• Aucun résultat trouvé

Robust and accurate inference for generalized linear models

N/A
N/A
Protected

Academic year: 2022

Partager "Robust and accurate inference for generalized linear models"

Copied!
10
0
0

Texte intégral

(1)

Robust and Accurate Inference for Generalized Linear Models:

Complete Computations

by

Serigne N. Lˆo

The George Institute University of Sydney, Australia

and

Elvezio Ronchetti

Department of Econometrics University of Geneva, Switzerland

February 2008 / Revised: May 2009

(2)

APPENDIX A

To determine λ(β), we calculate

−n∂Kψ(λ;β)

∂λ = −

n

X

i=1

∂Kψi(λ;β)

∂λ

=

n

X

i=1

∂(µiλTxi+ b(θ0i)−b(θa(φ)0iTxia(φ)))

∂λ

=

n

X

i=1

ixi−b00iTxia(φ))·xi]

= 0

Sinceg(·) is the canonical link,θi =xTi β, and −n2∂λ∂λKψ(λ;β)T is negative definite, this equation has a unique solution given by λ(β) = β−βa(φ)0.

Then, by replacing this expression for λ in Kψ and after simplification we obtain

−Kψi(λ(β);β) = (θi−θ0ii−(b(θi)−b(θ0i))

a(φ) ,

and

h(β) = −Kψ(λ(β);β)

= 1

n

n

X

i=1

−Kψi(λ(β);β)

= 1

n

n

X

i=1

i−θ0ii−(b(θi)−b(θ0i)) a(φ)

= 1

n

n

X

i=1

b0(xTi β)xTi (β−β0)−(b(xTi β)−b(xTi β0))

a(φ) .

APPENDIX B Calculation of the integrals Ii1, Ii2, Ii3

(3)

(i)

Ii1 = Z

ri<−c

e−λ

Tc w(xi)

V1/2(µi)µ0i−λT˜a(β)

·e0ia(φ)−b(θ0i) ·ed(y;φ)·dy

= e−λ

Tc w(xi)

V1/2(µi)µ0i−λT˜a(β)

· Z

ri<−c

e0ia(φ)−b(θ0i) ·ed(y;φ)·dy

= e−λ

Tc w(xi)

V1/2(µi)µ0i−λT˜a(β)

· Z

y<−cV1/2i)+µi

e0ia(φ)−b(θ0i) ·ed(y;φ)·dy

= e−λ

Tc w(xi)

V1/2(µi)µ0i−λT˜a(β)

·P(Zi ≤ −cV1/2i) +µi)

whereZiis a random variable distributed according to the exponential family (2) with parameter θ0i.

(ii)

Ii2 = Z

|ri|<c

e

yλT µ0 i V1/2(µi)

w(xi) V1/2(µi) ·e

λT µiµ0 i V1/2(µi)

w(xi)

V1/2(µi) ·e−λT˜a(β)·e

0i−b(θ0i)

a(φ) ·ed(y;φ)·dy

= Z

|ri|<c

e

yλT µ0 iw(xi) V(µi) ·e

λT µiµ0 iw(xi)

V(µi) ·e−λT˜a(β)·e

−b(θ0i) a(φ) ·e

0i

a(φ) ·ed(y;φ)·dy

= Z

|ri|<c

e

λT µiµ0 iw(xi)

V(µi) ·e−λT˜a(β)·e−b(θa(φ)0i) ·e

y(θ0i+λT µ0

iw(xi)a(φ) V(µi) )

a(φ) ·ed(y;φ)·dy

= Z

|ri|<c

e

λT µiµ0 iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

·e

y(θ0i+λT µ0

iw(xi)a(φ)

V(µi) )−b(θ0i+λT µ0 iw(xi)a(φ) V(µi) )

a(φ) ·ed(y;φ)·dy

= e

λT µiµ0 iw(xi)

V(µi) ·e−λTa(β)˜ ·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

· Z

|ri|<c

e

y(θ0i+λT µ0 iw(xi)a(φ)

V(µi) )−b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )

a(φ) ·ed(y;φ)·dy

= e

λT µiµ0 iw(xi)

V(µi) ·e−λTa(β)˜ ·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

. P(−cV1/2i) +µi < Zλi < cV1/2i) +µi)

(4)

whereZλi is a random variable distributed according to the exponential family (2) with parameter [θ0i+λTµ0iVw(xii))a(φ)].

(iii) This result can be easily derived as in (i).

We obtain:

Ii3 =eλ

Tc w(xi)

V1/2(µi)µ0i−λT˜a(β)

·P(Zi ≥cV1/2i) +µi).

APPENDIX C

For i= 1, ..., n, we have from Appendix B:

∂Ii1

∂λ + ∂Ii2

∂λ +∂Ii3

∂λ = −hcw(xi0i

V1/2i) + ˜a(β)i

·Ii1

− hµiµ0iw(xi)

V(µi) + ˜a(β)− µ0iw(xi)

V(µi) b00iTµ0iw(xi)a(φ) V(µi) )i

·Ii2 + e

λT µiµ0 iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

h ∂

∂λP(|Zλi |< c)i + hcw(xi0i

V1/2i) −a(β)˜ i

·Ii3

= −hcw(xi0i

V1/2i) + ˜a(β)i

·Ii1

− hµiµ0iw(xi)

V(µi) + ˜a(β)− µ0iw(xi)

V(µi) b00iTµ0iw(xi)a(φ) V(µi) )i

·Ii2

+ e

λT µiµ0 iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ) · µ0iw(xi) V(µi) EZ|rλi

i|<c[Y]

− µ0iw(xi)

V(µi) b00iTµ0iw(xi)a(φ) V(µi) )·Ii2

+ hcw(xi0i

V1/2i) −a(β)˜ i

·Ii3

(5)

= −hcw(xi0i

V1/2i) + ˜a(β)i

·Ii1

− hµiµ0iw(xi)

V(µi) + ˜a(β)i

·Ii2

+ e

λT µiµ0 iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0 iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ) · µ0iw(xi)

V(µi) EZ|riλi|<c[Y] + hcw(xi0i

V1/2i) −˜a(β)i

·Ii3.

Furthermore,

∂s(λ;β)

∂λ =

n

X

i=1

∂h∂Ii1

∂λ +∂Ii2∂λ +∂Ii3∂λ Ii1+Ii2+Ii3

i

∂λ

=

n

X

i=1

2(Ii1+Ii2+Ii3)

∂λ∂λT ·(Ii1+Ii2+Ii3)−[∂I∂λi1 + ∂I∂λi2 +∂I∂λi3]·[∂I∂λi1 + ∂I∂λi2 +∂I∂λi3]T

(Ii1+Ii2+Ii3)2 .

LetS1i and S2i such that:

S1i : = ∂2(Ii1+Ii2+Ii3)

∂λ∂λT ·(Ii1+Ii2+Ii3)

= ∂(∂I∂λi1 + ∂I∂λi2 +∂I∂λi3)

∂λT ·(Ii1+Ii2+Ii3)

= (Ii1+Ii2+Ii3)·n

£cw(xi0i

V1/2i) + ˜a(β)¤

·£cw(xi0i

V1/2i) + ˜a(β)¤T

·Ii1

+ £µiµ0iw(xi)

V(µi) + ˜a(β)¤

·£µiµ0iw(xi)

V(µi) + ˜a(β)¤T

·Ii2

− £µiµ0iw(xi)

V(µi) + ˜a(β)¤

·hµiw(xi) V(µi)

iT

E|rZλi

i|<c[Y] . eλT µiµ

0iw(xi)

V(µi) ·e−λTa(β)˜ ·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

− hµiµ0iw(xi)

V(µi) + ˜a(β)i

.hµ0iw(xi) V(µi)

iT

.E|rZλi

i|<c[Y] . eλT µiµ

0iw(xi)

V(µi) ·e−λTa(β)˜ ·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

+ eλT µiµ

0iw(xi)

V(µi) ·e−λTa(β)˜ ·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ) .hµ0iw(xi)i

·hµ0iw(xi)iT

.E|rZλi|<c[Y2]

(6)

+ £cw(xi0i

V1/2i) −˜a(β)¤

·£cw(xi0i

V1/2i) −˜a(β)¤T

·Ii3o

and

S2i : = £∂Ii1

∂λ +∂Ii2

∂λ + ∂Ii3

∂λ

¤·£∂Ii1

∂λ +∂Ii2

∂λ + ∂Ii3

∂λ

¤T

= £cw(xi0i

V1/2i) + ˜a(β)¤

·£cw(xi0i

V1/2i) + ˜a(β)¤T

·Ii12 + £µiµ0iw(xi)

V(µi) + ˜a(β)¤

·£µiµ0iw(xi)

V(µi) + ˜a(β)¤T

·Ii22 + e−2λT µiµ

0iw(xi)

V(µi) ·e−2λT˜a(β)·e

2b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−2b(θ0i)

a(φ)

·£µ0iw(xi) V(µi)

¤·£µ0iw(xi) V(µi)

¤T

E|rZiλi|<c[Y]¤2

+ £cw(xi0i

V1/2i) −a(β)˜ ¤

·£cw(xi0i

V1/2i) −a(β)˜ ¤T

·Ii32 + 2·£cw(xi0i

V1/2i) + ˜a(β)¤£µiµ0iw(xi)

V(µi) + ˜a(β)¤T

Ii1Ii2

− 2·£cw(xi0i

V1/2i) + ˜a(β)¤

·hµ0iw(xi) V(µi)

iT

.E|rZλi

i|<c[Y]·Ii1 . eλT µiµ

0iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

− 2£cw(xi0i

V1/2 −˜a(β)¤

·£cw(xi0i

V1/2 + ˜a(β)¤T

·Ii1·Ii3

− 2·£µiµ0iw(xi)

V(µi) + ˜a(β)¤

·hµ0iw(xi) V(µi)

iT

.E|rZλi

i|<c[Y]·Ii2

·eλT µiµ

0iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ)

− 2·£µiµ0iw(xi)

V(µi) + ˜a(β)¤£cw(xi0i

V1/2i) −˜a(β)¤T

Ii2·Ii3 + 2·£cw(xi0i

V1/2i) −˜a(β)¤

·hµ0iw(xi) V(µi)

iT

.E|rZiiλ|<c[Y]·Ii3

·eλT µiµ

0iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ) .

(7)

Then,

∂s(λ;β)

∂λ =

n

X

i=1

[S1i −S2i] (Ii1+Ii2 +Ii3)2

=

n

X

i=1

0i·µ0Ti ]

(Ii1+Ii2 +Ii3)2w2(xi)n

£ c

V1/2i)− µi

V(µi)

¤2

·Ii1Ii2

2 c

V1/2i)

¤2

Ii1Ii3

+ £ c

V1/2i)+ µi V(µi)

¤2

·Ii2Ii3 + 2·eλT µiµ

0iw(xi)

V(µi) ·e−λTa(β)˜ ·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ) · 1

V(µi) . £

( c

V1/2i)− µi

V(µi))·Ii1−( c

V1/2i) + µi

V(µi))·Ii3

¤·E|rZλii|<c[Y]

+ eλT µiµ

0iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0 iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ) · 1

V2i) . £

Ii1+Ii2+Ii3¤

·E|rZλi

i|<c[Y2]

− £

eλT µiµ

0iw(xi)

V(µi) ·e−λT˜a(β)·e

b(θ0i+λT µ0

iw(xi)a(φ) V(µi) )−b(θ0i)

a(φ) ¤2

· 1

V2i)·£ E|rZλi

i|<c[Y]¤2o .

APPENDIX D

Special cases

(i) Yi ∼N(µi, σ2)

b(θi) = θ22i a(φ) =σ2

and in this case ˜a(β) = 0. Then, we have :

∂s(λ;β)

∂λ =

n

X

i=1

xixTi · w2(xi) (Ii1+Ii2+Ii3)2

c−xTi β¢2

·Ii1Ii2 + ¡

2.c¢2

·Ii1Ii3

c+xTi β¢2

·Ii2Ii3

+ 2·exTiλw(xi)xTi(2β0−β)+(xTiλw(xi)σ)2 . £

(c−xTi β)Ii1−(c+xTi β)·Ii3¤

·E|rZλi

i|<c[Y]

+ exTiλw(xi)xTi(2β0−β)+(xTiλw(xi)σ)2 ·[Ii1+Ii2+Ii3]·E|rZiλ

i|<c[Y2]

(8)

− £

exTiλw(xi)xTi(2β0−β)+(xTiλw(xi)σ)2¤2

·£ E|rZiλ

i|<c[Y]¤2o

=

n

X

i=1

xixTi ·Ai(λ),

where Ai(λ) is scalar function defined by

Ai(λ) = w(xi)

(Ii1+Ii2+Ii3)2 · n ¡

c−xTi β¢2

·Ii1Ii2+¡ 2.c¢2

·Ii1Ii3

c+xTi β¢2

·Ii2Ii3 + 2·exTiλw(xi)xTi(2β0−β)+(xTiλw(xi)σ)2

. £

(c−xTi β)Ii1−(c+xTi β)·Ii3

¤·E|rZiλi|<c[Y]

+ exTiλw(xi)xTi(2β0−β)+(xTiλw(xi)σ)2 ·[Ii1+Ii2+Ii3]·E|rZλii|<c[Y2]

− £

exTiλw(xi)xTi(2β0−β)+(xTiλw(xi)σ)2¤2

·£ E|rZλi

i|<c[Y]¤2o .

(ii) Yi ∼P(µi)

b(θ) = eθ, a(φ) = 1 Then, we have :

∂s(λ;β)

∂λ =

n

X

i=1

xixTi · w2(xi)·e2xTiβ (Ii1+Ii2+Ii3)2

ce12xTiβ−1¢2

·Ii1Ii2

+ ¡

2.ce12xTiβ¢2

·Ii1Ii3

ce12xTiβ+ 1¢2

·Ii2Ii3

+ 2·e−xTiλwexTi β−λT˜a(β).e[exTi0+w(xi)λ)−exTi β0]·e−xTiβ ·E|rZλi

i|<c[Y]

·£¡

ce12xTiβ −1¢

·Ii1−¡

ce12xTiβ + 1¢

·Ii3

¤

+ e−xTiλwexTi β−λT˜a(β).e[exTi0+w(xi)λ)−exTi β0].e−2xTiβ[Ii1+Ii2+Ii3]·E|rZiiλ|<c[Y2]

− £

e−xTiλwexTi β−λT˜a(β).e[exTi0+w(xi)λ)−exTi β0]¤2

.e−2xTiβ·³

E|rZiλi|<c[Y]´2o

=

n

X

i=1

xixTi ·Ai(λ),

(9)

where Ai(λ) is scalar function defined by

Ai(λ) = w2(xi)·e2xTiβ (Ii1+Ii2+Ii3)2 n ¡

ce12xTiβ −1¢2

·Ii1Ii2

2.ce12xTiβ¢2

·Ii1Ii3

ce12xTiβ + 1¢2

·Ii2Ii3 + 2·e−xTiλwexTi β−λT˜a(β).e[exTi0+w(xi)λ)−exTi β0]·e−xTiβ·E|rZλi

i|<c[Y]

·£¡

ce12xTiβ−1¢

·Ii1−¡

ce12xTiβ+ 1¢

·Ii3

¤

+ e−xTiλwexTi β−λTa(β)˜ .e[exTi0+w(xi)λ)−exTi β0].e−2xTiβ[Ii1+Ii2 +Ii3]·E|rZiλi|<c[Y2]

− £

e−xTiλwexTi β−λT˜a(β).e[exTi0+w(xi)λ)−exTi β0]¤2

.e−2xTiβ ·³

E|rZλii|<c[Y]´2 o .

(iii) Yi ∼Bin(m, πi)

b(θ) = m·log(1 +eθ), a(φ) = 1 Then, we have :

∂s(λ;β)

∂λ =

n

X

i=1

xixTi w2(xi)e2xTiβ (Ii1 +Ii2+Ii3)2

n¡ c−√

m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢2

Ii1Ii2

+ ¡ 2c

√m·e12xTiβ(1 +exTiβ)

¢2

Ii1Ii3+¡ c+√

m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢2

Ii2Ii3

+ 2.¡1 +xTi β0+xTiλw(xi) 1 +β0Txi

¢m

.e

−mxTi λw(xi)exTi β 1+exTiβ

.e−λT˜a(β)· 1 mexTiβ . £¡ c−√m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢Ii1−¡ c+√m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢Ii3¤

·E|rZλi

i|<c[Y] + ¡1 +xTi β0+xTi λw(xi)

1 +xTi β0

¢m

.e

−mxTi λw(xi)exTi β 1+exTiβ

.e−λT˜a(β)· 1 m2e2xTiβ . £

Ii1 +Ii2+Ii3

¤·E|rZλii|<c[Y2]

− ¡1 +xTi β0+xTi λw(xi) 1 +xTi β0

¢m

.e

−mxTi λw(xi)exTi β 1+exTiβ

.e−λT˜a(β)· 1 m2e2xTiβ

£E|rZλii|<c[Y]¤2o

=

n

XxixTi ·Ai(λ),

(10)

where Ai(λ) is scalar function defined by

Ai(λ) = w2(xi)e2xTiβ (Ii1+Ii2+Ii3)2

n¡ c−√m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢2

Ii1Ii2

+ ¡ 2c

√m·e12xTiβ(1 +exTiβ)

¢2

Ii1Ii3+¡ c+√m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢2

Ii2Ii3

+ 2.¡1 +xTi β0+xTi λw(xi) 1 +xTi β0

¢m

.e

−mxTi λw(xi)exTi β 1+exTiβ

.e−λT˜a(β)· 1 mexTiβ . £¡ c−√

m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢Ii1−¡ c+√

m·e12xTiβ

√m·e12xTiβ(1 +exTiβ)

¢Ii3

¤·E|rZλii|<c[Y]

+ ¡1 +xTi β0+xTi λw(xi) 1 +xTi β0

¢m

.e

−mxTi λw(xi)exTi β 1+exTiβ

.e−λT˜a(β)· 1 m2e2xTiβ . £

Ii1+Ii2+Ii3¤

·E|rZiλ

i|<c[Y2]

− ¡1 +xTi β0+xTi λw(xi) 1 +xTi β0

¢m

.e

−mxTi λw(xi)exTi β 1+exTiβ

.e−λT˜a(β)· 1 m2e2xTiβ

£E|rZλi

i|<c[Y]¤2o .

Références

Documents relatifs

In the presence of small deviations from the model (Figure 2), the χ 2 approximation of the classical test is extremely inaccurate (even for n = 100), its saddlepoint version and

By starting from a natural class of robust estimators for generalized linear models based on the notion of quasi-likelihood, we define robust deviances that can be used for

In spite of their second order accuracy, standard classical inference based on saddlepoint technique can be drastically affected by small deviations from the underlying assumptions

By starting from a natural class of robust estimators for generalized linear models based on the notion of quasi-likelihood, we define robust deviances that can be used for

The IF of this globally weighted robust (GWR) estimator with bounding constant c = 4 for the parameters of the mixed GLT model with two binary and three normal manifest variables,

The robustness properties of the approximate ML estimator are analysed by means of the IF. For the simulations we will consider the case of a mixture of normal and binary variables

In this work, we propose a new estimator for the parameters of a GLLVM, the LAMLE, based on a Laplace approximation to the likelihood function and which can be computed even for

Laird, Lange, and Stram (1987) consider the EM algorithm (Dempster, Laird and Rubin, 1977) for both maximum likelihood and restricted maximum like- lihood estimation in a