Abstract
The skew t-distribution includes both the skew normal and the normal distributions as special cases. Inference for the skew t-model becomes problematic in these cases because the expected information matrix is singular and the parameter corresponding to the degrees of freedom takes a value at the boundary of its parameter space. In particular, the distributions of the likelihood ratio statistics for testing the null hypotheses of skew normality and normality are not asymptotically \(\chi ^2\). The asymptotic distributions of the likelihood ratio statistics are considered by applying the results of Self and Liang (J Am Stat Assoc 82:605–610, 1987) for boundary-parameter inference in terms of reparameterizations designed to remove the singularity of the information matrix. The Self–Liang asymptotic distributions are mixtures, and it is shown that their accuracy can be improved substantially by correcting the mixing probabilities. Furthermore, although the asymptotic distributions are non-standard, versions of Bartlett correction are developed that afford additional accuracy. Bootstrap procedures for estimating the mixing probabilities and the Bartlett adjustment factors are shown to produce excellent approximations, even for small sample sizes.
Similar content being viewed by others
References
Arellano-Valle RB (2010) On the information matrix of the multivariate skew-\(t\) model. Metron 68:371–386
Arellano-Valle RB, Azzalini A (2008) The centred parametrization for the multivariate skew-normal distribution. J Multivar Anal 99:1362–1382
Arellano-Valle RB, Azzalini A (2013) The centred parametrization and related quantities of the skew-\(t\) distribution. J Multivar Anal 113:73–90
Azzalini A (1985) A class of distributions which includes the normal ones. Scand J Stat 12:171–178
Azzalini A (2005) The skew-normal distribution and related multivariate families (with discussion). Scand J Stat 32:159–200
Azzalini A, Capitanio A (1999) Statistical applications of the multivariate skew normal distribution. J R Stat Soc Ser B 61:579–602
Azzalini A, Capitanio A (2003) Distributions generated by perturbation of symmetry with emphasis on a multivariate skew \(t\) distribution. J R Stat Soc Ser B 65:367–389
Azzalini A, Capitanio A (2014) The skew-normal distribution and related families. IMS monographs series. Cambridge University Press, Cambridge
Azzalini A, Genton MG (2008) Robust likelihood methods based on the skew-\(t\) and related distributions. Int Stat Rev 76:106–129
Bickel PJ, Ghosh JK (1990) A decomposition for the likelihood ratio statistic and the Bartlett correction: a Bayesian argument. Ann Stat 18:1070–1090
Branco MD, Dey DK (2001) A general class of multivariate skew-elliptical distributions. J Multivar Anal 79:99–113
Chambers JM, Cleveland WS, Kleiner B, Tukey PA (1983) Graphical methods for data analysis. Wadsworth, Belmont
Chiogna M (2005) A note on the asymptotic distribution of the maximum likelihood estimator for the scalar skew-normal distribution. Stat Methods Appl 14:331–341
Cook RD, Weisberg S (1994) An introduction to regression graphics. Wiley, New York
DiCiccio TJ, Monti AC (2011) Inferential aspects of the skew \(t\)-distribution. Quad Stat 13:1–21
Gupta AK (2003) Multivariate skew \(t\)-distribution. Statistics 37:359–363
Hallin M, Ley C (2012) Skew-symmetric distributions and Fisher information—a tale of two densities. Bernoulli 18:747–763
Hallin M, Ley C (2014) Skew-symmetric distributions and Fisher information: the double sin of the skew-normal. Bernoulli 20:1432–1453
Ley C, Paindavein D (2010) On Fisher information matrices and profile log-likelihood functions in generalized skew-elliptical models. Metron 68:235–250
Ley C, Paindaveine D (2010) On the singularity of multivariate skew-symmetric models. J Multivar Anal 101:1434–1444
Pace L, Salvan A (1997) Principles of statistical inference from a neo-Fisherian perspective. World Scientific, Singapore
Self SG, Liang KY (1987) Asymptotic properties of maximum likelihood estimators and likelihood ratio test under nonstandard conditions. J Am Stat Assoc 82:605–610
Acknowledgements
This research was partially supported by MIUR, Italy, with grant PRIN 2006132978. The numerical work was undertaken on the R package sn written by Azzalini and available at http://cran.r-project.org/src/contrib/PACKAGES.html. The Authors thank an anonymous referee for his constructive comments.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
Let \(S_{(\xi ,\omega ,\alpha ,\nu )}(y)=\big \{ S_\xi (y), S_\omega (y), S_\alpha (y), S_\nu (y) \big \}'\) be the score function of the skew t-model based on a single observation y; thus, \(S_{\xi }(y)=\partial \ln f\big ( y; \xi , \omega , \alpha , \nu \big ) / \partial \xi \), and so forth. The components of S(y) are
where \(\varsigma =\alpha z \tau \), \(w=t(\varsigma ; \nu +1) / T(\varsigma ;\nu +1)\),
and \(\varPsi (x)=\partial \ln \big \{\varGamma (x)\big \}/\partial x\). The expected information matrix for \(\big (\xi ,\omega ,\alpha ,\nu \big )\) is singular under the skew normal distribution since \(S_\nu (y)\) can be shown to be of order \(O\big (\nu ^{-2}\big )\), and hence, it vanishes as \(\nu \rightarrow \infty \). The situation is further complicated under the normal distribution: when \(\alpha =0\), as \(\nu \rightarrow \infty \), the components of the score vector for the location parameter \(\xi \) and the skewness parameter \(\alpha \) tend to \(S^N_\xi (y)= {{ z}/\omega }\) and \(S^N_\alpha (y)= z(2 / \pi )^{1/2}\), which are linearly dependent.
The problem of the score component \(S_{\nu }(y)\) tending to 0 as \(\nu \rightarrow \infty \) can be remedied by using the inverse degrees of freedom, \(\kappa =1/\nu \), in place of \(\nu \). Thus, the parameter now becomes \((\xi ,\omega ,\alpha ,\kappa )\), and the skew normal distribution corresponds to the boundary case \(\kappa =0\).
The component of the score function corresponding to \(\kappa \) is \(S_{\kappa }(y)=-\kappa ^{-2}S_{\nu }(y)\). As \(\kappa \rightarrow 0\), the components of the score function for \(\big (\xi ,\omega ,\alpha ,\kappa \big )\) tend to
where \(w_\phi =\phi (\alpha z) /\varPhi (\alpha z)\), and \(\phi (z)\) and \(\varPhi (z)\) are the density function and cumulative distribution function, respectively, of the standard normal distribution. In particular, \(S^{SN}_\kappa (y)\) is non-zero, and the components \(S_\xi ^{SN} (y)\), \(S_\omega ^{SN} (y)\), \(S_\alpha ^{SN} (y)\), and \(S^{SN}_\kappa (y)\) are linearly independent provided \(\alpha \ne 0\). Consequently, the information matrix for the parameterization \(\big (\xi ,\omega ,\alpha ,\kappa \big )\) is non-singular under the skew normal distribution when \(\alpha \ne 0\).
The problem that \(S_\xi (y)\) and \(S_\alpha (y)\) tend to linear dependency as \(\kappa \rightarrow 0\) when \(\alpha =0\) persists for the \((\xi ,\omega ,\alpha ,\kappa )\) parameterization; this parameterization does, however, serve as a useful intermediate step for defining the centered parameterization \(\big (\mu ,\sigma ^2,\gamma _1,\gamma _2\big )\), introduced in Sect. 2.1. The expected information matrix for the parameter \(\big (\mu ,\sigma ^2,\gamma _1,\gamma _2\big )\) is non-singular as \(\kappa \rightarrow 0\) for all \(\alpha \).
An expression for the Jacobian \(\partial \big (\mu ,\sigma ^2,\gamma _1,\gamma _2\big )/\partial \big (\xi ,\omega ,\alpha ,\nu \big )\) given by Azzalini (personal communication, see also Arellano-Valle and Azzalini 2013) yields
where
\(b=(2/\pi )^{1/2}\), \(\delta =\alpha /(1+\alpha ^2)^{1/2}\), \(\lambda _2=1-b^2\delta ^2\), \(\delta '=\big (1+\alpha ^2\big )^{-3/2}\), \(a_1=4 \pi - 12 \delta ^2- \delta ^2 \pi + 4 \delta ^4 \), \(a_2=18 \delta ^2 \pi - 36 \delta ^4- 2 \delta ^4 \pi + 12 \delta ^6 -3\pi ^2\), \(a_3=6\delta ^4+8 \delta ^2 \pi - 36 \delta ^2+3\pi \), and \(c=12-3\pi -4\delta ^2+2\delta ^4\).
Let \(S_{(\mu ,\sigma ^2,\gamma _1,\gamma _2)}(y)=\big \{ S_{\mu }(y), S_{\sigma ^2}(y), S_{\gamma _1}(y), S_{\gamma _2}(y) \big \}'\) be the score function of the skew t-model for the centered parameterization \(\big (\mu ,\sigma ^2,\gamma _1,\gamma _2\big )\) based on a single observation y. As \(\kappa \rightarrow 0\), the score function tends to
and hence, it follows from the preceding calculations that the components of the score vector tend to
As \(\alpha \rightarrow 0\), these expressions tend to
which are the components of the score function for \(\big (\mu ,\sigma ^2,\gamma _1,\gamma _2\big )\) under the normal model. In particular, under the normal model, the expected information matrix of the skew t-model for \(\big (\mu ,\sigma ^2,\gamma _1,\gamma _2\big )\) is
which is invertible. Thus, as \(\kappa \rightarrow 0\), the expected information matrix for the centered parameterization is non-singular, even for the case \(\alpha =0\).
Rights and permissions
About this article
Cite this article
DiCiccio, T.J., Monti, A.C. Testing for sub-models of the skew t-distribution. Stat Methods Appl 27, 25–44 (2018). https://doi.org/10.1007/s10260-017-0387-x
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10260-017-0387-x