103
Views
1
CrossRef citations to date
0
Altmetric
Original Articles

Improved score tests for exponential family nonlinear models

, ORCID Icon, , ORCID Icon & ORCID Icon
Pages 3731-3745 | Received 20 Jun 2019, Accepted 17 Dec 2019, Published online: 07 Feb 2020
 

Abstract

This paper focuses on the corrections to the score test statistic under the exponential family nonlinear model. We use Monte Carlo simulations to compare the corrected statistics and their uncorrected versions and to examine the impact of the number of nuisance parameters on their finite-sample behaviors for the normal nonlinear regression model. The numerical results have shown that the corrected score statistic performs better than the uncorrected version. Finally, we perform a statistical analysis with real data by using the approach proposed in the article.

Acknowledgments

This paper is based on part of Cavalcanti’s doctoral thesis submitted to the Instituto de Matemática e Estatística, Universidade de São Paulo, under the supervision of the second author. The authors thank the Editors and referees for their constructive comments on an earlier version of this manuscript which resulted in this improved version. Also, we thank Prof. Dr. Michelli Barros (UFCG) for valuable help in simulation program.

Appendix

A. Cumulants of log-likelihood derivatives in exponential family nonlinear models

Let L=L(β,δ,γ) be the log-likelihood (Equation4). We hall use the standard notation for joint cumulants of log-likelihood derivatives, i.e., κrs=E(2L/βrβs),κr,s=E(L/βr L/βs),κrst=E(3L/βrβsβt),κr,s,t=E(L/βr L/βs L/βt),κr,st=E(L/βr 2L/βsβt),κrs(t)=κrs/βt, and so on. We obtain the following cumulants: κrs=1γ=1nxr*qwxs*,κRS=12=1nqRSq,κγγ=n2γ2,κrS=κrγ=κRγ=0 where xr*=η/βr and qRS=2q/δRδS, for r=1,,p and R,S=1,,q. The third-order cumulants are κrst=1γ=1n{q(f+2g)xr*xs*xt*+qw(xrs**xt*+xrt**xs*+xst**xr*)},κγγγ=2nγ3,κRST=12=1nqRSTq,κrsT=1γ=1nxr*qTwxs*,κrsγ=1γ2=1nxr*qwxs*κRSγ=12γ=1nqRSq,κγ,γγ=nγ3,κγ,γ,γ=nγ3,κγ,RS=12γ=1nqRSqκRγγ=κrγγ=κRSt=κrSγ=0,κr,s,t=1γ=1nq(fg)xr*xs*xt*,κR,S,T=12=1n{qRSTqqRSqTq2qRTqSq2qSTqRq2},κr,s,T=1γ=1nqTxr*xs*w,κr,s,γ=1γ2=1nqxr*xs*w,κR,S,γ=1γ=1nqRSq,κr,S,γ=κr,γ,γ=κr,S,T=κR,γ,γ=0,κr,st=1γ=1nq(gxr*xs*xt*+wxr*xst**),κR,ST=12=1nqRqSTq2,κr,sT=1γ=1nxr*qTwxs*,κr,sγ=1γ2=1nqxr*xs*w,κR,Sγ=12γ=1nqRqSq2,κr,ST=κr,Sγ=κr,γγ=κR,st=κR,St=κR,sγ=κR,γγ=κγ,rs=κγ,rS=κγ,γr=κγ,γR=0 where xrs**=2η/βrβs,qRST=3q/δRδSδT, for r,s,t=1,,p and R,S,T=1,q and the quantities q,f and g, =1,,n, are defined by EquationEquations (4) and Equation(9), respectively. Further, the fourth-order cumulants are given by κrstu=1γ=1nq{[(2wη2)+(2μη2)(2θη2)+2(μη)(3θη3)]xr*xs*xt*xu*+w(xrst***xu*+xrsu***xt*+xrut***xs*+xust***xr*+xrs**xtu**+xrt**xsu**+xru**xst**)+(xrs**xt*xu*+xts**xr*xu*+xus**xr*xt*+xru**xs*xt*+xtu**xr*xs*+xrt**xs*xu*)×[2(μη)(2θη2)+(θη)(2μη2)]},κRSTU=12=1nqRSTUq,κRSγγ=1γ2=1nqRSqκr,s,t,u=1γ=1nqbxr*xs*xt*xu*,κR,S,T,U=12=1n(qRSTUqqRSTqUq2qRSUqTq2qRTUqSq2qSTUqRq2+2qRqSqTUq3+2qRqTqSUq3+2qRqUqSTq3+2qSqTqRUq3+2qSqUqRTq3+2qTqUqRSq3qRSqTUq2qRTqSUq2qRUqSTq2), κr,s,T,U=2γ=1nqwxr*xs*qTqUq2,κr,s,γ,γ=2γ3=1nqwxr*xs*,κR,S,γ,γ=3γ2=1nqRSq,κr,s,tu=1γ=1nq{(hb)xr*xs*xt*xu*+(fg)xr*xs*xtu**},κR,S,γγ=2γ2=1nqRSq,κR,S,TU==1nqRqSqTUq3,κr,s,TU=1γ=1nwxr*xs*qTU,κr,s,γγ=2γ3=1nqwxr*xs* where xrst***=3η/βrβsβt,qRSTU=4q/δRδSδTδU, for r,s,t,u=1,,p and R,S,T,U=1,,q, and the quantities b and h defined by EquationEquations (9) and Equation(10), respectively.

The equation m=exp(zδ) gives q=exp{(zz¯)δ},qR=(zz¯)Rexp{(zz¯)δ}=(zz¯)Rq,qRS=(zz¯)RSq,qRST=(zz¯)RSTq and qRSTU=(zz¯)RSTUq, where (zz¯)R=zRz¯R, (zz¯)RS=(zRz¯R)(zSz¯S) and so on, gives z¯=(z¯1,,z¯q) and z¯R=(1/n)=1nzR, for R=1,,q. Thus we obtain the following cumulants κRS=12=1n(zz¯)RSκRST=12=1n(zz¯)RST,κrsT=1γ=1nxr*wxs*(zz¯)Tq,κRSγ=12γ=1n(zz¯)RS,κR,S,T==1n(zz¯)RST,κr,s,T=1γ=1n(zz¯)Rqxr*xs*w,κR,S,γ=1γ=1n(zz¯)RS,κr,sT=1γ=1nxr*wxs*(zz¯)Tq,κR,ST=12=1n(zz¯)RST,κR,Sγ=12γ=1n(zz¯)RS,κγ,RS=12γ=1n(zz¯)RS,κγ,RS=12γ=1n(zz¯)RS,κRSTU=12=1n(zz¯)RSTU,κRSγγ=1γ2=1n(zz¯)RS,κR,S,T,U=3=1n(zz¯)RSTU,κr,s,T,U=2γ=1nqwxr*xs*(zz¯)TU,κR,S,γ,γ=3γ2=1n(zz¯)RS,κR,S,TU==1n(zz¯)RSTU,κr,s,TU=1γ=1nwxr*xs*(zz¯)TUq,κR,S,γγ=2γ2=1n(zz¯)RS

B. Cumulants for the Bartlett correction factor

B.1. Tests for components of β and δ

Suppose that we want to test the null hypothesis H0:β1=β1(0),δ1=δ1(0) versus H1:β1β1(0)orδ1δ1(0). Then, the coefficients for the Bartlett-type corrections can be expressed terms A11=31ΦFZβ2d(ZβZβ2)Zβ2dFΦ1+61ΦWP2(ZβZβ2)Zβ2dFΦ1+31ΦWP2(ZβZβ2)P2WΦ1+31ΦWZβ2d(ZδZδ2)Zδ2d1+31ΦWZβ2d(ZδZδ2)Zβ2dWΦ1+341Zδ2d(ZδZδ2)Zδ2d1,A12=61ΦFZβ2dZβ2(ZβZβ2)d(FG)Φ1+61ΦWP2Zβ2(ZβZβ2)d(FG)Φ1+61ΦWZβ2dZδ2(ZβZβ2)dWΦ1+61ΦWZβ2dZδ2(ZδZδ2)d1+12n1ΦWZβ2d(ZβZβ2)dWΦ1+12n1ΦWZβ2d(ZδZδ2)d1+31Zδ2dZδ2(ZβZβ2)dWΦ1+31Zδ2dZδ2(ZδZδ2)d1+6n1Zδ2d(ZβZβ2)dWΦ1+6n1Zδ2d(ZδZδ2)d1,A13=61Φ(2GF)[Zβ2(2)(ZβZβ2)]FΦ161ΦW[(ZβZβ2)J2]WΦ161ΦW[(ZβZβ2)C2]FΦ1+921[Zδ2(2)(ZδZδ2)]161Φ(2GF)[(ZβZβ2)C2]WΦ1+18n1[Zδ2(ZδZδ2)]1+61ΦW[(ZδZδ2)Zβ2(2)]WΦ1+121ΦW[(ZβZβ2)Zβ2Zδ2]WΦ1,A14=121ΦWZβ2d(ZδZδ2)d161ΦH(ZβZβ2)dZβ2d1121(ZδZδ2)dZδ2d161ΦW(ZβZβ2)dZδ2d161Φ(FG)P2(ZβZβ2)d112n1(ZδZδ2)d1,A21=31Φ(FG)(ZβZβ2)dZβ2(ZβZβ2)d(FG)Φ16n1[(ZδZδ2)d]2131ΦW(ZβZβ2)dZδ2(ZδZδ2)d131(ZδZδ2)dZδ2(ZβZβ2)dWΦ112n1ΦW(ZβZβ2)d(ZδZδ2)d131(ZδZδ2)dZδ2(ZδZδ2)d131ΦW(ZβZβ2)dZδ2(ZβZβ2)d16n1[ΦW(ZβZβ2)d]21,A22=61ΦFZβ2d(ZβZβ2)(ZβZβ2)d(FG)Φ161ΦWZβ2d(ZδZδ2)(ZβZβ2)dWΦ1+61ΦWZβ2d(ZδZδ2)(ZδZδ2)d131ΦWZδ2d(ZδZδ2)(ZβZβ2)d131Zδ2d(ZδZδ2)(ZδZδ2)d161ΦWP2(ZβZβ2)(ZβZβ2)d(FG)Φ1,A23=61Φ(FG)[Zβ2(ZβZβ2)(2)](FG)Φ112n1[(ZδZδ2)(2)]161ΦW[Zδ2(ZβZβ2)(2)]WΦ161[Zδ2(ZδZδ2)(2)]1121ΦW[Zβ2(ZβZβ2)(ZδZδ2)]WΦ112n1ΦW[(ZβZβ2)(2)]WΦ1,A24=31ΦB(ZβZβ2)d21+121ΦW(ZδZδ2)d(ZβZβ2)d1+91(ZδZδ2)d21, A31=31Φ(FG)(ZβZβ2)d(ZβZβ2)(ZβZβ2)d(FG)Φ1+31(ZδZδ2)d(ZδZδ2)(ZδZδ2)d1+61ΦW(ZβZβ2)d(ZδZδ2)(ZδZδ2)d1+31ΦW(ZβZβ2)d(ZδZδ2)(ZβZβ2)dWΦ1,A32=21Φ(FG)[(ZβZβ2)(3)](FG)Φ1+21[(ZδZδ2)(3)]1+61ΦW[(ZδZδ2)(ZβZβ2)(2)]WΦ1

B.2. Tests for components of δ

Here, suppose that we want to test the null hypothesis H0:δ1=δ1(0), versus H1:δ1δ1(0). In this situation, just consider Zβ=Zβ2 and then the coefficients of the corrected score reduce to A11=31ΦWZβd(ZδZδ2)Zδ2d1+31ΦWZβd(ZδZδ2)ZβdWΦ1+341Zδ2d(ZδZδ2)Zδ2d1,A12=61ΦWZβdZδ2(ZδZδ2)d1+12n1ΦWZβd(ZδZδ2)d1+31Zδ2dZδ2(ZδZδ2)d1+6n1Zδ2d(ZδZδ2)d1,A13=921[Zδ2(2)(ZδZδ2)]1+18n1[Zδ2(ZδZδ2)]1+61ΦW[(ZδZδ2)Zβ(2)]WΦ1,A14=121ΦWZβd(ZδZδ2)d1121(ZδZδ2)dZδ2d112n1(ZδZδ2)d1,A21=6n1[(ZδZδ2)d]2131(ZδZδ2)dZδ2(ZδZδ2)d1,A22=61ΦWZβd(ZδZδ2)(ZδZδ2)d131Zδ2d(ZδZδ2)(ZδZδ2)d1,A23=12n1[(ZδZδ2)(2)]161[Zδ2(ZδZδ2)(2)]1,A24=91(ZδZδ2)d21,A31=31(ZδZδ2)d(ZδZδ2)(ZδZδ2)d1,A32=21[(ZδZδ2)(3)]1 Further, for testing all the components of δ, we have Zδ2=0. Hence, A11=31ΦWZβdZδZβdWΦ1,A12=12n1ΦWZβdZδd1,A13=61ΦW[ZδZβ(2)]WΦ1,A14=121ΦWZβdZδd112n1Zδd1,A21=6n1[Zδd]21,A22=61ΦWZβdZδZδd1,A23=12n1[Zδ(2)]1,A24=91Zδd21,A31=31ZδdZδZδd1,A32=21[Zδ(3)]1

B.3. Tests for components of β

Finally, suppose that we want to test the null hypothesis H0:β1=β1(0) versus H1:β1β1(0). If we consider Zδ=Zδ2, then the coefficients of the corrected score statistic are A11=31ΦFZβ2d(ZβZβ2)Zβ2dFΦ1+61ΦWP2(ZβZβ2)Zβ2dFΦ1+31ΦWP2(ZβZβ2)P2WΦ1,A12=61ΦFZβ2dZβ2(ZβZβ2)d(FG)Φ1+61ΦWP2Zβ2(ZβZβ2)d(FG)Φ1+61ΦWZβ2dZδ(ZβZβ2)dWΦ1+12n1ΦWZβ2d(ZβZβ2)dWΦ1+31ZδdZδ(ZβZβ2)dWΦ1+6n1Zδd(ZβZβ2)dWΦ1,A13=61Φ(2GF)[Zβ2(2)(ZβZβ2)]FΦ161ΦW[(ZβZβ2)J2]WΦ161ΦW[(ZβZβ2)C2]FΦ161Φ(2GF)[(ZβZβ2)C2]WΦ1+121ΦW[(ZβZβ2)Zβ2Zδ]WΦ1,A14=61ΦH(ZβZβ2)dZβ2d161ΦW(ZβZβ2)dZδd161Φ(FG)P2(ZβZβ2)d1,A21=31Φ(FG)(ZβZβ2)dZβ2(ZβZβ2)d(FG)Φ131ΦW(ZβZβ2)dZδ(ZβZβ2)d16n1[ΦW(ZβZβ2)d]21,A22=61ΦFZβ2d(ZβZβ2)(ZβZβ2)d(FG)Φ161ΦWP2(ZβZβ2)(ZβZβ2)d(FG)Φ1,A23=61Φ(FG)[Zβ2(ZβZβ2)(2)](FG)Φ161ΦW[Zδ(ZβZβ2)(2)]WΦ112n1ΦW[(ZβZβ2)(2)]WΦ1,A24=31ΦB(ZβZβ2)d21,A31=31Φ(FG)(ZβZβ2)d(ZβZβ2)(ZβZβ2)d(FG)Φ1,A32=21Φ(FG)[(ZβZβ2)(3)](FG)Φ1 Finally, for testing all the components of β, we have Zβ2=P2=C2=J2=0. Hence, A11=A13=A22=0,A12=3 1 Zδd Zδ Zβd W Φ 1+6n 1 Zδd Zβd W Φ 1,A14=6 1 Φ W Zβd Zδd 1,A21=3 1 Φ W Zβd Zδ Zβd 16n 1 [Φ W Zβd]2 1,A23=6 1 Φ W [ZδZβ(2)] W Φ 112n 1 Φ W [Zβ(2)] W Φ 1,A24=3 1 Φ B Zβd2 1,A31=3 1 Φ (FG) Zβd Zβ Zβd (FG) Φ 1,A32=2 1 Φ (FG) Zβ(3) (FG) Φ 1

Additional information

Funding

This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001 and Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) [grant number 303133/2014-7].

Log in via your institution

Log in to Taylor & Francis Online

PDF download + Online access

  • 48 hours access to article PDF & online version
  • Article PDF can be downloaded
  • Article PDF can be printed
USD 61.00 Add to cart

Issue Purchase

  • 30 days online access to complete issue
  • Article PDFs can be downloaded
  • Article PDFs can be printed
USD 1,069.00 Add to cart

* Local tax will be added as applicable

Related Research

People also read lists articles that other readers of this article have read.

Recommended articles lists articles that we recommend and is powered by our AI driven recommendation engine.

Cited by lists all citing articles based on Crossref citations.
Articles with the Crossref icon will open in a new tab.