.010
Density
lifetime
140
120
0.005
100
80
0.000
Adjusted Boxplot Original Boxplot 100 150 200

lifetime
(a) (b)
Figure 2.4: Boxplots (a) and histogram with the estimated density of the RBS (b) for the lifetimes of coupons.
Point estimates for the parameters using the proposed methods and 90% and 95% CIs are pre-
sented in Table 2.11. From this table, we note that less accurate CIs are obtained by the GMM.
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
22
Table 2.11: Point estimates and CIs for the indicated parameter and method using S2.
µ δ
Method Estimate CI(90%) CI(95%) Estimate CI(90%) CI(95%)
ML 133.73 [129.99;137.47] [129.27;138.19] 68.89 [52.95; 84.84] [49.89;87.89]

Moment 133.73 [130.09;137.37] [129.39;138.07] 72.76 [55.24; 90.27] [51.88;93.63]
MM 133.73 [129.99;137.47] [129.27;138.19] 68.89 [52.95; 84.84] [49.89;87.89]
GMM 137.69 [129.62;145.76] [128.08;147.31] 75.36 [33.88;116.85] [25.93;124.80]
By means of the function specTest() of the gmm package, which computes the J -test, we check
whether moment conditions of the GMM are satisfied or not for fatigue data. This test can also be used
to verify goodness-of-fit, because, when the null hypothesis is rejected, this is an indication that moment
conditions are defined in a wrong way or the model does not fit the data well. For S2 data, we obtain
a p-value = 0.720, thus the null hypothesis is not rejected at significance levels of 1%, 5% and 10%.
Therefore, we do not have evidence to conclude that the moment conditions are incorrect or that the
RBS(µ , δ ) distribution does not fit the data well.
7 Conclusion
In this chapter, we have provided some novel results on moments and generation of random num-
bers from a reparameterized version of the Birnbaum-Saunders distribution. In addition, we have studied
several estimation methods for this parameterization. We have considered the maximum likelihood, mo-
ment, modified moment and generalized moment methods to estimate the corresponding parameters.
Furthermore, we have conducted a Monte Carlo study to evaluate the performance of each of these
estimators. From this study, we can conclude that the maximum likelihood and modified moment es-
timators present similar statistical properties and better that those of the other estimators considered.
Therefore, due to the modified moment estimators are easier to compute, we recommend their use for
the RBS distribution. In addition, we have obtained moment estimators in a closed-form, which is not
possible with the original parameterization of the Birnbaum-Saunders distribution. However, the param-
eter estimators obtained by the moment method, as well as those obtained by the generalized method
of moments, are underperformed with respect to their statistical properties. Nevertheless, for the case
of large sample sizes, all the studied estimators have similar statistical properties. We have discussed
applications of the BS distribution in different scientific fields and taken advantage of the computational
implementation in the R software for carrying an application with two real data sets, which allowed us to
illustrate the obtained results.
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
23
Table 2.2: Empirical mean of the estimator of the indicated parameter, method, n and δ , with µ = 1.0.
µ δ
n δ
ML Moment MM GMM ML Moment MM GMM
0.5 1.004 1.002 1.002 0.869 0.561 0.772 0.561 0.633
2.0 1.001 1.001 1.001 0.929 2.232 2.526 2.232 2.508
30 8.0 1.000 1.000 1.000 0.978 8.886 9.285 8.886 9.949
32.0 1.000 1.000 1.000 1.005 35.477 35.920 35.477 40.352
200.0 1.000 1.000 1.000 1.003 221.734 222.150 221.734 245.663
0.5 0.999 0.998 0.998 0.896 0.536 0.668 0.536 0.578

2.0 0.999 0.999 0.999 0.946 2.137 2.321 2.137 2.319
50 8.0 1.000 1.000 1.000 0.981 8.522 8.775 8.522 9.174
32.0 1.000 1.000 1.000 1.002 34.058 34.339 34.058 37.064
200.0 1.000 1.000 1.000 1.002 212.782 213.040 212.782 227.794
0.5 0.998 0.996 0.996 0.916 0.524 0.610 0.524 0.552

2.0 0.999 0.998 0.998 0.958 2.092 2.210 2.092 2.220
75 8.0 0.999 0.999 0.999 0.985 8.355 8.518 8.355 8.835
32.0 1.000 1.000 1.000 1.000 33.385 33.559 33.385 35.463
200.0 1.000 1.000 1.000 1.001 208.676 208.810 208.676 219.441
0.5 0.999 0.998 0.998 0.933 0.518 0.581 0.518 0.539

2.0 0.999 0.999 0.999 0.967 2.068 2.150 2.068 2.163
100 8.0 1.000 1.000 1.000 0.988 8.261 8.377 8.261 8.634
32.0 1.000 1.000 1.000 0.999 33.022 33.148 33.022 34.590
200.0 1.000 1.000 1.000 1.001 206.366 206.453 206.366 214.828
0.5 0.998 0.997 0.997 0.960 0.509 0.541 0.509 0.521

2.0 0.999 0.999 0.999 0.980 2.036 2.077 2.036 2.085
200 8.0 1.000 0.999 0.999 0.993 8.137 8.195 8.137 8.338
32.0 1.000 1.000 1.000 0.998 32.529 32.600 32.529 33.313
200.0 1.000 1.000 1.000 1.001 203.274 203.362 203.274 207.927
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
24
Table 2.3: Empirical bias of the estimator of the indicated parameter, method, n and δ , with µ = 1.0.
µ δ
n δ
0.5 0.004 0.002 0.002 −0.131 0.061 0.272 0.061 0.133

2.0 0.001 0.001 0.001 −0.071 0.232 0.526 0.232 0.508
30 8.0 0.000 0.000 0.000 −0.022 0.886 1.285 0.886 1.949
32.0 0.000 0.000 0.000 0.005 3.477 3.920 3.477 8.352
200.0 0.000 0.000 0.000 0.003 21.734 22.150 21.734 45.663
0.5 −0.001 −0.002 −0.002 −0.104 0.036 0.168 0.036 0.078

2.0 −0.001 −0.001 −0.001 −0.054 0.137 0.321 0.137 0.319
50 8.0 0.000 0.000 0.000 −0.019 0.522 0.775 0.522 1.174
32.0 0.000 0.000 0.000 0.002 2.058 2.339 2.058 5.064
200.0 0.000 0.000 0.000 0.002 12.782 13.040 12.782 27.794
0.5 −0.002 −0.004 −0.004 −0.084 0.024 0.110 0.024 0.052

2.0 −0.001 −0.002 −0.002 −0.042 0.092 0.210 0.092 0.220
75 8.0 −0.001 −0.001 −0.001 −0.015 0.355 0.518 0.355 0.835
32.0 0.000 0.000 0.000 0.000 1.385 1.559 1.385 3.463
200.0 0.000 0.000 0.000 0.001 8.676 8.810 8.676 19.441
0.5 −0.001 −0.002 −0.002 −0.067 0.018 0.081 0.018 0.039

2.0 −0.001 −0.001 −0.001 −0.033 0.068 0.150 0.068 0.163
100 8.0 0.000 0.000 0.000 −0.012 0.261 0.377 0.261 0.634
32.0 0.000 0.000 0.000 −0.001 1.022 1.148 1.022 2.590
200.0 0.000 0.000 0.000 0.001 6.366 6.453 6.366 14.828
0.5 −0.002 −0.003 −0.003 −0.040 0.009 0.041 0.009 0.021

2.0 −0.001 −0.001 −0.001 −0.020 0.036 0.077 0.036 0.085
200 8.0 0.000 −0.001 −0.001 −0.007 0.137 0.195 0.137 0.338
32.0 0.000 0.000 0.000 −0.002 0.529 0.600 0.529 1.313
200.0 0.000 0.000 0.000 0.001 3.274 3.362 3.274 7.927
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
25
Table 2.4: Empirical SE of the estimator of the indicated parameter, method, n and δ , with µ = 1.0.
µ δ
n δ
0.5 0.296 0.298 0.298 0.308 0.162 0.440 0.162 0.257

2.0 0.182 0.182 0.182 0.195 0.638 0.986 0.638 0.925
30 8.0 0.092 0.092 0.092 0.102 2.532 2.993 2.532 3.533
32.0 0.046 0.046 0.046 0.051 10.100 10.627 10.100 13.359
200.0 0.018 0.018 0.018 0.020 63.122 63.636 63.122 78.065
0.5 0.226 0.228 0.228 0.237 0.113 0.340 0.113 0.153

2.0 0.139 0.139 0.139 0.148 0.448 0.733 0.448 0.582
50 8.0 0.071 0.071 0.071 0.078 1.786 2.186 1.786 2.212
32.0 0.035 0.035 0.035 0.040 7.134 7.624 7.134 8.809
200.0 0.014 0.014 0.014 0.015 44.580 45.136 44.580 51.508
0.5 0.185 0.187 0.187 0.193 0.089 0.276 0.089 0.104

2.0 0.114 0.114 0.114 0.121 0.353 0.591 0.353 0.432
75 8.0 0.058 0.058 0.058 0.063 1.404 1.744 1.404 1.663
32.0 0.029 0.029 0.029 0.032 5.609 6.025 5.609 6.693
200.0 0.012 0.012 0.012 0.013 35.043 35.502 35.043 39.398
0.5 0.159 0.160 0.160 0.166 0.075 0.240 0.075 0.084

2.0 0.099 0.099 0.099 0.104 0.299 0.504 0.299 0.347
100 8.0 0.051 0.051 0.051 0.055 1.191 1.484 1.191 1.372
32.0 0.025 0.025 0.025 0.028 4.764 5.128 4.764 5.535
200.0 0.010 0.010 0.010 0.011 29.733 30.126 29.733 32.884
0.5 0.114 0.115 0.115 0.118 0.051 0.172 0.051 0.055

2.0 0.070 0.070 0.070 0.073 0.206 0.354 0.206 0.221
200 8.0 0.036 0.036 0.036 0.037 0.820 1.028 0.820 0.884
32.0 0.018 0.018 0.018 0.019 3.283 3.538 3.283 3.563
200.0 0.007 0.007 0.007 0.008 20.510 20.790 20.510 21.865
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
26
√
Table 2.5: Empirical MSE of the estimator of the indicated parameter, method, n and δ , with µ = 1.0.
µ δ
n δ
0.5 0.296 0.298 0.298 0.334 0.173 0.517 0.173 0.290

2.0 0.182 0.182 0.182 0.208 0.679 1.117 0.679 1.055
30 8.0 0.092 0.092 0.092 0.104 2.683 3.257 2.683 4.035
32.0 0.046 0.046 0.046 0.052 10.682 11.327 10.682 15.755
200.0 0.018 0.018 0.018 0.020 66.759 67.380 66.759 90.440
0.5 0.226 0.228 0.228 0.259 0.119 0.379 0.119 0.172

2.0 0.139 0.139 0.139 0.158 0.469 0.800 0.469 0.663
50 8.0 0.071 0.071 0.071 0.080 1.861 2.320 1.861 2.505
32.0 0.035 0.035 0.035 0.040 7.425 7.975 7.425 10.161
200.0 0.014 0.014 0.014 0.016 46.376 46.981 46.376 58.528
0.5 0.185 0.187 0.187 0.210 0.092 0.297 0.092 0.116

2.0 0.114 0.114 0.114 0.128 0.365 0.627 0.365 0.485
75 8.0 0.058 0.058 0.058 0.065 1.448 1.819 1.448 1.861
32.0 0.029 0.029 0.029 0.032 5.777 6.223 5.777 7.536
200.0 0.012 0.012 0.012 0.013 36.101 36.578 36.101 43.933
0.5 0.159 0.161 0.161 0.179 0.077 0.253 0.077 0.093

2.0 0.099 0.099 0.099 0.109 0.307 0.526 0.307 0.383
100 8.0 0.051 0.051 0.051 0.056 1.219 1.531 1.219 1.511
32.0 0.025 0.025 0.025 0.028 4.873 5.255 4.873 6.111
200.0 0.010 0.010 0.010 0.011 30.407 30.809 30.407 36.072
0.5 0.114 0.115 0.115 0.125 0.052 0.177 0.052 0.059

2.0 0.070 0.070 0.070 0.075 0.209 0.362 0.209 0.237
200 8.0 0.036 0.036 0.036 0.038 0.832 1.046 0.832 0.947
32.0 0.018 0.018 0.018 0.019 3.325 3.589 3.325 3.797
200.0 0.007 0.007 0.007 0.008 20.770 21.060 20.770 23.258
Chapter
___________2___
-_Estimation
_______________in
___the
_____RBS
_______distribution
_______________________________________________________________________________________
27
Table 2.6: CP of 95% CIs for the indicated parameter, method, n and δ , with µ = 1.0.
µ δ
n δ
0.5 0.899 0.884 0.896 0.622 0.956 0.993 0.957 0.864
2.0 0.917 0.906 0.916 0.707 0.956 0.983 0.956 0.861
30 8.0 0.930 0.924 0.930 0.785 0.956 0.970 0.956 0.858
32.0 0.937 0.935 0.937 0.826 0.956 0.961 0.956 0.836
200.0 0.942 0.940 0.942 0.815 0.956 0.958 0.956 0.880
0.5 0.915 0.903 0.914 0.703 0.955 0.984 0.955 0.886

2.0 0.929 0.921 0.930 0.779 0.954 0.978 0.954 0.878
50 8.0 0.939 0.934 0.938 0.826 0.954 0.967 0.954 0.886
32.0 0.943 0.941 0.942 0.857 0.954 0.960 0.953 0.864
200.0 0.943 0.943 0.943 0.843 0.953 0.954 0.953 0.896
0.5 0.928 0.920 0.926 0.757 0.954 0.982 0.953 0.904

2.0 0.936 0.930 0.936 0.820 0.954 0.973 0.953 0.899
75 8.0 0.941 0.938 0.940 0.862 0.954 0.964 0.954 0.901
32.0 0.943 0.942 0.942 0.880 0.954 0.957 0.953 0.887
200.0 0.944 0.944 0.944 0.862 0.954 0.953 0.954 0.906
0.5 0.935 0.929 0.933 0.794 0.952 0.978 0.952 0.913

2.0 0.942 0.939 0.942 0.848 0.952 0.972 0.952 0.910
100 8.0 0.944 0.942 0.944 0.879 0.953 0.961 0.953 0.911
32.0 0.944 0.940 0.944 0.888 0.952 0.955 0.952 0.897
200.0 0.944 0.944 0.943 0.869 0.952 0.953 0.952 0.912
0.5 0.940 0.935 0.938 0.851 0.952 0.978 0.952 0.926

2.0 0.944 0.942 0.943 0.888 0.951 0.969 0.951 0.926
200 8.0 0.949 0.947 0.948 0.916 0.950 0.958 0.950 0.926
32.0 0.948 0.949 0.948 0.916 0.950 0.952 0.950 0.925
200.0 0.947 0.947 0.947 0.894 0.950 0.950 0.950 0.927
28
CHAPTER 3
Reparameterized Birnbaum-Saunders regression models with varying

precision
1 Resumo
Neste capítulo, introduzimos um modelo de regressão com precisão variável baseada na dis-
tribuição Birnbaum-Saunders reparametrizada, que generaliza o modelo de Leiva et al. (2014). Além
disso, estudamos as estatíticas da razão de verossimilhanças, score, Wald e gradiente utilizadas para
testar se o parâmetro de precisão é constante e, em seguida, propomos alguns resíduos para este
modelo e estudamos seus comportamentos empíricos. Também, calculamos a alavancagem general-
izada e as curvaturas no estudo de influência local. Por fim, usamos um conjunto de dados reais para
ilustrar a metodologia proposta.
2 Introduction
Different strategies have been proposed for modeling with best using the Birnbaum-Saunders dis-
tribution in the past few years. However, there does not exist a BS model that allows data to be modeled
in their original scale. Leiva et al. (2014) proposed a RBS regression model in which the mean is related
to a linear predictor through a link function. This model also includes a precision parameter whose recip-
rocal can be viewed as a dispersion measure. Such a model is flexible, because one can use different
link functions. In the standard formulation of the RBS regression model, it is assumed that the precision
is constant across observations. However, in many practical situations this assumption does not hold.
Thus, we consider an extension of the model proposed by Leiva et al. (2014) in that both the mean and
precision parameters are related to explanatory variables.
The gols of this chapter are: (i) to generalize the model proposed by Leiva et al. (2014) for including
variable precision, (ii) to propose diagnostics for the precision parameter and residuals for this model,
(iii) to develop influence diagnostics for these regression models, (iv) to study the performance of the
proposed methodology, and (v) to apply the obtained results to real data.
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
29
The outline of the chapter is as follows. In Section 3, we introduce a regression model with varying
precision based in the RBS distribution, which generalizes the model of Leiva et al. (2014). In Section 4,
we discuss the likelihood ratio, Wald, score and gradient tests for the precision parameter. In Section 5,
we propose some residuals for this model and study their empirical behavior. Also, we obtain an ex-
pression for the generalized leverage and the curvatures in the study of local influence. In Section 6, we
use a real data set to show the flexibility and potentially of the proposed methodology. Some concluding
remarks are given in Section 7.
3 The model with varying precision

Consider a random sample Y = [Y1 , . . . ,Yn ]⊤ , where Yi ∼ RBS(µi , δi ), for i = 1, . . . , n, and consequently
Var[Yi ] = µi2 [2δi +5]/[δi +1]2 . In this chapter the corresponding mean and the precision parameters satisfy
the functional relations
g(µi ) = ηi = x ⊤
i β, and h(δi ) = τi = z ⊤
i α i = 1, . . . , n, (3.1)
where µi = g−1 (xx⊤ ⊤

i β ), with β = [β1 , . . . , β p ] being a vector of unknown parameters to be estimated,
for p < n, and x i = [1, x21 , . . . , xip ]⊤ a vector that contains the values of p explanatory variables, δi =
h−1 (zz⊤ ⊤
i α ), with α = [α1 , . . . , αq ] being a vector of unknown parameters to be estimated, for q < n, and
z i = [1, z21 , . . . , ziq ]⊤ a vector of explanatory variables. In this model, the link functions g: R → R+ and
h: R → R+ are strictly monotone, positive, and at least twice differentiable.
Table 3.1 list the most common link functions for g and h, respectively, along with their first and
second derivatives.
Table 3.1: Results of d µ /dη , dδ /dτ ,d2 µ /dη 2 and d2 δ /dτ 2 .
Link functions g h d µ /dη dδ /dτ d2 µ /dη 2 d2 δ /dτ 2
Identity µ =η δ =τ 1 1 0 0
Log log(µ ) = η log(δ ) = τ µ δ µ δ
√ √
Square root µ =η δ =τ 2µ 2δ 2 2
The log-likelihood function for RBS regression models with varying precision has is
n
n
ℓ(θ ) = ∑ ℓi (δi , µi ) − log(16π ), (3.2)
i=1 2
where
δi log(δi + 1) log(µi ) 3 log(yi ) yi [δi + 1] δi2 µi
ℓi (µi , δi ) = − − − + log(δi yi + yi + δi µi ) − − .
2 2 2 2 4µi 4yi [δi + 1]
The components of the score vector, obtained by differentiating the log-likelihood function with re-
spect to the parameters, are given, for j = 1, . . . , p and r = 1, . . . , q, as
n
∂ ℓ(β , α )
Uβ j (θ ) = = ∑ [y∗i − µi∗ ] ai xi j ,
∂βj i=1
where
δi yi [δi + 1] δi2 1 d µi 1 ∂ ηi
y∗i = + − , µ∗ = , ai = = ′ , xi j = ,
[δi yi + yi + δi µi ] 4 µi 2 4yi [δi + 1] i 2 µi d ηi g ( µi ) ∂βj
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
30
and
n
∂ ℓ(β , α )
Uαr (θ ) = = ∑ [y⋆i − δi⋆ ] bi zir ,
∂ αr i=1
where
[yi + µi ] yi δi [δi + 2]µi ⋆ δi d δi 1 ∂ τi
y⋆i = − − 2
, δi = − , bi = = ′ and zik = .
[δi yi + yi + δi µi ] 4µi 4[δi + 1] yi 2 [δi + 1] d τi h (δi ) ∂ αk
Therefore, we can write the [p + q] × 1 dimensional score vector U(θ ) in the form [Uβ (θ )⊤ ,Uα (θ )⊤ ]⊤ ,
with
Uβ (θ ) = X ⊤ A [yy∗ − µ ∗ ],
Uα (θ ) = Z ⊤ B[yy⋆ − δ ⋆ ],
where A = ai δinj and B = bi δinj , with δinj being the Kronecker delta. Thus, A and B are two diagonal
matrices.
The ML estimates of β and α are obtained as the solution of the non-linear system U(θ ) = 0 .
In practice, the ML estimates can be obtained through a numerical maximization of the log-likelihood
function using a non-linear optimization algorithm. The BS distribution is inserted into the gamlss pack-
age; for more details, see Stasinopoulos et al. (2006); Stasinopoulos and Rigby (2007). Thus, using the
gamlss() function, we obtain the ML estimates of the parameters of the proposed regression model.
The Hessian matrix is given by
⎡ ⎤ / 0
∂ 2 ℓ(β ,α ) ∂ 2 ℓ(β ,α )
∂β ∂β ⊤ ∂β ∂α⊤ X ⊤C X X ⊤M Z
h (θ ) = ⎣ ∂ 2 ℓ(β ,α ) ∂ 2 ℓ(β ,α )
⎦= . (3.3)
Z ⊤M X Z ⊤W Z
∂α∂β ⊤ ∂α∂α⊤
The matrix h β is the first element of the block matrix h (θ ), and it is obtained by the partial derivatives
n
∂ 2 ℓ(β , α )
∂ β j ∂ βk
= ∑ ci xi j xik ,
i=1
which can be written in matrix form as ḧhβ = X ⊤C X , with C = ci δinj , and
(i) (i) ′
ci = dµ 2 (ai )2 + d µ ai ai ,
where
y∗i µ∗
: ;8 9 :;89
i
∂ ℓ(θ ) δi yi [δi + 1] δi2 1

(i)
dµ = = + − − = y∗i − µi∗ ,
∂ µi [δi yi + yi + δi µi ] 4µi 2 4yi [δi + 1] 2µi
(i) ∂ 2 ℓ(θ ) 1 δi2 yi [δi + 1]
dµ 2 = 2
= 2
− 2
− ,
∂ µi 2µi [δi yi + yi + δi µi ] 2µi 3
′′
′ ∂ d µi g (µi )
ai = =− ′ .
µi ηi {g (µi )}2
The second element of the block matrix h (θ ), h β α say, is given by

n
∂ ℓ(θ )
hβ α = = ∑ mi xi j zir ,
∂ β j ∂ αr i=1
where
(i)
mi = dµδ ai bi
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
31
and
(i) ∂ 2 ℓ(θ ) yi yi δi [δi + 2]
d µδ = = + − .
∂ µi ∂ δi [δi yi + yi + δi µi ]2 4µi2 4[δi + 1]2 yi
In matrix form, we have h β α = X ⊤ M Z , with M = mi δinj . Consequently, we obtain the third element of the
matrix block h (θ ), since this is the matrix transpose of h β α . Finally, the last element of h (θ ), h α say, is
given by
n
∂ 2 ℓ(β , α )
hα = = ∑ wi zil zis ,
∂ αl ∂ αs i=1
(i) (i) ′
where wi = dδ 2 (bi )2 + dδ bi bi , with
y⋆i −δi⋆
: ;8 9 : ;8 9
(i) ∂ ℓ(θ ) [yi + µi ] yi δi [δi + 2]µi δi
dδ = = − − + = y⋆i − δi⋆ ,
∂ δi [δi yi + yi + δi µi ] 4µi 4[δi + 1]2yi 2[δi + 1]
(i) ∂ 2 ℓ(θ ) 1 [yi + µi ]2 µi
dδ 2 = 2
= 2
− 2
− ,
∂ δi 2[δi + 1] [δi yi + yi + δi µi ] 2[δi + 1]3 yi
′′
′ ∂ d δi h (δi )
bi = =− ′ ,
δi τi {h (δi )}2
and its matrix form can be expressed by h α = Z ⊤W Z , with W = wi δinj .

The inverse matrix h (θ )−1 can be expressed as
/ 0
−1 hβ hβ α
h (θ ) = , (3.4)
h αβ hα
where
hβ X ⊤W 1 X ]−1 ,
= [X
hα Z ⊤W 2 Z ]−1 ,
= [Z
(hhαβ )⊤ X ⊤W 1 X ]−1 X ⊤ M Z [Z
= h β α = −[X Z ⊤W Z ]−1 ,
with
W1 Z ⊤W Z ]−1 Z ⊤ M ,
= C − M Z [Z
W2 X ⊤C X ]−1 X ⊤ M .
= W − M X [X
The Fisher information matrix, i (θ ), is given by

/ 0
X ⊤V X X ⊤ SZ
i (θ ) = , (3.5)
Z ⊤SX Z ⊤U Z
where i β = X ⊤V X , i β α = (iiαβ )⊤ = X ⊤ S Z and i α = Z ⊤U Z with V , S and U given by
V = Eci δinj = diag{Ec1 , . . . , Ecn },

S = Emi δinj = diag{Em1 , . . . , Emn },
U = Ewi δinj = diag{Ew1 , . . . , Ewn },
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
32
and
# $
δi δi2
Eci = E[ci ] = + I(θ ) a2i ,
2µi2 {δi + 1}2
# $
1 δi µi
Emi = E[mi ] = + I(θ ) ai bi ,
2µi {δi + 1} {δi + 1}3
# 2 $
{δi + 3δi + 1} µi2
E wi = E[wi ] = + I( θ ) b2i ,
2δi2 {δi + 1}2 {δi + 1}4
where
B ∞√ # $−1 " # $%
δi + 1 exp{δi /2} δi µi δi {δi + 1}yi δi µi
I(θ ) = √ 3/2
yi + exp − + dyi . (3.6)
0 4 π µ i yi δi + 1 4 δi µi {δi + 1}yi
The integral defined in (3.6) can be calculated, for example, numerically using the integrate() function of
the R software; see R Development Core Team (2011).
The inverse Fisher information matrix is given by
/ 0
iβ iβ α
i (θ )−1 = , (3.7)
i αβ iα
where
iβ X ⊤W 3 X ]−1 ,
= [X
iα Z ⊤W 4 Z ]−1 ,
= [Z
(iiαβ )⊤ X ⊤W 3 X ]−1 X ⊤ S Z [Z
= i β α = −[X Z ⊤U Z ]−1 ,
with
W3 Z ⊤U Z ]−1 Z ⊤ S ,
= V − S Z [Z
W4 X ⊤V X ]−1 X ⊤ S .
= U − S X [X
Consider W β β = V , W β α = W ⊤ W say, has

β α = S and W αα = U . Therefore, the augmented 2n × 2n matrix, W̃
the form / 0 / 0
W ββ W βα V S
W=
W̃ = .
W βα W αα S U
X as being another augmented 2n × [p + q] matrix of the form

Also, consider X̃
/ 0
X 0
X=
X̃ .
0 Z
⊤
X W̃
Therefore, the Fisher information matrix can be rewritten as i (θ ) = X̃ X and its inverse as i (θ )−1 =
W X̃
⊤
X W̃
[X̃ X ]−1 . We present below the Fisher scoring method, whose estimation algorithm for θ = [β ⊤ , α ⊤ ]⊤
W X̃
is given by
θ (m+1) = θ (m) + i (θ )−1U (θ )(m)

/ 0
⊤ (m) X ⊤ A (m) [yy∗ − µ ∗ ](m)
= θ (m) + [X̃
X W̃
W X ]−1
X̃
Z ⊤ B (m) [yy⋆ − δ ⋆ ](m)
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
33
/ 0/ 0
⊤ (m) ⊤ A (m) 0 [yy∗ − µ ∗ ](m)
= θ (m) + [X̃
X W̃
W X ]−1 X̃
X̃ X
0 B (m) B[yy⋆ − δ ⋆ ](m)
⊤ (m) ⊤ (m) ∗(m)
X W̃
= [X̃ W X ]−1 X̃
X̃ X W̃
W z ,
where / 0/ 0
∗(m) (m) (m) −1 A (m) 0 [yy∗ − µ ∗ ](m)
z Xθ
= X̃ W
+ {W̃ } .
0 B(m) B[yy⋆ − δ ⋆ ](m)
Note that θ (m+1) has the form of an estimated of reweighted least squares, where z ∗ is a modified
X is the model matrix. In the gamlss() function, the use of the option method =
response variable and X̃
CG() provides the Cole and Green algorithm, which can be better for distributions with potentially highly
correlated parameter estimates and equivalent to the Fisher scoring method, when we are in the fully
parametric case.
Under the usual regularity conditions Cox and Hinkley (1974, p.107), the ML estimators of θ and
i (θ ), θ! and i (θ! ), respectively, are consistent and efficient. Suppose that j (θ ) = limn→∞ (1/n)iin (θ ) exists
and is non-singular, then
√ ! D
n[θ − θ ] → N p+q (00, j (θ )−1 ).
Consequently, if θk denotes the kth element of θ , we have
D
[θ!k − θk ]{iikk }−1/2 → N (0, 1),
where i kk is kth diagonal element of i (θ )−1 , for k = 1, . . . , p + q. Let i j j be the jth diagonal element
and i rr the [p + r]th diagonal element of this matrix. We have that z(γ ) denotes the quantiles γ of the
N (0, 1) distribution. The asymptotic intervals of β!j and α
!r , for j = 1, . . . , p and r = 1, . . . , q, with asymptotic
coverage of 100[1 − q]%, are given by β!j ± z(1 − γ /2){ii j j }−1/2 and α
!r ± z(1 − γ /2){iirr }−1/2, respectively. It
is also possible to obtain an approximate 100 × [1 − q]% confidence region for θ as
[θ! − θ ]⊤ i (θ! )−1 [θ! − θ ] ≤ χ1−

2
γ (p + q), θ ∈ R p+q ,
2 (p + q) is the [1 − γ ]th quantile of the chi-squared distribution with p + q degree of freedom.

where χ1− γ
In addition, we can obtain 100 × [1 − γ ]% confidence bands for the linear predictor µ (xx pred ) = g−1 (xx⊤
pred
β ),
p !
for all x pred ∈ R , where x pred is an arbitrary p × 1 vector. The asymptotic distribution of β is given by
β! ∼ N p (β , i β ). Then, an approximate 100 × [1 − γ ]% confidence region for µ (xxpred ) can be obtained by
·
/ 0
3 7 ? A 21 6 3 7 ? A 12 6
⊤ β! ⊤ β!
g −1
x pred β! −
⊤ 2 (p)
χ1− γ x pred i x pred ,g −1 ⊤ ! 2
x pred β + χ1−γ (p) x pred i x pred ,
!
X ⊤W
where i β = [X L3 = V! − !
L3 X ]−1 , with W Z ⊤U
S Z [Z S evaluated at β! ; for more details, see Liu et al.
! Z ]−1 Z ⊤ !
(2008).
4 Diagnostics for varying precision

Below, we study some classical methods (for large samples) based on the likelihood function. The
likelihood ratio test proposed by Wilks (1938); see also Neyman and Pearson (1928), the Wald test,
proposed by Wald (1943), the Rao test proposed by Rao (1948) and gradient test proposed by Terrell
(2002), are studied.
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
34
In what follows, we consider tests based on the likelihood ratio (ζR ), Wald (ζW ), Rao Score (ζS )
and gradient (ζG ) statistics for testing the null hypothesis H0 : α ∗ ∈ Ω 0 against the alternative hypothesis
/ Ω0 , where Ω0 = {θ = [β , α ]: αi = 0, i = 2, . . . , q}, Ω = Ω0 ∪ Ωc0 and α ∗ = [α2 , . . . , αq ]⊤ .
HA : α ∗ ∈
The likelihood ratio statistic for testing H0 can be expressed by
n
ζR = 2 ∑ Λi , (3.8)
i=1
where " % 1 2 / 0 / 0
[δ!i − δ̃i ] µ̃i δ!i yi + yi + δ!i µ
!i yi {δ̃i + 1} {δ!i + 1} 1 δ̃i µ̃i δ!i µ̃i
Λi = + log + log + − + − .
2 !i
µ δ̃i yi + yi + δ̃i µ̃i 4 µ̃i !i
µ 4yi {δ̃i + 1} {δ!i + 1}
Define now Uα∗ (θ ) as a column q − 1 vector containing the final elements of the score vector and
∗
Var[α ∗ ] = i α is a [q − 1] × [q − 1] matrix composed of the q − 1 last lines and of the q − 1 last columns of
∗
the inverse matrix i (θ ). It is possible to show that the matrix i α can be expressed as
4 5−1
Var(α ∗ ) = Z ∗⊤W 4 Z ∗ , (3.9)
where W 4 is defined in (3.7) and Z ∗ is composed of the elements of the q − 1 last columns of Z . Thus,
the score statistic for testing H0 can be expresses in the general form
ζS = Uα∗⊤ (θC )Var(

Mα C ∗ )Uα∗ (θC ), (3.10)
where Uα∗ (θC ) and Var[

Mα C ∗ ] are evaluated under H0 . Based on (3.9), (3.10) and (3.3), we have
∗ 4 5−1 ∗
ζS = [yy∗ − δC ]⊤ B C 4Z ∗
C Z ∗ Z ∗⊤W C [yy∗ − δC ],
Z ∗⊤ B (3.11)
∗
where δC , B
C and W
C 4 are evaluated under H0 .
The Wald statistic for testing H0 is given by
ζW ! ∗⊤ Var(
= α Nα ! ∗ )−1 α
!∗
! ∗⊤ Z ∗⊤W
= α ! ∗,
! 4Z ∗α (3.12)
N α
where Var( ! ∗ ) is evaluated at the unrestricted ML estimator and α
! ∗ is the ML estimator of α ∗ . Finally,
the gradient statistic for testing H0 is given by
∗
ζG = [yy∗ − δC ]⊤ B ! ∗,
C Z ∗α (3.13)
∗
where δC and B
C are evaluated under H0 and α ∗
! is the ML estimator of α ∗ . There are some recent studies
on this test statistic, see eg, Lemonte (2013) and Vargas et al. (2013).
For n large, supposing the usual regularity conditions and under H0 the statistics defined in (3.8),
2
(3.11), (3.12) and (3.13), have a χq−1 distribution. Thus, the tests above can be performed using ap-
2
proximate critical values of the χq−1 distribution.
4.1 Simulation
Below, we report the result of Monte Carlo simulations carried out to compare the performances
of the ζR , ζS , ζW and ζG tests in small- and moderate-sized samples. The estimates of the model
coefficients are obtained using the gamlss() function, contained in the package of the same name. To
use the gamlss() function for our goals, we introduce RBS distribution defined in (1.2) in the same
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
35
structure used by the distributions contained in the gamlss.dist library. Thereafter, we implement four
functions based on the statistics defined in (3.8), (3.11), (3.12) and (3.13). These functions are named
of lr.test, wald.test, score.test and grad.test. We define a measure for the degree of heterogeneity of
precision of the data as
max δi
λ= , i = 1, . . . , n.
min δi
When λ = 1 the precision is constant.
We consider the RBS regression model with varying precision. The simulations are based on the
model
log(µi ) = 2.0 − 1.7xi1 and log(δi ) = α0 + α1 zi1 ,
where i = 1, . . . , n. The values of the explanatory variable values, xi1 and zi1 , are selected as random
draws from the U (0, 1) distribution and for fixed n those values are kept constant throughout the exper-
iment. To perform the simulations, we implement in R language a program that calculates the rejection
rates of all the tests (varying the values of δ ) and the finite-sample power properties considering dif-
ferent sample sizes n ∈ {30, 50, 100}. Through this function, we obtain the descriptive levels of the test
and compared with the nominal levels ξ ∈ {1%, 5%, 10%}. In this study, the number of Monte Carlo
replications is 5000. The null hypothesis is H0 : α1 = 0, which is tested against a two-sided alternative.
Considering α0 ∈ {1.5, 2.0, 3.0}, we have δ ∈ {4.5, 7.4, 20.1}. The results of this study are shown in Table
3.2. Observing Table 3.2 we note that the likelihood ratio, Wald and gradient tests are markedly liberal.
Overall, the score test is conservative. It is also the lesst size distorted test. For instance, when δ = 20.1,
n = 100, and ξ = 1%, the rejection rates are 1.38 (ζR ), 0.96 (ζS ), 1.76 (ζW ) and 1.26 (ζG ). As expected, the
rejection rates of all the tests approach the corresponding nominal levels as the sample size increases.
Note that the score and gradient tests presented the best performances.
In Figure 3.1, we observe the fit of the empirical distributions of the test statistics, of the four test
studied with the asymptotic distribution of these statistics for n = 50. Note that the empirical distributions
of the score and gradient tests, for this scenario, present behavior closer to the theoretical distribution
(χ12 ).
The simulations for evaluating the power of these tests are performed using empirical critical values
obtained from simulations, thus ensuring that they all have the correct size. We computed the rejection
rates under the alternative hypothesis α1 = −1.0, −2.0 and − 3.0 and α0 = 1.5 who implies on values of
λ ≈ 2.6, 6.8 and18.4. Table 3.3 presents the power of the tests for different values of sample sizes and
degree of heterogeneity of the dispersion (λ ).
In relation to power the score and gradient tests present lower values, however next, to the likelihood
ratio and Wald tests. For instance, when λ = 18.4, n = 50 and ξ = 5% the powers of tests are 99.20 (ζR ),
98.62 (ζS ), 99.44 (ζW ) and 99.20 (ζG ). As expected, the powers of all the four tests increase with λ .
In small to moderate-sized samples the best performing tests are the score and gradient tests.
These tests are less size distorted than the other two and are as powerful as the others. Hence, these
tests may be recommended for testing hypotheses on the precision parameter in the RBS regression
models. The gradient test has a slight advantage over the score test because the gradient statistic is
simpler to calculate.
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
36
Table 3.2: Rejection rates of, ζR , ζS , ζW and ζG .
n statistics 4.5 7.4 20.1
1% 5% 10% 1% 5% 10% 1% 5% 10%

ζR 1.52 6.62 12.34 1.42 6.60 12.34 1.44 6.58 12.46
ζS 0.72 4.38 9.60 0.72 4.40 9.74 0.72 4.34 9.82
30
ζW 2.76 8.72 15.34 2.76 8.60 15.38 2.76 8.58 15.24
ζG 1.24 6.12 11.74 1.22 6.08 11.70 1.20 6.22 11.82
ζR 1.28 5.90 10.64 1.30 5.80 10.72 1.26 5.80 10.76

ζS 0.78 4.64 9.38 0.78 4.66 9.46 0.76 4.70 9.48
50
ζW 1.90 7.06 12.40 1.88 7.02 12.38 1.86 6.96 12.42
ζG 1.10 5.70 10.40 1.12 5.66 10.40 1.14 5.70 10.40
ζR 1.36 5.54 10.80 1.36 5.54 10.72 1.38 5.64 10.72

ζS 0.92 4.96 9.62 0.94 5.02 9.76 0.96 5.02 9.78
100
ζW 1.80 6.34 11.62 1.84 6.36 11.78 1.76 6.30 11.60
ζG 1.26 5.38 10.62 1.28 5.40 10.58 1.26 5.38 10.62
χ12 χ12 χ12

ζR ζR ζR
ζW ζW ζW
ζS ζS ζS
ζG ζG ζG
3
3
Empirical Quantiles
Empirical Quantiles
Empirical Quantiles
2
2
1
1
0
0 1 2 3 0 1 2 3 0 1 2 3
Theorical Quantiles Theorical Quantiles Theorical Quantiles
(a) (b) (c)
Figure 3.1: Quantiles for δ = 4.5-(a), δ = 7.4(b) and δ = 20.1-(c) when n = 50.
5 Diagnostic techniques
5.1 Residuals
Below, we introduce several residuals, thus extending the results in Leiva et al. (2014). The re-
sponse or ordinary residuals are given by
!i .
ri = yi − µ
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
37
The first residual we introduce is the so-called Pearson residual or standardized residual given by
1/2
!i ]
[yi − µ φ! [yi − µ !i ]
r p,i = 7 = i& ,
N i)
Var(y !i )
V (µ
!i = g−1
where V (µi ) = µi2 and φi = [δi + 1]2/[2δi + 5], with µ ! ! −1 !
1 (ηi ) and δi = g2 (τi ) being the ML estimators
of µi and δi , respectively.
Table 3.3: Powers of the ζR , ζS , ζW and ζG tests.
n Statistic 2.6 6.8 18.4
1% 5% 10% 1% 5% 10% 1% 5% 10%

ζR 7.40 20.82 32.26 38.20 63.38 75.00 76.68 91.78 95.48
ζS 5.80 19.66 30.24 30.24 57.70 70.96 61.82 86.32 92.34
30
ζW 7.36 20.78 31.92 36.50 63.80 75.60 77.44 92.82 96.20
ζG 7.30 20.46 31.84 37.48 62.38 74.72 75.72 91.40 95.42
ζR 13.24 32.36 45.46 65.56 85.24 92.00 96.22 99.20 99.68

ζS 11.92 31.66 43.80 60.00 82.38 89.92 92.22 98.62 99.34
50
ζW 14.20 32.72 45.58 68.06 86.18 92.50 97.40 99.44 99.80
ζG 13.48 31.88 45.60 65.94 85.12 92.00 96.34 99.20 99.70
ζR 28.38 54.86 68.58 92.90 98.18 99.18 99.94 100.00 100.00

ζS 29.72 54.14 67.00 91.58 97.56 98.98 99.80 100.00 100.00
100
ζW 27.36 54.78 68.80 93.20 98.44 99.30 100.00 100.00 100.00
ζG 28.12 54.80 68.50 92.82 98.20 99.16 99.94 100.00 100.00
Considering the Fisher Scoring iterative process for β given by
β (m+1) = β (m) + [X
X ⊤V (m) X ]−1 X ⊤ A (m) [yy∗ − µ ∗(m) ],
and we can rewrite the above procedure as an iterative process of reweighted least squares given by
X ⊤V (m) X ]−1 X ⊤V (m) z ∗(m)

β (m+1) = [X 1 ,
∗(m)
where z 1 V (m) ]−1 A (m) [yy∗ − µ ∗(m) ]. With the convergence of the process we have
= η (m) + [V
β! ≈ [X
X ⊤V! X ]−1 X ⊤V!!z ∗1 ,
! + [V! ]−1 A
with !z ∗1 = η ! ∗ ]. This result can be interpreted as least squares solution of the linear
! [yy∗ − µ
regression of !z ∗1 on X with weight matrix V! . Based on this result we present the second proposal of
residuals that is the score residual they are obtained by performing a weighted least squares of !z ∗1 on X ,
1/2 −1/2
being defined as follows r = V! [!z ∗ − η
! ] = V!
1 A ! ∗ ). Thus, the score residuals can be expressed
! (yy∗ − µ
as
[y∗i − µ! ∗]
rs,i = & i ,
ν!i
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
38
2
δ!i δ!i
where ν!i = + I(µ̂i , δ̂i ).
2µ!i2 [δ!i + 1]2
We also propose a residual type deviance component, where we assume that the dispersion pa-
rameter has a known regression structure. This residuals are given by
/ 1 20 1
√ δ!i {δ!i + 1}yi δ!i2 µ
!i 1 δ!i {δ!i + 1}µ
!i yi 2
!i ) 2 log(2) − +
rd,i = sgn(yi − µ + + log ,
2 !i
4µ 4{δ!i + 1}yi 2 {δ!i yi + yi + δ!i µ
!i }2
where the signal function sgn (or signum), is defined by

⎧
⎪
⎨ −1, if x < 0,
⎪
sgn(x) = 0, if x = 0,
⎪
⎪
⎩ 1, if x > 0.
Finally, if FY (y; µ , δ ) is the CDF of the RBS(µ , δ ), then the quantile residual is defined by
3 6
!i , δ!i ) ,
rq,i = Φ−1 FY (y; µ
where Φ(·) is the cumulative distribution function of the standard normal distribution. For more details
this residual, see Dunn and Smyth (1996).
Residual plots can be used to verify the adequacy of the model, investigate the non-linearity, outliers,
!i .
!i or η
autocorrelation, varying precision plotting against explanatory variables individually, or against µ
5.1.1 Simulation
The basic idea is to generate simulated values, by Monte Carlo simulation, form the RBS regression
model with varying precision and obtain the residuals proposed in the previous section to study their
empirical distributions. The structures adopted for µi and δi are
log(µi ) = β0 + β1xi and h(δi ) = α0 + α1 zi , i = 1, . . . , 20,
respectively. The simulations are performed considering the following two cases: Case I: we consider
the case where h(·) = log(·), the vectors x = [x1 , x2 , . . . , x20 ]⊤ and z = [z1 , z2 , . . . , z20 ]⊤ are equal and are
obtained from a uniform distribution on the interval (30,60). We consider β = [2.14, 0.10]⊤ and α =
[3.30, −0.01]⊤ as the true values of the coefficients of model. Case II: we consider the link function
&
h(·) = (·).
Software for performing the analysis in this section is available as code in R. The Monte Carlo
experiment is conducted considering 5000 replicates. To study the behavior of these residuals, we
computed some descriptive measures: the mean (R̄), the SD, the CS and the estimator of Pearson’s
measure of kurtosis (KT). These results are presented in Tables 3.4 and 3.5.
Notice that all the residuals have mean approximately equal to zero for both cases. The results
show that, in case I and case II, the residuals have SD’s closer to one. Note also that the empirical
distributions of the residuals r p,i have positive asymmetry, see Figures 3.2(a)-3.3(a). In addition, notice
that the residuals rs,i , rd,i and rq,i present asymmetry close to zero. The residuals r p,i and rs,i presents,
generally, KT closer to three for both cases. Notice that the residuals rd,i and rq,i display a similar
behavior.
We perform a comparison between the empirical distributions of the residuals with the normal dis-
tribution. For doing this, we use a quantile versus quantile (QQ) plot, which displays sample quantiles
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
39
of the residuals versus theoretical quantiles of the normal distribution with mean zero and variance one.
Figures 3.2-3.3 present four QQ plots one for each 5000 residuals from r p,i , rs,i , rd,i and rq,i . Notice that
the QQ plots of the residuals rs,i , rd,i and rq,i are more distant from the diagonal line. However, notice
that the residual rd,i presents behavior approximately linear, although some problems are presented in
central part and in the tails of the sample distribution.
5.2 Generalized leverage and local influence
The generalized leverage (GL) of an estimator is defined in regression models as a measure of the
importance of individual observations. The GL can be obtained in general form as follows
∂ !y S @ >>
= D θ [−hh(θ )]−1 L θ y > ! ,
∂y θ =θ
where D θ = ∂ µ /∂ θ , h (θ ) = ∂ 2 ℓ(θ )/∂ θ ∂ θ and L θ y = ∂ 2 ℓ(θ )/∂ θ ∂ y ⊤ . In the specific case of the RBS
regression model with varying precision we have that the matrix of generalized leverages is given by
β βα ⊤
! X [−!
GL(θ! ) = A !E
X ⊤A
h ]X ! X [−!
! +A Z B
h ]Z !L
!, (3.14)
(i) (i)
where E = dyµ δinj and L = dyδ δinj with
(i) [δi + 1] δ2 δi [δi + 1] (i) µi δi [δi + 2] 1 µi

dyµ = 2
+ 2 i − and dyδ = − − .
4 µi 4 yi [δi + 1] [δi yi + yi + δi µi ]2 4 y2i [δi + 1]2 4 µi [δi yi + yi + δi µi ]2
Considering the results presented in (3.14) and (3.4) we have that the GL can be rewritten as
? A
GL(θ! ) = GL(β! ) [A
!E! ]−1 M
! Z [Z ! Z ]−1 Z ⊤ B
Z ⊤W !L!−I , (3.15)
where GL(β! ) = A
! X (X !E
! 1 X )−1 X ⊤ A
X ⊤W ! is the GL considering only the vector of parameter β .
The definition of local influence is due to Cook (1986) and is based on the curvature of the plane
of the log-likelihood function. In case of the BS model given in Eq. (3.1), let θ = [β ⊤ , α ⊤ ]⊤ and ℓ(θ |ω )
be the parameter vector and the log-likelihood function corresponding to this model perturbed by ω ,
respectively. The vector of perturbations ω belongs to a subset Ω ∈ Rn and ω 0 is an non-perturbed
n-vector, such that ℓ(θ |ω 0 ) = ℓ(θ ), for all θ . The likelihood displacement (LD) given by LD(ω ) = 2[ℓ(θ! ) −
ℓ(θ! ω )], where θ! ω denotes the ML estimate of θ upon the perturbed BS model can be used to assess
the influence of the perturbation on the ML estimate. Cook (1986), showed that the normal curvature for
θ in the direction l , with ||ll || = 1, is expressed as
Cl (θ ) = 2|ll ⊤ ∆ ⊤ h (θ )−1 ∆ l |, (3.16)
where ∆ is a [p + q] × n matrix of perturbations with elements ∆ ji = ∂ 2 ℓ(θ |ω )/∂ θ j ∂ ωi evaluated at θ = θ!

and ω = ω 0 , for j = 1, . . . , p + 1 and i = 1, . . . , n, and h(θ ) is given in Eq. (3.3). A local influence diagnostic
is generally based on index plots. For example, the index graph of the eigenvector l max corresponding
to the maximum eigenvalue of
F (θ ) = ∆ ⊤ h (θ )−1 ∆ , (3.17)
say Clmax (θ ), evaluated at θ = θ! , can repeal those observations that under small perturbations have a
great influence on LD(ω ).
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
40
Table 3.4: Descriptive summary of the indicated residual in case I.
r(·),i 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
r p,i 0.01 0.01 −0.00 −0.02 0.02 −0.00 0.00 −0.01 −0.00 −0.00 0.02 −0.00 −0.01 0.00 0.00 −0.01 0.02 0.00 −0.02 0.00
rs,i 0.01 0.03 0.02 0.00 0.00 0.00 −0.01 0.01 −0.02 0.02 0.02 −0.00 −0.02 −0.01 −0.00 −0.02 −0.00 −0.01 −0.01 0.00
R̄
rd,i 0.00 0.02 0.01 −0.01 −0.01 −0.01 −0.02 −0.00 −0.03 0.01 0.01 −0.01 −0.03 −0.02 −0.01 −0.03 −0.01 −0.02 −0.02 −0.01
rq,i 0.01 0.03 0.02 0.00 0.00 0.00 −0.01 0.01 −0.02 0.02 0.02 −0.00 −0.02 −0.01 −0.01 −0.02 −0.00 −0.01 −0.01 0.00
r p,i 0.98 0.92 0.91 0.93 1.05 0.96 1.05 0.93 1.06 0.91 0.97 0.98 1.00 1.04 1.02 1.01 1.06 1.03 0.96 1.01
rs,i 1.00 0.94 0.93 0.94 1.05 0.98 1.03 0.95 1.03 0.93 0.98 1.00 1.03 1.02 1.03 1.04 1.05 1.03 0.97 1.01
SD
rd,i 0.99 0.92 0.92 0.93 1.03 0.97 1.04 0.93 1.05 0.92 0.97 0.98 1.01 1.04 1.01 1.01 1.04 1.03 0.96 0.99
rq,i 1.00 0.93 0.93 0.94 1.05 0.98 1.05 0.95 1.06 0.93 0.98 0.99 1.02 1.05 1.02 1.03 1.05 1.04 0.97 1.01
r p,i 0.59 0.60 0.50 0.55 0.86 0.56 0.73 0.61 0.85 0.60 0.57 0.70 0.67 0.75 0.76 0.73 0.84 0.81 0.62 0.76
rs,i −0.03 −0.01 −0.05 −0.00 0.06 0.01 0.00 0.01 0.04 −0.01 −0.03 0.05 −0.05 0.03 0.04 −0.04 0.05 0.01 0.04 0.08
CS
rd,i −0.01 0.02 −0.02 0.02 0.08 0.04 0.02 0.05 0.07 0.02 0.00 0.07 −0.01 0.05 0.06 −0.01 0.06 0.04 0.06 0.10
rq,i −0.03 −0.02 −0.05 −0.00 0.05 0.01 0.00 0.01 0.04 −0.02 −0.03 0.04 −0.04 0.03 0.04 −0.03 0.04 0.02 0.04 0.07
r p,i 2.90 2.87 2.72 2.82 3.75 2.73 3.33 2.80 3.70 2.83 2.84 3.06 3.13 3.31 3.30 3.36 3.75 3.62 2.92 3.26
rs,i 2.52 2.48 2.44 2.44 2.85 2.38 2.73 2.41 2.87 2.45 2.53 2.53 2.67 2.75 2.67 2.80 2.84 2.89 2.48 2.62
KT
rd,i 2.40 2.35 2.34 2.34 2.66 2.29 2.58 2.29 2.69 2.32 2.42 2.40 2.51 2.60 2.52 2.62 2.65 2.68 2.39 2.48
rq,i 2.40 2.35 2.34 2.34 2.66 2.28 2.57 2.29 2.68 2.32 2.42 2.40 2.51 2.60 2.51 2.62 2.64 2.68 2.38 2.48
Table 3.5: Descriptive summary of the indicated residual in case II.
r(·),i 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
r p,i 0.00 0.01 −0.01 −0.02 0.02 0.00 0.01 −0.01 0.00 0.00 0.01 0.00 -0.02 0.00 0.00 −0.01 0.02 0.01 −0.02 0.00
rs,i 0.01 0.04 0.03 0.01 0.00 0.01 −0.02 0.02 −0.03 0.04 0.02 0.00 −0.03 −0.02 −0.01 −0.03 −0.01 −0.02 −0.01 0.00
R̄
rd,i −0.01 0.02 0.01 −0.01 −0.02 −0.01 −0.03 0.00 −0.04 0.01 0.00 −0.02 −0.05 −0.04 −0.03 −0.04 −0.03 −0.03 −0.03 −0.02
rq,i 0.01 0.04 0.03 0.01 −0.01 0.01 −0.02 0.02 −0.03 0.03 0.02 0.00 −0.03 −0.02 −0.01 −0.03 −0.01 −0.02 −0.01 0.00
r p,i 0.96 0.90 0.88 0.90 1.07 0.94 1.04 0.91 1.06 0.88 0.95 0.98 0.99 1.04 1.02 1.01 1.07 1.04 0.93 1.00
rs,i 0.97 0.89 0.89 0.90 1.06 0.94 1.06 0.91 1.07 0.88 0.96 0.98 1.02 1.05 1.02 1.03 1.07 1.05 0.94 1.00
SD
rd,i 0.96 0.90 0.89 0.90 1.03 0.94 1.03 0.91 1.04 0.88 0.95 0.97 1.00 1.03 1.00 1.01 1.04 1.02 0.93 0.98
rq,i 0.99 0.92 0.91 0.93 1.06 0.97 1.06 0.94 1.07 0.91 0.98 1.00 1.03 1.06 1.03 1.04 1.07 1.05 0.96 1.01
r p,i 0.87 0.84 0.76 0.82 1.20 0.82 1.04 0.84 1.18 0.82 0.83 0.96 0.96 1.05 1.05 1.04 1.17 1.14 0.89 1.03
rs,i −0.03 0.01 −0.04 0.01 0.08 0.03 0.00 0.03 0.05 0.01 −0.03 0.05 −0.06 0.03 0.05 −0.05 0.07 0.02 0.05 0.08
CS
rd,i 0.02 0.06 0.01 0.05 0.11 0.07 0.05 0.09 0.09 0.07 0.02 0.09 0.01 0.08 0.09 0.02 0.09 0.07 0.09 0.12
rq,i −0.03 −0.01 −0.05 0.00 0.06 0.02 0.01 0.02 0.05 0.00 −0.03 0.04 −0.04 0.04 0.04 −0.03 0.05 0.02 0.04 0.07
r p,i 3.46 3.33 3.20 3.33 4.71 3.24 4.07 3.22 4.68 3.24 3.34 3.63 3.80 3.97 3.99 4.10 4.67 4.49 3.48 3.87
rs,i 2.68 2.61 2.60 2.60 3.07 2.52 2.91 2.51 3.09 2.56 2.68 2.67 2.84 2.90 2.84 3.00 3.05 3.12 2.63 2.77
KT
rd,i 2.42 2.36 2.38 2.38 2.66 2.31 2.58 2.30 2.68 2.32 2.45 2.40 2.51 2.59 2.52 2.62 2.65 2.67 2.41 2.48
rq,i 2.42 2.35 2.37 2.37 2.64 2.30 2.56 2.28 2.66 2.31 2.44 2.39 2.50 2.57 2.51 2.61 2.63 2.66 2.39 2.46
5.3 Total local influence
In addition to the direction vector of maximum normal curvature, l max , another direction of interest
is l i = e in , which corresponds to the direction of the ith observation, where e in is an n × 1 vector of
zeros with a value equal to one at the ith position, i.e., {eein , 1 ≤ i ≤ n} is the canonical basis of Rn .
In this case, the normal curvature is given by Ci (θ ) = 2| fii |, where fii is the ith diagonal element of
F (θ ) given in Eq. (3.17), for i = 1, . . . , n, evaluated at θ = θ! . Observations such that Ci (θ! ) > 2C(θ! ),
where C(θ! ) = ∑ni=1 Ci (θ! )/n, are considered as potentially influential. This procedure is called total local
influence of the ith case; see Lesaffre and Verbeke (1998).
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
41
4
2
2
r p,i
rs,i
0
0
-2
-2
-4
-4
-4 -2 0 2 4 -4 -2 0 2 4
Index Index
(a) (b)
4
4
2
0 2
rd,i
rq,i
0
-2
-2
-4
-4
-4 -2 0 2 4 -4 -2 0 2 4
Index Index
(c) (d)
Figure 3.2: Normal probability plots for the residuals r p,i (a), rs,i (b), rd,i (c) and rq,i (d) in case I.
5.4 Calculation of the curvatures
Consider the log-likelihood function given in Eq. (3.2). Next, for the indicated scheme, we obtain
the respective perturbation matrix ∆ given in general in Eq. (3.16), which is already evaluated at the
non-perturbed vector ω 0 .
Scheme 1: Case-weight perturbation– Under this scheme, we want to determine whether the contribu-
tions of the observations with different weights impact the ML estimate of θ . This scheme is the most
used on for evaluating the local influence in a model. In this scheme, the log-likelihood function of the
perturbed BS model is ℓ(θ |ω ) = ∑ni=1 ℓi (θ |ωi ) = ∑ni=1 ωi ℓi (θ ). Then, considering its derivative with respect
to ω ⊤ , we obtain / 0
∆β
∆= ,
∆α
where the elements of ∆ β and ∆ α are respectively, given by
( ji) (i)
∆ β = d µ a i xi j , i = 1, . . . , n, j = 1, . . . , p,
(ri) (i)
∆ α = dδ bi zir , i = 1, . . . , n, r = 1, . . . , q,
(i) (i)
which must be evaluated at θ = θ! . In matrix form we have to ∆ β = X ⊤ ai dµ δinj and ∆ α = Z ⊤ bi dδ δinj .
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
42
4
4
2
2
0
r p,i
rs,i
0
-2
-2
-4
-4
-4 -2 0 2 4 -4 -2 0 2 4
Index Index
(a) (b)
4
4
2
2
rd,i
rq,i
0
0
-2
-2
-4
-4
-4 -2 0 2 4 -4 -2 0 2 4
Index Index
(c) (d)
Figure 3.3: Normal probability plots for the residuals r p,i (a), rs,i (b), rd,i (c) and rq,i (d) in case II.
Scheme 2: Response perturbation – This scheme can be applied in several ways. Here, we use an
additive perturbation in the BS model given by
Yiω = Yi + ωi Syi , i = 1, . . . , n, (3.18)
where Syi is a scale factor that could represent the standard deviation of the response. In this case, the
perturbed log-likelihood function for the BS model is ℓ(θ |ω ) = ∑ni=1 ℓi (θ |ωi ), ω 0 = [0, . . . , 0]⊤ and
/ 0
∆β
∆= ,
∆α
where the elements of ∆ β and ∆ α , are respectively, given by
( ji) (i)
∆ β = dyµ ai xi j Syi , i = 1, . . . , n, j = 1, . . . , p,
(ri) (i)
∆ α = dyδ bi zir Syi , i = 1, . . . , n, r = 1, . . . , q,
(i) (i)
In matrix we have to ∆ β = X ⊤ ai dyµ Syi δinj and ∆ α = Z ⊤ bi dyδ Syi δinj evaluated at θ = θ! .
It is possible to establish a relationship between the generalized leverage and the local influence.
We have under the scheme of perturbation studied that
>
>
F (θ ) = ∆ ⊤ h (θ )−1 ∆ > !
θ =θ
>
>
= ∆ β⊤ h β ∆ β + ∆⊤
αh
αβ
∆β + ∆⊤
β h βα
∆ α + ∆ ⊤ α
α h ∆ α >
θ =θ!
? 4 5 A
!y L
= Σ !B
! Z (Z ! 2 Z )−1 Z ⊤ B
Z ⊤W !L! − (Z ! Z )−1 Z ⊤ M
Z ⊤W !A! −1 GLβ − E L θ) Σ
! GL( !y ,
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
43
! y = SL
where Σ n
yi δi j . From this result it is natural to make a comparison between the results obtained by
analyzing the local influence and generalized leverage.
Scheme 3: Predictor perturbation (x⊤

i ) – The third perturbation scheme involves the perturbation of a
continuos explanatory variable, say xl . We replace xl by xl + ω Sxl , where Sxl is the standard deviation
of xl . In this case, the ith linear predictor component is ηi (ω ) = β1 + · · · + βl [xil + ωi Sxl ] + · · · + β p xip and
ω 0 = [0, . . . , 0]⊤ . The perturbed log-likelihood function for the BS model is ℓ(θ |ω ) = ∑ni=1 ℓi (µi (ωi ), δi ) and
/ 0
∆β
∆= ,
∆α
where the elements of ∆ β and ∆ α , are respectively, given by

⎧
⎪
⎨ βl ci xi j Sxl , for j ̸= l
∆(βji) =
⎪
⎩ βl ci xil Sx + d (i) ai Sx , for j = l
l µ l
(ri)
∆ α = βl mi zir Sxl , i = 1, . . . , n, r = 1, . . . .
(i)
In matrix form we have ∆ β = βl Sxl X ⊤ ci δinj + Sxl X ⊤ n
0 d µ ai δi j , where X 0 is a n × p matrix of zeros
except for the lth column, which is a column vector of ones and ∆ α = βl Sx Z ⊤ mi δ n evaluated at θ = θ! .
l ij
Scheme 4: Predictor perturbation (z⊤

i ) – The fourth perturbation scheme also involves the perturbation
of a continuos explanatory variable. Now we substitute zk by zk + ω Szk , where Szk is the standard deviation
of zk . In this case, the ith linear predictor component is τi (ω ) = α1 + · · · + αk [zik + ωi Szk ] + · · · + αq ziq and
ω 0 = [0, . . . , 0]⊤ . The perturbed log-likelihood function for the BS model is ℓ(θ |ω ) = ∑ni=1 ℓi (µi , δi (ωi )) and
/ 0
∆β
∆= ,
∆α
where the elements of ∆ β and ∆ α are, respectively, given by
( ji)
∆ β = αk mi xi j Szk , i = 1, . . . , n, j = 1, . . . , p,
⎧
⎪
⎨ αk wi zir Szk , for r ̸= k;
(ri)
∆α =
⎪
⎩ αk wi zik Sz + d (i) bi Sz , for r = k.
k δ k
(i)
In matrix form we have ∆ α = αk Szk Z ⊤ wi δinj + Sxl Z ⊤ n
0 dδ bi δi j , where Z 0 is an n × q matrix of zeros
except for the kth column, which is a column vector of ones and ∆ = αk Sz X ⊤ mi δ n evaluated at θ = θ! .
β k ij
Scheme 5: Predictors perturbation (x⊤ ⊤

i , zi ) – In the fifth and last perturbation scheme presented, we
considered when some perturbed explanatory variables in X are also present in Z. We consider that xil =
zik and in this case, the ith linear predictor component is τi (ω ) = α1 + · · · + αk [xil + ωi Sxl ] + · · · + αq ziq and
ω 0 = [0, . . . , 0]⊤ . The perturbed log-likelihood function for the BS model is ℓ(θ |ω ) = ∑ni=1 ℓi (µi (ωi ), δi (ωi ))
and / 0
∆β
∆= ,
∆α
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
44
where the elements of ∆ β and ∆ α are, respectively, given by

⎧
⎪
⎨ βl ci xi j Sxl + αk mi xi j Sxl , for j ̸= l;
( ji)
∆β =
⎪
⎩ βl ci xil Sx + αk mi xil Sx + d (i) ai Sx , for j = l.
l l µ l
⎧
⎪
⎨ αk wi zir Sxl + βl mi zir Sxl , for r ̸= k;
(ri)
∆α =
⎪
⎩ αk wi xil Sx + βl mi xil Sx + d (i) bi Sx , for r = k.
l l δ l
(i) n
X ⊤ [βl ci δinj + αk mi δinj ] + X ⊤
We can formulate the above results in the matrix form as ∆ β = Sxl {X 0 ai d µ δi j }
(i) n
Z ⊤ [αk wi δinj + βl mi δinj ] + Z ⊤
and ∆ α = Sxl {Z 0 bi dδ δi j }, where X 0 and Z 0 are, respectively, n × p and n × q
matrices of zeros except for the (l = k)th columns, which are columns vector of ones. The matrices ∆ β
and ∆ are evaluated at θ = θ! .
β
6 Application
We consider the data set presented in the alr3 package use the command data(landrent) of the R
software or in http://www.statsci.org/data/general/landrent.txt, for more details, see Weisberg
(2005, Problem 9.10, p. 208). The data were collected by Douglas Tiffany to study the variation in
rent paid in 1977 for agricultural land planted with alfalfa. One of the objectives of this study was to
investigate the relationship between the rent for agricultural land planted with alfalfa crops and density
of cows from 67 counties of Minnesota.
The data correspond to the average rent (Y ) per acre planted with alfalfa in 67 counties of Min-
nesota and also include measurements on each counties for four predictors: average rent paid for all
tillable land (Xi1 ); density of dairy cows in number per square mile (Xi2 ); proportion of farmland used as
pasture (Xi3 ); and the dummy variable Xi4 = 1 if liming is required to grow alfalfa or Xi4 = 0, otherwise.
This data set has been analyzed by Tsai et al. (1998), Taylor and Verbyla (2004), Lin et al. (2009) and
Wu et al. (2012), among others. Tsai et al. (1998) showed that linear residual plots are useful for diag-
nosing non-linearity, whereas squared residual plots are useful for detecting non-constant variance in
the linear regression model with errors normals. Taylor and Verbyla (2004) discussed the joint modelling
of location and scale parameters and a methodology was derived to detect heteroscedasticity, when the
response is t distributed. Lin et al. (2009) considered tests for heteroscedasticity in t linear regression
models. Wu et al. (2012) investigated the simultaneous variable selection on the mean and dispersion of
the lognormal distribution. Previous analyses (see Lin et al., 2009; Taylor and Verbyla, 2004; Tsai et al.,
1998; Wu et al., 2012) suggested that exist an evidence of heteroscedasticity in these data. We will use
as response variable the ratio (Y /X1 ) and only the land where liming is required to grow alfalfa. Accord-
ing to Taylor and Verbyla (2004), the density of dairy cows is measured by area and then it is necessary
a square root transformation for this variable. Thus, for the location model, we consider the structure
√
µi = β0 + β2 xi2 , i = 1 . . . , 67,
where Yi ∼ RBS(µi , δi ). For the precision parameter, we propose the functional relationship
√
log(δi ) = α0 + α2 xi2 , i = 1 . . . , 67. (3.19)
Using the test proposed on Section 4, we observe that the functional relationship given in (3.19) seems
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
45
to be adequate. For this functional relationship, the gradient test have p-value 0.0086, which show that
the hypothesis of varying precision is strongly significant.
We fit the RBS model with varying precision by using the gamlss() function. The values of the
estimates of the model parameters with approximate SEs are shown in Table 3.6. All estimates are
significant at the 5% of significance level.
Table 3.6: Estimates, SE, t-value and p-value for the coefficients of the model.
Coefficients Estimate SE t-value Pr(> |t|)

β0 0.5388 0.0533 10.1100 <0.0001
β2 0.1053 0.0117 9.0280 <0.0001
α0 2.2925 0.8280 2.769 0.0097
α2 0.7751 0.2495 3.107 0.0042
The model assumptions given in are verified using a residual analysis based on Land Rent data.
The normal probability plot with envelope for the rd,i provided in Figure 3.4(a) is used to verify the distri-
butional assumption of the model. This figure does not show unusual features, so that the assumption
that the response variable follows a RBS distribution does not seem to be unsuitable. In addition, the
independence assumption also is verified by the normal probability with envelope and by the plot of
residuals displayed in Figure 3.4(b), from which a outlying observation (#33) is detected.
We also verified whether the identity link function used in modeling of the average is correct or not.
∗
To this link function we have !z ∗1 is given by !z ∗1 = η v ∗ ◦ [yy∗ − µ
! +! v ∗ = [v∗1 , . . . , v∗n ]⊤ with v∗i being
! ], where !
the ith diagonal element of the inverse matrix of V defined in (3.5) and “◦” representing the Hadamard
product. From Figure 3.4(c), note that the identity link function seems to be suitable for modeling the
mean.
2
3
1.2
2
1.0
1
0.8
0
0
rd,i
rd,i
1i
z∗
!
0.6
-1
-1
0.4
-2
-2
0.2
-3
-3
33
-2 -1 0 1 2 0.7 0.8 0.9 1.0 1.1 1.2 0.7 0.8 0.9 1.0 1.1 1.2
Percentiles of N(0,1) fitted values !i
η
(a) (b) (c)
Figure 3.4: Normal probability plot with simulated envelope for rd,i (a), plot index values versus rd,i (b) and plot of
z∗1i against η
! !i for the model fit identity link (c) based on land rent data and the RBS regression model with varying
precision.
6.1 Diagnostic analysis

Influence diagnostics for the RBS regression model with varying precision are presented in Figure
3.5. Figure 3.5(j)-3.5(l) displays index plots of Ci (·), from where observation #26 is detected as potentially
influential. We investigate the impact on the model inference when the cases detected as potentially
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
46
influential in the diagnostic analysis are removed. Then, we again estimate the model after removing the
observation #26. To study the impact of the removed observations define the following relative changes
> θ! −θ! > > SE(θ! ) −SE(θ! ) >

> j > > j (i) j (i) >
RCθ j(i) = > ! j(i) > × 100% and RCSE(θ! ) = > > × 100%,
θj j (i) SE(θ!j )(i)
where θ!j(i) and SE(θ!j )(i) denote the ML estimate of θ j and the SE of θ!j , respectively, obtained after
removing the ith observation, for j = 1 and i = 1, . . . , 67.
The relative changes (RCs) in the parameter estimates are 6.7538 (β!0 ), 10.4338 (β!2 ), 9.1557 (α
!0 )
!2 ). Whereas that the relative changes in the corresponding SEs are 17.5659, 40.0764,
and 8.8063 (α
3.8158 and 8.5395. There no inferential changes are found at level significance of 5%. The results
presented in this table show that the diagnostic measures presented in this article are useful for identify-
ing potentially influential points. In summary, these diagnostic analysis confirm that the RBS regression
model with varying precision is quite suitable for modeling the Land Rent data. We investigate the ob-
servation #26 and conclude that it represents a county in that the average rent per acre planted with
alfalfa, 23% large that the average rent paid for all tillable land, having a density of dairy cows larger
than the other countries. Finally we interpret the adjusted model. We have that the average spending
food can be represented by
√
! (x2 ) = 0.5388 + 0.1053 x2 .
µ
In this model, 0.6441 is the expected ratio if we increase in one unit in square root of the dairy cow
density per square mile.
7 Conclusion
In this chapter, we proposed a RBS regression model with varying precision generalizing the model
proposed in Leiva et al. (2014). We dealt with the issue of performing hypothesis testing on RBS regres-
sion models with varying precision. We considered three classic tests, such as likelihood ratio, score
and Wald tests, and a recently proposed test, the gradient test. Additionally, we presented Monte Carlo
simulations to compare the finite-sample performance of these tests. The simulation results indicated
that the score and gradient tests are preferred. We presented four types of residuals when the mean
and the precision are modeled simultaneously. Again using Monte Carlo simulation, we evaluated the
performance of presented residuals and determined their distributions empirically. We developed diag-
nostic tools for this wider class of BS regression models. We also derived a local influence analysis
by considering different perturbation schemes, such as case-weight, explanatory variable and response
perturbations. We presented a application with real data which made clearer that the graphical diag-
nostic tools we proposed can be a useful tool for assessing model adequacy and also for identifying
influential observations. The plot of the residuals against the index of the observations and the nor-
mal probability plot suggested that there is no clear evidence of model misspecification, i.e., the RBS
regression model with varying precision fitted the data well.
Chapter
___________3___
-_RBS
_______regression
_______________models
___________with
______varying
__________precision
____________________________________________________________________
47
1.0
1.0
1.0
0.8
0.8
0.8
0.6
0.6
0.6
Ci (θ )
Ci (α )
Ci (β )
0.4
0.4
0.4
0.2
0.2
0.2
0.0
0.0
0.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30
Index Index Index
(a) (b) (c)
1.0
1.0
1.0
0.8
0.8
0.8
0.6
0.6
0.6
Ci (θ )
Ci (α )
Ci (β )
0.4
0.4
0.4
0.2
0.2
0.2
0.0
0.0
0.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30
Index Index Index
(d) (e) (f)
1.0
1.0
1.0
0.8
0.8
0.8
26
0.6
0.6
0.6
26 26
Ci (θ )
Ci (α )
Ci (β )
0.4
0.4
0.4
0.2
0.2
0.2
0.0
0.0
0.0
2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7
xi2 xi2 xi2
(g) (h) (i)
1.0
1.0
1.0
0.8
0.8
0.8
0.6
0.6
0.6
Ci (θ )
Ci (α )
Ci (β )
0.4
0.4
0.4
0.2
0.2
0.2
0.0
0.0
0.0
2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7
zi1 zi1 zi1
(j) (k) (l)
Figure 3.5: Index plots of Ci (·) for θ (a), β (b) and α (c), under case-weight perturbation, Ci (·) for θ (d), β (e) and α
(f), under response perturbation, Ci (·) for θ (g), β (h) and α (i), under predictor perturbation (xi2 ), Ci (·) for θ (j), β (k)
and α (l), under predictor perturbation (zi1 ) for the RBS regression model with varying precision.
48
CHAPTER 4
Zero-inflated reparameterized Birnbaum-Saunders distribution
1 Resumo
Neste capítulo, propomos um novo modelo probabilístico baseado na distribuição Birnbaum-Saunders
reparametrizada. Esse novo modelo é útil para descrever dados positivos na presença de zeros. Es-
pecificamente, propomos a distribuição ZIRBS (“Zero-Inflated Reparameterized Birnbaum-Saunders”)
e apresentamos algumas de suas propriedades, considerando o problema de estimação associado.
Realizamos simulações de Monte Carlo para avaliar o comportamento dos estimadores de máxima
verossimilhança na presença de diferentes proporções de zeros. Por fim, aplicamos a metodologia
proposta para dois conjuntos de dados reais de demanda, respondendo as perguntas habituais de um
modelo de gestão de estoque, ou seja, qual é a quantidade de estoque necessária em cada ordem
e, quando tal ordem deve ser colocada. Esta aplicação mostra a importância de usar a metodologia
proposta através da melhoria das margens de contribuição de uma empresa chilena.
2 Introduction and Motivation

Logistics and inventories are two important areas of applied operational research. A scientific man-
agement of supply systems and inventory policies can affect positively the logistics of the companies.
Specifically, it is well-known that the total cost of an inventory involves purchase, ordering and storing
costs. The minimization of this total cost is possible by using an appropriate inventory model, which can
improve the contribution margins and reduce inefficiencies in logistics management of a company.
The inventory management model requires a suitable study of the demand behavior of the assort-
ment of products that are stored in the physical inventory of a company. Because demand for products
is occurring in a random fashion, the corresponding demanded amount is assumed to be a RV. Thus,
if this demand is modeled by an adequate distribution, and a suitable inventory management model
is considered, these two aspects may help to substantially improvement the contribution margins of a
company.
Most common inventory management models are planned by setting an optimum stock level of prod-
Chapter
___________4___
ZIRBS
_________distribution
_____________________________________________________________________________________________________________
49
ucts. In general, the products that form the inventory assortments of the companies can be perishable
or non-perishable (stable). Levels of their stocks are reduced over the time, but they are again supplied
with the arrivals of new units of these products. This is the most usual situation of a storage of products
that face manufacturers, distributors and dealers. For stable products, generally, multiple periods are
considered and the inventory management is based on the economic order quantity (EOQ) model. This
model is used altogether with the reorder point (ROP) to reach a fixed service level and to have safety
stocks, under demand uncertainty and lead time (LT). As it is known, the ROP corresponds to the opti-
mum level of the inventory quantity when a purchase order must be placed. To calculate the EOQ and
ROP, it is necessary to know the demand statistical distribution during the LT. Then, a safety factor is
obtained using some percentile of this distribution to be protected against any unexpected situation of lo-
gistics. Usually, a service level related to the 95th percentile is considered. This inventory management
model for multiple periods is known as “reorder-point, order-quantity policy” or “(r, Q) policy”, in short,
where r is the ROP and Q the order quantity. For more details about this model, see Keaton (1995),
Hillier and Lieberman (2001, pp. 956-961), Strijbosch and Moors (2006), Porras and Dekker (2008), and
Wanke (2008, 2009). For perishable products, generally, a single period is considered and the inventory
management is based on the critical ratio (CR). This ratio is determined by the cost of ordering one unit
less (generating a temporary shortage), in contrast to ordering one unit more (generating a temporary
overstock). This must be considered to establish the optimum quantity of units in stock. For more details
of this single period model, see Hillier and Lieberman (2001, pp. 961-975). In both of these models, the
optimum quantities and the ROP are based on the involved costs and on a suitable study of the demand
behavior for the products stored.
Usually, product demand has been modeled by a normal distribution; see Silver and Peterson
(1985) and Strijbosch and Moors (2006). However, some studies on the topic criticize the use of this
distribution; see Lau (1989), Rojas et al. (2013) and reference therein, for example. This is because
the demanded amount for products in inventory only admits non-negative values, and it follows often an
asymmetric probabilistic model, with positive skewness. Also, if the product is packaged by a volume
unit, so that it can be measured by its volumetric content, the demand is a continuous RV. In any case,
in general, continuous probabilistic models are often used for describing the demand of products. Some
of these models are the gamma (see Burgin, 1975; Das, 1976; Gallego et al, 2007; Snyder, 1984), in-
verse Gaussian (IG, see Tadikamalla, 1981), log-normal (LN, see Cobb et al, 2013; Tadikamalla, 1979),
uniform (see Chen et al, 2002; Wanke, 2008), and Weibull (see Lau, 1989) distributions.
For the purpose of modeling demand for products in inventory management, we proposed the
ZIRBS distribution. Thus, the main objective of this chapter is to introduce a new methodology useful for
describing positive data in presence of zeros based on the ZIRBS distribution. Presence of zero-values
occurs with frequency when daily demand data of a product are analyzed. We propose estimators for
the parameters of the ZIRBS distribution and study their behavior with different proportions of zeros.
This chapter is organized as follows. In Sections 3 and 4, we propose a new methodology useful for
describing positive data in presence of zeros based on the ZIRBS distribution. Specifically, in Section
3, we introduce the ZIRBS demand distribution and present some of its properties, whereas, in Section
4, we consider the estimation problem associated. In Section 5, we conduct MC simulations to evaluate
the behavior of the considered estimators. In Section 6, we apply the proposed methodology to two real
demand data sets, responding the usual questions of an inventory management model, that is, what is
the quantity of stock required in each order and when such an order should be placed. This application
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
50
shows the convenience of using the proposed methodology by improving the contribution margins of a
Chilean company. In the final section, we provide the conclusions of this chapter.
3 The demand statistical model

Let Y be a mixture between two RVs Y1 and Y2 , with continuous PDF and a Dirac mass at zero,
respectively. Thus, if Y1 ∼ RBS(µ , δ ), then the PDF of Y is given by
fY (y; µ , δ , p) = [1 − p] fY1 (y)I(0,∞) (y) + p fY2 (y)I{0} (y), y ≥ 0, (4.1)
where 0 < p < 1 is the mixture parameter, fY1 (·) is the PDF given in (1.2), and IB (·) is the indicator
function of the set B. Explicitly, PDF in (4.1) can be written as
√ # $
exp( δ2 ) δ + 1 δµ
fY (y; µ , δ , p) = [1 − p] √ y+ ×
4 y3/2 π µ δ +1
" # $%
δ {δ + 1}y δµ
exp − + I(0,∞) (y) + pI{0}(y),
4 δµ {δ + 1}y
with, δ > 0, µ > 0 and 0 < p < 1. Alternatively, the PDF given in (4.1), expressed with respect to the
measure generated by the mixture, can be written as
<
p, if y = 0;
fY (y; µ , δ , p) = (4.2)
[1 − p] fY1 (y; µ , δ ), if y > 0.
Thus, from (4.2), the CDF of the ZIRBS distribution is given by

<
p, if y = 0;
FY (y; µ , δ , p) = (4.3)
p + [1 − p] FY1 (y; µ , δ ), if y > 0,
where FY1 (·) is the CDF of the RBS distribution given in (1.4). We have use the following notation:
Y ∼ ZIRBS(µ , δ , p). From (4.3), we have: (P3) if Y ∼ ZIRBS(µ , δ , p), then V = aY ∼ ZIRBS(a µ , δ , p), for
a > 0. Property (P3) is obtained from P(V ≤ v) = P(aY ≤ v) = P(Y ≤ v/a). Hence, for Y = 0, P(V ≤ v) = p
and, for Y > 0, P[V ≤ v] = p + [1 − p] FY1 (v/a; µ , δ ), where FY1 (v/a; µ , δ ) is the RBS CDF of parameters aµ
and δ . The QF of Y ∼ ZIRBS(µ , δ , p) can be obtained by the inverse CDF defined in (4.3) as
<
0, if p ≥ u;
yY (u; µ , δ , p) = FY−1 (u; µ , δ , p) = (4.4)
y([u − p]/[1 − p]; µ , δ ), if p < u,
where 0 < u < 1 and y(·) is the QF of the RBS distribution given in (1.5). Expression provided in (4.4)
can be obtained from Castellacci (2012, Proposition 2.1).
Figure 4.1 presents some shapes of ZIRBS PDFs for different combinations of µ , δ and p. In
Figure 4.1(a), note that, as the parameter µ increases, the scale of the distribution decreases, whereas
the variability increases, and a flattening of the PDF is observed. From Figure 4.1(b), we detect that
the parameter δ controls the shape of the PDF, making it more platykurtic as δ increases. Figure 4.1(c)
shows how the parameter p only controls the scale not altering the shape of the distribution.
A random number generator from Y ∼ ZIRBS(µ , δ , p) is given by Algorithm 2.
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
51
1.0
2.0
1.5
0.8
µ = 0.25
1.5
p = 0.0
µ = 0.50 p = 0.1
µ = 0.75
1.0
p = 0.4
µ = 1.00 δ =2 p = 0.6
0.6
µ = 1.50
fY (y; µ , δ , p)
fY (y; µ , δ , p)
fY (y; µ , δ , p)
δ =5 p = 0.8
µ = 2.00 δ = 10
1.0
δ = 25
δ = 50
0.4
δ = 100
0.5
0.5
0.2
0.0
0.0
0.0
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
y y y
(a) (b) (c)
Figure 4.1: Plots of the ZIRBS PDF for the indicated values of µ (a), δ (b) and p (c).
Algorithm 2 Random number generator for the ZIRBS distribution

1: Generate a random number u from U ∼ U(0, 1);
2: Obtain a random number z from Z ∼ N (0, 1);
3: Set values for µ , δ and p of Y ∼ ZIRBS( µ , δ , p);
4: Compute a random number y = y1 or y = y2 from Y ∼ ZIRBS( µ , δ , p) according to the following criterion:
4.1: If u ≤ p, then y1 = 0;
√ √
4.2: Else, using (1.3), y2 = [δ µ /{δ + 1}][z/ 2 δ + ({z/ 2 δ }2 + 1)1/2 ]2 ; and
5: Repeat Steps 1 to 4 until the required amount of random numbers to be completed.
An alternative way to Algorithm 2 that is computationally more efficient to generate random numbers
from Y ∼ ZIRBS(µ , δ , p), is given by Algorithm 3.
Algorithm 3 Random number generator for the ZIRBS distribution

1: Generate uniform values u from U ∼ U(0, 1).
2: Set values for µ , δ and p of Y ∼ ZIRBS( µ , δ , p);
3: Compute a random number y = y1 or y = y2 from Y ∼ ZIRBS( µ , δ , p) using (4.4), that is,
3.1: If u ≤ p, then y1 = 0;
3.2: Else, y2 = y([u − p]/[1 − p]; µ , δ );
4: Repeat Steps 1 to 3 until the required amount of random numbers to be completed.
The kth moment about zero of Y ∼ ZIRBS(µ , δ , p) is given by
E[Y k ] = [1 − p]µk , (4.5)
where µk is the kth moment about zero of RBS(µ ,δ ), obtained by Corollary 1 and Kv (·) being the modified
Bessel function of second type; see Abramowitz and Stegun (1972). Table 2.1 displays expressions for
the function Kδ /2 (v) for some values of v, which are useful for calculating the moments around zero of
the ZIRBS distribution.
From (4.5), we have the mean and variance of Y ∼ ZIRBS(µ , δ , p) are, respectively, given by
E[Y ] = [1 − p] µ and Var[Y ] = p[1 − p]µ 2 + [1 − p]µ 2[2 δ + 5]/[δ + 1]2 . (4.6)
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
52
4 Estimation of ZIRBS parameters

Let Y = [Y1 , . . . ,Yn ]⊤ be a random sample from Y ∼ ZIRBS(µ , δ , p). Then, the corresponding likelihood
function for θ = [µ , δ , p]⊤ is given by
n n
L(µ , δ , p) = ∏ fY (yi ; µ , δ , p) = pN0 [1 − p]n−N0 ∏ fY1 (yi ; µ , δ )1−I{0} (yi ) , (4.7)
i=1 i=1
where N0 = ∑ni=1 I{0} (yi ) is the number of zeros in the sample. Hence, the respective log-likelihood
function obtained from (4.7) can be expressed as ℓ(µ , δ , p) = ℓ(p) + ℓ(µ , δ ), where
ℓ(p) =N0 log(p) + [n − N0] log(1 − p), (4.8)

3 [δ + 1]
2 y∑ 4µ y∑
ℓ(µ , δ ) =[n − N0]c(µ , δ ) − log(yi ) − yi
i >0 i >0
" %
µδ 2 µδ
−∑ + ∑ log yi + ,
yi >0 4[δ + 1]yi yi >0 [δ + 1]
with c(µ , δ ) = −[1/2] log(16π ) + [δ /2] − [1/2] log(µ ) + [1/2] log(δ + 1). As can be observed in (4.8), the
log-likelihood function factorizes in a term that depends only on p and another that depends on δ and µ .
Consequently, the ML method can be performed separately; see Pace and Salvan (1997, p. 128). We
obtain the corresponding score function by taking derivatives of the log-likelihood function given in (4.8)
with respect to the unknown parameters as
[n − N0] [δ + 1] δ2 1 δ 1
Uµ (θ ) = − + 2 ∑ y i − ∑ + ∑ 4 5, (4.9)
2µ 4µ yi >0 4[δ + 1] yi >0 yi [δ + 1] yi >0 yi + µδ
δ +1
[n − N0][δ + 2] 1 µδ [δ + 2] 1 µ 1
∑ yi − 4 5,
4[δ + 1]2 y∑ 2 ∑
Uδ (θ ) = − + (4.10)
2[δ + 1] 4µ yi >0 i >0
y i [ δ + 1] yi >0 yi +
µδ
{δ +1}
N0 n − N0
U p (θ ) = − . (4.11)
p 1− p
From (4.11), the ML estimate of p is p! = N0 /n, which represents the proportion of zeros in the
sample. As the solution of the equations Uµ (θ ) = 0 and Uδ (θ ) = 0 obtained from (4.9) and (4.10) does
not have a closed-form, then we maximize the log-likelihood function given in (4.7) on µ and δ by using a
non-linear optimization algorithm for determining the ML estimates of µ and δ . A non-linear optimization
algorithm called CG can be used in this case; see Cole and Green (1992). This algorithm works well for
distributions with correlated parameter estimates, as is the case of the ZIRBS distribution.
To obtain starting values for the numerical optimization algorithm, consider the following. Note that,
if we have only the continuous component of the distribution of Y , that is, the RBS distribution, E[Y1 ] = µ
and Var[Y1 ] = µ 2 [2δ + 5]/[δ + 1]2 . Then, the moment estimates of µ and δ are, respectively, given by
7
µ̆ = ȳ∗ and δ̆ = [ȳ2∗ − s2∗ + ȳ4∗ + 3 ȳ2∗ s2∗ ]/s2∗ , (4.12)
where ȳ∗ = [1/{n − N0}] ∑yi >0 yi and s2∗ = [1/{n − N0}] ∑yi >0 {yi − ȳ∗ }2 . In addition, we have that δ̆ is well-
√
defined when the sample coefficient of variation is less that 5; see Santos-Neto et al. (2013a). Moment
estimates given in (4.12) can be used as starting values for the CG algorithm.
Now, we obtain the respective expected Fisher information matrix by taking derivatives of the ele-
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
53
ments of the score vector given in (4.9)-(4.11) with respect to the unknown parameters as
⎛ ⎞
# $ iµ µ iµδ 0
∂ ℓ(θ ) ⎜ ⎟
i (θ ) = −E =⎜
⎝ iδ µ iδ δ 0⎟ ⎠, (4.13)
∂ θi ∂ θ j
0 0 i pp
where θ1 = µ , θ2 = δ , θ3 = p, and
# $
δ δ2
iµ µ = n[1 − p] + I(µ , δ ) ,
2µ 2 {δ + 1}2
# $
1 µδ
iµδ = iδ µ = n[1 − p] + I(µ , δ ) ,
2µ {δ + 1} {δ + 1}3
# 2 $
[δ + 3δ + 1] µ2
iδ δ = n[1 − p] + I(µ , δ ) ,
2δ 2 {δ + 1}2 {δ + 1}4
n
i pp = ,
p[1 − p]
with I(µ , δ ) being given by

B ∞
1
I(µ , δ ) = fY1 (y; µ , δ ) dy,
0 [y + {µδ }/{δ + 1}]2
and fY1 (·) being the PDF of Y1 ∼ RBS(µ , δ ). From (4.13), note that the parameter p is orthogonal to the
parameters µ and δ . Under some regularity conditions (see Cox and Hinkley, 1974), θ! is a consistent
estimator of θ and it has a distribution that is asymptotically normal. Then,
√ ! D
n [θ − θ ] → N3 (0, j (θ )−1 ), as n → ∞, (4.14)
where j (θ ) = limn→∞ 1n i (θ ), with i (θ ) being the inverse expected Fisher information matrix given in (4.13).
Note that i (θ! ) is a consistent estimator of the asymptotic variance-covariance matrix of θ! . Asymptotic
CIs for µ , δ and p, with level of 100 × [1 − ξ ]%, can be obtained from (4.14) as
T U V U VW
L θ!1 ) , exp θ!1 + z(1 − ξ /2) SE(
CI(θ ; [1 − ξ ] × 100%) = exp θ!1 − z(1 − ξ /2) SE( L θ!1 ) , (4.15)
T S U V@W
L p!1 ) ,
CI(p; [1 − ξ ] × 100%) = 1/ 1 + exp − p!1 ± z(1 − ξ /2) SE( (4.16)
L θ!) is the estimated asymptotic SE of the ML estimator of

where θ! is the ML estimate of θ = µ or δ , SE(
θ . We also define θ1 = log(µ ) or log(δ ) and p1 = logit(p), with the corresponding SEs of these estimators
being obtained by using the delta method. Confidence bounds (CBs) can also be obtained in a similar
way to the CIs given in (4.15) and (4.16).
5 Simulation study
We use MC simulations to evaluate the behavior of the ML estimators of the parameters of the
ZIRBS distribution. Algorithm 3 proposed in Section 3 is employed to generate pseudo-random numbers
of a ZIRBS RV.
The scenario of simulation considers a sample size n ∈ {25, 50, 100}, the shape parameter δ ∈
{0.5, 2.0, 8.0, 22, 200} and the proportion of zeros p ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6}. The parameter µ is fixed
at 1.0, without loss of any generality, because µ is, besides the mean, a scale parameter. All the results
are based on B = 10000 MC experiments.
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
54
The simulations are performed using the R software. The ML estimates of the parameters of the
ZIRBS distribution are obtained using the gamlss() function of an R package of the same name, which
uses the CG algorithm; see Stasinopoulos and Rigby (2007) and www.gamlss.org. We have incorpo-
rated the ZIRBS family in this package. Rcodes used in this study are available under request from the
author.
To analyze the point estimation results, we compute, for each sample size and for each parameter,
√
the empirical mean, relative bias (RB), SE and MSE of the corresponding estimators. Concerning to
interval estimation, we present the empirical CPs obtained from the relative frequencies, at which the
true parameter value belongs to the CI. The CIs for µ , δ and p, given in (4.15) and (4.16), aim to ensure
that the estimates of the limits are within of the parameter space that index the ZIRBS distribution.
Furthermore, we compute the percentages of the 10000 intervals in which the true parameter value is
smaller than the lower limit of the intervals (in short %L) and greater than its upper limit (in short %U).
We consider confidence levels of 90%, 95% and 99%.
√
! , δ! and p!. From Table 4.1,
Tables 4.1 and 4.2 present the empirical RB and MSE of the estimators µ
! decreases as δ increases. Observe that, for a large sample, for example,
note that the empirical RB of µ
n = 100, the empirical RB of µ! does not seem to be affected by an increase of the proportion p of zeros
in the sample. Also, notice that the empirical RB of δ! increases as p increases, but it decreases when n
increases. In addition, note that the empirical RB of p! is practically not affected by changes in δ . From
√
Table 4.2, observe that, as n increases, the empirical MSE decreases, for all the estimators. Also, note
√ √
that the empirical MSE of p! is practically not affected by changes in δ . For n fixed, the empirical MSE
√
! decreases as δ increases, whereas the empirical MSE of µ
of µ ! increases in this case.
Tables 4.3, 4.4 and 4.5 display empirical CPs of 90%, 95% and 99% CIs for µ , δ and p, respectively.
In addition, we calculate empirical CPs at the tails for determining the balancing of the proposed CIs.
From these tables, observe that CIs for µ and p have a better balancing than for the case of δ , which
! and p!, shown in the empirical study
can be justified by means of the symmetry of the distributions of µ
displayed in Figures 4.2 and 4.3. CIs for µ and δ present a CP below fixed confidence level, whereas the
CIs for p present a CP lightly greater than the fixed level. Note also that an increase in δ does not affect
the behavior of the CIs for p. A similar behavior can be detected in the CIs for µ and δ . Furthermore,
notice that the CIs for µ turn to be more balanced as δ increases. This behavior is justified by the fact
! turn to be more symmetrical; see Figures 4.2
that an increase on δ does the empirical distribution of µ
and 4.3. Thus, we conclude that the worst results for the CPs are obtained for large values of p, that is,
when we have a large number of zeros in the sample. Hence, as the sample size increases, we detect
that the influence of the proportion of zeros in the sample tends to decrease.
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
55
Table 4.1: Empirical RB of the indicated parameter estimator, δ , p and n, based on MC simulations with µ = 1.0.
!
µ δ! p!
δ p n = 25 50 100 25 50 100 25 50 100
0.5 0.1 −0.0681 −0.0444 −0.0228 0.2788 0.1416 0.0670 −0.0140 −0.0090 −0.0070
0.2 −0.0738 −0.0474 −0.0267 0.3330 0.1556 0.0778 −0.0035 −0.0030 −0.0035
0.3 −0.0776 −0.0561 −0.0268 0.3820 0.1816 0.0898 0.0017 0.0007 −0.0013
0.4 −0.0893 −0.0607 −0.0334 0.4546 0.2130 0.1054 −0.0030 −0.0025 0.0020
0.5 −0.1000 −0.0663 −0.0415 0.5830 0.2566 0.1300 −0.0012 0.0010 −0.0014
0.6 −0.0953 −0.0746 −0.0461 0.8410 0.3284 0.1598 0.0030 0.0010 0.0002
2.0 0.1 0.0016 0.0016 0.0010 0.1648 0.0730 0.0368 −0.0130 −0.0080 −0.0020
0.2 0.0046 −0.0004 0.0009 0.1906 0.0838 0.0403 −0.0010 0.0000 −0.0020
0.3 0.0037 −0.0001 0.0008 0.2221 0.1001 0.0483 −0.0050 0.0007 −0.0020
0.4 0.0043 −0.0003 0.0006 0.2662 0.1198 0.0540 −0.0052 0.0007 −0.0017
0.5 0.0055 −0.0007 0.0008 0.3486 0.1451 0.0659 −0.0024 0.0004 −0.0010
0.6 0.0077 0.0003 0.0010 0.5190 0.1952 0.0819 −0.0005 0.0010 −0.0007
8.0 0.1 0.0004 0.0006 0.0005 0.1587 0.0703 0.0357 −0.0130 −0.0080 −0.0020
0.2 0.0020 −0.0005 0.0003 0.1834 0.0808 0.0390 −0.0010 0.0000 −0.0020
0.3 0.0012 −0.0002 0.0002 0.2135 0.0968 0.0468 −0.0050 0.0007 −0.0020
0.4 0.0014 −0.0003 0.0001 0.2557 0.1158 0.0523 −0.0052 0.0007 −0.0017
0.5 0.0019 −0.0007 0.0001 0.3341 0.1398 0.0636 −0.0024 0.0004 −0.0010
0.6 0.0034 0.0000 0.0002 0.4952 0.1880 0.0790 −0.0005 0.0010 −0.0007
22 0.1 0.0002 0.0003 0.0003 0.1571 0.0696 0.0354 −0.0130 −0.0080 −0.0020
0.2 0.0011 −0.0003 0.0002 0.1815 0.0800 0.0387 −0.0010 0.0000 −0.0020
0.3 0.0005 −0.0001 0.0001 0.2113 0.0958 0.0464 −0.0050 0.0007 −0.0020
0.4 0.0006 −0.0002 0.0000 0.2530 0.1147 0.0518 −0.0052 0.0007 −0.0017
0.5 0.0009 −0.0005 0.0000 0.3305 0.1384 0.0630 −0.0024 0.0004 −0.0010
0.6 0.0018 −0.0001 0.0001 0.4896 0.1861 0.0782 −0.0005 0.0010 −0.0007
200 0.1 0.0000 0.0001 0.0001 0.1560 0.0692 0.0352 −0.0130 −0.0080 −0.0020
0.2 0.0004 −0.0001 0.0000 0.1802 0.0795 0.0384 −0.0010 0.0000 −0.0020
0.3 0.0001 −0.0001 0.0000 0.2098 0.0952 0.0462 −0.0050 0.0007 −0.0020
0.4 0.0001 −0.0001 0.0000 0.2512 0.1140 0.0515 −0.0052 0.0007 −0.0017
0.5 0.0002 −0.0002 0.0000 0.3281 0.1374 0.0626 −0.0024 0.0004 −0.0010
0.6 0.0005 0.0000 0.0000 0.4859 0.1849 0.0777 −0.0005 0.0010 −0.0007
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
56
√
Table 4.2: Empirical MSE of the indicated parameter estimator, δ , p and n, based on MC simulations with µ = 1.0.
!
µ δ! p!
δ p n = 25 50 100 25 50 100 25 50 100
0.5 0.1 0.3135 0.2236 0.1615 0.2349 0.1279 0.0784 0.0603 0.0417 0.0295
0.2 0.3385 0.2384 0.1705 0.2774 0.1408 0.0861 0.0802 0.0573 0.0398
0.3 0.3671 0.2548 0.1822 0.3216 0.1585 0.0953 0.0939 0.0654 0.0464
0.4 0.3864 0.2728 0.1957 0.3904 0.1844 0.1057 0.0985 0.0690 0.0501
0.5 0.4283 0.3012 0.2151 0.5119 0.2157 0.1233 0.1001 0.0710 0.0501
0.6 0.4896 0.3377 0.2412 1.0316 0.2756 0.1447 0.0978 0.0689 0.0493
2.0 0.1 0.2119 0.1509 0.1059 0.8745 0.4991 0.3283 0.0599 0.0427 0.0299
0.2 0.2282 0.1600 0.1126 0.9612 0.5407 0.3501 0.0801 0.0570 0.0402
0.3 0.2453 0.1703 0.1200 1.1141 0.5958 0.3870 0.0917 0.0657 0.0466
0.4 0.2645 0.1832 0.1301 1.2962 0.6877 0.4208 0.0983 0.0697 0.0496
0.5 0.2917 0.2013 0.1436 1.6668 0.7943 0.4747 0.0997 0.0707 0.0502
0.6 0.3294 0.2266 0.1610 2.7887 0.9890 0.5422 0.0979 0.0695 0.0499
8.0 0.1 0.1078 0.0768 0.0540 3.4561 1.9824 1.3084 0.0599 0.0427 0.0299
0.2 0.1159 0.0818 0.0574 3.7886 2.1462 1.3949 0.0801 0.0570 0.0402
0.3 0.1243 0.0868 0.0611 4.3833 2.3637 1.5412 0.0917 0.0657 0.0466
0.4 0.1343 0.0933 0.0662 5.0912 2.7266 1.6752 0.0983 0.0697 0.0496
0.5 0.1484 0.1030 0.0733 6.5154 3.1410 1.8861 0.0997 0.0707 0.0502
0.6 0.1669 0.1156 0.0821 10.8153 3.9019 2.1525 0.0979 0.0690 0.0499
22 0.1 0.0641 0.0457 0.0321 9.5715 5.4979 3.6322 0.0599 0.0427 0.0299
0.2 0.0688 0.0487 0.0342 10.4854 5.9507 3.8718 0.0801 0.0570 0.0402
0.3 0.0737 0.0516 0.0363 12.1265 6.5528 4.2772 0.0917 0.0657 0.0466
0.4 0.0797 0.0555 0.0394 14.0798 7.5576 4.6487 0.0983 0.0697 0.0496
0.5 0.0883 0.0614 0.0436 18.0000 8.7003 5.2312 0.0997 0.0707 0.0502
0.6 0.0991 0.0689 0.0488 29.8370 10.8013 5.9685 0.0979 0.0695 0.0499
200 0.1 0.0212 0.0151 0.0106 86.0233 49.4463 32.6806 0.0599 0.0427 0.0299
0.2 0.0227 0.0161 0.0113 94.2050 53.5127 34.8356 0.0801 0.0570 0.0402
0.3 0.0243 0.0171 0.0120 108.9280 58.9226 38.4801 0.0917 0.0657 0.0466
0.4 0.0263 0.0183 0.0130 126.4506 67.9535 41.8212 0.0983 0.0697 0.0496
0.5 0.0292 0.0204 0.0144 161.5761 78.2001 47.0489 0.0997 0.0707 0.0502
0.6 0.0327 0.0228 0.0161 267.6505 97.0553 53.6745 0.0979 0.0695 0.0499
Table 4.3: Empirical CP of 90% CIs based on MC simulation for the indicated δ , p and n with µ = 1.0.
CP of CI(µ ;90%) CP of CI(δ ;90%) CP of CI(p;90%)

Chapter
n = 25 50 100 25 50 100 25 50 100

___________4___
δ p %L CP %U %L CP %U %L CP %U %L CP %U %L CP %U %L CP %U %L CP %U %L CP %U %L CP %U
0.5 0.1 1.31 85.85 12.84 1.67 88.02 10.31 2.08 89.72 8.20 14.53 85.47 0.00 10.81 89.19 0.00 8.46 91.54 0.00 3.33 96.67 0.00 5.09 91.94 2.97 6.55 91.09 2.36
ZIRBS
0.2 1.38 84.73 13.89 1.64 87.96 10.40 2.13 89.61 8.26 15.01 84.99 0.00 10.77 89.23 0.00 9.33 90.67 0.00 4.48 93.17 2.35 6.09 88.64 5.27 5.12 90.04 4.84
0.3 1.74 83.31 14.95 1.37 87.03 11.60 1.95 89.39 8.66 16.34 83.66 0.00 11.66 88.34 0.00 9.45 90.55 0.00 5.10 91.22 3.68 5.43 90.30 4.27 5.58 89.13 5.29
0.4 1.29 82.96 15.75 1.35 86.77 11.88 1.80 89.22 8.98 17.52 82.48 0.00 12.69 87.31 0.00 9.71 90.29 0.00 3.39 93.34 3.27 5.64 88.72 5.64 4.65 91.19 4.16
0.5 1.57 81.18 17.25 1.41 86.29 12.30 1.70 88.31 9.99 20.10 79.90 0.00 14.02 85.98 0.00 11.26 88.74 0.00 5.27 89.37 5.36 6.07 87.83 6.10 4.09 91.31 4.60
0.6 1.77 79.59 18.64 1.43 84.82 13.75 2.00 87.47 10.53 22.05 77.95 0.00 15.66 84.34 0.00 11.38 88.62 0.00 2.97 93.51 3.52 5.68 88.54 5.78 4.12 91.67 4.21
2.0 0.1 3.31 87.09 9.60 3.53 88.45 8.02 3.94 89.28 6.78 11.08 86.82 2.10 8.53 88.66 2.81 7.52 88.81 3.67 2.97 97.03 0.00 5.35 91.36 3.29 7.11 90.40 2.49
0.2 3.33 87.05 9.62 3.44 88.28 8.28 3.95 89.01 7.04 10.91 87.13 1.96 8.63 88.66 2.71 7.65 89.24 3.11 4.58 93.07 2.35 6.15 88.76 5.09 5.61 89.62 4.77
0.3 3.53 86.27 10.20 3.16 88.49 8.35 3.64 89.27 7.09 11.98 86.19 1.83 9.10 88.46 2.44 7.97 88.72 3.31 4.40 92.12 3.48 5.14 90.54 4.32 5.46 89.42 5.12
0.4 3.45 85.65 10.90 3.29 88.10 8.61 3.72 89.20 7.08 12.67 85.73 1.60 9.98 87.83 2.19 8.13 88.69 3.18 3.49 93.30 3.21 6.12 88.56 5.32 4.29 91.39 4.32
0.5 3.26 84.97 11.77 3.00 88.13 8.87 3.70 88.86 7.44 13.97 84.55 1.48 10.57 87.18 2.25 8.69 88.34 2.97 5.11 89.31 5.58 5.97 88.00 6.03 4.27 90.90 4.83
0.6 3.47 83.68 12.85 3.17 87.29 9.54 3.41 88.41 8.18 15.77 82.86 1.37 11.99 86.15 1.86 8.45 89.00 2.55 2.75 93.78 3.47 5.78 88.27 5.95 3.91 91.31 4.78
8.0 0.1 4.42 87.61 7.97 4.28 88.57 7.15 4.44 89.15 6.41 10.79 87.06 2.15 8.28 88.83 2.89 7.33 88.96 3.71 2.97 97.03 0.00 5.35 91.36 3.29 7.11 90.40 2.49
0.2 4.62 86.96 8.42 4.41 88.28 7.31 4.56 88.97 6.47 10.65 87.33 2.02 8.49 88.74 2.77 7.50 89.39 3.11 4.58 93.07 2.35 6.15 88.76 5.09 5.61 89.62 4.77
0.3 4.68 86.45 8.87 4.11 88.41 7.48 4.51 89.12 6.37 11.55 86.52 1.93 8.91 88.60 2.49 7.83 88.77 3.40 4.40 92.12 3.48 5.14 90.54 4.32 5.46 89.42 5.12
0.4 4.85 85.92 9.23 4.17 88.48 7.35 4.43 88.90 6.67 12.32 86.04 1.64 9.78 87.96 2.26 7.99 88.80 3.21 3.49 93.30 3.21 6.12 88.56 5.32 4.29 91.39 4.32
0.5 4.94 85.32 9.74 4.19 87.99 7.82 4.66 88.60 6.74 13.52 84.89 1.59 10.23 87.43 2.34 8.49 88.50 3.01 5.11 89.31 5.58 5.97 88.00 6.03 4.27 90.90 4.83
0.6 5.15 84.13 10.72 4.39 87.33 8.28 4.55 88.63 6.82 15.22 83.39 1.39 11.62 86.41 1.97 8.22 89.17 2.61 2.75 93.78 3.47 5.78 88.27 5.95 3.91 91.31 4.78
22 0.1 5.11 87.44 7.45 4.63 88.81 6.56 4.83 89.13 6.04 14.59 81.79 3.62 8.58 88.30 3.12 7.31 88.95 3.74 2.97 97.03 0.00 5.35 91.36 3.29 7.11 90.40 2.49
0.2 5.34 87.12 7.54 5.10 87.88 7.02 4.88 89.08 6.04 10.80 87.08 2.12 8.45 88.79 2.76 7.49 89.34 3.17 4.58 93.07 2.35 6.15 88.76 5.09 5.61 89.62 4.77
0.3 5.42 86.51 8.07 4.62 88.57 6.81 4.89 89.08 6.03 11.38 86.66 1.96 8.83 88.66 2.51 7.81 88.76 3.43 4.40 92.12 3.48 5.14 90.54 4.32 5.46 89.42 5.12
0.4 5.63 86.08 8.29 4.63 88.63 6.74 5.02 88.69 6.29 12.20 86.10 1.70 9.74 88.00 2.26 7.99 88.76 3.25 3.49 93.30 3.21 6.12 88.56 5.32 4.29 91.39 4.32
0.5 5.81 85.52 8.67 5.07 87.65 7.28 5.07 88.59 6.34 13.37 85.02 1.61 10.12 87.52 2.36 8.44 88.49 3.07 5.11 89.31 4.41 5.97 88.00 6.03 4.27 90.90 4.83
0.6 6.19 84.12 9.69 5.16 87.42 7.42 4.96 88.85 6.19 15.06 83.52 1.42 11.50 86.51 1.99 8.16 89.21 2.63 2.75 93.78 3.47 5.78 88.27 5.95 3.91 91.31 4.78
200 0.1 5.78 87.52 6.70 5.41 88.45 6.14 5.23 89.04 5.73 14.60 80.42 4.98 8.54 88.32 3.14 7.28 88.99 3.73 2.97 97.03 0.00 5.35 91.36 3.29 7.11 90.40 2.49
0.2 6.26 86.84 6.90 5.66 87.94 6.40 5.45 88.95 5.60 10.77 87.04 2.19 8.46 88.76 2.78 7.47 89.34 3.19 4.58 93.07 2.35 6.15 88.76 5.09 5.61 89.62 4.77
0.3 6.07 86.76 7.17 5.37 88.57 6.06 5.30 89.07 5.63 11.35 86.67 1.98 8.80 88.68 2.52 7.79 88.76 3.45 4.40 92.12 3.48 5.14 90.54 4.32 5.46 89.42 5.12
0.4 6.36 86.34 7.30 5.35 88.56 6.09 5.56 88.65 5.79 12.22 86.07 1.71 9.74 88.01 2.25 7.99 88.71 3.30 3.49 93.30 3.21 6.12 88.56 5.32 4.29 91.39 4.32
0.5 6.68 85.47 7.85 5.72 87.69 6.59 5.45 88.75 5.80 13.91 84.47 1.62 10.09 87.54 2.37 8.42 88.49 3.09 5.11 89.31 5.58 5.97 88.00 6.03 4.27 90.90 4.83
0.6 7.41 84.35 8.24 6.09 87.25 6.66 5.38 88.81 5.81 17.73 80.83 1.44 11.43 86.58 1.99 8.15 89.23 2.62 2.75 93.78 3.47 5.78 88.27 5.95 3.91 91.31 4.78
_____________________________________________________________________________________________________________
57
Table 4.4: Empirical CP of 95% CIs based on MC simulations for the indicated δ , p and n with µ = 1.0.

Chapter
n = 25 50 100 25 50 100 25 50 100

___________4___
0.5 0.1 0.42 90.87 8.71 0.61 92.98 6.41 0.94 94.23 4.83 8.88 91.12 0.00 5.99 94.01 0.00 4.58 95.42 0.00 3.33 96.67 0.00 1.94 98.06 0.00 3.62 95.66 0.72
ZIRBS
0.2 0.49 89.52 9.99 0.59 92.92 6.49 0.92 94.09 4.99 9.92 90.08 0.00 6.17 93.83 0.00 4.94 95.06 0.00 1.76 98.24 0.00 3.29 94.75 1.96 3.01 94.41 2.58
0.3 0.63 88.73 10.64 0.48 92.49 7.03 0.77 94.11 5.12 10.64 89.36 0.00 7.13 92.87 0.00 5.39 94.61 0.00 2.12 96.87 1.01 2.73 95.17 2.10 2.39 96.05 1.56
0.4 0.59 88.00 11.41 0.59 91.58 7.83 0.66 94.14 5.20 11.44 88.56 0.00 7.99 92.01 0.00 5.78 94.22 0.00 3.39 93.34 3.27 3.01 94.00 2.99 2.96 94.42 2.62
0.5 0.59 87.03 12.38 0.55 91.45 8.00 0.83 93.24 5.93 13.65 86.35 0.00 8.82 91.18 0.00 6.52 93.48 0.00 1.98 95.68 2.34 1.66 96.68 1.66 2.64 94.39 2.97
0.6 0.82 85.28 13.90 0.58 89.96 9.46 0.73 92.46 6.81 15.44 84.56 0.00 10.06 89.94 0.00 6.80 93.20 0.00 2.97 93.51 3.52 2.97 93.84 3.19 2.62 94.56 2.82
2.0 0.1 1.46 92.50 6.04 1.45 93.74 4.81 1.75 94.30 3.95 6.95 92.32 0.73 4.75 94.10 1.15 4.32 94.09 1.59 2.97 97.03 0.00 2.18 97.82 0.00 4.04 95.10 0.86
0.2 1.30 92.20 6.50 1.31 93.61 5.08 1.76 94.24 4.00 6.97 92.24 0.79 5.13 93.88 0.99 4.26 94.39 1.35 1.78 98.22 0.00 3.32 94.74 1.94 3.50 94.07 2.43
0.3 1.46 91.62 6.92 1.33 93.50 5.17 1.66 94.16 4.18 7.53 91.88 0.59 5.35 93.70 0.95 4.50 94.09 1.41 1.83 97.20 0.97 2.75 95.25 2.00 2.29 95.83 1.88
0.4 1.29 91.40 7.31 1.31 93.31 5.38 1.54 94.12 4.34 8.28 91.14 0.58 6.00 93.15 0.85 4.85 93.76 1.39 3.49 93.30 3.21 3.35 93.82 2.83 2.75 94.47 2.78
0.5 1.31 90.68 8.01 1.25 93.03 5.72 1.58 93.83 4.59 9.40 90.13 0.47 6.39 92.85 0.76 4.96 93.74 1.30 1.85 95.74 2.41 1.62 96.89 1.49 2.74 94.16 3.10
0.6 1.62 88.93 9.45 1.32 92.57 6.11 1.46 93.72 4.82 10.65 88.93 0.42 7.29 91.91 0.80 4.90 94.18 0.92 2.75 93.78 3.47 3.08 93.75 3.17 2.42 94.46 3.12
8.0 0.1 2.31 92.74 4.95 2.12 93.68 4.20 2.25 94.34 3.41 6.77 92.50 0.73 4.67 94.16 1.17 4.26 94.09 1.65 2.97 97.03 0.00 2.18 97.82 0.00 4.04 95.10 0.86
0.2 2.15 92.57 5.28 2.01 93.71 4.28 2.39 94.20 3.41 6.74 92.43 0.83 4.96 94.01 1.03 4.23 94.41 1.36 1.78 98.22 0.00 3.32 94.74 1.94 3.50 94.07 2.43
0.3 2.30 91.98 5.72 1.84 93.81 4.35 2.10 94.23 3.67 7.35 92.05 0.60 5.21 93.71 1.08 4.46 94.07 1.47 1.83 97.20 0.97 2.75 95.25 2.00 2.29 95.83 1.88
0.4 2.46 91.48 6.06 1.98 93.52 4.50 2.16 94.03 3.81 7.92 91.49 0.59 5.84 93.25 0.91 4.76 93.82 1.42 3.49 93.30 3.21 3.35 93.82 2.83 2.75 94.47 2.78
0.5 2.58 90.82 6.60 2.10 92.90 5.00 2.08 94.12 3.80 8.97 90.54 0.49 6.17 93.03 0.80 4.80 93.88 1.32 1.85 95.74 2.41 1.62 96.89 1.49 2.74 94.16 3.10
0.6 2.82 89.47 7.71 2.27 92.71 5.02 1.95 93.85 4.20 10.24 89.32 0.44 7.09 92.10 0.81 4.81 94.28 0.91 2.75 93.78 3.47 3.08 93.75 3.17 2.42 94.46 3.12
22 0.1 2.78 92.87 4.35 2.46 93.74 3.80 2.43 94.42 3.15 10.95 86.86 2.19 4.99 93.60 1.41 4.23 94.12 1.65 2.97 97.03 0.00 2.18 97.82 0.00 4.04 95.1 0.86
0.2 2.70 92.60 4.70 2.51 93.60 3.89 2.56 94.15 3.29 6.89 92.17 0.94 4.95 93.99 1.06 4.17 94.47 1.36 1.78 98.22 0.00 3.32 94.74 1.94 3.50 94.07 2.43
0.3 2.75 92.32 4.93 2.24 93.83 3.93 2.44 94.20 3.36 7.27 92.11 0.62 5.20 93.73 1.07 4.43 94.08 1.49 1.83 97.20 0.97 2.75 95.25 2.00 2.29 95.83 1.88
0.4 3.11 91.69 5.20 2.42 93.64 3.94 2.47 94.06 3.47 7.87 91.52 0.61 5.76 93.34 0.90 4.75 93.82 1.43 3.49 93.30 3.21 3.35 93.82 2.83 2.75 94.47 2.78
0.5 3.27 90.98 5.75 2.58 92.91 4.51 2.59 93.76 3.65 8.92 90.57 0.51 6.11 93.10 0.79 4.74 93.93 1.33 1.85 95.74 2.41 1.62 96.89 1.49 2.74 94.16 3.10
0.6 3.65 89.91 6.44 2.88 92.70 4.42 2.47 93.86 3.67 10.15 89.41 0.44 7.00 92.18 0.82 4.80 94.29 0.91 2.75 93.78 3.47 3.08 93.75 3.17 2.42 94.46 3.12
200 0.1 3.32 92.98 3.70 2.84 93.83 3.33 2.82 94.31 2.87 10.93 85.40 3.67 4.99 93.58 1.43 4.22 94.11 1.67 2.97 97.03 0.00 2.18 97.82 0.00 4.04 95.10 0.86
0.2 3.33 92.57 4.10 2.97 93.59 3.44 2.87 94.18 2.95 6.89 92.12 0.99 4.93 94.01 1.06 4.17 94.48 1.35 1.78 98.22 0.00 3.32 94.74 1.94 3.50 94.07 2.43
0.3 3.47 92.30 4.23 2.76 93.74 3.50 2.79 94.24 2.97 7.49 91.89 0.62 5.19 93.73 1.08 4.42 94.09 1.49 1.83 97.20 0.97 2.75 95.25 2.00 2.29 95.83 1.88
0.4 3.67 92.03 4.30 2.92 93.81 3.27 2.78 93.97 3.25 9.06 90.34 0.60 5.72 93.38 0.90 4.74 93.82 1.44 3.49 93.30 3.21 3.35 93.82 2.83 2.75 94.47 2.78
0.5 4.05 91.31 4.64 3.37 92.87 3.76 2.96 93.68 3.36 11.90 87.58 0.52 6.12 93.07 0.81 4.72 93.96 1.32 1.85 95.74 2.41 1.62 96.89 1.49 2.74 94.16 3.10
0.6 4.57 89.95 5.48 3.50 92.71 3.79 2.59 94.04 3.37 17.00 82.56 0.44 7.12 92.05 0.83 4.77 94.31 0.92 2.75 93.78 3.47 3.08 93.75 3.17 2.42 94.46 3.12
_____________________________________________________________________________________________________________
58
Table 4.5: Empirical CP of 99% CIs based on MC simulations for the indicated δ , p and n with µ = 1.0.

Chapter
n = 25 50 100 25 50 100 25 50 100

___________4___
0.5 0.1 0.03 96.45 3.52 0.01 97.84 2.15 0.08 98.46 1.46 3.18 96.82 0.00 1.66 98.34 0.00 1.18 98.82 0.00 1.05 98.95 0.00 0.71 99.29 0.00 0.79 99.21 0.00
ZIRBS
0.2 0.03 95.67 4.30 0.04 97.76 2.20 0.07 98.39 1.54 3.96 96.04 0.00 2.11 97.89 0.00 1.43 98.57 0.00 0.46 99.54 0.00 0.76 99.24 0.00 0.68 99.11 0.21
0.3 0.05 94.71 5.24 0.02 97.16 2.82 0.04 98.42 1.54 4.25 95.75 0.00 2.26 97.74 0.00 1.57 98.43 0.00 0.94 99.06 0.00 0.65 99.01 0.34 0.80 99.01 0.19
0.4 0.01 94.09 5.90 0.02 97.00 2.98 0.06 98.16 1.78 4.89 95.11 0.00 2.57 97.43 0.00 1.75 98.25 0.00 0.34 99.50 0.16 0.21 99.55 0.24 0.63 98.88 0.49
0.5 0.06 92.96 6.98 0.08 96.42 3.50 0.05 97.73 2.22 5.80 94.20 0.00 3.04 96.96 0.00 2.07 97.93 0.00 0.20 99.59 0.21 0.34 99.35 0.31 0.38 99.23 0.39
0.6 0.22 91.96 7.82 0.04 95.96 4.00 0.08 97.32 2.60 7.09 92.91 0.00 3.95 96.05 0.00 2.32 97.68 0.00 0.27 99.32 0.41 0.21 99.59 0.20 0.49 98.85 0.66
2.0 0.1 0.14 97.38 2.48 0.23 98.08 1.69 0.26 98.50 1.24 2.47 97.44 0.09 1.48 98.42 0.10 1.17 98.64 0.19 0.92 99.08 0.00 1.02 98.98 0.00 1.00 99.00 0.00
0.2 0.19 97.05 2.76 0.17 98.00 1.83 0.19 98.57 1.24 2.52 97.41 0.07 1.48 98.39 0.13 1.22 98.62 0.16 0.58 99.42 0.00 0.74 99.26 0.00 0.68 99.03 0.29
0.3 0.16 96.86 2.98 0.14 97.99 1.87 0.15 98.55 1.30 2.78 97.20 0.02 1.63 98.29 0.08 1.33 98.47 0.20 0.69 99.31 0.00 0.79 98.92 0.29 0.91 98.86 0.23
0.4 0.15 96.32 3.53 0.16 97.95 1.89 0.18 98.31 1.51 3.15 96.80 0.05 1.79 98.17 0.04 1.29 98.47 0.24 0.39 99.42 0.19 0.39 99.37 0.24 0.59 98.77 0.64
0.5 0.16 95.71 4.13 0.15 97.37 2.48 0.20 98.19 1.61 3.79 96.18 0.03 2.28 97.66 0.06 1.55 98.28 0.17 0.14 99.65 0.21 0.28 99.42 0.30 0.34 99.30 0.36
0.6 0.39 94.41 5.20 0.17 97.05 2.78 0.13 98.12 1.75 4.72 95.28 0.00 2.88 97.07 0.05 1.70 98.17 0.13 0.20 99.41 0.39 0.23 99.47 0.30 0.47 98.82 0.71
8.0 0.1 0.37 97.86 1.77 0.38 98.26 1.36 0.44 98.58 0.98 2.30 97.61 0.09 1.44 98.46 0.10 1.13 98.69 0.18 0.92 99.08 0.00 1.02 98.98 0.00 1.00 99.00 0.00
0.2 0.44 97.51 2.05 0.36 98.25 1.39 0.34 98.57 1.09 2.41 97.52 0.07 1.43 98.44 0.13 1.20 98.63 0.17 0.58 99.42 0.00 0.74 99.26 0.00 0.68 99.03 0.29
0.3 0.42 97.50 2.08 0.29 98.21 1.50 0.32 98.63 1.05 2.67 97.32 0.01 1.59 98.33 0.08 1.31 98.49 0.20 0.69 99.31 0.00 0.79 98.92 0.29 0.91 98.86 0.23
0.4 0.51 97.17 2.32 0.29 98.33 1.38 0.32 98.49 1.19 2.98 96.97 0.05 1.76 98.20 0.04 1.26 98.50 0.24 0.39 99.42 0.19 0.39 99.37 0.24 0.59 98.77 0.64
0.5 0.53 96.50 2.97 0.33 97.93 1.74 0.38 98.30 1.32 3.67 96.31 0.02 2.23 97.71 0.06 1.51 98.32 0.17 0.14 99.65 0.21 0.28 99.42 0.30 0.34 99.30 0.36
0.6 0.85 95.65 3.50 0.38 97.72 1.90 0.43 98.25 1.32 4.49 95.50 0.01 2.76 97.19 0.05 1.63 98.24 0.13 0.20 99.41 0.39 0.23 99.47 0.30 0.47 98.82 0.71
22 0.1 0.58 97.93 1.49 0.51 98.35 1.14 0.50 98.66 0.84 6.79 91.68 1.53 1.81 97.85 0.34 1.13 98.69 0.18 0.92 99.08 0.00 1.02 98.98 0.00 1.00 99.00 0.00
0.2 0.64 97.72 1.64 0.47 98.35 1.18 0.46 98.61 0.93 2.67 97.17 0.16 1.43 98.44 0.13 1.21 98.62 0.17 0.58 99.42 0.00 0.74 99.26 0.00 0.68 99.03 0.29
0.3 0.75 97.49 1.76 0.43 98.29 1.28 0.43 98.67 0.90 2.64 97.35 0.01 1.59 98.33 0.08 1.29 98.51 0.20 0.69 99.31 0.00 0.79 98.92 0.29 0.91 98.86 0.23
0.4 0.80 97.30 1.90 0.45 98.38 1.17 0.42 98.59 0.99 2.96 96.99 0.05 1.73 98.23 0.04 1.23 98.53 0.24 0.39 99.42 0.19 0.39 99.37 0.24 0.59 98.77 0.64
0.5 0.90 96.68 2.42 0.59 98.02 1.39 0.50 98.36 1.14 3.60 96.38 0.02 2.21 97.74 0.05 1.50 98.33 0.17 0.14 99.65 0.21 0.28 99.42 0.30 0.34 99.30 0.36
0.6 1.14 95.94 2.92 0.55 97.86 1.59 0.55 98.36 1.09 4.45 95.54 0.01 2.72 97.23 0.05 1.64 98.23 0.13 0.20 99.41 0.39 0.23 99.47 0.3 0.47 98.82 0.71
200 0.1 0.89 98.00 1.11 0.65 98.41 0.94 0.62 98.71 0.67 7.92 89.03 3.05 1.79 97.86 0.35 1.12 98.70 0.18 0.92 99.08 0.00 1.02 98.98 0.00 1.00 99.00 0.00
0.2 0.91 97.90 1.19 0.70 98.36 0.94 0.58 98.74 0.68 4.59 95.20 0.21 1.42 98.45 0.13 1.21 98.62 0.17 0.58 99.42 0.00 0.74 99.26 0.00 0.68 99.03 0.29
0.3 1.06 97.61 1.33 0.52 98.52 0.96 0.53 98.68 0.79 5.98 94.01 0.01 1.59 98.32 0.09 1.29 98.51 0.20 0.69 99.31 0.00 0.79 98.92 0.29 0.91 98.86 0.23
0.4 1.30 97.22 1.48 0.69 98.34 0.97 0.60 98.60 0.80 8.52 91.43 0.05 1.80 98.16 0.04 1.22 98.53 0.25 0.39 99.42 0.19 0.39 99.37 0.24 0.59 98.77 0.64
0.5 1.42 96.79 1.79 0.79 98.05 1.16 0.68 98.45 0.87 11.76 88.23 0.01 2.80 97.15 0.05 1.48 98.35 0.17 0.14 99.65 0.21 0.28 99.42 0.30 0.34 99.30 0.36
0.6 1.62 96.14 2.24 0.82 97.93 1.25 0.71 98.30 0.99 17.00 82.99 0.01 4.93 95.02 0.05 1.64 98.23 0.13 0.20 99.41 0.39 0.23 99.47 0.30 0.47 98.82 0.71
_____________________________________________________________________________________________________________
59
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
60
p
µ δ
CI CI CI
n = 25 n = 25 n = 25
n = 50 n = 50 n = 50
n = 100 n = 100 n = 100
0.4236
2.039
0.422
0.745
0.438
1.974
0.392
0.484
0.706
0.525
1.128
1.637
90%
90%
90%
0.840
0.634
0.400
0.518
0.677
1.438
2.297
2.381
0.407
0.366
0.768
0.366
0.354
0.472
0.462
0.725
1.826
1.250
95%
95%
95%
0.373
0.586
0.902
0.502
0.691
1.555
0.276
3.234
3.096
0.275
0.348
0.809
0.291
1.525
2.264
0.382
0.418
0.758
99%
99%
99%
0.503
0.325
0.717
0.471
1.036
1.814
(a) (b) (c)
Figure 4.2: Empirical distributions of µ ! (a), δ! (b) and p! (c) and CI (average of the confidence intervals), for the
indicated values of n, with µ = 1.0, δ = 0.5 and p = 0.6, respectively.
6 Application to inventory management

We apply the methodology developed in this chapter to two real demand data sets. One of them
associated with perishable products and other one with non-perishable products. These data sets are
novel and have not been previously analyzed. The application is divided in three parts. First, we carry
out the statistical analysis of the two data sets, describing the computational implementation developed
for the proposed methodology, and providing details of this analysis for the demand of the products under
study. Here, we perform an exploratory data analysis and, based on it, we show the good fitting of the
ZIRBS distribution to the analyzed data. Second, once the ZIRBS distribution parameters are estimated
and it is fitted to the data, we use inventory management models for responding the questions: (i) what
the quantity of stock should be required in each order, and (ii) when such an order should be placed.
Third and finally, we carry out a financial analysis, which can be helpful in making decisions of inventory
management and demonstrates the importance of the proposed methodology.
6.1 Description of the problem
Currently, most Chilean food service companies are not optimizing their supply of raw materials.
These materials form the inventory assortment of such food companies, which are divided in perish-
able products under single period models (suitable for fruits, meats and vegetables), and non-perishable
products under multiple period models (with greater storage capacity, suitable for grocery). These com-
panies have administrative deficiencies, because they do not consider an optimized inventory policy in
their organizational practices. They rather rely the management of these raw materials on the monthly
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
61
µ δ p
CI CI CI
n = 25 n = 25 n = 25
n = 50 n = 50 n = 50
n = 100 n = 100 n = 100
507.357
198.936
0.953
0.743
0.436
1.051
90%
150.095
384.75
0.484
0.706
0.965
1.036
90%
90%
311.259
149.512
0.975
0.676
0.517
1.026
569.519
187.181
0.766
0.944
0.405
1.061
423.208
138.089
0.958
0.462
0.725
1.044
95%
95%
95%
334.018
139.427
1.031
0.690
0.970
0.501
721.191
168.538
0.927
0.807
0.347
1.080
510.743
117.857
0.418
0.758
0.946
1.058
99%
99%
99%
121.670
383.468
1.041
0.960
0.470
0.716
(a) (b) (c)
Figure 4.3: Empirical distributions of µ ! (a), δ! (b) and p! (c) and CI (average of the confidence intervals), for the
indicated values of n, with µ = 1.0, δ = 200 and p = 0.6, respectively.
planning of a food menu. This menu takes into account technical specifications for procurement con-
tracts, which include certain specifications related to nutritional contributions and adequacy, frequency
and amounts of food consumption, with respect to the people who consume this menu. In this way,
due to technical regulations, several nutritional recommendations must be met, but the administrative
deficiencies aforementioned remain. These deficiencies negatively affect the contribution margins of
the company, which can be improved by a scientific management of logistics and inventories. Thus,
in this application, we carry out a statistical analysis of the demanded amount of two products utilized
for preparing a food menu of an anonymous Chilean food service company, which allows an optimized
inventory system for this company to be established and its contribution margins to be improved.
6.2 Statistical analysis

6.2.1 Computational implementation
The BS distribution is implemented in the R software through the gbs package, with capabilities
related to probabilistic aspects, parameter estimation and goodness-of-fit methods for this distribution.
The estimation of parameters and fit of the ZIRBS distribution to the data under analysis is done by using
the gamlss package, with the ZIRBS family being added to this R package by the author of the present
thesis.
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
62
6.2.2 Data sets
The data set I corresponds to the demanded amount of a non-perishable food product (oil), which
was collected since 20-Nov-2011 and until 26-May-2012, during a study of inventory management con-
ducted by Fernando Rojas and Victor Leiva in the University of Valparaíso, Chile. There were 189 days,
of which 7 (3%) days the product was not demanded, so that during 182 days (97%) at least one unit of
the product was requested in the period of observation. The data set II corresponds to the demanded
amount of a perishable product (potatoes), collected during the same period and for the same study than
oil. In this case, 45 (23,8%) of 189 days, the product was not demanded and, during 144 days (76,2%)
at least one product was requested. Data of the set I are divided by 10000 for facilitating the analysis.
The quantity of each product used in the preparation of a food menu has its respective measuring unit.
This unit possesses an equivalence with respect to the other products of the inventory assortment. In
our case, for oil, 1 unit = 1 package of 900 ml, whereas for potatoes, 1 unit = 1 kg. Data sets I and II are
available under request from the authors.
6.2.3 Exploratory data analysis and model selection
Table 4.6 presents some descriptive measures of the data sets under analysis, such as the sample
size (n), minimum and maximum values, Ȳ , Ỹ , SD, CV, CS and KT. Note that the distributions of the
demanded amounts of the both products have positive skewness and moderate kurtosis.
Table 4.6: Descriptive measures for the indicated demand data set.
Data set With zeros n Min Ỹ Ȳ SD CV CS KT Max

I Yes 189 0.000 7.000 8.614 6.502 0.755 1.456 4.453 29.000
No 182 1.000 7.000 8.945 6.398 0.715 1.527 4.475 29.000
II Yes 189 0.000 5.000 6.294 6.827 1.085 1.244 3.855 28.800
No 144 0.019 6.000 8.261 6.701 0.811 1.086 3.410 28.800
Figures 4.4(a) and 4.4(b) show the histogram of the oil and potatoes demand data with estimated
ZIRBS PDF. From these figures, note that the ZIRBS distribution reproduces the shape of the empirical
distributions of the oil and potatoes demanded amounts very well.
0.10
0.20
0.08
0.15
0.06
PDF
PDF
ZIRBS distribution ZIRBS distribution

0.10
0.04
0.05
0.02
0.00
0.00
0 5 10 15 20 25 30 0 5 10 15 20 25 30
oil demanded amount potatoes demanded amount
(a) (b)
Figure 4.4: Histograms with the estimated ZIRBS PDF for oil (a) and potatoes (b) demand data.
Demand data are usually observed over time. Then, one should ask whether or not these data
display time a dependence. Table 4.7 presents the Box-Pierce, Durbin-Watson and Ljung-Box test
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
63
statistics for examining independence in the oil and potatoes demand data. From this table, it is possible
to see through the corresponding p-values of the tests that the null hypothesis of independence is not
rejected. We generate the plots of the estimated autocorrelations and partial autocorrelations for oil and
potatoes demand data (omitted here), which corroborate the results presented in Table 4.7.
Table 4.7: Values for the indicated statistics to test the null hypothesis of independence in the indicated demand
data set.
Data set Test Test Statistic p-value

I Box-Pierce 1.905 0.168
Durbin-Watson 2.158 0.274
Ljung-Box 1.935 0.164
II Box-Pierce 2.739 0.098
Durbin-Watson 2.241 0.096
Ljung-Box 2.783 0.095
6.2.4 Estimation of the demand rate
Table 4.8 provides point and interval estimates (at confidence level of 95%) of the ZIRBS demand
distribution parameters and the corresponding estimated SEs for oil and potatoes demand data sets.
Table 4.8: Point and interval estimates of the parameters of the ZIRBS demand distribution for the indicated data
set.
Data set Parameter (θ ) θ! L θ!)

SE( CI(θ )
I µ 8.937 0.484 [8.037;9.937]
δ 3.897 0.409 [3.173;4.786]
p 0.037 0.014 [0.036;0.038]
II µ 7.265 0.862 [5.758;9.166]
δ 0.790 0.093 [0.627;0.995]
p 0.238 0.031 [0.227;0.249]
With the results in Table 4.8, we estimate the demand rate (mean of the demanded amount Y , λ
say) of the product oil per time unit, whose point estimate and corresponding 95% CI and CB, using the
delta method and asymptotic independence of the estimators, are given by
N] = !
E[Y ! = [1 − 0.037] × 8.937 = 8.606 units/day,
λ = [1 − p!]µ
CI(λ ; 95%) = [! L !
λ ± z(0.975) × SE(λ )] = [8.606 ± 1.960 × 0.482] = [7.662; 9.550] units/day,
CB(λ ; 95%) = ! L !
λ + z(0.95) × SE( λ ) = 8.606 + 1.645 × 0.482 = 9.399 units/day,
7
L !
respectively, where SE( λ) = µ N p!] + [1 − !
! 2 Var[ Nµ
p]2 Var[ ! ]. Proceeding in an analogous way to the prod-
uct oil, we estimate the demand rate of the potatoes product per time unit, whose point estimate and
corresponding 95% CI and CB are given by ! λ = 5.535 units/day, CI(λ ; 95%) = [4.175; 6.895] units/day and
CB(λ ; 95%) = 6.677 units/day, respectively.
6.3 Inventory management analysis

Once we fit the ZIRBS distribution to the demanded quantity data of the oil and potatoes products,
we then use the adequate inventory management model to determine the optimum stock and minimize
the total cost of the inventory, depending whether the product is stable (oil) or perishable (potatoes).
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
64
6.3.1 Inventory management model for a stable product
Here, we apply the (r, Q) policy as stochastic continuous review model for multiple periods; see
details about this model and it assumptions in Hillier and Lieberman (2001, pp. 956-961). Note that the
ROP is given by
r = λ × l = λDLT , (4.17)
where λDLT is the average demand during the LT, denoted by l. However, as it is known, demand for
products occurs in a random fashion, so that, to be protected from this randomness, it is necessary that
safety stocks are included in the ROP given in (4.17), becoming as
r′ = λDLT + κq × σDLT , (4.18)
where κq × σDLT is the safety stock, which corresponds to the safety factor (κq ) multiplied by the SD of the
demand during the LT (σDLT ). The safety factor κq is associated with a determinate cycle service level
(or number of SDs σDLT ), which corresponds to the qth standardized percentile and is often fixed at the
95th position for assuring a service level of 95%. In general, the mean and SD of the demand during the
LT are
& √
λDLT = E[DLT ] = λ × l and σDLT = Var[DLT ] = l × σ , (4.19)
respectively, where σ is the SD of the demanded amount of the product.

Hence, to calculate the ROP based on the ZIRBS distribution for the product oil, we need the mean
and variance of Y ∼ ZIRBS(µ , δ , p) given in (4.6). For a LT of l = 3 days, and by using expressions in
(4.6), (4.18) and (4.19), and the ZIRBS QF provided in (4.4), we estimate the ROP for the product oil as
√
r′ = !
! λ × 3 + !k95 × 3 × σ! = 48.1984 units.
The optimum quantity of stock required in each order is estimated as

7
Q! = 2×! λ × co/cs = 2113.366 units, (4.20)
where co is the ordering cost and cs the storing cost per unit/day, which in our study are co = US$63.959
per order placed and cs = US$0.00024648 per product unit/day. Thus, we must place an order of Q !=
r′ = 48.1984 units, for assuring a service level of 95%. Hence, the
2113.366 units, when the stock level is !
average stock level per cycle is
!a = Q/2
Q ! +κ!95 × σ
!DLT = 1056.683 + 22.3804 = 1079.063 units, (4.21)
where we recall κ95 × σDLT is the safety stock for assuring a service level of 95%.
6.3.2 Inventory management model for a perishable product
For the potatoes product, we consider the stochastic single period model based on the critical ratio
given by
CR = [cu − c p]/[cu + ch ], (4.22)
where cu is the unsatisfied demand shortage cost per unit, that could include lost revenue and cost of
loss of customer goodwill; c p is the purchase unit cost of the product; and ch is the holding cost per
unit/day, that could include the storing cost minus a salvage value of the unit of product. On the one
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
65
hand, the expression of the numerator cu − c p results in the decrease in profit, due to not ordering a unit
that could have been sold during that period. On the other hand, the denominator cu + cs is the unit cost
of ordering more, which refers to the decrease in profit, due to ordering a unit that could not be sold
during that period. Thus, the single period model for perishable products allows the optimum stored
quantity of units (y0 ) to be obtained from the following expression known as optimum service level:
FY (y0 ) = CR, (4.23)
where FY (·) is the CDF of the demanded quantity, and CR is defined in (4.22). For more details about
this model, see Hillier and Lieberman (2001, pp. 961-975). Then, to calculate the optimum quantity
of units to be stored, based on the ZIRBS distribution for the potatoes product, we need the CDF of
Y ∼ ZIRBS(µ , δ , p) given in (4.3). By using expressions (4.22) and (4.23), cu = US$1.08, c p = US$0.032,
and ch = US$0.088, we estimate y0 for the potatoes product from FY (y0 ) = 0.8968, that is, we need to
determinate the estimated 89.68th quantile of the ZIRBS distribution. Then, the optimum quantity of
units of potatoes to be stored, according to the QF of the ZIRBS distribution given in (4.4), is y!0 =
! , δ!) = 12.704 units. Thus, at the beginning of each single period (one week, in our case), the
y!(0.8968; µ
stock level must be checked, and then a quantity of
12.704 − l j , j = 1, . . . , 27, (4.24)
must be ordered, where l j is the stock level at the beginning of the jth week. Hence, in this case, the
average stock level is
B 12.704
y!a = [12.704 − x] fY (x) dx = 5.812 units, (4.25)
0
where fY (·) is the PDF of the demanded quantity.
6.4 Financial analysis
Now, to demonstrate the economic profits of using our methodology, we financially compare the
optimized system with the current (non-optimized) system used to date by the food service company.
To state this, we determine the total contribution margin of the oil and potatoes products used in the
preparation of the food menu, during the time period under study (27 weeks). Note these products are
only two from 89 products of the inventory assortment of the food service company, and we use them
as illustration; see details in Rojas et al. (2013). Optimized and non-optimized margins are obtained
according to the following.
6.4.1 Calculation of prices and incomes
Below, we calculate the sale price and the unit contribution margin for each product of the inventory
assortment (89 types of products used for the preparation of food). This is carried out based on the
incomes of the company during the period under study (27 weeks), corresponding to the food menu.
As mentioned, the quantity of each product used in the preparation of a food menu has its respective
measuring criterion (we recall that, for oil, 1 unit = 1 package of 900 ml, and for potatoes, 1 unit = 1
kg). This criterion is needed for establishing the corresponding comparison among products. Firstly,
the prorated demand (in “units”) of the ith product in the jth week can be obtained by means of the
proportion consumed for this product, in relation to the demand (in “units”) for all the menus produced
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
66
by the company during such a week, calculated according to
PDi, j = DQi, j /DQ j , i = 1, . . . , 89, j = 1, . . . , 27, (4.26)
where DQi, j is the demanded quantity of the ith product in the jth week and DQ j is the demanded
quantity for all the products during that week. The income of the company for all the portions of the food
menu sold during the jth week is given by
I j = N j × Sj, j = 1 . . . , 27, (4.27)
where N j is the number of menus sold and S j is the price of the food portion of the menu, both of them in
the jth week of study. Thus, the prorated income due to the ith product during the jth week is obtained
as
PIi, j = I j × PDi, j , i = 1, . . . , 89, j = 1, . . . , 27, (4.28)
where PDi, j and I j are defined in (4.26) and (4.27), respectively. Hence, the unit sale price of the ith
product in the jth week is given by
SPi, j = PIi, j /DQi, j , i = 1, . . . , 89, j = 1, . . . , 27, (4.29)
where DQi, j is defined in (4.26) and PIi, j in (4.28). The purchase cost for the ith product in the jth week
is
Ci, j = NCi, j × PQi, j , i = 1, . . . , 89, j = 1, . . . , 27, (4.30)
where NCi, j and PQi, j are the unit net cost and the purchased quantity of the ith product during the jth
week, respectively. Note that, for the optimized system with the inventory model for stable products,
! given in (4.20), whereas that, in the case of perishable products, PQi, j
PQi, j must be estimated from Q
must be estimated from y!0 − l j given in (4.24). For the non-optimized system, this value is empirically
calculated.
6.4.2 Calculation of contribution margins
Once financial indicators defined in (4.28) and (4.30) are obtained, we compute the variable contri-
bution margin of the ith product in the inventory assortment during the jth week as
VCMi, j = PIi, j − Ci, j , i = 1, . . . , 89, j = 1, . . . , 27, (4.31)
where PIi, j and C j are defined in (4.28) and (4.30), respectively. Thus, the ordering cost for the ith
product during the jth week can be obtained as
OCi, j = OCi /52, i = 1, . . . , 89, j = 1, . . . , 27, (4.32)
where OCi is the annual ordering cost of the ith product given by
3
OCi = ∑ OCi,h × Oi , i = 1, . . . , 89,
h=1
with Oi being the annual number of orders and Oi,h is the hth type of cost given in Table 4.9, both for the
ith product. Note that, for the optimized system with the inventory model for stable products, Oi must be
estimated from Q! given in (4.20) using the expression λ /Q for each product (with λ being expressed as
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
67
a demand rate per year), whereas that, in the case of perishable products, whereas that, in the case of
perishable products, Oi = 52, for all i = 1, . . . , 89. For the non-optimized system, this value is empirically
calculated.
Table 4.9: Costs involved in generating a purchase order (OCk ).
OC1 Administrative costs associated with the order movements (input and general service costs.
with respect to order generation)
OC2 Inspection and receiving costs (social security contributions and warehouseman wages)
of movements associated with an order.
OC3 Transportation costs related solely to order generation.
The storing cost for the ith product during the jth week is given by
SCi, j = [SCi /52] × SQi, j , i = 1, . . . , 89, j = 1, . . . , 27, (4.33)
where SQi, j is the stored quantity of the ith product in the jth week, and SCi is the annual storing unit
cost of the ith product given by
5
SCi = ∑ SCi,k /Ri , i = 1, . . . , 89,
k=1
with SCi,k being the kth type of annual cost associated with the storage of Ri = ∑52
j=1 SQi, j units of the
ith product during one year, defined in Table 4.10. Note that, for the optimized system with the inventory
management model for stable products, SQi, j must be estimated from the average stock level Q !a given
in (4.21), whereas that, in the case of perishable products, SQi, j must be estimated from y!a given in
(4.25). For the non-optimized system, this value is empirically calculated.
Table 4.10: Annual costs involved in the storage of a product (SCk ).
SC1 Annual cost of amortization of buildings and networks for air conditioning, handling equipment,
information processing, receiving, storage media, storage and offices, and weighing.
SC2 Annual cost of damage, losses, obsolescence and product losses incurred in the storage period.
SC3 Annual cost of cleaning materials and storehouse, containers, packaging, and printed matter.
SC4 Annual cost of energy spent on the storehouse, including battery charging necessary for
handling, data processing equipment, lighting.
SC5 Annual cost of rental of equipment and facilities, during insurance, storage and communications,
and taxes.
Then, we obtain the contribution margin of the ith product during the jth week as
CMi, j = VCMi, j − [OCi, j + SCi, j ], i = 1, . . . , 89, j = 1, . . . , 27, (4.34)
where VCMi, j , OCi, j and SCi, j are given in (4.31), (4.32) and (4.33), respectively. Therefore, with this
financial methodology, we collect a series of values for 27 weeks and 89 types of products, with contribu-
tion margins for each product of the inventory assortment during the period under study. The contribution
margin of all the products of the inventory assortment during the jth week is given by CM j = ∑89
i=1 CMi, j ,
for j = 1 . . . , 27, where CMi, j is given in (4.34). Thus, the total contribution margin of the inventory system
is CM = ∑27
j=1 CM j , where CM j is given before.
Chapter
___________4___
ZIRBS
_____________________________________________________________________________________________________________
68
6.5 Results of the financial analysis

As illustration, for oil and potatoes products, we have used the financial analysis described in Sub-
sections 6.4.1 and 6.4.2, obtaining the following results. First, an improving in the contribution margins
of the oil product is reached, when the optimized inventory management system proposed in this article
is used, in relation to the current system (non-optimized) employed to date by the food service company.
However, we must report that, still using the optimized system, negative contribution margins are ob-
tained, such as in the non-optimized case. Second, in the case of the potatoes product, once again, an
improvement in the contribution margins is reached with the optimized system that we propose, but we
report that the contribution margins are now positive. We recall that the results reported here are only
an illustration with two of 89 products of the inventory assortment of the food service company studied in
this paper. Thus, the profits in the contribution margins of each product must be totalized. In any case,
the overall contribution margin for all the inventory assortment should be positive and improved with the
optimized inventory management system, in relation to the non-optimized system used to date by the
food service company.
7 Conclusion
In this chapter, we have proposed a methodology based on a new zero-inflated Birnbaum-Saunders
distribution. This methodology can be used for modeling data that contain values equal to zero. A direct
application of this methodology is in the demand data analysis for inventory management models, where
this type of values (zeros) is presented with frequency. We have obtained several properties of the new
distribution and conducted a simulation study for evaluating the behavior of the proposed estimators
for its parameters. The simulation study provided coverage probabilities of confidence intervals for the
parameters that index the zero-inflated Birnbaum-Saunders distribution. This has shown the good prop-
erties of the proposed estimators. We have developed a function that incorporates the new methodology
in the family of generalized additive models of location, shape and scale. This implementation in the R
statistical software has allowed a real demand data analysis to be carried out. Such a analysis has
shown the convenience of using the new proposed methodology by improving the contribution margins
of a Chilean company.
69
CHAPTER 5
Conclusions and Future Research
1 Resumo
Neste trabalho podemos destacar três importantes resultados a saber: (i) obtenção dos esti-
madores de momentos para os parâmetros da distribuição Birnbaum-Saunders reparametrizada pro-
posta por Santos-Neto et al. (2012) e suas distribuições assintóticas. Bem como, obtenção da função
característica que nos permitiu encontrar os momentos não centrais de ordem k dessa distribuição. (ii)
Extensão do modelo de regressão proposto por Leiva et al. (2014) considerando uma estrutura de re-
gressão para o parâmetro de precisão. Estimação e testes de hipóteses para os parâmetros do modelo
foram apresentados. Em particular, para verificar se o parâmetro de precisão no modelo de regressão
BS com precisão variável é constante. Estudos de simulação de Monte Carlo revelaram que dentre
as estatísticas de testes propostas, as estatísticas escore e gradiente apresentaram um melhor de-
sempenho. Resíduos foram desenvolvidos e propriedades empíricas foram apresentadas. Medidas de
diagnóstico sob o enfoque de influência local foram desenvolvidas e ilustradas através de um exemplo
com dados reais. (iii) um novo modelo probabilístico baseado na Birnbaum-Saunders reparametrizada
capaz de modelar conjuntos de dados contínuos positivos que contenham zeros foi proposto. Além
disso, dados em gestão de estoque de produtos foram ajustados e obtemos bons resultados.
Para trabalhos futuros podemos destacar a possibilidade de construir um modelo de regressão
baseado na distribuição Birnbaum-Saunders reparametrizada inflacionada no zero. Utilizar a distribuição
BS reparametrizada em modelagens com enfoque bayesiano. Além disso, é possível acrescentar uma
estrutura semi-paramétrica no modelo de regressão Birnbaum-Saunders reparametrizada estudado
neste trabalho.
2 Conclusions
In Chapter 2, we provided some novel results on moments and generation of random numbers from
the RBS distribution. In addition, we have studied several estimation methods for this distribution. We
have considered the maximum likelihood, moment, modified moment and generalized moment methods
Chapter
___________5___
Conclusions
_________________and
______
Future
_________Research
______________________________________________________________________________________
70
to estimate the corresponding parameters. Furthermore, we have conducted a MC study to evaluate the
performance of each of these estimators. Therefore, due to the fact that the modified moment estimators
are easier to compute, we recommend their use for the RBS distribution. In addition, we have obtained
moment estimators in a closed-form, which is not possible with the original parameterization of the BS
distribution. However, the parameter estimators obtained by the moment method, as well as those
obtained by the GMM, present underperformed with respect to their statistical properties. Nevertheless,
for the case of large sample sizes, all the studied estimators have similar statistical properties.
In Chapter 3, we proposed a RBS regression model that supposes a structure of regression for
the precision parameter, which is a generalization of the model proposed in Leiva et al. (2014). We
also have proposed four tests for varying the precision in the framework of RBS regression model. We
show through of numerical examples and simulation study that the gradient test is quite effective. Other
important step in regression models is a residual analysis. Thus, there were proposed four residuals
and display through of simulation study that the rd,i and rq,i residuals presented better performance.
Moreover, we have developed methods of local influence to assess the potential influence of some ob-
servations on the model by using several perturbation schemes. Finally, we have performed a statistical
modeling with real data by using the new approach proposed in this chapter, which have shown the
importance of our proposal.
Finally, in Chapter 4, we proposed the ZIRBS distribution. This methodology can be used for mod-
eling data that contain values equal to zero. A direct application of this methodology is in the demand
data analysis for inventory management models, where this type of values (zeros) is presented with fre-
quency. We have obtained several properties of the new distribution and conducted a simulation study
for evaluating the behavior of the proposed estimators for its parameters. This has shown the good
properties of the proposed estimators. This implementation in the R statistical software has allowed a
real demand data analysis to be carried out. Such a analysis has shown the convenience of using the
new proposed methodology by improving the contribution margins of a Chilean company.
3 Future Research
We list below some issues that can be addressed (or being developed) in the future. (i) A RBS
regression model with semi-parametric structure; (ii) A RBS regression model with bayesian approach;
(ii) A ZIRBS regression model among others.
71
REFERENCES
Abramowitz, M., Stegun, I. (1972) Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical
Tables. Dover Publications, New York.
Achcar, J.A. (1993) Inference for the Birnbaum-Saunders fatigue life model using Bayesian methods. Comp. Statist.
Data Anal., 15, 367–380.
Ahmad, L.A. (1988) Jackknife estimation for a family of life distributions. J. Statist. Comp. Simul., 29, 211–223.
Ahmed, S.E., Budsaba, K., Lisawadi, S., Volodin, A. (2008) Parametric estimation for the Birnbaum-saunders
lifetime distribution based on a new parametrization. Thailand Statist., 6, 213–240.
Ahmed, S.E., Castro, C., Flores, E., Leiva, V., Sanhueza, A. (2010) A truncated version of the Birnbaum-Saunders
distribution with an application in financial risk. Pak. J. Stat., 26, 293–311.
Aitchison J. (1955) On the distribution of a positive random variable having a discrete probability mass at the origin.
J. Amer. Statist. Soc., 50, 901–908.
Aitkin M., Anderson D., Francis B., Hinde J. (1989) Statistical Modelling in GLIM. Oxford University Press.
Atkinson, A.C. (1985). Plots, Transformation and Regression. Clarendon Press, Oxford.
Azevedo, C., Leiva, V., Athayde, E., Balakrishnan, N. (2012) Shape and change point analyses of the Birnbaum-
Saunders-t hazard rate and associated estimation. Comp. Statist. Data Anal., 56, 3887–3897.
Azzalini, A. (1996) Statistical Inference: Based on the Likelihood. Chapman & Hall, New York.
Balakrishnan, N., Leiva, V., López, J. (2007) Acceptance sampling plans from truncated life tests based on the
generalized Birnbaum-Saunders distribution. Comm. Statist. Simul. Comp., 36, 643–656.
Balakrishnan, N., Leiva, V., Sanhueza, A., Cabrera, E. (2009a) Mixture inverse Gaussian distribution and its
transformations, moments and applications. Statistics, 43, 91–104.
Balakrishnan, N., Leiva, V., Sanhueza, A., Vilca, F. (2009b) Estimation in the Birnbaum-Saunders distribution based
on scale-mixture of normals and the EM-algorithm. Stat. Oper. Res. Trans. 33, 171–192.
Balakrishnan, N., Gupta, R. Kundu, D., Leiva, V., Sanhueza, A., (2011) On some mixture models based on the
Birnbaum-Saunders distribution and associated inference. J. Statist. Plann. Infer., 141, 2175–2190.
Barros, M., Paula, G.A., Leiva, V. (2008) A new class of survival regression models with heavy-tailed errors:
robustness and diagnostics. Lifetime Data Anal., 14, 316–332.
REFERENCES
____________________________________________________________________________________________________________________________________
72
Barros, M., Paula, G.A., Leiva, V. (2009) An R implementation for generalized Birnbaum-Saunders distributions.
Comp. Statist. Data Anal., 53, 1511–1528.
Bhattacharyya, G.K., Fries, A. (1982) Fatigue failure models: Birnbaum-Saunders versus inverse Gaussian. IEEE
Trans. Rel., 31, 439–440.
Bhatti, C.R. (2010) The Birnbaum-Saunders autoregressive conditional duration model. Math. Comp. Simul., 80,
2062–2078.
Birnbaum, Z.W., Saunders, S.C. (1969a) A new family of life distributions. J. Appl. Prob., 6, 319–327.
Birnbaum, Z.W., Saunders, S.C. (1969b) Estimation for a family of life distributions with applications to fatigue. J.
Appl. Prob., 6, 328–347.
Burgin, T.A., (1975) The gamma distribution in inventory control. Oper. Res. Q., 23, 507–525.
Cao, C.Z , Lin, J.G, Zhu, L.X. (2010) Heteroscedasticity and/or autocorrelation diagnostics in nonlinear models with
AR(1) and symmetrical errors. Stat Papers, 51, 813–836.
Caro-Lopera, F.J., Leiva, V., Balakrishnan, N. (2012) Connection between the Hadamard and matrix products with
an application to a matrix-variate Birnbaum-Saunders distribution. Journal of Multivariate Analysis, 104, 126–139.
Castellacci, G. (2012) A formula for the quantiles of mixtures of distributions with disjoint supports. Working paper
available at http://dx.doi.org/10.2139/ssrn.2055022.
Chang, D.S., Tang, L.C. (1994) Graphical analysis for Birnbaum-Saunders distribution. Microelect. Rel., 34, 17–22.
Chaussé, P. (2010) Computing generalized method of moments and generalized empirical likelihood with R. J.
Statist. Soft., 34, 1–35.
Chen, G., Balakrishnan, N. (1995). A general purpose approximate goodness-of-fit test. Journal of Quality
Technology, 27, 154–161.
Chen, F.Y., Feng, Y., Simchi-Levi, D. (2002) Uniform distribution of inventory positions in two-echelon periodic
review systems with batch-ordering policies and interdependent demands. Eur. J. Oper. Res., 140, 648–654.
Cobb, B.R., Rumí, R., Salmerón, A. (2013) Inventory management with log-normal demand per unit time. Comp.
Oper. Res., 40, 1842–1851.
Cole, T. and Green, P. (1992) Smoothing reference centile curves: The LMS method and penalized likelihood.
Statist. in Med, 11, 1305–1319.
Cook, R.D., Weisberg, S. (1983). Diagnostics for heteroscedasticity in regression. Biometrika. 70, 1–10.
Cook, R.D. (1986) Assessment of local influence. J. R. Statist. Soc. B, 48, 133–169.
Cook D.O., Kieschnick R., McCullough B.D. (2004) On the heterogeneity of corporate capital structures and its
implications. SSRN working paper available at http://ssrn.com/abstract=671061.
Cox, D.R., Hinkley, D.V. (1974) Theoretical Statistics. Chapman & Hall, London.
Crawley, M.J. (2007) The R book. John Wiley & Sons, London.
Cysneiros, A.H.M.A., Cribari-Neto, F., Araujo, C.G.J. (2008) On Birnbaum-Saunders inference. Comp. Statist. Data
Anal., 52, 4939–4950.
Cysneiros, F.J.A., Paula, G.A., Galea, M. (2007). Heteroscedastic symmetrical linear models. Statist. Probab. Lett.
77, 1084–1090.
Das, C. (1976) Approximate solutions to the (Q, r) inventory model for gamma lead time demand. Manag. Sc., 22,
REFERENCES
____________________________________________________________________________________________________________________________________
73
1043–1047.
De Jong, P., Heller, G.Z. (2008) Generalized Linear Models for Insurance Data. Cambridge University Press, New
York.
Díaz-García, J.A., Leiva, V. (2005) A new family of life distributions based on the elliptically contoured distributions.
Journal of Statistical Planning and Inference, 128, 445–457.
Dunn, P. K., Smyth, G. K. (1996) Randomized Quantile Residuals. Journal of Computational and Graphical
Statistics, 5, 236–244.
Dupuis, D.J., Mills, J.E. (1998) Robust estimation of the Birnbaum-Saunders distribution. IEEE Trans. Rel., 47,
88–95.
Engelhardt, M., Bain, L.J., Wright, F.T. (1981) Inferences on the parameters of the Birnbaum-Saunders fatigue life
distribution based on maximum likelihood estimation. Technometrics, 23, 251–256.
Ferrari, S.L.P., Cribari-Neto, F. (2004) Beta regression for modelling rates and proportions. J. Appl. Statist., 31,
799–815.
Ferrari, S.L.P., Espinheira, P.L., Cribari-Neto, F. (2011) Diagnostic tools in beta regression with varying dispersion.
Statistica Neerlandica, 65, 337–351.
Ferreira, M., Gomes, M.I., Leiva, V. (2012) On an extreme value version of the Birnbaum-Saunders distribution.
RevStat Statist. J., 10, 181–210.
Feuerverger, A. (1979) On some methods of analysis for weather experiments. Biometrika. 66, 665–668.
Freudenthal, A.M., Shinozuka, M. (1961). Structural safety under conditions of ultimate load failure and fatigue.
Technical Report WADD-TR-61-177, Wright Air Development Division, Wright Air Force Base, Dayton, Ohio, USA.
From, S.G., Li, L.X. (2006) Estimation of the parameters of the Birnbaum-Saunders distribution. Comm. Statist.
Theor. Meth., 35, 2157–2169.
Galea, M., Paula, G.A., Leiva, V. (2004) Influence diagnostics in log-Birnbaum-Saunders regression models. J.
Appl. Statist., 31, 1049–1064.
Gallego, G., Jin, Y., Murie, A., Zhang, G., Yildiz, V.T. (2007) Optimal policies with convertible lead times. Eur. J.
Oper. Res., 176, 892–910.
Gómez, H.W., Olivares-Pacheco, J.F., Bolfarine, H. (2009) An extension of the generalized Birnbaum-Saunders
distribution. Statist. Prob. Letters, 79, 331–338.
Griffiths, W., Hill, R., Judge, G. (1993) Learning and Practicing Econometrics. John Wiley & Sons, New York.
Hansen, L.P. (1982) Large sample properties of generalized method of moments estimators. Econometrica, 50,
1029–1054.
Heller, G., Stasinopoulos, M., Rigby, B. (2006) The zero-adjusted inverse Gaussian distribution as a model for
insurance claims.In: Hinde J, Einbeck J, Newell J (eds) Proceedings of the 21th International Workshop on
Statistical Modelling, 226–233, Ireland, Galway.
Hillier, F., Lieberman, G.J. (2001) Introduction to Operational Research. 7th ed. McGraw Hill, New York.
Hogg, R.V., Klugman, S.A. (1984) Loss Distributions. John Wiley & Sons , New York .
Huang, S, Qu, Y. (2006) The loss in power when the test of differential expression is performed under a wrong
scale. J. Comp. Biol., 13, 786–797.
REFERENCES
____________________________________________________________________________________________________________________________________
74
Jin, X., Kawczak, J. (2003) Birnbaum-Saunders and lognormal kernel estimators for modelling durations in high
frequency financial data. Ann. Econ. Fin., 4, 103–124.
Jørgensen, B. (1984) The delta algorithm and GLIM. Inter. Stat. Rev., 52, 283–300.
Jøgensen, B., de Souza, M.C.P. (1994). Fitting Tweedie’s compound Poisson model to insurance claims data.
Scand. Actual. J., 69–93.
Johnson, N., Kotz, S., Balakrishnan, N. (1995) Continuous Univariate Distributions. Vol. 2. John Wiley & Sons, New
York.
Keaton, M. (1995) Using the gamma distribution to model demand when lead time is random. J. Bus. Logist., 16,
107–131.
Kotz, S., Leiva, V., Sanhueza, A. (2010) Two new mixture models related to the inverse Gaussian distribution. Meth.
Comp. Appl. Prob., 12, 199–212.
Lau, H-S. (1989) Toward an inventory control system under non-normal demand and lead time uncertainty. J. Bus.
Logist., 10, 88–103.
Leiva, V., Barros, M., Paula, G.A., Galea, M. (2007) Influence diagnostics in log Birnbaum-Saunders regression
models with censored data. Comp. Statist. Data Anal., 51, 5694–5707.
Leiva, V., Barros, M., Paula, G.A., Sanhueza, A. (2008a) Generalized Birnbaum-Saunders distributions applied to
air pollutant concentration. Environmetrics, 19, 235–249.
Leiva, V., Sanhueza, A., Sen, P.K., Paula, G.A. (2008b) Random number generators for the generalized Birnbaum-
Saunders distribution. J. Statist. Comp. Simul., 78, 1105–1118.
Leiva, V., Sanhueza, A., Silva, A., Galea, M. (2008c) A new three-parameter extension of the inverse Gaussian
distribution. Statist. Prob. Letters, 78, 1266–1273.
Leiva, V., Sanhueza, A., Angulo, J.M. (2009) A length-biased version of the Birnbaum-Saunders distribution with
application in water quality.Stoch. Envir. Res. Risk Ass., 23, 299–307.
Leiva, V., Sanhueza, A. Kotz, S., Araneda, N. (2010) A unified mixture model based on the inverse gaussian
distribution. Pak. J. Stat., 26, 445–460.
Leiva, V., Athayde, E., Azevedo, C., Marchant, C. (2011) Modeling wind energy flux by a Birnbaum-Saunders
distribution with unknown shift parameter. J. Appl. Statist., 38, 2819–2838.
Leiva, V., Soto, G., Cabrera, E.G., (2011b) New control charts based on the Birnbaum-Saunders distribution and
their implementation. Col. J. Statist., 34, 147–176.
Leiva, V., Ponce, M.G., Marchant, C., Bustos, O. (2012) Fatigue statistical distributions useful for modeling diameter
and mortality of trees. Col. J. Statist., 35, 349–367.
Leiva, V., Rojas, E., Galea, M., Sanhueza, A. (2013) Diagnostics in Birnbaum-Saunders accelerated life models with
an application to fatigue data. Appl. Stoch. Mode. Bus. Ind. (in press) http://dx.doi.org/10.1002/asmb.1944.
Leiva, V. Santos-Neto, M., Cysneiros, F.J.A. , Barros, M. (2014) Birnbaum-Saunders statistical modelling: a new
approach. Statist. Model.(in print)
Lemonte, A.J., Cribari-Neto, F., Vasconcellos, K.L.P. (2007) Improved statistical inference for the two-parameter
Birnbaum-Saunders distribution. Comp. Statist. Data Anal., 51, 4656–4681.
Lemonte, A.J, Simas, A., Cribari-Neto, F. (2008) Bootstrap-based improved estimators for the two-parameter
Birnbaum-Saunders distribution. J. Statist. Comp. Simul., 78, 37–49.
REFERENCES
____________________________________________________________________________________________________________________________________
75
Lemonte, A.J., Cordeiro, G.M. (2009) Birnbaum-saunders nonlinear regression models. Comp. Statist. Data Anal.,
53, 4441–4452.
Lemonte, A.J., Patriota, A.G. (2011) Influence diagnostics in Birnbaum-Saunders nonlinear regression models. J.
Appl. Statist., 38, 871–884.
Lemonte, A.J. (2013) On the gradient statistic under model misspecification. Statist. Prob. Letters, 83, 390–398.
Lesaffre, E., Rizoupoulus, D., Tsonaka, S. (2012) The logistic-transform for bounded outcome scores. Biostatistics,
8, 72–85.
Lesaffre, F., Verbeke, G. (1998) Local influence in linear mixed models. Biometrics, 38, 963–974.
Li, W.K. (1991) Testing model adequacy for some markov regression models for time series. Biometrika, 78, 83–89.
Li, A.P., Chen, Z.X, Xie, F.C. (2012) Diagnostic analysis for heterogeneous log-Birnbaum-Saunders regression
models. Statist. Prob. Letters, 89, 1690–1698.
Lin, J.G., Zhu, L.X., Xie, F.C. (2009) Heteroscedasticity diagnostics for t linear regression models. Metrika, 70,
59–77.
Lio, Y.L., Tsai, T-R, Wu, S-J (2010) Acceptance sampling plans from truncated life tests based on the Birnbaum-
Saunders distribution for percentiles. Comm. Statist. Simul. Comp., 39, 119–136.
Liu, W., Lin, S., Piegorsch, W.W. (2008) Construction of exact simultaneous confidence bands for a simple linear
regression model. Inter. Stat. Rev., 76, 39–57.
Marchant, C., Bertin, K., Leiva, V., Saulo, G. (2013a) Generalized Birnbaum-Saunders kernel density estimators
and an analysis of financial data. Comp. Statist. Data Anal., 63, 1–15.
Marchant, C., Leiva, V., Cavieres, M.F., Sanhueza, A. (2013b) Air contaminant statistical distributions with
application to PM10 in Santiago, Chile. Rev. Environ. Contam. Tox., 223, 1–31.
Mátyás, L. (1999) Generalized Methods of Moments Estimation. Cambridge University Press, New York.
McCullagh, P., Nelder, J.A. (1989) Generalized Linear Models. Chapman & Hall, London.
McLachlan, G.J., Peel, D. (2000) Finite Mixture Models. John Wiley & Sons, New York.
Neyman, J., Pearson, E.S. (1928) On the use and interpretation of certain test criteria for purposes of statistical
inference. Biometrika, 20 A(1/2), 175–240.
Ng, H.K., Kundu D., Balakrishnan N. (2003) Modified moment estimation for the two-parameter Birnbaum-Saunders
distribution. Comp. Statist. Data Anal., 43, 283–298.
Ospina, R., Ferrari, S. L. P. (2010) Inflated beta distributions. Statist. Papers, 51, 111–126.
Pace, L., Salvan, A. (1997) Principles of statistical inference from a Neo-Fisherian perspective. World Scientific
Publishing, Singapore.
Paula, G.A., Leiva, V., Barros, M., Liu, S. (2012) Robust statistical modeling using the Birnbaum-Saunders-t
distribution applied to insurance. Appl. Stoch. Mod. Bus. Ind., 28, 16–34.
Podlaski, R. (2008) Characterization of diameter distribution data in near-natural forests using the Birnbaum-
Saunders distribution. Can. J. For. Res., 18, 518–527.
Porras, E., Dekker, R. (2008) An inventory control system for spare parts at a refinery: an empirical comparison of
different re-order point methods. Eur. J. Oper. Res., 184, 101–132.
Qu, H., Xie, F.C. (2011) Diagnostics analysis for log-Birnbaum-Saunders regression models with censored data.
REFERENCES
____________________________________________________________________________________________________________________________________
76
Statist. Neerl., 65, 1–21.
R Development Core Team (2011) R: A Language and Environment for Statistical Computing. R Foundation for
Statistical Computing, Vienna, Austria. www.R-project.org
Rao, C.R. (1948) Large sample tests of statistical hypotheses concerning several parameters with applications to
problems of estimation. Math. Proc. Camb. Phil. Soc., 44, 50–57.
Rao, C.R. (1965) Linear Statistical Inference and its Applications. John Wiley & Sons, New York.
Rieck, J. R., Nedelman, J.R. (1991) A log-linear model for the Birnbaum-Saunders distribution. Technometrics, 33,
51–60.
Rieck, J.R. (1999) A moment-generating function with application to the Birnbaum-Saunders distribution. Comm.
Statist. Theor. Meth., 28, 2213–2222.
Rigby, R.A., Stasinopoulos, D.M. (1996a). A semi-parametric additive model for variance heterogeneity. Statist.
Comput., 6, 57–65.
Rigby, R.A., Stasinopoulos, D.M. (1996b). Mean and dispersion additive models. In: Hardle, W. and Schimek, M.G.
(eds.), Statist. Theo. Comp. Asp. Smooth., 215–230. Physica, Heidelberg.
Rigby, R.A., Stasinopoulos, D.M. (2005). Generalized additive models for location, scale and shape, (with discus-
sion). Appl. Statist., 54, 507–554.
Rocha, A.V., Cribari-Neto, F. (2009) Beta autoregressive moving average models. Test, 18, 529–545.
Rocha, A.V., Simas, A.B. (2011) Influence diagnostics in a general class of beta regression models. Test, 20,
95–119.
Rojas, F., Leiva, V., Wanke, P., Marchant, C. (2013) Improving contribution margins in a
chilean food service SME by using inventory management models. Working paper available at
http://www.deuv.cl/leiva/working_papers/art_RLWM.pdf.
Sanhueza, A., Leiva, V., Lopez-Kleine, L. (2011) On the Student-t mixture inverse Gaussian model with an
application to protein production. Col. J. Stat., 34, 177–195.
Santos-Neto, M., Cysneiros, F.J.A. Leiva, V., Ahmed, S.E. (2012) On new parameterizations of the Birnbaum-
Saunders distribution. Pak. J. Statist., 28, 1–26.
Santos-Neto, M., Cysneiros, F.J.A., Leiva, V., Barros, M. (2013a) A reparameterized Birnbaum-Saunders distribu-
tion and its moments, estimation and applications. Revstat - Statistical Journal. (in print)
Santos-Neto, M., Cysneiros, F.J.A., Leiva, V., Barros, M. (2013b) A zero-inflated Birnbaum-Saunders dis-
tribution useful for modeling demand applied to inventory managements. Working paper available at
http://www.deuv.cl/leiva/archivos/leiva_art/article_SCLB.pdf
Santos-Neto, M., Cysneiros, F.J.A., Leiva, V., Barros, M. (2013c) Birnbaum-Saunders regression model varying
precision. Working paper available at http://www.deuv.cl/leiva/archivos/leiva_art/article_SCLB1.pdf
Silver, E.A., Peterson R. (1985) Decision Systems for Inventory Management and Production Planning. Wiley, New
York.
Smithson, M., Verkuilen, J. (2006), A better lemon-squeezer? Maximum likelihood regression with beta distribuited
dependent variables. Psychological Methods, 11, 54–71.
Smyth, G.K. (1989) Generalized linear models with varying dispersion. J. R. Statist. Soc. B, 51, 47–60.
Smyth, G.K., Verbyla, A.P. (1999) Adjusted likelihood methods for modelling dispersion in generalized linear
REFERENCES
____________________________________________________________________________________________________________________________________
77
models. Environmetrics, 10, 695–709.
Smyth, G.K., Jørgensen, B. (2002). Fitting Tweedie’s compound Poisson model to insurance claims data: dispersion
modelling. ASTIN Bulletin, 32(1), 143–157.
Snyder, R.D. (1984) Inventory control with the gamma probability distribution. Eur. J. Oper. Res., 17, 373–381.
Speh, T.W., Wagenheim, G.D. (1978) Demand and lead-time uncertainty: the impacts on physical distribution
performance and management. J. Bus. Logist., 1, 95–113.
Stasinopoulos D.M., Rigby R.A., Akantziliotou C. (2006) gamlss: A collection of functions to fit Generalized Additive
Models for Location Scale and Shape, R package version 1.1-0, url = http://www.londonmet.ac.uk/gamlss/.
Stasinopoulos, D.M., Rigby, R.A. (2007) Generalized additive models for location scale and shape (GAMLSS). J.
Statist. Soft., 23, Issue 7, December 2007.
Strijbosch, L.W.G., Moors, J.J.A. (2006) Modified normal demand distributions in (R, S)-inventory control. Eur. J.
Oper. Res., 172, 201–212.
Tadikamalla, P.R. (1979) The lognormal approximation to the lead time demand in inventory control. Omega, 7,
553–556.
Tadikamalla, P.R. (1981) The inverse Gaussian approximation to the lead time demand in inventory control. Int. J.
Prod. Res., 19, 213–219.
Taylor, J., Varbyla, A. (2004) Joint modelling of location and scale parameters of the t distribution. Statist. Model.,
4, 91–112.
Terrell, G.R. (2002) The Gradient Statistic. Comp. Scie. Statist., 34, 206–215.
Tsai, C.L., Cai, Z., Wu, X. (1998) The examination fo residual plots. Statistica Sinica, 8, 445–465.
Vanegas, L.H., Rondon, L.M., Cysneiros, F.J.A. (2012) Diagnostic procedures in Birnbaum-Saunders nonlinear
regression models. Comp. Statist. Data Anal., 56, 1662–1680.
Vargas, T.M., Ferrari, S.L.P, Lemonte, A.J. (2013) Gradient statistic: Higher-order asymptotics and Bartlett-type
correction. Electron J Stat, 7, 43–61.
Vilca, F., Sanhueza, A., Leiva, V., Christakos, G. (2010) An extended Birnbaum-Saunders model and its application
in the study of environmental quality in Santiago, Chile. Stoch. Environ. Res. Risk. Assess., 24, 771–782.
Vilca, F., Santana, L., Leiva, V., Balakrishnan, N. (2011) Estimation of extreme percentiles in Birnbaum-Saunders
distributions. Comp. Statist. Data Anal., 55, 1665–1678.
Villegas, C., Paula, G.A., Leiva, V. (2011) Birnbaum-Saunders mixed models for censored reliability data analysis.
IEEE Trans. Rel., 60, 748–758.
Volodin, I.N., Dzhungurova, O.A. (2000) On limit distribution emerging in the generalized Birnbaum-Saunders
model. J. Math. Sc., 99, 1348–1366.
Xie, F.C., Wei, B.C. (2007) Diagnostics analysis for log-Birnbaum-Saunders regression models. Comp. Statist. Data
Anal., 51, 4692–4706.
Wald, A. (1943) Tests of statistical hypotheses concerning several parameters when the number of observations is
large. Trans. Amer. Mathem. Soc., 54, 426–482.
Wanke, P.F. (2008) The uniform distribution as a first practical approach to new product inventory management. Int.
J. Prod. Econ., 114, 811–819.
REFERENCES
____________________________________________________________________________________________________________________________________
78
Wanke, P.F. (2009) Consolidation effects and inventory portfolios. Transportation research: Part E. Logist. Transp.
Rev., 45, 107–124.
Wei, B.C., Hu, Y.Q., Fung, W.K. (1998) Generalized leverage and its applications. Scand. J. Statist., 25, 25–37.
Weisberg, S. (2005) Applied Linear Regression, 3rd edition. New York, Wiley.
Wilks, S.S. (1938) The large-sample distribution of the likelihood ratio for testing composite hypotheses. The Ann.
Math. Statist., 9, 60–62.
Whitmore, G. (1986) Inverse Gaussian ratio estimation. Appl. Statist., 35, 8–15.
Wu, L.C, Zhang, Z.Z, Xu, D.K. (2012) Variable selection for joint mean and dispersion models of the lognormal
distribution. Hacettepe Journal of Mathematics and Statistics, 41, 307–320.
Yoo, S. (2004) A note on an approximation of the mobile communications expenditures distribution function using a
mixture model. J. Appl. Statist., 31, 747–775.

2013 DissT Reparameterized Birnbaum Saunders Distr.

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

2013 DissT Reparameterized Birnbaum Saunders Distr.

Caricato da

Copyright:

Formati disponibili

Universidade Federal de Pernambuco

Centro de Ciências Exatas e da Natureza

Reparameterized Birnbaum-Saunders distribution: Estimation,

Manoel Ferreira dos Santos Neto

Manoel Ferreira dos Santos Neto

Reparameterized Birnbaum-Saunders distribution: Estimation,

Trabalho apresentado ao Programa de Pós-graduação em Es-

Orientador: Prof. Dr. Francisco José de Azevêdo Cysneiros

Palavras-chave: Distribuição Birnbaum-Saunders reparametrizada, métodos de estimação, méto-

In this work, we derive methodologies based on a new parameterization of the Birnbaum-Saunders

Keywords: Reparameterized Birnbaum-Saunders distribution, estimation methods, Monte Carlo

1.1 The main packages used in this thesis. p. 8

LIST OF TABLES p. vii

2 Estimation in the RBS distribution p. 9

3 Reparameterized Birnbaum-Saunders regression models with varying precision p. 28

4 Zero-inflated reparameterized Birnbaum-Saunders distribution p. 48

6.4.2 Calculation of contribution margins p. 66

5 Conclusions and Future Research p. 69

2 Introduction and bibliographical review

3.2 A new parameterization for the BS distribution

3.3 Regression models based on BS distributions

where β = [β1 , . . . , β p ]⊤ , for p < n, is a vector of unknown parameters to be estimated, and x i ⊤ =

where g′ (·) is the derivative of g(·). In matrix form, we have

Uβ (θ ) = X ⊤ A [yy∗ − µ ∗ ] and Uδ (θ ) = 1 ⊤ [yy⋆ − µ ⋆ ], (1.8)

δ [δ +1]yi δ2 1 [yi + µi ] yi δ [δ +2] µi

4 Motivation of the thesis

5 Objectives of the thesis

• The BS distribution is an asymmetric probabilistic model that is receiving considerable attention

7 Products of the thesis

8 Organization of the thesis

9 Computing platform for implementation of results

Table 1.1: The main packages used in this thesis.

Package Authors Used Functions

Estimation in the RBS distribution

3 Moments of the RBS distribution

Theorem 1. Let Y ∼ RBS(µ , δ ). Then, the CF ϕ : R → C of Y is

dr ϕY (t) >> >

where Kv is the modified Bessel function of second type.

Table 2.1: Values of Kv (δ /2) for the indicated values of v.

v Value of the function

[δ + 1]2 [2δ + 5][δ + 1]2

Algorithm 1 Generator of BS random numbers

4.1 Maximum likelihood estimation

1.0 1.2 1.4 1.6 1.8 2.0

where iδ µ = iµδ and

4.2 Moment estimation

Consider the following definition.

which, after some algebraic manipulations, results to be

of the parameters µ and δ , respectively. Given by that the random vector

We want to determine the asymptotic joint distribution of the estimators

C = f1 (Ȳ , S2 ) and δC = f2 (Ȳ , S2 ).

These estimators can be expressed as

4.3 Modified moment estimation

where Ȳh = [{1/n} ∑nj=1 {1/Y j }]−1 .

Recall the moment conditions are E[ f (Y j , θ 0 )] = 0 . We have that fn (θ̆θ ) = 0 implies

Hence, solving (4.3), we obtain the MM estimators

However, our interest is to find the asymptotic joint distribution of

µ̆ = f1 (Ȳ , Ȳh−1 ) and δ̆ = f2 (Ȳ , Ȳh−1 ).

For these estimators, consider

two continuous functions and the delta method. Then,

4.4 Generalized Method of Moments estimation

where the gradient function of fn (Y j , θ ) is given by

6.1 Data set I (S1): Griffiths et al. (1993)

is slightly positively skewed.