11.07.2015 Views

Bayesian Nonlinear Regression Models with Scale Mixtures of Skew ...

Bayesian Nonlinear Regression Models with Scale Mixtures of Skew ...

Bayesian Nonlinear Regression Models with Scale Mixtures of Skew ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Bayesian</strong> <strong>Nonlinear</strong> <strong>Regression</strong> <strong>Models</strong> <strong>with</strong> <strong>Scale</strong><strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> Normal Distributions: Estimation andCase Influence DiagnosticsVicente G. Cancho (ICMC-USP), Victor H. Lachos(IMECC-UNICAMP) and Marinho Andrade (ICMC-USP)Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:1 / 35


MotivationMotivationThe routine use <strong>of</strong> the normality in nonlinear models (N-NLM) has beenrecently questioned by many authors (see Cysneiros & Vanegas, 2008;Cordeiro et al., 2009, among others).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:2 / 35


MotivationMotivationThe routine use <strong>of</strong> the normality in nonlinear models (N-NLM) has beenrecently questioned by many authors (see Cysneiros & Vanegas, 2008;Cordeiro et al., 2009, among others).Thus, it is <strong>of</strong> practical interest to develop statistical model <strong>with</strong> considerableflexibility in the distributional assumptions <strong>of</strong> the random error term in NLM.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:2 / 35


MotivationMotivationThe routine use <strong>of</strong> the normality in nonlinear models (N-NLM) has beenrecently questioned by many authors (see Cysneiros & Vanegas, 2008;Cordeiro et al., 2009, among others).Thus, it is <strong>of</strong> practical interest to develop statistical model <strong>with</strong> considerableflexibility in the distributional assumptions <strong>of</strong> the random error term in NLM.Cancho et al. (2009) and Xie et al. (2009b,a) has shown the advantage <strong>of</strong>using the skew-normal distribution in the context <strong>of</strong> nonlinear regressionmodels (SN-NLM).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:2 / 35


MotivationMotivationThe routine use <strong>of</strong> the normality in nonlinear models (N-NLM) has beenrecently questioned by many authors (see Cysneiros & Vanegas, 2008;Cordeiro et al., 2009, among others).Thus, it is <strong>of</strong> practical interest to develop statistical model <strong>with</strong> considerableflexibility in the distributional assumptions <strong>of</strong> the random error term in NLM.Cancho et al. (2009) and Xie et al. (2009b,a) has shown the advantage <strong>of</strong>using the skew-normal distribution in the context <strong>of</strong> nonlinear regressionmodels (SN-NLM).We extend the SN-NLM by assuming that the model errors follow scalemixtures <strong>of</strong> skew-normal distributions (SMSN, Branco and Dey, 2001).(skew-normal,skew-t, skew-slash, skew-contaminated normal).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:2 / 35


SMSN DistributionsThe <strong>Skew</strong>-Normal distributionZ ∼ SN(µ, σ 2 , λ) , <strong>with</strong> pdf( ) λ(z − µ)f (z) = 2φ(z; µ, σ 2 )Φ, (Azzalini, 1985)σicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:3 / 35


SMSN DistributionsThe <strong>Skew</strong>-Normal distributionZ ∼ SN(µ, σ 2 , λ) , <strong>with</strong> pdf( ) λ(z − µ)f (z) = 2φ(z; µ, σ 2 )Φ, (Azzalini, 1985)σMarginal stochastic representation:where ∆ = σδ, δ =Z = µ + ∆|T 0 | + Γ 1/2 T 1 ,λ √1+λ 2 , Γ = (1 − δ2 )σ 2 , T 0 and T 1 areindependent standard normal random variables.icente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:3 / 35


SMSN Distributions<strong>Scale</strong> <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong>-Normal distributions(SMSN)Y = µ + κ 1/2 (U)Z,Z ∼ SN(0, σ 2 , λ)U is a positive random variable <strong>with</strong> cdf H(·; ν) (pdf h(·; ν))κ(.) is weight function, this paper we restrict κ(u) = 1/uThe pdf <strong>of</strong> Y is given by:f (y) = 2∫ ∞0φ(y; µ, u −1 σ 2 )ΦNotation: Y ∼ SMSN(µ, σ 2 , λ; H)()u 1/2 λ(y − µ)dH(u; ν),σVicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:4 / 35


SMSN DistributionsExamples <strong>of</strong> SMSN distributionsThe skew–t: Y ∼ ST (µ, σ 2 , λ; ν), U ∼ Gamma(ν/2, ν/2).f (y) =ν+1Γ(2 ) (Γ( ν 2 )√ 1 + d ) −ν+1(√ )2 v + 1Tπνσ νd + ν A; ν + 1 , y ∈ R,icente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:5 / 35


SMSN DistributionsExamples <strong>of</strong> SMSN distributionsThe skew–t: Y ∼ ST (µ, σ 2 , λ; ν), U ∼ Gamma(ν/2, ν/2).f (y) =ν+1Γ(2 ) (Γ( ν 2 )√ 1 + d ) −ν+1(√ )2 v + 1Tπνσ νd + ν A; ν + 1 , y ∈ R,The skew–slash:Y ∼ SSL(µ, σ 2 , λ; ν),U ∼ Beta(ν, 1).f (y) = 2ν∫ 10u ν−1 φ(y; µ, u −1 σ 2 )Φ(u 1/2 A)du, y ∈ R,icente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:5 / 35


SMSN DistributionsExamples <strong>of</strong> SMSN distributionsThe skew–t: Y ∼ ST (µ, σ 2 , λ; ν), U ∼ Gamma(ν/2, ν/2).f (y) =ν+1Γ(2 ) (Γ( ν 2 )√ 1 + d ) −ν+1(√ )2 v + 1Tπνσ νd + ν A; ν + 1 , y ∈ R,The skew–slash:Y ∼ SSL(µ, σ 2 , λ; ν),U ∼ Beta(ν, 1).f (y) = 2ν∫ 10u ν−1 φ(y; µ, u −1 σ 2 )Φ(u 1/2 A)du, y ∈ R,Applications <strong>of</strong> the skew–slash distribution can be found in Wang &Genton (2006).icente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:5 / 35


SMSN DistributionsExamples <strong>of</strong> SMSN distributionsThe skew contaminated normal:Y ∼ SCN(µ, σ 2 , λ; ν, γ). The pdf<strong>of</strong> U is given byh(u|ν) = νI (u=γ) + (1 − ν)I (u=1) , 0 < ν < 1, 0 < γ ≤ 1,f (y) = 2{νφ(y; µ, γ −1 σ 2 )Φ(γ 1/2 A) + (1 − ν)φ(y; µ, σ 2 )Φ(A)}.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:6 / 35


The SMSN nonlinear regression modelThe SMSN nonlinear regression modelY i = η(β, x i ) + ε i , i = 1, . . . , n,Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:7 / 35


The SMSN nonlinear regression modelThe SMSN nonlinear regression modelY i = η(β, x i ) + ε i , i = 1, . . . , n,η(.) is injective and twice continuously differentiable,Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:7 / 35


The SMSN nonlinear regression modelThe SMSN nonlinear regression modelY i = η(β, x i ) + ε i , i = 1, . . . , n,η(.) is injective and twice continuously differentiable,√2ε i ∼ SMSN(−π k 1∆, σ 2 , λ; H) where k r = E[U r |y],Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:7 / 35


The SMSN nonlinear regression modelThe SMSN nonlinear regression modelY i = η(β, x i ) + ε i , i = 1, . . . , n,η(.) is injective and twice continuously differentiable,√2ε i ∼ SMSN(−π k 1∆, σ 2 , λ; H) where k r = E[U r |y],√Y i ∼ SMSN(η(β, x i ) + b∆, σ 2 2, λ; H), <strong>with</strong> b = −π k 1,Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:7 / 35


The SMSN nonlinear regression modelThe SMSN nonlinear regression modelY i = η(β, x i ) + ε i , i = 1, . . . , n,η(.) is injective and twice continuously differentiable,√2ε i ∼ SMSN(−π k 1∆, σ 2 , λ; H) where k r = E[U r |y],√Y i ∼ SMSN(η(β, x i ) + b∆, σ 2 2, λ; H), <strong>with</strong> b = −π k 1,E[Y i ] = η( β, x i ), Var[Y i ] = k 2 σ 2 − b 2 ∆ 2 .Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:7 / 35


<strong>Bayesian</strong> inference<strong>Bayesian</strong> inferenceThe hierarchical representation:Y i |T i = t i ∼ N 1 (η(β, x i ) + ∆t i , U −1iΓ)T i |U i ∼ TN 1 (b, u −1i)I (b, ∞)U i ∼ H(.; ν),where i = 1, . . . , n Γ = (1 − δ 2 )σ 2 , and ∆ = σδ.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:8 / 35


<strong>Bayesian</strong> inference<strong>Bayesian</strong> inferenceLet y = (y 1 , . . . , y n ) ⊤ , x = (x 1 , . . . , x n ) ⊤ , t = (t 1 , . . . , t n ) ⊤ andu = (u 1 , . . . , u n ) ⊤ . It follows that the complete likelihood function <strong>of</strong> θassociated <strong>with</strong> (y, x, t, u) is given byL c (θ|y, x, t, u) ∝n∏[φ 1 (y i ; η(β, x i )+∆t i , u −1 τ)φ 1 (t i ; b, u −1 )I (b,∞) h(u i |ν)].i=1ii(1)Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> <strong>Skew</strong> março (ICMC-USP) Normal 2010Distributions:9 / 35


<strong>Bayesian</strong> inferencePrior distributionβ j = N(µ βj , σ 2 β j), j = 1, . . . , p,∆ ∼ N 1 (µ ∆ , σ 2 ∆ ) andτ −1 ∼ Gamma( ρ 2 , ϱ 2 ).For the skew-t model: ν ∼ exp( ς 2 )I (2,∞),For the skew-slash model: ν ∼ Gamma(a, b) (b ≪ a),For the skew-contaminated normal model: ν ∼ U(0, 1) andγ Beta(a, b) (ν ⊥ γ).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 10 / 35


<strong>Bayesian</strong> inferenceThe posterior distributionThe joint posterior density <strong>of</strong> all unobservable is is given byπ(θ, t, u|y, x)∝ ∏ ni=1 [φ 1(y i ; η(β, x i ) + ∆t i , u −1iτ) (2)φ 1 (t i ; b, u −1i)I (b,∞) h(u i |ν)]π(θ).Distribution (2) is not tractable analytically but MCMC methods such asthe Gibbs sampler, can be used to draw samples, from which features <strong>of</strong>marginal posterior distributions <strong>of</strong> interest can be inferred.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 11 / 35


<strong>Bayesian</strong> inferenceThe full conditional distributionsT i |β, ∆, τ, ν, y, u ∼ TN 1 (µ Ti + b, u −1iMT 2 )I (b, ∞), i = 1, . . . , n;where MT 2 = τ∆ 2 + τ , µ T i= ∆∆ 2 + τ (y i − η(β, x i ) − ∆b),( Bσ2∆|β, τ, ν, y, u, t ∼ N ∆+ τµ ∆1Aσ∆ 2 + τ , Aσ2 ∆ + τ )τσ∆2 ,where A = ∑ ni=1 u it i and B = ∑ ni=1 u it i (y i − η(β, x i )),Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 12 / 35


<strong>Bayesian</strong> inferenceThe full conditional distributionsτ −1 |β, ∆, ν, y, u, t ∼ Gamma( 1 2 (ρ + n), 1 n∑2 (ϱ + u i (y i − η(β, x i ) − ∆t i ) 2 ));i=1β|∆, τν, y, u, t ∝ φ p (β; µ β , D)n∏i=1φ 1 (η(β, x i ); (y i − ∆t i ), u −1iτ),where µ β = (µ β1 , . . . , µ βp ) ⊤ , D = diagonal(σ 2 β 1, . . . , σ 2 β p).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 13 / 35


<strong>Bayesian</strong> inferenceThe full conditional distributionsFor each element <strong>of</strong> u, the density is:π(u i |β, ∆, τ, ν, y, x, t) ∝ u i exp{− 12τ u i(y i − η(β, x i ) − ∆t i ) 2 (3)− 1 2 u i(t i − b) 2 }h(u i |ν),for i = 1, . . . , n. For ν, the density is:n∏π(ν|β, ∆, τ, y, u, t) ∝ π(ν) h(u i |ν). (4)i=1Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 14 / 35


<strong>Bayesian</strong> inferenceThe full conditional distributions(i) <strong>Skew</strong>-t. The density <strong>of</strong> the conditional posterior distribution in (3)takes the form:u i |θ (−ν) , y, t ∼ Gamma(2; ν/2 + C i /2),where C i = 1 τ (y i − η(β, x i ) − ∆t i ) 2 + (t i − b) 2 and the full conditionalposterior density <strong>of</strong> ν isπ(ν|θ (−ν) , y, t, u) ∝ π 1 (ν) × Gamma( nν 2 − 1, 1 ∑(ui − log u i ))I2 (2,∞) ,where π 1 (ν) = (2 ν/2 Γ(ν/2)) −n .Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 15 / 35


<strong>Bayesian</strong> inferenceThe full conditional distributions• <strong>Skew</strong>-slash.In this case, the fully conditional posterior density <strong>of</strong> each u i is:u i |θ (−ν) , y, t ∼ Gamma((nu + 1)/2; C i /2)I (0,1) .Further, the conditional posterior density <strong>of</strong> ν isν|θ (−ν) , y, b, u, t ∼ Gamma(n + a, b − ∑ log u i ).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 16 / 35


<strong>Bayesian</strong> inferenceThe full conditional distributions• <strong>Skew</strong>-contaminated normal distributionThe full conditional posterior density <strong>of</strong> the proportion <strong>of</strong> outliers ν is:(ν|θ (−ν) , y, b, u, t ∼ Beta a + n − ∑ ∑ )u i1 − γ ; b + ui − nγ.1 − γThe conditional posterior density <strong>of</strong> γ is:π(γ|θ (−γ) , y, b, u, t) ∝ ν (n − ∑ ∑u i1 − γ ) ui − nγ× (1 − ν) ( )1 − γ .An interesting Metropolis–Hastings method to update from γ is describedin Rosa et al. (2003).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 17 / 35


<strong>Bayesian</strong> case influence diagnostics<strong>Bayesian</strong> case influence diagnosticsLet K(P, P (−i) ) denote the Kullback-Leibler (K–L) divergence between Pand P (−i) , where P denotes the posterior distribution <strong>of</strong> θ for full data,and P (−i) denotes the posterior distribution <strong>of</strong> θ <strong>with</strong>out the ith case.Specifically,∫[ ]π(θ|D)K(P, P (−i) ) = π(θ|D) logπ(θ|D (−i) dθ. (5))Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 18 / 35


<strong>Bayesian</strong> case influence diagnostics<strong>Bayesian</strong> case influence diagnosticsAs pointed by Peng & Dey (1995) and Cho et al. (2009), calibration <strong>of</strong>K(P, P (−i) ) can be done by solving for p i the equationK(P, P (−i) ) = K [ B(0.5), B(p i ) ] = − log [ 4p i (1 − p i ) ] /2,where B(p) denotes the Bernoulli distribution <strong>with</strong> success probability p.After some algebra it can be shown thatp i = 1 { √1 + 1 − exp [ − 2K(P, P2(−i) ) ]} .This equation implies that 0.5 ≤ p i ≤ 1.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 19 / 35


<strong>Bayesian</strong> case influence diagnostics<strong>Bayesian</strong> case influence diagnosticsFor our model in (7) it can be shown that (5) can be expressed as aposterior expectationK(P, P (−i) ) = log E θ|D{[g(yi |θ)] −1} + E θ|D {log [g(y i |θ)]}= − log(CPO i ) + E θ|D {log [g(y i |θ)]} ,(6)where E θ|D (·) denotes the expectation <strong>with</strong> respect to the joint posteriorπ(θ|D).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 20 / 35


<strong>Bayesian</strong> case influence diagnostics<strong>Bayesian</strong> case influence diagnosticsThus (6) can be computed by sampling from the posterior distribution <strong>of</strong>θ via MCMC methods. Let θ 1 , . . . , θ Q be a sample <strong>of</strong> size Q <strong>of</strong> π(θ|D).Then, a Monte Carlo estimate <strong>of</strong> K(P, P (−i) ) is given bŷ K(P, P (−i) ) = − log(ĈPO i ) + 1 QQ∑log [g(y i |θ q )] . (7)q=1Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 21 / 35


ApplicationsSimulated dataInfluence <strong>of</strong> outlying observationsWe consider the following logistic model:Y i =β 11 + β 2 exp(−β 3 x i ) + ε i, i = 1, · · · , 50, (8)where ε i ∼ SN(− √ 2/π∆, σ 2 , λ), the variable x i ranging from 1 to 50 andwe fixed the parameter values at: β 1 = 30, β 2 = 5, β 3 = .7, σ 2 = 2 andλ = −3.We selected cases 4, 13 and 45 for perturbation. To create influentialobservation in the dataset, we choose one, two or tree <strong>of</strong> these selectedcases and perturbed the response variable as followsỹ i = y i + 4S y , i = 4, 13 and 45, where S y is the standard deviations <strong>of</strong> they ′i s.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 22 / 35


ApplicationsSimulated dataSimulated data1 The following independent priors were adopted in the <strong>Bayesian</strong>computations. β k ∼ log − N 1 (0, 10 3 ), k = 1, 2, 3, ∆ ∼ N 1 (0, 10 3 ),and 1/τ ∼ Gamma(1, 0.01).2 we generated two parallel independent runs <strong>of</strong> the Gibbs sampler <strong>with</strong>size 60000 for each parameter, disregarding the first 10000 iterationsto eliminate the effect <strong>of</strong> the initial values and, to avoid correlationproblems, we considered a spacing <strong>of</strong> size 20, obtaining a sample <strong>of</strong>size 2500.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 23 / 35


ApplicationsSimulated dataInfluence <strong>of</strong> outlying observationsTable: Simulated SN data. Posterior means and standard deviations <strong>of</strong> theparameters from fitting a SN-NLM.Data Perturbed β 1 = 30 β 2 = 5 β 3 = 0.7 σ 2 = 2 λ = −3names case Mean SD Mean SD Mean SD Mean SD Mean SDa None 29.86 0.17 5.05 0.74 0.69 0.05 2.59 0.84 -2.81 1.67b 4 30.24 0.35 10.56 6.93 0.99 0.21 10.65 3.07 1.14 1.18c 13 30.48 0.32 5.91 2.74 0.73 0.12 11.04 2.55 3.60 1.11d 45 30.51 0.32 6.13 3.11 0.74 0.12 11.20 2.67 3.59 1.06e {4,45} 30.93 0.43 10.06 7.25 0.94 0.21 19.60 4.33 4.41 1.34f {4,13,45} 31.55 0.50 10.68 8.74 0.93 0.24 30.18 6.86 5.07 1.41Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 24 / 35


ApplicationsSimulated dataSimulated SN data. Index plots <strong>of</strong> K(P, P (−i) ) from fittinga SN-NLM.(a)(b)(c)K−L divergence0 1 2 3 4 5 6K−L divergence0 1 2 3 4 5 64K−L divergence0 1 2 3 4 5 6130 10 20 30 40 50Index(d)0 10 20 30 40 50Index(e)0 10 20 30 40 50Index(f)K−L divergence0 1 2 3 4 5 645K−L divergence0 1 2 3 4 5 6445K−L divergence0 1 2 3 4 5 6413450 10 20 30 40 50Index0 10 20 30 40 50Index0 10 20 30 40 50IndexVicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 25 / 35


ApplicationsSimulated dataInfluence <strong>of</strong> outlying observationsTable: Simulated SN data. Posterior means and standard deviations <strong>of</strong> theparameters from fitting a ST–NLM.Data β 1 β 2 β 3 σ 2 λ νnames Mean SD Mean SD Mean SD Mean SD Mean SD Mean SDa 29.87 0.17 5.03 0.76 0.69 0.05 1.88 0.76 -2.26 1.36 12.31 8.95b 29.89 0.24 5.43 1.41 0.72 0.09 1.07 0.42 -0.55 0.96 2.88 0.79c 29.92 0.23 5.090 0.85 0.69 0.05 0.99 0.43 -0.77 0.98 2.87 0.80d 29.92 0.23 5.06 0.84 0.69 0.05 0.99 0.42 -0.77 0.98 2.87 0.80e 30.05 0.26 5.31 1.29 0.71 0.08 1.01 0.47 0.04 0.78 2.43 0.42f 30.22 0.28 5.19 1.32 0.70 0.08 1.17 0.58 0.47 0.72 2.31 0.31Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 26 / 35


ApplicationsSimulated dataSimulated SN data. Index plots <strong>of</strong> K(P, P (−i) ) from fittinga ST-NLM.(a)(b)(c)K−L divergence0 1 2 3 4 5 6K−L divergence0 1 2 3 4 5 64K−L divergence0 1 2 3 4 5 6130 10 20 30 40 50Index(d)0 10 20 30 40 50Index(e)0 10 20 30 40 50Index(f)K−L divergence0 1 2 3 4 5 645K−L divergence0 1 2 3 4 5 6445K−L divergence0 1 2 3 4 5 64 13 450 10 20 30 40 50Index0 10 20 30 40 50Index0 10 20 30 40 50IndexVicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 27 / 35


ApplicationsSimulated dataInfluence <strong>of</strong> outlying observationsTable: Simulated data. Comparison between SN–NLM and ST–NLM fitting byusing different <strong>Bayesian</strong> criteria.Data SN–NLM ST–NLMnames B DIC EAIC EBIC B DIC EAIC EBICa -76,802 153.338 157.928 167.4881 -77.141 153.949 159.539 171.011b -116.794 218.119 223.519 233.079 -93.142 176.988 182.389 193.861c -112.506 216.516 221.35 230.910 -90.133 176.508 182.578 194.050d -113.861 216.51 221.594 231.154 -90.766 176.834 182.824 194.296e -126.484 240.758 245.334 254.894 -99.465 195.338 201.348 212.820f -134.148 259.939 264.85 274.4101 -106.483 210.077 215.897 227.369∑where B = n log(ĈPO i ).i=1Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 28 / 35


ApplicationsThe oil palm datasetThe oil palm datasetWe consider a <strong>Bayesian</strong> analysis <strong>of</strong> the data set presented in Foong (1999)that describe the oil palm yield. Assuming a nonlinear growth-curvemodel, we fit a NLM to the data as specified by Cancho et al. (2009)Y i =β 11 + β 2 exp(−β 3 x i ) + ε i, (9)where ε iiid ∼ SMSN(−√2π k 1∆, σ 2 , λ, H), for i = 1, . . . , 19. In our analysiswe assume SN-NLM, ST-NLM, SCN-NLM and SSL-NLM from the SMSNclass for comparative purposes.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 29 / 35


ApplicationsThe oil palm datasetThe oil palm datasetTable: Oil palm yield data set. Summary results from the posterior distribution,mean and standard deviation (SD) for parameter under SMSN distributions.SN-NLM ST-NLM SSL-NLM SCN-NLMParameter Mean SD Mean SD Mean SD Mean SDβ 1 37.251 0.491 37.565 0.469 37.461 0.538 37.421 0.478β 2 42.025 14.538 40.343 10.739 40.144 12.407 40.328 12.284β 3 0.730 0.065 0.717 0.050 0.712 0.057 0.720 0.058σ 2 6.447 2.890 1.966 1.482 2.674 1.907 2.484 2.022λ -2.603 2.070 -1.823 1.369 -3.423 2.108 -2.241 1.710ν - - 3.415 1.267 1.967 0.539 0.562 0.197γ - - - - - - 0.389 0.114Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 30 / 35


ApplicationsThe oil palm datasetThe oil palm datasetTable: Oil palm yield data set. Comparison between SMSN–NLM by usingdifferent <strong>Bayesian</strong> criteria.criterion SN–NLM ST–NLM SSL–NLM SCN–NLMB -40.007 -38.949 -39.459 -39.146DIC 75.469 72.626 73.561 73.183EAIC 85.595 83.349 84.561 84.5612EBIC 90.317 88.015 90.227 90.172Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 31 / 35


ApplicationsThe oil palm datasetThe oil palm datasetThe K-L divergence and related calibration are computed. We do not findhighly influential cases. The K(P, P (−i) ) is smaller that 0.71 and thecorresponding calibrations are smaller than 1. However, for SN-NLM,cases 13,18, 10 and 15 had larger K(P, P (−i) ) when compared <strong>with</strong> theother cases.Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 32 / 35


ApplicationsThe oil palm datasetThe oil palm datasetTable: Oil palm yield data. Case influence diagnostics.SN–NLM ST–NLM SSL-NLM SCN–NLMCase K(P, P (−i) ) Cal. K(P, P (−i) ) Cal. K(P, P (−i) ) Cal. K(P, P (−i) ) Cal.13 0.709 0.935 0.549 0.908 0.699 0.934 0.571 0.91218 0.703 0.934 0.290 0.832 0.307 0.818 0.561 0.91110 0.306 0.838 0.304 0.836 0.297 0.835 0.214 0.79515 0.272 0.823 0.244 0.810 0.258 0.817 0.256 0.816Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 33 / 35


ApplicationsThe oil palm dataset(a)(b)K−L divergence0.0 0.5 1.0 1.5 2.010131518K−L divergence0.0 0.5 1.0 1.5 2.0101315 185 10 15Index(c)5 10 15Index(d)K−L divergence0.0 0.5 1.0 1.5 2.0101315 18K−L divergence0.0 0.5 1.0 1.5 2.0101315185 10 15Index5 10 15IndexVicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 34 / 35


ConclusionsConclusions• We propose a interesting, useful and novel class <strong>of</strong> asymmetric heavy-tailed NLM.• We discus the use <strong>of</strong> Markov Chain Monte Carlo methods as an alternative way to get<strong>Bayesian</strong> inference for the proposed model• <strong>Bayesian</strong> case influence diagnostics based on the Kullback-Leibler divergence in order tostudy the sensitivity <strong>of</strong> the <strong>Bayesian</strong> estimates under perturbations in the model/data.• Our analysis indicates that a ST-NLMVicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 35 / 35


ReferencesBranco, M. D., Dey, D. K., 2001. A general class <strong>of</strong> multivariate skew-elliptical distributions. Journal <strong>of</strong> Multivariate Analysis 79,99–113.Cancho, V. C., Lachos, V. H. & Ortega, E. M. M. (2009). A nonlinear regression model <strong>with</strong> skew-normal errors. StatisticalPapers, doi:10.1007/s00362-008-0139-y.Cho, H., Ibrahim, J. G., Sinha, D. & Zhu, H. (2009). <strong>Bayesian</strong> case influence diagnostics for survival models. Biometrics, 65,116–124.Cordeiro, G. M., Cysneiros, A. H. M. A., Cysneiros, F. J. A., 2009. Corrected maximum likelihood estimators im heteroscedasticsymmetric nonlinear models. Journal <strong>of</strong> Statistical Computation and Simulation doi:10.1080/00949650802706420.Cysneiros, F. J. A., Vanegas, L. H., (2008). Residuals and their statistical properties in symmetrical nonlinear models. Statistics& Probability Letters 78, 3269–3273.Foong, F. S. (1999). Impact <strong>of</strong> mixture on potential evapotranspiration, growth and yield <strong>of</strong> palm oil. PORIM Interl. Palm OilCong. (Agric.), pages 265–287.Peng, F. & Dey, D. (1995). <strong>Bayesian</strong> analysis <strong>of</strong> outlier problems using divergence measures. The Canadian Journal <strong>of</strong>Statistics/La Revue Canadienne de Statistique, 23(2), 199–213.Rosa, G. J. M., Padovani, C. R. & Gianola, D. (2003). Robust linear mixed models <strong>with</strong> normal/independent distributions andbayesian mcmc implementation. Biometrical Journal, 45, 573–590.Spiegelhalter, D. J., Best, N. G., Carlin, B. P. & van der Linde, A. (2002). <strong>Bayesian</strong> measures <strong>of</strong> model complexity and fit.64(4), 583–639.Wang, J. & Genton, M. G. (2006). The multivariate skew-slash distribution. Journal <strong>of</strong> Statistical Planning and Inference, 136,209–220.Xie, F. C., Wei, B. C., Lin, J. G., 2009a. Homogeneity diagnostics for skew-normal nonlinear regression models. Statistics &Probability Letters 79, 821–827.Xie, F. C., Lin, J. G., Wei, B. C., 2009b. Diagnostics for skew-normal nonlinear regression models <strong>with</strong> ar(1) errors.Computational Statistics & Data Analysis, (in press).Vicente G. Cancho (ICMC-USP), Victor H. Lachos <strong>Bayesian</strong> (IMECC-UNICAMP) <strong>Nonlinear</strong> <strong>Regression</strong> and <strong>Models</strong> Marinho<strong>with</strong> Andrade <strong>Scale</strong>(ICMC-USP) <strong>Mixtures</strong> <strong>of</strong> março <strong>Skew</strong> (ICMC-USP) Normal - 2010 Distributions: 35 / 35

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!