Article
Reference
Robust Inference for Generalized Linear Models
CANTONI, Eva, RONCHETTI, Elvezio
Abstract
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 stepwise model selection as in the classical framework. We derive the asymptotic distribution of tests based on robust deviances, and we investigate the stability of their asymptotic level under contamination. The binomial and Poisson models are treated in detail. Two applications to real data and a sensitivity analysis show that the inference obtained by means of the new techniques is more reliable than that obtained by classical estimation and testing procedures.
CANTONI, Eva, RONCHETTI, Elvezio. Robust Inference for Generalized Linear Models.
Journal of the American Statistical Association , 2001, vol. 96, no. 455, p. 1022-1030
DOI : 10.1198/016214501753209004
Available at:
http://archive-ouverte.unige.ch/unige:22899
Disclaimer: layout of this document may differ from the published version.
Models
Eva Cantoni and Elvezio Ronhetti
Department of Eonometris
University of Geneva
CH - 1211 Geneva 4, Switzerland
May 1999
Revised January 2001
Abstrat
Bystartingfromanaturallassofrobustestimatorsforgeneralizedlinear
models based on the notion of quasi-likelihood, we dene robust devianes
that an be used for stepwise model seletion as in the lassialframework.
We derivetheasymptotidistributionof testsbasedonrobustdevianesand
weinvestigatethestabilityoftheirasymptotilevelunderontamination. The
binomial and Poisson modelsare treated indetail. Two appliations to real
data anda sensitivityanalysis show thatthe inferene obtainedby meansof
thenewtehniquesismorereliablethanthatobtainedbylassialestimation
and testing proedures.
Generalizedlinearmodels(MCullaghandNelder,1989)areapowerfulandpopular
tehnique for modeling a large variety of data. In partiular, generalized linear
models allowtomodeltherelationshipbetween thepreditors andafuntionof the
mean of the response for ontinuous and disrete response variables. The response
variablesY
i
, for i=1;:::;n are supposed toome from adistribution belonging to
the exponential family, suh that E[Y
i
℄=
i
and V[Y
i
℄=V(
i
) for i=1;:::;n and
i
=g(
i )=x
T
i
; i=1;:::;n; (1)
where 2IR p
isthe vetor of parameters, x
i 2IR
p
, and g(:)is the link funtion.
Thenon-robustnessof themaximumlikelihoodestimatorforhas beenstudied
extensively in the literature: f. for instane the early work of Pregibon (1982) on
logisti regression, Stefanski, Carroll, and Ruppert (1986), Kunsh, Stefanski, and
Carroll (1989), Morgenthaler (1992), and Rukstuhl and Welsh (1999). In more
reent work, Preisserand Qaqish(1999) onsider alass of robustestimators inthe
general framework of generalized estimating equations.
The quasi-likelihood estimatorof the parameter of model (1)(see Wedderburn,
1974,MCullaghandNelder,1989,andHeyde,1997)sharesthesamenon-robustness
properties. This estimatoristhe solutionof the system of estimating equations
n
X
i=1
Q(y
i
;
i )=
n
X
i=1 (y
i
i )
V(
i )
0
i
=0; (2)
where 0
i
=
i
,andQ(y
i
;
i
)isthequasi-likelihoodfuntion. Thesolutionof (2)is
an M-estimator(see Huber, 1981, and Hampel, Ronhetti, Rousseeuw, and Stahel,
1986) dened by the sore funtion
~
(y
i
;
i ) =
(y
i
i )
V(
i )
0
i
. Its inuene funtion
(Hampel, 1974 and Hampel et al., 1986) is proportional to
~
and is unbounded.
explanatory variables x
i
an have a large inuene on the estimator. Thus, the
quasi-likelihood estimator { as well as the maximum likelihood estimator { is not
robust. Several robust alternatives have been proposed in the literature; see the
referenes given above.
However, in spite of the fair amount of existing literature, robust inferene for
generalized linear models seems to be very limited. Moreover, only the logisti
regression situation is usually onsidered in detail, and the problem of developing
robust alternatives tolassial tests is not addressed globally forthe whole lass of
generalized linear models.
In this paper we propose a robust approah to inferene based on robust de-
vianes whih are natural generalizationsof quasi-likelihoodfuntions. Our robust
devianes are based on the same lass of robust estimators as that proposed by
Preisser and Qaqish (1999) in the more general setup of generalized estimating
equations. Although these estimators are not optimally robust, they form a lass
of M-estimators easy to deal with, and whih admits handy inferene not only for
logisti regression but for the whole lass of generalized linear models.
One ould argue that two alternative approahes ould be onsidered. A rst
possibility would be to view variable seletion as a parametri hypothesis and to
use Wald, sore orlikelihoodratio tests for whih robustversions are available;see
e.g. Heritier and Ronhetti (1994) and Markatou and He (1994). Whilethis would
in priniple be feasible, Wald and sore tests do not seem to be used muh in the
lassialanalysisofgeneralizedlinearmodels. Moreover,robustlikelihoodratiotests
annot be proposed in this ase, beause the optimal robust sore funtion does
not admit an analyti primitive funtion and numerial integration in the spae
approahwould betorelyonthe robustmodelseletionbasedonAkaikeCriterion,
Mallows' C
p
or similar tehniques; see e.g. Ronhetti and Staudte (1994), Sommer
and Huggins (1996) and Ronhetti (1997) for a review. This approah has the
advantage to perform a full model searh. However, when the number of variables
is moderate to large suh a full searh is impossible and a stepwise seletion is the
only feasible alternative.
For these reasons and in view of the importane of the notion of deviane for
model buildingin generalized linear models, we propose robust devianesbased on
generalizations of quasi-likelihood funtions. The general struture of the lassial
approah by quasi-likelihood is preserved, whih oers the advantage of having ro-
bust tools playing the same role as devianes, anova tables, stepwise proedures,
and so on.
The paper isorganized asfollows. In the next setionwe disuss robustestima-
tors of a generalized linear modelbased on quasi-likelihood. As an illustration, we
fous inpartiular on the estimationof binomialand Poissonmodels. In Setion3,
we disuss inferene and propose afamily of test statistis for modelseletion. We
derive their asymptotidistribution through the development of an asymptotially
equivalent quadratiformand we study their robustness propertiesthrough the in-
uene funtion. Setion4presentssomeomputationalaspetsand Setion5gives
twoappliations. Finally,inSetion6wedisusssomepotentialresearhdiretions.
2.1 General Denition
We onsider a general lass of M-estimators of Mallows's type, where the inuene
of deviations ony and on xare bounded separately. The estimator is the solution
of the estimating equations:
n
X
i=1 (y
i
;
i
)=0; (3)
where (y;) = (y;)w(x) 0
a(), a() = 1
n P
n
i=1 E[(y
i
;
i )℄w(x
i )
0
i
with the
expetation taken with respet to the onditional distribution of yjx, (;), w(x)
are weight funtions dened below, and
i
=
i
() = g 1
(x T
i
). The onstant
a() ensures the Fisher onsisteny of the estimator. The estimating equation (3)
for generalized linear models is a speial ase of equation (1)p. 575 for generalized
estimating equationsinPreisserand Qaqish(1999),where ourfuntion (y;)w(x)
is (intheir notation) V 1
()w(x;y; )(y ) and a()= 0
V 1
() .
Let y = (y
1
;:::;y
n )
T
and = (
1
;:::;
n )
T
. The estimating equation (3)
orresponds tothe minimization of the quantity
Q
M
(y;)= n
X
i=1 Q
M (y
i
;
i
); (4)
with respet to , where the funtions Q
M (y
i
;
i
)an bewritten as
Q
M (y
i
;
i )=
Z
i
~ s
(y
i
;t)w(x
i )dt
1
n n
X
j=1 Z
j
~
t E
(y
j
;t)w(x
j )
dt; (5)
with ~s suh that (y
i
;s)~ =0, and
~
t suh that E[(y
i
;
~
t)℄ =0. Note that dierenes
of devianes, asthe test statisti(8), are independent of s~and
~
t.
Thestrutureof (3) issuggested bythe lassialquasi-likelihoodequations. The
estimatordened byequation(3)isanM-estimatorharaterizedbythesorefun-
tion (y
i
;
i
)=(y
i
;
i )w(x
i )
i
a(). Its inuene funtionis then IF(y; ;F)=
M( ;F) 1
(y;), where M( ;F) = E[
(y;)℄; f. Hampel et al. (1986).
Moreover, the estimator has an asymptoti normal distribution with asymptoti
variane =M( ;F) 1
Q( ;F)M( ;F) 1
, where Q( ;F) =E[ (y;) (y;) T
℄.
Itisthenlearthatthehoieofaboundedfuntion ensuresrobustnessbyputting
a bound on the inuene funtion. Therefore, a bounded funtion (y;) is intro-
duedtoontroldeviationsinthey-spae,andleveragepointsaredown-weightedby
the weightsw(x). Simplehoies for(;)and w()suggested by robustestimators
in linear models are (y
i
;
i ) =
(r
i )
1
V 1=2
(i)
(see (6) below) and w(x
i )=
p
1 h
i ,
where h
i
is the i-th diagonal element of the hat matrix H =X(X T
X) 1
X T
. More
sophistiated hoiesfor w()are available(see Staudte and Sheather, 1990, p.258,
for a disussion in linear regression or Carroll and Welsh, 1988). Weights dened
on H do not have high breakdown properties, and from this point of view, other
hoies of w(x
i
) aremore suitable. Forexample, w(x
i
)an behosen asthe inverse
of the Mahalanobisdistane dened through ahigh breakdown estimateof the en-
ter and of the ovariane matrix of the x
i
(see, for example, the minimum volume
ellipsoidestimatororthe minimum ovarianedeterminantestimatorinRousseeuw
and Leroy, 1987, p. 258 .). Finally notie that the hoie of (y
i
;
i ) =
y
i
i
V(
i )
and
w(x
i
) = 1 for all i, reovers the lassial quasi-likelihood estimator, so that for a
judiious hoie of (y
i
;
i
) and ofthe weights w(x
i
), the funtionQ
M
(y;)an be
seen as the robust ounterpart of the lassial quasi-likelihoodfuntion.
The form of this estimator is attrative beause the estimating equation (3)
orresponds to the minimization of (4) and this leads to a natural denition of
robust deviane; see Setion3.1.
We onsider here the partiular ase of (3), dened by (y
i
;
i ) =
(r
i )
1
V 1=2
(i) ,
where r
i
= y
i
i
V 1=2
(
i )
are the Pearson residuals and
is the Huber funtion dened
by
(r)=
8
<
:
r jrj;
sign(r) jrj>:
(6)
We all the estimator dened in this way, the Mallows quasi-likelihood estimator.
It solvesthe set of estimating equations
n
X
i=1 h
(r
i )w(x
i )
1
V 1=2
(
i )
0
i
a() i
=0; (7)
where a() = 1
n P
n
i=1 E[
(r
i )℄w(x
i )
1
V 1=2
(
i )
0
i
. Using the same notation as in the
linear regressionase, whenw(x
i
)=1 weallthis estimatorHuberquasi-likelihood
estimator.
The tuning onstant is typially hosen to ensure a given level of asymptoti
eÆieny. In Setion 3.2 we propose an alternative proedure for the hoie of
the tuning onstant. a() is a orretion term to ensure Fisher onsisteny; see
Hampel et al. (1986) for general parametri models and He and Simpson (1993),
Setion4.1,forpowerseriesdistributions. Notethata()anbeomputedexpliitly
for binomial and Poisson models and does not require numerial integration; f.
Appendix A. ThematriesM(
;F)andQ(
;F)an alsobeeasilyomputedfor
the Mallows quasi-likelihoodestimator:
Q(
;F)= 1
n X
T
AX a()a()
T
;
where A is adiagonalmatrix with elementsa
i
=E[
(r
i )
2
℄w 2
(x
i )
1
V(
i )
(
i
i )
2
,and
M(
;F)= 1
n X
T
BX;
whereBisadiagonalmatrixwithelementsb
i
=E [
(r
i )
i
logh(y
i jx
i
;
i )℄
1
V 1=2
(
i )
w(x
i )(
i )
2
,
andh()istheonditionaldensityorprobabilityofy
i jx
i
. WerefertoAppendixBfor
further details and for the omputationof these matriesfor binomialand Poisson
models.
3 Robust Inferene
3.1 Model Seletion Based on Robust Devianes
The funtion Q
M
(y;) dened in (4) and (5) allows to develop robust tools for
inferene and modelseletion basedon robust quasi-devianes.
Denote by a=(a T
(1)
;a T
(2) )
T
the partitionof avetor a into(p q) and q ompo-
nents and the orresponding partitionof a matrix A by
A= 0
A
11 A
12
A
21 A
22 1
A
;
where A
11 2 IR
(p q)(p q)
, A
12 2IR
(p q)q
, A
21 2IR
q(p q)
and A
22 2 IR
.
Toevaluatethe adequayofamodel,wedenearobustgoodness-of-tmeasure
| alled robust quasi-deviane | based on the notion of robust quasi-likelihood
funtion, i.e.
D
QM
(y;)= 2Q
M
(y;)= 2 n
X
i=1 Q
M (y
i
;
i );
where Q
M
is dened by (4)and (5).
D
QM
(y;) desribes the quality of a t and will be used to dene a statisti
for model seletion. Let us onsider the model M
p
, with p parameters. Suppose
that the orresponding set of parameters is = (
1
;:::;
p )
T
= ( T
(1)
; T
(2) )
T
. We
are interested in testing the null hypothesis H
0 :
(2)
= 0. This is equivalent to
p q p
the sub-model M
p q
holds.
Weestimatethe vetor ofparametersbysolving (3)fortheompletemodel,and
weobtainanestimator
^
of. Underthenullhypothesis,thesameproedureyields
an estimator _
of (
(1)
;0). We write
^
and _
for the estimated linear preditors
assoiated to the estimate
^
and _
respetively. Then, we dene a robust measure
of disrepany between twonested models by
QM
= h
D
QM
(y;)_ D
QM (y;)^
i
= 2 h
n
X
i=1 Q
M (y
i
;^
i )
n
X
i=1 Q
M (y
i
;_
i )
i
; (8)
where the funtion Q
M (y
i
;
i
) isdened by (5).
Thestatisti(8) isinfatageneralizationof thequasi-deviane testforgeneral-
ized linear models, whih isreovered by taking Q
M (y
i
;
i )=
R
i
yi y
i t
V(t)
dt. Moreover,
when the link funtion is the identity (linear regression), (8) beomes the -test
statisti dened in Hampelet al.(1986),Chapter 7.
Thesameformsforthefuntions(y
i
;
i
)andw(x
i
)asintheestimationproblem
an be onsidered here. In partiular, a Mallows quasi-deviane statisti an be
dened by taking (y
i
;
i )=
(r
i )=V
1=2
(
i ).
The following Proposition establishes the asymptoti distribution of the test
statisti (8). We assume the onditions for the existene, onsisteny, and asymp-
toti normality of M-estimators as given by (A.1)-(A.9) in Heritier and Ronhetti
(1994), p. 902. These onditions have been studied by Huber (1967, 1981), Clarke
(1986) and Bednarski (1993).
Proposition 1 Under onditions (A.1)-(A.9) in Heritier and Ronhetti (1994),
[C1℄, [C2℄ of Appendix C, and under H
0 :
(2)
= 0, the test statisti
QM
dened
nL T
n
C( ;F)L
n +o
P
(1)=nR T
n(2)
M( ;F)
22:1 R
n(2) +o
P
(1); (9)
whereC( ;F)=M 1
( ;F)
~
M +
( ;F), p
n L
n
isnormallydistributedN 0;Q( ;F)
,
M( ;F)
22:1
= M( ;F)
22
M( ;F) T
12
M( ;F) 1
11
M( ;F)
12 , and
p
n R
n
is nor-
mally distributed N 0;M 1
( ;F)Q( ;F)M 1
( ;F)
.
Moreover,
QM
is asymptotially distributed as
q
X
i=1 d
i N
2
i
;
where N
1
;:::;N
q
are independent standard normal variables, d
1
;:::;d
q
are the q
positiveeigenvaluesofthematrixQ( ;F) M 1
( ;F)
~
M +
( ;F)
,and
~
M +
( ;F)
is suh that
~
M +
( ;F)
11
= M( ;F) 1
11 and
~
M +
( ;F)
12
= 0,
~
M +
( ;F)
21
= 0,
~
M +
( ;F)
22
=0.
The proof is given in Appendix D. A similar result an be obtained for the dis-
tribution of
QM
under ontiguous alternatives
(2)
= n 1=2
. In suh a ase
QM
is asymptotially distributed as P
q
i=1 (d
1=2
i N
i +S
T
) 2
, where S is suh that
SS T
= M
22:1
and S T
(M 1
( ;F
0
)Q( ;F
0 )M
1
( ;F
0 ))
22
S = D and D is the
diagonalmatrix with elementsd
1
;:::;d
q .
3.2 Robustness Properties and Choie of the Tuning Con-
stant
The robustness properties of the test based on (8) an be investigated by showing
that a small amount of ontamination at a point z has bounded inuene on the
asymptoti level and power of the test. This ensures the loalstability of the test.
the breakdown point as dened in He, Simpson, and Portnoy (1990). However, we
fous hereonsmalldeviationswhihareprobablythemain onernattheinferene
stage of a statistialanalysis.
We onsider the sequene of -ontaminations
F
;n
= 1
p
n
F
0 +
p
n
G; (10)
where Gis anarbitrary distribution(see Heritier and Ronhetti,1994) and investi-
gate the asymptoti level of the test under (10).
Proposition 2 Consider a parametri model F
0
and the null hypothesis H
0 :
(2)
=0. DenotebyF (n)
theempirialdistributionandbyU
n
thefuntionalU(F (n)
)
suh that U(F
0
)=0, IF(z;U;F
0
) is bounded and
p
n(U
n
U(F
;n
))N(0;) (11)
uniformly over the -ontamination F
;n
. Let (F) be the level of the test based on
the quadrati form nU T
n AU
n
when the underlying distribution is F. The nominal
level is (F
0 )=
0 .
Then, under the -ontamination F
;n
, we have
lim
n!1 (F
;n )=
0 +
2
T
diag
P
Z
IF(z;U;F
0
)dG(z)
Z
IF(z;U;F
0
)dG(z)
T
P T
+o(
2
);
where=
H
d
1
;:::;d
q (
1
0
;)
=0
,=(
1
;:::;
q )
T
=( 2
1
;:::; 2
q )
T
,H
d
1
;:::;d
q (:;)
is the .d.f. of the random variable P
q
i=1 d
i
2
1 (
2
i ),
1
0
is the (1
0
)-quantile of
P
q
i=1 d
i
2
1
(0), P is an orthogonal matrix suh that P T
DP = A, and D is the di-
agonal matrix with elements d
1
;:::;d
q
, the eigenvalues of A. Moreover, diag(R )
indiates the vetor with omponents the diagonal elements of the matrix R .
under ontamination is also bounded. The proof of this proposition is presented
in Appendix E. A similar result an be obtained for the power, showing that the
asymptoti power is stableunder ontamination.
Notethatthis propositiongeneralizesthe result ofProposition4inHeritier and
Ronhetti (1994),whih an be reovered by taking =
1
=Æ() and A=I
q .
Thegeneralresultof Proposition2anbeappliedtotherobustquasi-likelihood
test statisti(8) and in the speial ase of a pointmass ontamination G(z)=
z .
This givesthe following Corollary.
Corollary 1 Under onditions (A.1)-(A.9) in Heritier and Ronhetti (1994), and
foranyM-estimator
^
(2)
withboundedinuenefuntion,theasymptoti levelofthe
robust quasi-likelihood test statisti (8) under a point mass ontamination is given
by
lim
n!1 (F
;n )=
0 +
2
T
diag
P IF(z;
^
(2)
;F
0 )IF(z;
^
(2)
;F
0 )
T
P T
+o(
2
);
where P is an orthogonal matrix suh that P T
DP =
22 M
22:1
, is the asymptoti
variane of
^
dened in Setion 2.1, and D is the diagonal matrix with elements
d
1
;:::;d
q
dened in Proposition 1.
The result is obtained by applying Proposition 2 with G(z) =
z
, U =
^
(2) ,
=
22
, A =M
22:1
, and by using the Frehet dierentiability of
^
(2)
; see Heritier
and Ronhetti (1994).
Hene,aboundedinuene M-estimator
^
(2)
ensures aboundonthe asymptoti
levelof the robust quasi-likelihoodtest under ontamination.
the estimation of parameters an be performed via M-estimation aording to (3),
and the test statisti(8) allows usto make inferene and modelhoie.
Thefuntion(y
i
;
i
)whihappearsinthedenitionofQ
M (y
i
;
i
),isoftentuned
by aonstant;f.forinstane(6). AssuggestedinRonhettiandTrojani(2001),we
anonsidertheproblemfromthepointofviewofinfereneandhoosetheonstant
thatontrolsthe maximalbias ontheasymptotilevelofthetestinaneighborhood
of the model. To serve this last purpose, one an use the Corollary above. The
maximallevelof therobustquasi-likelihoodteststatistiinaneighborhoodofthe
model of radius is given by
=
0 +
2
(
^
(2)
;F
0 )
2
T
diag
P11 T
P T
; (12)
where (
^
(2)
;F
0
)=sup
z jjIF(z;
^
(2)
;F
0
)jj and 1=(1;:::;1) T
.
By(12),we an write
b= 1
r
0
T
diag(P11 T
P T
)
; (13)
where b is the bound on the inuene funtion of the estimator
^
(2)
. Then, for a
xedamountofontaminationandbyimposingamaximalerroronthelevelofthe
test
0
,oneandeterminetheboundbontheinuenefuntionoftheestimator,
and hene the tuning onstant by solving b = (
^
(2)
;F
0 ) =
with respet to .
For example, if q = 1 we have P = 1, diag(P11 T
P T
) = 1, and = 0:1145, see
Ronhetti and Trojani (2001). In pratie, the supremum onz =(y;x) is taken as
the maximum overthe sampleof thesupremum onyjx. Notealsothat the solution
depends onthe unknown parameter
0
; our experieneshows that itdoes not vary
muh for dierent values of , so that one an safely plug-ina reasonable (robust)
estimate. This is valid for a single test. However, in a stepwise proedure (as in
value of for eah test. Sine this is unreasonable from a pratial point of view,
we suggest to hoose a global value of by solving b = sup
z jjIF(z;
^
;F
0
)jj, based
on the fat that(
^
(2)
;F
0
)=sup
z jjIF(z;
^
(2)
;F
0
)jjjjsup
z IF(z;
^
;F
0 )jj.
4 Computational Aspets
The solution of equation (3) an be obtained numerially by a Newton-Raphson
proedure or by a Fisher soring proedure. In the latter ase, the algorithm is
alsoknown asthe inuene algorithm;f. for instane Hampeletal.(1986), p.263.
However, there is a potential problem with multiple roots of equation (3). In this
ase, wereommend touse abootstraprootsearhasproposed inMarkatou,Basu,
and Lindsay (1998), p.743-744, based on the objetive funtion Q
M
dened in (4)
as aseletion rule; see alsoHanfelt and Liang (1995).
The test statisti
QM
of equation (8), an be omputed diretly. It involves
n one-dimensional integrations, whih are performed numerially. Our experiene
shows that itworks wellforbinomialand Poisson models. Toavoidthesenumerial
integrations { espeially in the ase when n is large { one an onsider using the
asymptoti quadratiforms of Proposition1 given by (9)whihare asymptotially
equivalent to the test statisti
QM
. A systemati study on the omparison of (8)
withtheasymptotiequivalentquadratiforms(9)isleftforfurtherwork. Moreover,
ritialregionsorp-valuesfortheteststatisti
QM
areeasytoobtain. Infat,linear
ombinations of 2
1
variables have been well studied in the literature. Algorithms
for the omputation of these p-values have been proposed among others by Davies
(1980)and by Farebrother (1990). Analytial approximationsof thesedistributions
S-PLUS (MathSoft, Seattle) routines for estimation and inferene based on ro-
bust quasi-likelihoodare olleted in alibrary and are available from the authors.
5 Appliations
5.1 Binomial models
In this setion, we analyze the damaged arrots dataset. It is taken from Phelps
(1982) and is disussed by Williams (1987) and used in MCullagh and Nelder
(1989) to illustratetehniques for heking for isolateddepartures fromthe model,
beauseofthe presene of anoutlierinthe y-spae. Thedata are issuedfroma soil
experimentandgivetheproportionofarrotsshowing insetdamageina trialwith
three bloks and eight dose levels of insetiide. The logarithm of the dose ranges
from 1.52 to2.36 in anequally spaedgrid. The sample size is 24.
We assumea binomialmodel with logitlink
log
m
=
0 +
1
log(dose )+
2
blok2+
3
blok1 ;
where = E[Y℄ = E[number of damaged arrots℄, bloki; i = 1;2 are indiators
variablestaking the value of 1if measures are taken inblok i and 0otherwise.
Dierent tehniques |plot of devianeresiduals, plot of Pearson residuals and
Cook's distane | show that there is a single large outlier, namely observation 14
(dose level 6and blok2). On the other hand, this observation does not appear
as aleverage point beause itsh
i
value issmall.
In the followingwe ompare the lassialand the robust analysis. The lassial
estimates are obtained by maximum likelihood. The robust estimates are based on
i
tuning onstantof the Huber funtionis hosen tobe 1:2, whih isobtained by the
proedure desribed at the end of Setion 3.2 with
0
= 0:02, = 0:04 and
=0:1145.
[Table 1 about here.℄
Table 1 shows the eet of observation 14: it seems to inrease the value of
2
orresponding to the variable blok2 . The robust tehnique automatially takes
into aount the partiularity of observation 14: in the estimation proedure, most
oftheobservationsreeiveaweightequalto1,oratleastgreaterthan0.70,whereas
observation 14 reeivesa weight equalto0.26.
Also, the eet of observation 14 is lear on the value of the deviane. This
seems dangerous beause the deviane is used for assessing the signiane of the
variables used for modeling the response. This is onrmed by Table 2, where the
results of a lassialand robust stepwise proedure are ompared.
[Table 2 about here.℄
Thelassialanalysisshows that allthe variables,addedsequentially,are highly
signiant on the basis of their deviane value. Model seletion via a robust step-
wise proedure based on the Huber quasi-deviane dened by equation (8) with
(y
i
;
i )=
(r
i )=V
1=2
(
i
)and =1:2 shows that the variableblok1 is not signif-
iant.
5.2 Poisson models
We use a datasetissued from a study of the diversity of arboreal marsupials in the
Montane ash forest (Australia). This dataset was olleted in view of the man-
wood prodution, into aount. The study is fully desribed in Lindenmayer etal.
(1990, 1991). The numberof dierent speies of arboreal marsupials (possum) was
observed on 151 dierent 3ha sites with uniform vegetation. For eah site the fol-
lowingmeasures were reorded: numberof shrubs,numberof utstumps frompast
logging operations, number of stags (hollow-bearing trees), a bark index reeting
the quantity ofdeortiatingbark, ahabitatsore indiatingthe suitabilityof nest-
ing and foraging habitat for Leadbeater's possum, the basal area of aaia speies,
the speies of eualypt withthe greatest stand basal area (Eualyptus regnans,Eu-
alyptus delegatensis,Eualyptus nitens),and the aspet ofthe site. Theproblemis
to modelthe relationship between diversity and these other variables.
Weisberg and Welsh (1993) used these data to investigate by nonparametri
tehniques the shape of the link funtion. Their onlusion was that the anonial
linktsthisdatasetwell. Therefore,weonsideraPoissongeneralizedlinearmodels
with log-linkto desribediversity asa funtion of
shrubs +stumps +stags +bark+habitat +aaia +eualyptus +aspet,
where eualyptus is a fator with three levels and aspet is a fator with four
levels. Hene, the model involvesthe estimation of aparameter of dimension 12.
The robust estimation of parameters via a Mallows quasi-likelihood estimator
dened by (7) with tuning onstant = 1:6 and weights w(x
i ) =
p
1 h
i gives
the result of Table 3. In the same table, we report within parentheses the results
obtained by meansof lassialquasi-likelihood. It has tobenotiedthat 4observa-
tions, namely observations 59, 110, 133, 139, reeive a weight with respet to their
residualbetween 0.68and0.88. This showsthatthese4observationsare potentially
inuential not only for the estimation proedure, but also for inferene and model
many explanatoryvariablesdonot enter signiantly inthemodel,and aredution
of the numberof variablesin the modelis neessary.
[Table 3 about here.℄
We applied a forward stepwise proedure based on quasi-likelihood and on the
robust version of it. Starting from the null modelwhere only the onstant term is
tted, wetestedwhetheritisappropriatetoadd thenext explanatoryvariable. We
hose to retain a variable if the p-value was smaller than 5%. Table 4 shows the
p-value obtained ateahstep of the proedure. Boldp-valuesindiatethe variables
whihhave been retained inthe model.
[Table 4 about here.℄
Asone ansee fromthe table,themodelshosen bythe lassialandthe robust
analysis are essentially the same, even if the p-values involved are sometimes quite
dierent. The variable habitat is at the border of the deision rule and external
onsideration may be used to judge if it has to be kept in the model. It has to be
notied that the orrelationbetween habitat and aaia ishigh (0.54) and one of
these variablesan be dropped.
[Table 5 about here.℄
In the robust nal t, observations 59, 110, 133, 139 reeive a weights with
respet to their residuals between 0.68 and 0.86, as it was already the ase in the
full model. On the other hand, with respet to the inuene of position, the only
observations reeivingaweightless than 0.9isthe rst one. Therewere threeother
whole set of variables. Probably, this outlyingness was due to some explanatory
variables,whih were not retained in the nal model.
Forthenal modelaspresentedinTable5,weinvestigatethe sensitivityof Mal-
lows quasi-likelihood tests ompared to lassial tests by onsidering the following
proedure: weletthe response ofthe observationreeiving the lowest weight inthe
estimationofthenalmodel,namelyobservation110,spantherangeofvaluesfrom
0 to 6. These values over the range of the response in the sample. In eah situa-
tion, we test the null hypothesis that the oeÆient orresponding to the variable
habitat isequal to0. The p-values of these tests are represented in Figure1.
[Figure1 about here.℄
The p-value of the robust test ( = 1:6) is stable, irrespetive to the response
valuetaken by observation 110. Thisp-value rangesfrom2:6to3:3%. Onthe other
hand, thep-value ofthe lassialtest (=1),varies muhmore: from2:3 to6:5%,
givingrisetoadierentmodelhoie,ifthedeisionruleisset at5%. Moreover,by
lettingobservation110 takearbitrarilylargevalues,the p-valueof therobust testis
bounded, whereas the p-value of the lassialtest ontinues toinrease.
6 Conlusion
Inthispaperweproposedanaturallassofrobusttestingproeduresforgeneralized
linearmodels. Theyare avaluableomplementtolassialtehniquesand aremore
reliable in the presene of outlying points and other deviations from the assumed
model. Further researh inludes the extension of these proedures to generalized
estimatingequationsand tononparametri modelslikegeneralizedadditivemodels.
Wederive the onstant
a()= 1
n n
X
i=1 E
Y
i
i
V 1=2
(
i )
1
V 1=2
(
i )
0
i
;
forbinomialandPoissonmodels,whihreduestotheomputationofE
Yi i
V 1=2
(
i )
.
Letus denej
1
=b
i V
1=2
(
i
),and j
2
=b
i +V
1=2
(
i ).
The binomial model states that Y
i
B(m
i
;p
i
), so that E [Y
i
℄ =
i
= m
i p
i and
V[Y
i
℄=
i m
i
i
m
i
. Then we have
E
Y
i
i
V 1=2
(
i )
= 1
X
j= 1
j
i
V 1=2
(
i )
P(Y
i
=j)1I
fj2[0;m
i
℄g
= P(Y
i j
2
+1) P(Y
i j
1 )
+
i
V 1=2
(
i )
P(j
1
~
Y
i j
2
1) P(j
1
+1Y
i j
2 )
;
with
~
Y
i
B(m
i 1;p
i ).
ThePoissonmodelstates thatY
i
P(
i
),andhene E[Y
i
℄=V(
i )=
i
. Then,
E
Y
i
i
V 1=2
(
i )
= 1
X
j= 1
j
i
V 1=2
(
i )
P(Y
i
=j)1I
fj0g
= P(Y
i j
2
+1) P(Y
i j
1 )
+
i
V 1=2
(
i )
P(Y
i
=j
1
) P(Y
i
=j
2 )
:
B Asymptoti variane
Werst determinethe matrixQ(
;F)inthepartiularsituationofMallowsquasi-
likelihoodestimator. Using itsdenition, we have
Q(
;F) = E
(r)w(x) 1
V 1=2
()
0
a()
(r)w(x) 1
V 1=2
()
0
a()
T
= 1
n X
T
AX a()a()
T
;
whereA isthediagonalmatrixwithelementsa
i
=E[
(r
i )
2
℄w 2
(x
i )
1
V(
i )
(
i )
2
,sine
0
i
=
i
i
x
i
. Inthe same manner, writings(y;x;)=
logh(y
i jx
i
;
i
), wederive
the expression of M(
;F),
M(
;F) = E
(r)w(x) 1
V 1=2
()
0
a()
s(y;x; ) T
= 1
n n
X
i=1 E
(r
i )
i
logh(y
i jx
i
;
i )
1
V 1=2
(
i )
w(x
i )
0
i
0T
i
= 1
n X
T
BX;
whereBisthediagonalmatrixwithelementsb
i
=E[
(r
i )
i
logh(y
i jx
i
;
i )℄
1
V 1=2
(i) w(x
i )(
i
i )
2
.
So, the determinationof the asymptoti variane of a Mallows quasi-likelihood
estimatorinvolvesthe omputation of the diagonalterms of the matriesA and B.
Wedeterminethethreeterms:
i g
1
(
i ),E[
(r
i )
2
℄,andE[
(r
i )
i
logh(y
i jx
i
;
i )℄
for binomialand Poisson models.
Forthe binomialmodelwith logitlink
i g
1
(
i )=m
i
exp (
i )
(1+exp (
i ))
2
;
and
E
2
Y
i
i
V 1=2
(
i )
= 2
P(Y j
1
)+P(Y j
2 +1)
+
+ 1
V(
i )
h
2
i m
i (m
i
1)P(j
1
1
~
~
Y j
2
2)+
+ (
i 2
2
i )P(j
1
~
Y j
2
1)+
+
2
i P(j
1
+1Y j
2 )
i
;
with Y B(m
i
;
i ),
~
Y B(m
i 1;
i )and
~
~
Y B(m
i 2;
i ).
i
logh(y
i jx
i
;
i
) being equal to
V(
i )
, we have
E
(r
i )
i
logh(y
i jx
i
;
i )
=E
Y
i
i
V 1=2
(
i )
Y
i
i
V(
i )
=
=
i
V(
i )
h
P(Y
i j
1 ) P(
~
Y
i j
1
1)+P(
~
Y
i j
2
) P(Y
i j
2 +1)
i
+
+ 1
V 3=2
(
i )
h
2
i m
i (m
i
1)P(j
1
1
~
~
Y
i j
2 2)
+(
i 2
2
i )P(j
1
~
Y
i j
2
1)+ 2
i P(j
1
+1Y
i j
2 )
i
:
For the Poisson model, we use the log-link
i
= g(
i
) = log(
i
) whih leads to
i g
1
(
i
)=exp (
i
). We alsohave
E
2
Y
i
i
V 1=2
(
i )
= 2
P(Y
i j
1
)+P(Y
i j
2 +1)
+
+ 1
V() h
2
P(j
1
1Y
i j
2
2)+( 2 2
)P(j
1 Y
i j
2 1)
+
2
P(j
1
+1Y
i j
2 )
i
:
The sore funtionequals
i
logh(y
i jx
i
;
i )=
Y
i
i
i
= Y
i
i
V(
i )
,so that
E
(r
i )
i
logh(y
i jx
i
;
i )
=E
Y
i
i
V 1=2
(
i )
Y
i
i
V(
i )
=
= P(Y
i
=j
1
)+P(Y
i
=j
2 )
+
+
1
V 3=2
(
i )
2
i P(Y
i
=j
1
1) P(Y
i
=j
1
) P(Y
i
=j
2
1)+P(Y
i
=j
2 )
+
+
i P(j
1 Y
i j
2 1):
C Conditions for Robust Quasi-deviane Tests
[C1℄: DenotebyD
n
theset ofallsamplepointsz
i
,i=1;:::;n forwhihtheseond-
order derivatives 2
Q
M (z
i
;)=
j
k
, i = 1;:::;n; j;k = 1;:::;p exist and
are ontinuous funtions of . It is assumed that lim
n!1 P
(D
n )=1.
n
1
jk
1
leastupperboundandby
jk (z;
1
;Æ)thegreatestlowerboundof 2
Q
M
(z;)=
j
k ,
with respet to in the interval jj
1
jjÆ.
Moreover, assume that for any sequene fÆ
n
g for whih lim
n!1 Æ
n
=0,
lim
n!1 E
jk
(z;;Æ
n )
= lim
n!1 E
jk
(z;;Æ
n )
=E
2
Q
M
(z;)=
j
k
;
and that there exists a positive suh that the expetations E
2
jk
(z;;Æ)
and E
2
jk
(z;;Æ)
are bounded funtions of and Æ for all and Æ<.
Theseonditionsare obtained by replaing logf(z;)by Q
M
(z;)inthe orre-
sponding lassialresults for the likelihoodratio test; f. Rao (1973),Wald(1943).
D Proof of Proposition 1
First,wederivetheasymptotiequivalentquadratiformof
QM
. Theprooffollows
the same lines asin the lassialtheory.
Therststep oftheproofonsistsinapproximating
QM
underonditions[C1℄-
[C2℄ by
p
n (
^
_
) T
M( ;F) p
n(
^
_
); (14)
viaaTaylorexpansionandbymakinguseofSlutsky'stheorem. Then,underH
0 and
by the asymptoti properties of M-estimators whih hold under onditions (A.1)-
(A.9) of Heritier and Ronhetti (1994), the following distribution equality holds
asymptotially
p
n(
^
_
) D
p
n M 1
( ;F)
~
M +
( ;F)
L
n
; (15)
where L
n
= 1
n P
n
i=1 (y
i
;
i
) is suh that p
n L
n
N 0;Q( ;F) . Putting (15)
in (14), and taking into aount the symmetry of M( ;F), we nally have, as
n !1,
QM D
nL T
n
C( ;F)L
n
: (16)
(16) an berewritten as
QM D
nR
T
n(2)
M( ;F)
22:1 R
n(2)
;
where M( ;F)
22:1
= M( ;F)
22
M( ;F)
12
M( ;F) 1
11
M( ;F)
12 , and
p
nR
n is
distributed aording to N 0;M 1
( ;F)Q( ;F)M 1
( ;F)
.
Finally, from(16) we onlude that
QM
q
X
i=1 d
i N
2
i
;
where d
i
are the q positive eigenvalues of Q( ;F)C( ;F) and N
1
;:::;N
q
are in-
dependent standard normal variables. Thus, the distribution of
QM
is a linear
ombinationof 2
randomvariableswith 1 degree of freedom.
E Proof of Proposition 2
Byusing (11)andbystandard resultsonthedistributionofquadratiformsinnor-
mal variables, we an say that the statisti nU T
n AU
n
is asymptotially distributed
as P
q
i=1 d
i
2
1 (
2
i
), with () = (
1
();:::;
q ())
T
= p
nPU(F
;n
). Notie that the
distribution depends onlyon the 2
i
()(see Johnson and Kotz (1970), Chapter 29).
Moreover, up to O(1=n), we have that (F
;n
) =1 H
d1;:::;dq (
1 0
;()), with
()=diag (()() T
)=ndiag PU(F
;n )U(F
;n )
T
P T
.
d
1
;:::;d
q 1
0
(F
;n
)
0
=b() b(0)=b 0
(0)+ 1
2
2
b 00
(0)+o(
2
):
But
b 0
(0)= T
=0
=2n T
diag
P h
U(F
;n )
i
=0 U(F
0 )P
T
=0;
beause U(F
0 )=0.
We alsohave that
b 00
(0) = T
2
2
=0
= T
2ndiag
P h
U(F
;n )
U(F
;n )
T i
=0 P
T
= 2
T
diag
P
Z
IF(z;U;F
0
)dG(z)
Z
IF(z;U;F
0
)dG(z)
T
P T
;
byusingagainthefatthatU(F
0
)=0andbeause
U(F
;n )
=0
= R
IF(z;U;F
0 )
1
p
n dG(z)
(see Hampel etal., 1986,p. 83). This ompletes the proof.
Aknowledgments
EvaCantoni(Eva.Cantonimetri.unige. h)isma^tre-assistante andElvezioRon-
hetti (Elvezio.Ronhettimetri.un ige. h) is Professor, Department of Eono-
metris, CH-1211 Geneva 4, Switzerland. The authors would like to thank the
Editor, the Assoiate Editor, the referees, P. Rousseeuw and M.-P. Vitoria-Feser
for theirvaluable suggestions,and A.Welshfor providingthe data analyzedinSe-
tion 5.2. Thehospitality ofStanford(E.C.) and MIT(E.R.)duringtherevision of
the manusript isgratefully aknowledged.
Bednarski, T. (1993), \Frehet Dierentiability of Statistial Funtionals and Im-
pliations to Robust Statistis," in New Diretions in Statistial Data Analy-
sis and Robustness, eds. Morgenthaler, S., Ronhetti, E., and Stahel, W. A.,
Basel/Cambridge, MA:Birkhaeuser, pp. 25{34.
Carroll, R.J. and Welsh, A. H. (1988), \A Noteon Asymmetry and Robustness in
Linear Regression," TheAmerian Statistiian, 42,285{287.
Clarke, B. R. (1986), \Nonsmooth Analysis and Frehet Dierentiability of M-
funtionals," Probability Theory and Related Fields,73,197{209.
Davies,R.B.(1980),\[AlgorithmAS155℄TheDistributionofaLinearCombination
of 2
RandomVariables(ASR53: 84V33P366-369),"AppliedStatistis, 29,323{
333.
Farebrother, R. W. (1990), \[Algorithm AS 256℄ The Distribution of a Quadrati
Form inNormal Variables,"Applied Statistis, 39,294{309.
Hampel, F. R. (1974), \The Inuene Curve and Its Role in Robust Estimation,"
Journal of the Amerian StatistialAssoiation, 69,383{393.
Hampel,F.R.,Ronhetti,E.M.,Rousseeuw, P.J.,andStahel,W.A.(1986),Robust
Statistis: The Approah Based on Inuene Funtions, New York: Wiley.
Hanfelt,J.J.and Liang,K.-Y. (1995),\Approximate LikelihoodRatiosforGeneral
EstimatingFuntions," Biometrika,82, 461{477.
Minimax Versus Loally Linear Estimation," The Annals of Statistis, 21, 314{
337.
He, X., Simpson, D. G., and Portnoy, S. L. (1990), \Breakdown Robustness of
Tests," Journal of the Amerian StatistialAssoiation, 85, 446{452.
Heritier, S. and Ronhetti, E. (1994), \RobustBounded-inuene Tests inGeneral
ParametriModels,"JournaloftheAmerianStatistialAssoiation,89,897{904.
Heyde, C. C. (1997), Quasi-Likelihood and its Appliation, Berlin/New York:
Springer-Verlag.
Huber, P. J. (1967), \The Behavior of the Maximum Likelihood Estimates under
Nonstandard Conditions," in Proeedings of the Fifth Berkeley Symposium on
Mathematial Statistis and Probability, pp. 221{233.
| (1981),Robust Statistis, New York: Wiley.
Imhof, J. P. (1961), \Computing the Distribution of Quadrati Forms in Normal
Variables,"Biometrika,48,352{363.
Johnson, N. L. and Kotz, S. (1970), Continuous Univariate Distributions, vol. 2,
Boston; Geneva, IL: Houghton-Miin.
Kunsh,H. R.,Stefanski, L.A., andCarroll,R. J.(1989), \ConditionallyUnbiased
Bounded-inuene Estimation in General Regression Models, With Appliations
to Generalized Linear Models," Journal of the Amerian Statistial Assoiation,
84,460{466.
A. P. (1991), \The Conservation of Arboreal Marsupials in the Montane Ash
Forestsofthe CentralHighlandsofVitoria,South-EastAustralia: III.TheHabi-
tatRequirementsofLeadbeater'sPossumGymnobelideusleadbeateriandModels
oftheDiversityandAbundaneofArborealMarsupials,"BiologialConservation,
56,295{315.
Lindenmayer, D. B., Cunningham, R. B., Tanton, M. T., Smith, A. P., and Nix,
H. A. (1990), \The Conservation of Arboreal Marsupials in the Montane Ash
Forests of the Vitoria, South-East Australia, I. Fators Inuening the Ou-
pany of Trees with Hollows," BiologialConservation, 54,111{131.
Markatou,M.,Basu,A.,andLindsay,B.G.(1998),\WeightedLikelihoodEquations
With Bootstrap Root Searh," Journal of the Amerian Statistial Assoiation,
93,740{750.
Markatou, M. and He, X. (1994), \Bounded Inuene and High Breakdown Point
Testing Proedures in Linear Models," Journal of the Amerian Statistial Asso-
iation, 89,543{549.
MCullagh, P.and Nelder,J.A.(1989),GeneralizedLinear Models,London: Chap-
man &Hall, 2nd ed.
Morgenthaler, S. (1992), \Least-absolute-deviations Fits for Generalized Linear
Models," Biometrika,79, 747{754.
Pearson, E. S. (1959), \Note on the Distribution of Non-entral 2
," Biometrika,
46,364.
response Relationships in Insetiide Evaluation Field Trials," in Leture Notes
in Statistis, No. 14. GLIM.82: Proeedings of the International Conferene on
Generalized Linear Models, ed. Gilhrist, R.,Berlin/New York: Springer-Verlag.
Pregibon, D. (1982), \Resistant Fits for Some Commonly Used Logisti Models
WithMedial Appliations," Biometris,38, 485{498.
Preisser, J. S. and Qaqish, B. F. (1999), \Robust Regression for Clustered Data
with Appliations toBinary Regression," Biometris, 55,574{579.
Rao, C. R. (1973),Linear StatistialInferene, New York: Wiley, 2nd ed.
Ronhetti, E. (1997), \Robustness Aspets of Model Choie," Statistia Sinia, 7,
327{338.
Ronhetti,E. andStaudte, R.G.(1994),\ARobustVersionofMallows'C
p
,"Jour-
nal of the Amerian StatistialAssoiation, 89,550{559.
Ronhetti, E. and Trojani, F. (2001), \Robust Inferene with GMM Estimators,"
Journal of Eonometris, 101, 36{79.
Rousseeuw, P.J.and Leroy,A. M.(1987),Robust RegressionandOutlier Detetion,
New York: Wiley.
Rukstuhl, A.F.and Welsh, A.H.(1999),\RobustFittingofthe BinomialModel,"
Manusript.
Sommer,S. and Huggins,R. (1996),\VariableSeletion using the WaldTest and a
Robust C
p
," Applied Statistis, 45, 15{29.
York: Wiley.
Stefanski, L.A., Carroll,R.J.,and Ruppert, D. (1986),\OptimallyBoundedSore
Funtions for Generalized Linear Models With Appliations to Logisti Regres-
sion," Biometrika,73,413{424.
Wald, A. (1943), \Test for Statistial Hypotheses Conerning Several Parameters
whentheNumberofObservationsisLarge."Transations of theAmerianMath-
ematialSoiety, 54,426{482.
Wedderburn, R. W. M. (1974), \Quasi-likelihood Funtions, Generalized Linear
Models, and the Gauss-NewtonMethod," Biometrika, 61,439{447.
Weisberg, S. and Welsh, A. H. (1993), \Missing Links," in New Diretions in Sta-
tistial Data Analysis and Robustness, eds. Morgenthaler, S., Ronhetti, E., and
Stahel,W. A., Basel/Cambridge, MA: Birkhaeuser,pp. 275{284.
Williams, D.A. (1987),\GeneralizedLinearModelDiagnostisUsingtheDeviane
and Single Case Deletions,"Applied Statistis, 36, 181{191.
y110
pvalue
0 1 2 3 4 5 6
0.03 0.04 0.05 0.06
c=1.6 c=Inf
Figure 1: Sensitivity urves of the p-value for Mallows quasi-likelihood tests with
=1:6 (solid line) and =1 (dashed line).
Interept 1.480 (0.66) 1.939 (0.70)
logdose -1.817 (0.34) -2.049 (0.37)
blok2 0.843 (0.23) 0.685 (0.24)
blok1 0.542 (0.23) 0.450 (0.24)
Table 1: Estimationof by maximumlikelihoodandbythe Huberquasi-likelihood
estimatorwith =1:2. Standard errors are indiated within parentheses.
NULL 83.34 60.46
logdose 54.73 (0.000) 39.94 (0.000)
blok2 45.59 (0.003) 35.21 (0.017)
blok1 39.98 (0.018) 32.74 (0.085)
Table2: ResidualdevianeandresidualHuberquasi-devianewith=1:2. p-values
are indiated withinparentheses.
Interept -0.8978 (-0.947) 0.2682 (0.265)
shrubs 0.0099 (0.012) 0.0222 (0.022)
stumps -0.2514 (-0.272) 0.2876 (0.286)
stags 0.0402 (0.040) 0.0113 (0.011)
bark 0.0400 (0.040) 0.0145 (0.014)
aaia 0.0178 (0.018) 0.0107 (0.011)
habitat 0.0714 (0.072) 0.0385 (0.038)
eualyptus nitens 0 (0) - (-)
eualyptus regnans -0.020 (-0.015) 0.1938 (0.192)
eualyptus delegatensis 0.127 (0.115) 0.2738 (0.272)
aspet NW-NE 0 (0) - (-)
aspet NW-SE 0.0601 (0.067) 0.1913 (0.190)
aspet SE-SW 0.0949 (0.117) 0.1920 (0.190)
aspet SW-NW -0.5079 (-0.489) 0.2505 (0.247)
Table 3: CoeÆients estimation and orresponding standard errors for the Poisson
modelwith log-linkof the possum dataset.
shrubs 0.0871 0.3642
stumps 0.0646 0.2988
stags 0.0000 0.0000
bark 0.0035 0.0039
aaia 0.0002 0.0009
habitat 0.0500 0.0443
eualyptus regnans 0.8754 0.8030
eualyptus delegatensis 0.8591 0.8074
eualyptus nitens 0.5681 0.6461
aspet NW-NE 0.8336 0.9100
aspet NW-SE 0.2612 0.3462
aspet SE-SW 0.1996 0.1646
aspet SW-NW 0.0012 0.0023
Table4: p-valuesofaforwardstepwiseproedureforthePoissonmodelwithlog-link
of the possum dataset.
Interept -0.7981 (-0.8213) 0.2030 (0.2000)
stags 0.0406 (0.0410) 0.0104 (0.0103)
bark 0.0410 (0.0406) 0.0126 (0.0125)
habitat 0.0143 (0.0136) 0.0098 (0.0097)
aaia 0.0776 (0.0782) 0.0371 (0.0367)
aspet SW-NW -0.6044 (-0.5968) 0.2121 (0.2086)
Table 5: CoeÆients estimation and orresponding standard errors for the nal
Poisson model with log-linkof the possum dataset.