Beruflich Dokumente
Kultur Dokumente
OSVALDO VENEGAS
Universidad Católica de Temuco, Chile
DAVID ELAL-OLIVERO
Universidad de Atacama, Chile
and
HÉCTOR W. GÓMEZ
Universidad de Antofagasta, Chile
Received : December 2011. Accepted : March 2012
Abstract
In this article, we show that certain skew-normal parametric sta-
tistical models are a result of rescalings of the skew normal model
studied by Azzalini (1985). Using this procedure we define a class
of skew-normal distributions and we study its moment, skewness and
kurtosis coefficients. At the end of this article we will use this class
of distribution to make some extensions of the skew-normal model.
1. Introduction
where φ and Φ are the N (0, 1) probability density function and cumulative
distribution function, respectively. Some properties for this distribution
are: r
2 λ 2λ2
(1.2) E (X) = √ ; V ar (X) = 1 −
π 1 + λ2 π (1 + λ2 )
µ ¶
λt
(1.3) MX (t) = 2exp(t2 /2)Φ √
1 + λ2
à !3/2 à !2
p 1 E 2 (X) E 2 (X)
(1.4) β1 = (4 − π) ; β2 = 2 (π − 3) ,
2 V ar (X) V ar (X)
√
where β1 and β2 are the asymmetric and kurtosis coefficients, respectively.
From (4) we get
p
(1.5) − 0.9953 ≤ β1 ≤ 0.9953 ; 3.0000 ≤ β2 ≤ 3.8692.
The purpose in these notes, is to show how the rescalings affects the
representation of the skew-normal model, as well as provide some extensions
of this rescaled model.
This article is organized as follows. In section 2 we give the rescal-
ings of the skew-normal model, some properties, examples, moments and
asymmetric and kurtosis coefficients. In section 3 some extensions of this
rescaled model are presented. In section 4 presents the appendix.
2. Rescaling
Remark 1.
a) A distribution suitable for fitting positive data is the half-normal
distribution. We say that a random variable X follows a half-normal
distribution with scale parameter σ if its density function is given by:
µ ¶
2 x
f (x; σ) = φ I{x > 0}
σ σ
with σ > 0. We denote this by writing X ∼ HN (σ).
a) The pdf of the RSN (0, r, s) is identical to the pdf of the N (0, r2 ).
¡ ¢ ³ ´
Example 3. If fX (x) = σ2 φ σx Φ λδ x where σ and δ corresponds to the
standard deviations of the normal models, then X ∼ RSN (λ, σ, δ) since
r σ 2
s = δ > 0 and (σ, δ) ∈ Θ = ]0, ∞[ . Nadarajah and Kotz(2003) intro-
duced such model and they called Skew-Normal-Normal distribution.
⎧
⎨ 0, k odd
where bk = E(V k ) = √
⎩ 2k/2 Γ( k+1 )/ π, k even.
2
8 (π − 3) r4 λ4
β2 = + 3.
(πs2 + πr2 λ2 − 2λ2 r2 )2
Proposition 7. Let
µ X | W = w ∼√ N (w, 1) and W ERSN¶(0, 1, λ, r, s, 0).
√ s2 (1+r2 )2 +r2 λ2 (1+r2 )
Then X ∼ ERSN 0, 1, λ, 1 + r2 , r2 ,0 .
R∞
Proof By a direct computation in fX (x) = −∞ fX|W (x | w)fW (w)dw is
obtained the result. 2
Z ∞
Proof By a direct computation in fX (x) = fX|V (x | v)fV (v)dv the
−∞
result is obtained. 2
4. Appendix
P [X ≤ x]
= P [a |U| + bV ≤ x]
Z ∞ ∙ ¸ Z ∞ µ ¶
x − au x − au
= P V ≤ | |U | = u f|U| (u)du = Φ f|U | (u)du
0 b 0 b
x−r2 t
Taking y = r we have
" #Z Ã !
r2 t2 ∞ rλ r2 λt
E [exp (tX)] = 2 exp Φ y+ φ(y)dy
2 −∞ s s
" # " Ã !#
r2 t2 rλ r2 λt
= 2 exp E Φ T+ with T N (0, 1)
2 s s
A note on rescalings of the skew-normal distribution 205
h i ³ ´
r2 t2 2
√ λr t
Thus, E [exp (tX)] = 2 exp 2 Φ s2 +r2 λ2
.
Proof of the Proposition 5:
r r r r
H = √ X + √ Y =√ [a |U | + bV ] + √ Y,
2
r +σ 2 2
r +σ 2 2
r +σ 2 r + σ2
2
√
where a = √ λ rsr and b= √ rs . Thus,
s +r2 λ2
2 s2 +r2 λ2
ra r
H = √ |U | + √ (bV + Y )
2
r +σ 2 r + σ2
2
√
ra r b2 + σ 2
= √ |U | + √ W with W ∼ N (0, 1)
r2 + σ 2 r2 + σ2
Now, we have the hypotheses of Proposition 1, then letting A = √ ra
√ √ r2 +σ2
r√ b2 +σ2 A λ rsr
and B = and by using rA2 +sB 2 = r2 s and B
r2 +σ2
, = √
r2 s2 +s2 σ2 +r2 λ2 σ2
³ √ ´
follows that H ∼ SHN λ, r, s2 + r−2 s2 σ 2 + λ2 σ 2 .
P [X ≤ x] = P [aT + bV ≤ x]
Z ∞ ∙ ¸ Z ∞ µ ¶
x − at x − at
= P V ≤ | T = t fT (t)dt = Φ fT (t)dt
c b c b
Z ∞p µ ¶ p
dP [X ≤ x] s/r x − at φ( s/r)t
= φ sε dt
dx c b b Φ( √s2 +r 2 λ2 )
µ ¶ √ Z ∞ " µ ¶ #
1 x rs 1 rst − ax 2
= φ √ sε
exp − √ dt,
r r 2πbΦ( √s2 +r 2 λ2 ) c
2 rsb
rst−ax
taking y = √
rsb
, the result is obtained.
Acknowledgments
References
[4] Elal, D., Elal, S. and Gómez, H. W. La Distribución SN (α, β). Revista
de la Facultad de Ingeniería. 18, pp. 29-31, (2004).
[9] Sahu, S. K., Dey, D. K. and Branco, M. A new class of multivariate skew
distributions with application to Bayesian regression models. Canad. J.
Statist. 31, pp. 129-150, (2003).
Osvaldo Venegas
Departamento de Ciencias Matemáticas y Fı́sicas,
Facultad de Ingenierı́a,
Universidad Católica de Temuco,
Chile,
e-mail: ovenegas@uct.cl
A note on rescalings of the skew-normal distribution 207
David Elal-Olivero
Departamento de Matemática,
Facultad de Ingeniería,
Universidad de Atacama,
Avenida Copayapu 485,
Copiapó,
Chile,
e-mail: delal@mat.uda.cl
and
Héctor W. Gómez
Departamento de Matemáticas,
Facultad de Ciencias Básicas,
Universidad de Antofagasta,
Antofagasta,
Chile,
e-mail: hgomez@uantof.cl