Abstract
The paper discusses statistical inference dealing with the asymptotic theory of likelihood ratio tests when some parameters may lie on boundary of the parameter space. We derive a closed form solution for the case when one parameter of interest and one nuisance parameter lie on the boundary. The asymptotic distribution is not always a mixture of several chi-square distributions. For the cases when one parameter of interest and two nuisance parameters or two parameters of interest and one nuisance parameter are on the boundary, we provide an explicit solution which can be easily computed by simulation. These results can be used in many applications, e.g. testing for random effects in genetics. Contrary to the claim of some authors in the applied literature that use of chi-square distribution with degrees of freedom as in case of interior parameters will be too conservative when some parameters are on the boundary, we show that when nuisance parameters are on the boundary, that approach may often be anti-conservative.
Similar content being viewed by others
References
Chant, D. 1974. On asymptotic tests of composite hypotheses in nonstandard conditions. Biometrika 61:291–298.
Chernoff, H. 1954. On the distribution of the likelihood ratio. Annals of Mathematical Statistics 25:573–578.
Dominicus, A., A. Skrondal, H.K. Gjessing, N.L. Pedersenm, and J. Palmgren 2006. Likelihood ration tests in behavioral genetics: problems and solutions. Behavior Genetics 36:331–340.
Feder, P.I. 1968. On the distribution of the log likelihood ratio test statistic when the true parameter is near the boundaries of the hypothesis regions. Annals of Mathematical Statistics 39:2044–2055.
Kopylev, L., and B. Sinha. 2010. On the asymptotic distribution of the likelihood ratio test when parameters lie on the boundary. Technical Report, Department of Mathematics and Statistics, UMBC.
Meyer, K. 2008. Likelihood calculations to evaluate experimental designs to estimate genetic variances. Heredity 101:212-221.
Moran, P.A.P. 1971. Maximum likelihood estimators under nonstandard conditions. Proceedings of the Cambridge Philosophical Society 70:441–450.
Morris, N.J., R. Elston, and C.M. Stein. 2009. Calculating asymptotics significance level of the constrained likelihood ratio test with application to multivariate genetic linkage analysis. Statistical Applications in Genetics and Molecular Biology 8, Article 39
R Development Core Team. 2006. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, http://www.R-project.org.
Self, S.G., and K-Y. Liang. 1987 Asymptotic properties of maximum likelihood estimators and likelihood ratio tests under nonstandard conditions. Journal of the American Statistical Association 82:605–610.
Sinha, B., L. Kopylev, and J. Fox. 2007. Some new aspects of dose-response multistage models with applications. Technical Report, Department of Mathematics and Statistics, UMBC. http://www.math.umbc.edu/~kogan/technical_papers/2007/Sinha_Kopylev_Fox.pdf.
Stoel, R.D., F.G. Garre, C. Dolan, and G. van den Wittenboer. 2006. On the likelihhod ratio test in structural equatioon modeling when parameters are subject to boundary conditions. Psychological Methods 11:439–455.
Visscher, P.M. 2006. A note on the asymptotic distribution of likelihood ratio tests to test variance components. Twin Research and Human Genetics 9:490–495.
Acknowledgements
The views expressed in this article are those of the authors and do not necessarily reflect the views or policies of the U.S. Environmental Protection Agency. Research of the second author (Sinha) was supported under a NCEA/ORD/EPA visiting faculty fellowship program. The authors thank John Fox and Paul White of NCEA/ORD/EPA for encouragement and anonymous reviewers for helpful comments.
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix A
In the following, matrix R is defined as
Case I
MLE = (Z 1, Z 2, Z 3) over Region R I : Z 1 ≥ 0, Z 2 ≥ 0, Z 3 ≥ 0.
Obviously, the minimum value of the quadratic here is Q I = 0 and the region R I can be reexpressed in terms of U 1, U 2, U 3 as:
Case II
MLE = (0, Z 2.1, Z 3.1) over Region R II : Z 1 < 0, Z 2.1 ≥ 0, Z 3.1 ≥ 0.
Obviously, the minimum value of the quadratic here is Q II given by the following expression
and the region R II can be reexpressed in terms of U 1, U 2, U 3 as:
Case III
\(\mathit{MLE} = (Z_{1.2},0,Z_{3.2})\) over region \(R_{\it III}: Z_2 < 0, Z_{1.2} \geq 0\), Z 3.2 ≥ 0.
Obviously, the minimum value of the quadratic here is Q III given by the following expression
and the region R III can be reexpressed in terms of U 1, U 2, U 3 as
where \(c_{13} = \frac{ \rho_{13} - \rho_{12}\rho_{23}}{ \sqrt{(1-\rho^2_{12})(1-\rho^2_{13})}} \).
Case IV
\(\mathit{MLE} = (Z_{1.3},Z_{2.3}, 0)\) over region R VI : Z 3 < 0, Z 1.3 ≥ 0, Z 2.3 ≥ 0.
Similarly to the previous case
where \(c_{12} =\frac{\rho_{12} - \rho_{13}\rho_{23}}{\sqrt{(1-\rho^2_{12})(1-\rho^2_{13})}}\).
Case V
MLE = (0, 0, Z 3.12) over Region R V : Z 1.2 < 0, Z 2.1 < 0, Z 3.12 > 0.
Case VI
\(\mathit{MLE} (0, Z_{2.13}, 0)\) over the region \(R_{\it VI}\!:\! Z_{1.3}<, Z_{3.1}<, Z_{2.13}>\).
Case VII
\(\mathit{MLE}\! =\! (Z_{1.23}, 0, 0)\) over the region \(R_{\it VII}\!:\! Z_{2.3}<, Z_{3.2}<, Z_{1.23}>\).
To compute the quadratic \(Q_{VII}({\bf Z}|\hat {\boldsymbol \theta})\), note from Eq. B.6 that
This yields
Hence \(Q_{VII}({\bf Z}|\hat {\boldsymbol \theta}) = {\bf V}'{\bf R}^{-1}{\bf V}\).
From Eq. B.3, note that
Similarly, from Eq. B.4, we get
Finally, from Eq. B.6, we get
Case VIII
MLE = (0, 0, 0) over the region R VIII : Z 1.23 < 0, Z 2.13 < 0, Z 3.12 < 0.
To describe R VIII , from Eq. B.6
Also, it is easy to verify that Z 2.13 = U 2.3 and Z 3.12 = U 3.2. Hence R VIII can be expressed as
Appendix B
Here we express some standard residuals which are used in Appendix A in terms of Z 1, Y 2 and Y 3.
-
Z 1.2
$$\begin{array}{rll} Z_{1.2} = Z_1 -\sigma_1\rho_{12}\frac{Z_2}{\sigma_2} & =& Z_1 -\sigma_1\rho_{12}\left[\frac{Y_2}{\sigma_2}+\rho_{12}\frac{Z_1}{\sigma_1}\right] \\ & =& Z_1\big(1-\rho^2_{12}\big) - \sigma_1\rho_{12}\frac{Y_2}{\sigma_2}. \end{array} $$(B.1) -
Z 1.3 By symmetry with Z 1.2, we readily get
$$ Z_{1.3} = Z_1\big(1-\rho^2_{13}\big) - \sigma_1\rho_{13}\frac{Y_3}{\sigma_3}. $$(B.2) -
Z 2.3
$$\begin{array}{rll} Z_{2.3} = Z_{2} - \sigma_2\rho_{23}\frac{Z_3}{\sigma_3} & =& \left[Y_2 + \sigma_2\rho_{12}\frac{Z_1}{\sigma_1}\right] - \sigma_2\rho_{23}\left[\rho_{13}\frac{Z_1}{\sigma_1}+\frac{Y_3}{\sigma_3}\right]\\ & =& \sigma_2(\rho_{12}-\rho_{13}\rho_{23})\frac{Z_1}{\sigma_1}+Y_2-\sigma_2\rho_{23}\frac{Y_3}{\sigma_3}. \end{array} $$(B.3) -
Z 3.2 By symmetry with Z 2.3, we readily get
$$ Z_{3.2} = \sigma_3(\rho_{13}-\rho_{12}\rho_{23})\frac{Z_1}{\sigma_1}-\sigma_3\rho_{23}\frac{Y_2}{\sigma_2}+Y_3. $$(B.4) -
Z 1.23 This term which is the residual of Z 1, given Z 2 and Z 3, is defined as Z 1 − E(Z 1|Z 2, Z 3). Recalling that Z ~N[0, Σ] and writing \({\bf B}_{2 \times 2}= \mbox{dispersion matrix} \) of (Z 2, Z 3), we get
$$\begin{array}{rll} E(Z_1|Z_2,Z_3) & = & (\sigma_{12}, \sigma_{13}){\bf B}^{-1}(Z_2, Z_3)' \\ & = & \frac{(\sigma_{12}\sigma_{33}-\sigma_{13}\sigma_{23}, -\sigma_{12}\sigma_{23}+\sigma_{13}\sigma_{22})(Z_2, Z_3)'}{\sigma_{22}\sigma_{33}-\sigma^2_{23}} \\ & = & \frac{(\sigma_{12}\sigma_{33}-\sigma_{13}\sigma_{23})Z_2 + (-\sigma_{12}\sigma_{23}+\sigma_{13}\sigma_{22})Z_3 }{\sigma_{22}\sigma_{33}-\sigma^2_{23}} \\ & = & \sigma_1\frac{\rho_{12}-\rho_{13}\rho_{23}}{1-\rho^2_{23}}\frac{Z_2}{\sigma_2} + \sigma_1\frac{\rho_{13} - \rho_{12}\rho_{23}}{1-\rho^2_{23}} . \end{array} $$(B.5)Now replacing Z 2 by \(Y_2+\sigma_2\rho_{12}\frac{Z_1}{\sigma_1}\) and Z 3 by \(Y_3+\sigma_3\rho_{13}\frac{Z_1}{\sigma_1}\) and simplifying, we get
$$\begin{array}{rll} Z_{1.23} &= & Z_1\frac{1-\rho^2_{12}-\rho^2_{13}-\rho^2_{23}+2\rho_{12}\rho_{13}\rho_{23}}{1-\rho^2_{23}} \\ && -\sigma_1\frac{\rho_{12}-\rho_{13}\rho_{23}}{1-\rho^2_{23}}\frac{Y_2}{\sigma_2} -\sigma_1\frac{\rho_{13}-\rho_{12}\rho_{23}}{1-\rho^2_{23}}\frac{Y_3}{\sigma_3}. \end{array} $$(B.6) -
Z 2.13 This term which is the residual of Z 2, given Z 1 and Z 3, is defined as Z 2 − E(Z 2|Z 1, Z 3). Using symmetry with E(Z 1|Z 2, Z 3), we readily get
$$ E(Z_2|Z_1,Z_3) = \sigma_2\frac{\rho_{12}-\rho_{13}\rho_{23}}{1-\rho^2_{13}}\frac{Z_1}{\sigma_1} + \sigma_2\frac{\rho_{23} - \rho_{12}\rho_{13}}{1-\rho^2_{13}}\frac{Z_3}{\sigma_3}$$(B.7)Now replacing Z 2 by \(Y_2+\sigma_2\rho_{12}\frac{Z_1}{\sigma_1}\) and Z 3 by \(Y_3+\sigma_3\rho_{13}\frac{Z_1}{\sigma_1}\) and noting that the coefficient of Z 1 is 0, we get
$$ Z_{2.13} = Y_2 - \sigma_2 \times \frac{\rho_{23}-\rho_{12}\rho_{13}}{1-\rho^2_{13}}\times \frac{Y_3}{\sigma_3}. $$(B.8) -
Z 3.12 By symmetry with Z 2.13, we readily get
$$\label{58} Z_{3.12} = Y_3 - \sigma_3 \times \frac{\rho_{23}-\rho_{12}\rho_{13}}{1-\rho^2_{12}}\times \frac{Y_2}{\sigma_2}. $$(B.9)
Rights and permissions
About this article
Cite this article
Kopylev, L., Sinha, B. On the asymptotic distribution of likelihood ratio test when parameters lie on the boundary. Sankhya B 73, 20–41 (2011). https://doi.org/10.1007/s13571-011-0022-z
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13571-011-0022-z
Keywords
- Likelihood ratio test
- Nuisance parameters on the boundary
- One-sided tests
- Parameters of interest on the boundary
- Quadratic forms