Abstract
We propose and study the class of Box–Cox elliptical distributions. It provides alternative distributions for modeling multivariate positive, marginally skewed and possibly heavy-tailed data. This new class of distributions has as a special case the class of log-elliptical distributions, and reduces to the Box–Cox symmetric class of distributions in the univariate setting. The parameters are interpretable in terms of quantiles and relative dispersions of the marginal distributions and of associations between pairs of variables. The relation between the scale parameters and quantiles makes the Box–Cox elliptical distributions attractive for regression modeling purposes. Applications to data on vitamin intake are presented and discussed.
Similar content being viewed by others
References
Arellano-Valle RB, Branco MD, Genton MG (2006) A unified view on skewed distributions arising from selections. Can J Stat 34(4):581–601
Birnbaum Z, Meyer PL (1953) On the effect of truncation in some or all coordinates of a multi-normal population. J Indian Soc Agric Stat 5:17–28
Fang KT, Kotz S, Ng KW (1990) Symmetric multivariate and related distributions. Chapman and Hall, London
Ferrari SLP, Fumes G (2017) Box–Cox symmetric distributions and applications to nutritional data. Adv Stat Anal 101:321–344
Genz A, Bretz F (2009) Computation of multivariate normal and \(t\) probabilities. Springer, Heidelberg
Gómez-Sánchez-Manzano E, Gómez-Villegas MA, Marín JM (2008) Multivariate exponential power distributions as mixtures of normal distributions with Bayesian applications. Commun Stat Theory Methods 37(6):972–985
Gupta AK, Varga T, Bodnar T (2013) Elliptically contoured models in statistics and portfolio theory. Springer, New York
Ho HJ, Lin TI, Chen HY, Wang WL (2012) Some results on the truncated multivariate \(t\) distribution. J Stat Plan Inference 142:25–40
Horrace WC (2005) Some results on the multivariate truncated normal distribution. J Multivar Anal 94:209–221
Kan R, Robotti C (2017) On moments of folded and truncated multivariate normal distributions. J Comput Graph Stat 26(4):930–934
Kano Y (1994) Consistency property of elliptical probability density functions. J Multivar Anal 51:139–147
Kim H-J (2010) A class of weighted multivariate elliptical models useful for robust analysis of nonnormal and bimodal data. J Korean Stat Soc 39:83–92
Lange KL, Little RJA, Taylor JMG (1989) Robust statistical modeling using the \(t\) distribution. J Am Stat Assoc 84(408):881–896
Manjunath BG, Wilhelm S (2012). Moments calculation for the double truncated multivariate normal density. Working paper arXiv:1206.5387v1
Marchenko YV, Genton MG (2010) Multivariate log-skew-elliptical distributions with applications to precipitation data. Environmetrics 21(3–4):318–340
Nocedal J, Wright SJ (2006) Numerical optimization. Springer, New York
Quiroz AJ, Nakamura M, Pérez FJ (1996) Estimation of a multivariate Box–Cox transformation to elliptical symmetry via the empirical characteristic function. Ann Inst Stat Math 48(4):687–709
R Core Team (2016) R: a language and environment for statistical computing. R Foundation for Statistical Computing. Vienna https://www.R-project.org/
Rigby RA, Stasinopoulos DM (2004) Smooth centile curves for skew and kurtotic data modelled using the Box–Cox power exponential distribution. Stat Med 23(19):3053–3076
Rigby RA, Stasinopoulos DM (2006) Using the Box–Cox \(t\) distribution in GAMLSS to model skewness and kurtosis. Stat Model 6(3):209–229
Rousseeuw PJ, Ruts I, Tukey JW (1999) The Bagplot: a bivariate boxplot. Am Stat 53(4):382–387
Stasinopoulos DM, Rigby RA, Akantziliotou C (2008). Instructions on how to use the GAMLSS package in R. London. http://www.gamlss.org
Tallis GM (1961) The moment generating function of the truncated multi-normal distribution. J R Stat Soc 23:223–229
Tallis GM (1963) Elliptical and radial truncation in normal populations. Ann Math Stat 34(3):940–944
Tallis GM (1965) Plane truncation in normal populations. J R Stat Soc 27(2):301–307
Voudouris V, Gilchrist R, Rigby RA, Sedgwick J, Stasinopoulos DM (2012) Modelling skewness and kurtosis with the BCPE density in GAMLSS. J Appl Stat 39(6):1279–1293
Acknowledgements
We thank José Eduardo Corrente for providing the data used in this study. Funding was provided by Conselho Nacional de Desenvolvimento Científico e Tecnológico—CNPq (Grant No. 304388-2014-9) and Fundação de Amparo à Pesquisa do Estado de São Paulo—FAPESP (Grant No. 2012/21788-2). The first author received Ph.D. scholarships from Coordenação de Aperfeiçoamento de Pessoal de Nível Superior—CAPES—and CNPq. We are thankful to the Editor, Associate Editor and the anonymous reviewers for their helpful comments and suggestions.
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix A: Proof of the Theorem 1
The conditional PDF of \(W_k|{\varvec{W}}_{-k}\), \(k=1,\ldots ,p\), is given by
From the identity \(({\varvec{w}}-{\varvec{\mu }})'{\varvec{\varSigma }}^{-1}({\varvec{w}}-{\varvec{\mu }}) = [(w_k-\mu _{k.-k})/\sigma _{k.-k}]^2+q({\varvec{w}}_{-k})\), we get the result.
Appendix B: Proof of the Theorem 2
If \({\varvec{W}}=T_{{\varvec{\lambda }},{\varvec{\mu }}}({\varvec{Y}})\sim {\text{ TE }l}_p({\varvec{\xi }},{\varvec{\varSigma }};R({\varvec{\lambda }});g)\), then its PDF is given by
Let \(V:R({\varvec{\lambda }})\rightarrow R({\varvec{\lambda }})\) be the transformation defined as \(V({\varvec{w}})={\varvec{D}}_{{\varvec{\alpha }}}^{-1}({\varvec{w}}-{\varvec{\xi }})\), and let \({\varvec{U}}=V({\varvec{W}})\), with Jacobian \(J({\varvec{w}}\rightarrow {\varvec{u}})=\prod _{k=1}^{p}(1+\lambda _k \xi _k)\). The PDF of \({\varvec{U}}\) is
Hence, \({\varvec{U}}\sim {\text{ TE }l}_p({\varvec{0}},{\varvec{D}}_{{\varvec{\alpha }}}^{-1}{\varvec{\varSigma }}{\varvec{D}}_{{\varvec{\alpha }}}^{-1};R({\varvec{\lambda }});g)\). Because \({\varvec{U}}=V(T_{{\varvec{\lambda }},{\varvec{\mu }}}({\varvec{Y}}))=T_{{\varvec{\lambda }},{\varvec{\delta }}}({\varvec{Y}})\), where \({\varvec{\delta }}=T_{{\varvec{\lambda }},{\varvec{\mu }}}^{-1}({\varvec{\xi }})\), then from Definition 4 we have \({\varvec{Y}}=T_{{\varvec{\lambda }},{\varvec{\delta }}}^{-1}({\varvec{U}})\sim \text{ BCE }\ell _p({\varvec{\delta }},{\varvec{\lambda }},{\varvec{D}}_{{\varvec{\alpha }}}^{-1}{\varvec{\varSigma }}{\varvec{D}}_{{\varvec{\alpha }}}^{-1};g)\).
On the other hand, if \({\varvec{Y}}\sim \text{ BCE }\ell _p({\varvec{\delta }},{\varvec{\lambda }},{\varvec{D}}_{{\varvec{\alpha }}}^{-1}{\varvec{\varSigma }}{\varvec{D}}_{{\varvec{\alpha }}}^{-1};g)\), then its PDF is
Now, from the transformation \({\varvec{W}}=T_{{\varvec{\lambda }},{\varvec{\mu }}}({\varvec{Y}})\), with Jacobian \(J({\varvec{y}}\rightarrow {\varvec{w}})=\prod _{k=1}^{p}\mu _k(1+\lambda _k w_k)^{1/\lambda _k-1}\), in the PDF (17) we arrive at PDF (16).
Appendix C: Proof of Theorem 3
-
1.
From \({\varvec{T}}={\varvec{D_\alpha Y}}\), with Jacobian \(J({\varvec{y}}\rightarrow {\varvec{t}})=\prod _{k=1}^{p}\alpha _k^{-1}\), in (6), we get the PDF of \({\varvec{T}}\) as
$$\begin{aligned} f_{{\varvec{T}}}({\varvec{t}}) = \dfrac{g({\varvec{w}}'{\varvec{\varSigma }}^{-1}{\varvec{w}})\prod _{k=1}^{p}\frac{t_k^{\lambda _k-1}}{(\alpha _k\mu _k)^{\lambda _k}}}{\int _{R({\varvec{\lambda }})}g({\varvec{w}}'{\varvec{\varSigma }}^{-1}{\varvec{w}})\,\mathrm{d}{\varvec{w}}},\quad {\varvec{t}}\in \mathbb {R}_{+}^p, \end{aligned}$$where \({\varvec{w}}{=}T_{{\varvec{\lambda }},{\varvec{\mu }}}({\varvec{D}}_{{\varvec{\alpha }}}^{-1}{\varvec{t}}){=}T_{{\varvec{\lambda }},{\varvec{D_\alpha \mu }}}({\varvec{t}})\). Hence, \({\varvec{T}}={\varvec{D_{\alpha }Y}}\sim \text{ BCE }\ell _p({\varvec{D_\alpha \mu }},{\varvec{\lambda }},{\varvec{\varSigma }};g)\).
-
2.
Note that the PDF of \({\varvec{Y}}\), given in (6), can be expressed as
$$\begin{aligned} f_{{\varvec{Y}}}({\varvec{y}})=\dfrac{g({\varvec{v}}'({\varvec{D_\beta \varSigma D_\beta }})^{-1}{\varvec{v}})\prod _{k=1}^{p}\frac{|\beta _k|y_k^{\lambda _k-1}}{\mu _k^{\lambda _k}}}{\int _{R({\varvec{D}}_{{\varvec{\beta }}}^{-1}{\varvec{\lambda }})}g({\varvec{v}}'({\varvec{D_\beta \varSigma D_\beta }})^{-1}{\varvec{v}})\,\mathrm{d}{\varvec{v}}},\quad {\varvec{y}}\in \mathbb {R}_{+}^p, \end{aligned}$$where \({\varvec{v}}={\varvec{D_\beta }}T_{{\varvec{\lambda }},{\varvec{\mu }}}({\varvec{y}})\) has its kth component given by
$$\begin{aligned} v_k= {\left\{ \begin{array}{ll} \dfrac{[(y_k/\mu _k)^{\beta _k}]^{\lambda _k/\beta _k}-1}{\lambda _k/\beta _k}, &{}\quad \lambda _k\ne 0,\\ \quad \,\,\,\,\log (y_k/\mu _k)^{\beta _k}, &{}\quad \lambda _k=0, \end{array}\right. } \end{aligned}$$for \(k=1,\ldots ,p\). From \(U_k=(Y_k/\mu _k)^{\beta _k}\), \(k=1,\ldots ,p\), with Jacobian \(J({\varvec{y}}\rightarrow {\varvec{u}})=\prod _{k=1}^{p}\mu _k \beta _k^{-1}u_k^{1/\beta _k-1}\), we arrive at the desired result.
-
3.
Plugging \({\varvec{\lambda }}={\varvec{1}}\) in (6) we have that the PDF of \({\varvec{Y}}\) is
$$\begin{aligned} f_{{\varvec{Y}}}({\varvec{y}})=\dfrac{g(({\varvec{y}}-{\varvec{\mu }})'({\varvec{D_{\mu }\varSigma D_{\mu }}})^{-1}({\varvec{y}}-{\varvec{\mu }}))\prod _{k=1}^{p}\frac{1}{\mu _k}}{\int _{R({\varvec{1}})}g({\varvec{w}}'{\varvec{\varSigma }}^{-1}{\varvec{w}})\,\mathrm{d}{\varvec{w}}},\quad {\varvec{y}}\in \mathbb {R}_{+}^p. \end{aligned}$$From the change of variables \({\varvec{w}}={\varvec{D}}_{{\varvec{\mu }}}^{-1}({\varvec{y}}-{\varvec{\mu }})\) we arrive at the desired result.
Appendix D: Proof of Theorem 4
Plugging \({\varvec{\varSigma }}_{12}={\varvec{0}}\) in (8), and then making the change of variables \({\varvec{s}}=T({\varvec{w}}_2)={\varvec{\varSigma }}_{22}^{-1/2}{\varvec{w}}_2\), the marginal PDF of \({\varvec{Y}}_1\) is
Note that \(g_{1}(u) = \int _{T(R({\varvec{\lambda _2}}))}g(u + {\varvec{s}}'{\varvec{s}})\,\mathrm{d}{\varvec{s}} \le \int _{\mathbb {R}^{p-r}}g(u+{\varvec{s}}'{\varvec{s}})\,\mathrm{{d}}{\varvec{s}}=h_1(u)\), \(u\ge 0\), where \(h_1\) is such that \(\int _{0}^{\infty }t^{r-1}h_1(t^2)\,\text {d}t<\infty \) (Fang et al. 1990, Sec. 2.2). This completes the proof.
Appendix E: Proof of Theorem 5
Because \({\varvec{Y}}\sim \text{ LE }\ell _p({\varvec{\mu }},{\varvec{\varSigma }};g)\), with \(g(u)\propto \int _{0}^{\infty }t^{p/2}\exp (-ut/2)\,\mathrm{d}H(t)\), \(u\ge 0\), then \({\varvec{X}}=T_{{\varvec{0}},{\varvec{\mu }}}({\varvec{Y}})\sim \text{ E }\ell _p({\varvec{0}},{\varvec{\varSigma }};g)\). Thus, \({\varvec{X}}_1=T_{{\varvec{0}},{\varvec{\mu }}_1}({\varvec{Y}}_1)\sim \text{ E }\ell _p({\varvec{0}},{\varvec{\varSigma }}_{11};g)\) (Kano 1994). Hence, \({\varvec{Y}}_1\sim \text{ LE }\ell _r({\varvec{\mu }}_1,{\varvec{\varSigma }}_{11};g)\).
Appendix F: Proof of Theorem 6
The conditional PDF of \({\varvec{Y}}_1|{\varvec{Y}}_2\) is given by
Because \({\varvec{w}}'{\varvec{\varSigma }}^{-1}{\varvec{w}}={\varvec{u}}_1' \bigl ({\varvec{D}}_{{\varvec{\alpha }}({\varvec{w}}_2)}^{-1}{\varvec{\varSigma }}_{11\cdot 2}{\varvec{D}}_{{\varvec{\alpha }}({\varvec{w}}_2)}^{-1}\bigr )^{-1}{\varvec{u}}_1+q({\varvec{w}}_2)\), where \({\varvec{u}}_1={\varvec{D}}_{{\varvec{\alpha }}({\varvec{w}}_2)}^{-1} ({\varvec{w}}_1-{\varvec{\mu }}_1({\varvec{w}}_2))=T_{{\varvec{\lambda }}_1,{\varvec{\delta }}_1}({\varvec{y}}_1)\), (18) can be expressed as
where \({\varvec{u}}_1=T_{{\varvec{\lambda }}_1,{\varvec{\delta }}_1}({\varvec{y}}_1)\). Since \(\prod _{k=1}^{r}\mu _k^{\lambda _k}(1+\lambda _k\mu _{1k}({\varvec{w}}_2)) = \prod _{k=1}^{r}{\delta _k}^{\lambda _k}\) the proof is complete.
Appendix G: Proof of Theorem 7
\({\varvec{Y}}_1\) and \({\varvec{Y}}_2\) are independent if, and only if, the PDF of \({\varvec{Y}}\sim \text{ BCE }\ell _p({\varvec{\mu }},{\varvec{\lambda }},{\varvec{\varSigma }};g)\) given in (6) is such that \(f_{{\varvec{Y}}}({\varvec{y}})=f_{{\varvec{Y}}_1}({\varvec{y}}_1)f_{{\varvec{Y}}_2}({\varvec{y}}_2)\). This condition is satisfied if, and only if, \({\varvec{\varSigma }}_{12}={\varvec{0}}\) and the DGF g satisfies the functional equation \(g(u+v)=g(u)g(v)\), with \(u\ge 0\) and \(v\ge 0\), for which \(g(u)=\exp (-ku)\), for some \(k\ge 0\), is a solution (Gupta et al. 2013, Sec. 1.3). From \(\int _{0}^{\infty }t^{p-1}\exp (-kt^2)\,\text {d}t=2^{p/2-1}\varGamma (p/2)\), we find that \(k=1/2\). Hence, \({\varvec{Y}}_1\) and \({\varvec{Y}}_2\) are independent if, and only if, \({\varvec{\varSigma }}_{12}={\varvec{0}}\) and \({\varvec{Y}}\sim \text{ BCN }_p({\varvec{\mu }},{\varvec{\lambda }},{\varvec{\varSigma }})\).
Appendix H: Proof of Theorem 8
From (6) we have
where \({\varvec{w}}=T_{{\varvec{\lambda }},{\varvec{\mu }}}({\varvec{y}})\). By making the change of variables \({\varvec{u}} ={\varvec{D}}_{{\varvec{\mu }}}^{-1}{\varvec{y}}\) we arrive at the desired result.
Appendix I: Marginal PDF of \(Y_k\)
The function \(g_{{{\varvec{\varUpsilon }}_k}}\) given in (10) can be defined in \(\mathbb {R}\). Hence, we can define a random variable \(U_k\in \mathbb {R}\) from the PDF
where \(c_k^{-1} = \int _{-\infty }^{\infty }g_{{{\varvec{\varUpsilon }}_k}}(t)\,\mathrm{d}t\). The CDF of \(U_k\) is given by (11). We now define \(S_k\in I(\lambda _k\sqrt{\sigma _{kk}})\) as a random variable \(U_k\) truncated on \(I(\lambda _k\sqrt{\sigma _{kk}})\). The PDF of \(S_k\) is given by
From the transformation \(S_k=\sigma _{kk}^{-1/2}T_{\lambda _k,\mu _k}(Y_k)\), with Jacobian \(J(s_k\rightarrow y_k)=\sigma _{kk}^{-1/2}\mu _{k}^{-\lambda _k}y_k^{\lambda _k-1}\), we arrive at the PDF of \(Y_k\) given in (9).
Appendix J: Proof of Theorem 9
Because \({\varvec{Y}}\sim \text{ BCE }\ell _p({\varvec{\mu }},{\varvec{\lambda }},{\varvec{\varSigma }};g)\), the PDF of \(Y_k\), \(k=1,\ldots ,p\), is given by (9), where \(S_k=\sigma _{kk}^{-1/2}T_{\lambda _k,\mu _k}(Y_k)\) has PDF given in (19). The CDF of \(S_k\) is
from which we have that the \(\alpha \)-quantile \(y_{k,\alpha }\) of \(Y_k\), \(\alpha \in (0,1)\), is such that \({\text{ P }}(Y_k\le y_{k,\alpha })=\alpha \), or equivalently \({\text{ P }}[S_k \le \sigma _{kk}^{-1/2}T_{\lambda _k,\mu _k}(y_{k,\alpha })]=\alpha \). Hence,
where \(s_{k,\alpha }\) is such that \(F_{S_k}(s_{k,\alpha })=\alpha \), with \(F_{S_k}\) given in (20). Therefore, \(s_{k,\alpha }\) is given by
where \(F_{U_k}\) is the CDF given in (11).
Rights and permissions
About this article
Cite this article
Morán-Vásquez, R.A., Ferrari, S.L.P. Box–Cox elliptical distributions with application. Metrika 82, 547–571 (2019). https://doi.org/10.1007/s00184-018-0682-z
Received:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00184-018-0682-z