Abstract
We generalize the theory of linear models for doubly multivariate data from matrix-variate normally distributed errors to matrix-variate skew normally distributed errors. In addition, we assume that the covariance matrix \(\varvec{\varOmega }\) defining the location-scale matrix-variate skew normal distribution has block compound symmetry structure. We derive the maximum likelihood estimators of the model’s parameters; the Fisher information matrix for the direct, working, and centered parametrizations; Rao’s score tests and likelihood ratio tests for model building tests of hypotheses; and a hypothesis test for the centered intercept. A profiling argument is used to reduce the dimensionality of the optimization method used to obtain the maximum likelihood estimators and a comprehensive discussion of initial values is provided. Finally, we provide a real-world example to illuminate these derivations and apply a goodness-of-fit test to validate the distributional assumption.
Similar content being viewed by others
References
Akdemir D, Gupta AK (2010) A matrix variate skew distribution. Eur J Pure Appl Math 3(2):128–140
Arellano-Valle RB, Azzalini A (2006) On the unification of families of skew-normal distributions. Scand J Stat 33:561–574
Arellano-Valle RB, Azzalini A (2008) The centred parametrization for the multivariate skew-normal distribution. J Multivar Anal 99(7):1362–1382
Arnold SF (1973) Application of the theory of products of problems to certain patterned covariance matrices. Ann Stat 1(4):682–699
Arnold SF (1979) Linear models with exchangeably distributed errors. J Am Stat Assoc 74:194–199
Azzalini A, Dalla Valle A (1996) The multivariate skew-normal distribution. Biometrika 83:715–726
Azzalini A, Capitanio A (2014) The skew-normal and related families. Cambridge University Press, New York
Balakrishnan N, Capitanio A, Scarpa B (2014) A test for multivariate skew-normality based on its canonical form. J Multivar Anal 128:19–32
Chen JT, Gupta AK (2005) Matrix variate skew normal distributions. Stat: J Theor Appl Stat 39(3):247–253
Dutilleul P (1999) The MLE algorithm for the matrix normal distribution. J Stat Comput Simul 64:105–123
Gupta A, Gonzalez-Farias G, Dominguez-Molina J (2004) A multivariate skew normal distribution. J Multivar Anal 89(1):181–190
Harrar SW, Gupta AK (2008) On matrix variate skew-normal distributions. Stat: J Theor Appl Stat 42(2):179–194
Jana S, Balakrishnan N, Jemila H (2018) Estimation of the parameters of the extended growth curve model under multivariate skew normal distribution. J Multivar Anal 166:111–128
Kheradmandi A, Abdolrahman R (2015) Estimation in skew-normal linear mixed measurement error models. J Multivar Anal 136:1–11
Lachos VH, Bolfarine H, Arellano-Valle RB, Montenegro LC (2007) Likelihood-based inference for multivariate skew-normal regression models. Commun Stat-Theory Methods 36(9):1769–1786
Leiva R (2007) Linear discrimination with equicorrelated training vectors. J Multivar Anal 98(2):384–409
Lu N, Zimmerman DL (2005) The likelihood ratio test for a separable covariance matrix. Stat Probab Lett 73:449–457
Lin TI, Wang WL (2013) Multivariate skew-normal at linear mixed models for multi-outcome longitudinal data. Stat Model 13(3):199–221
Magnus JR, Neudecker H (1986) Symmetry, 0–1 matrices and Jacobians: a review. Econom Theory 2:157–190
Magnus JR, Neudecker H (1988) Matrix differential calculus with applications in statistics and econometrics. Wiley, New York
Opheim T, Roy A (2019) Revisiting the linear models with exchangeably distributed errors. Proc Am Stat Assoc 2677–2686
Opheim T, Roy A (2021) Linear models for multivariate repeated measures data with block exchangeable covariance structure. Comput Stat. https://doi.org/10.1007/s00180-021-01064-9
Rao CR (1945) Familial correlations or the multivariate generalizations of the intraclass correlation. Curr Sci 14:66–67
Rao CR (1948) Large sample tests of statistical hypotheses concerning several parameters with applications to problems of estimation. Math Proc Camb Philos Soc 44:50–57
Rao CR (1953) Discriminant functions for genetic differentiation and selection. Sankhya 12:229–246
Rao CR (2005) Score test: historical review and recent developments. In: Balakrishnan N, Nagaraja HN, Kannan N (eds) Advances in Ranking and Selection, Multiple Comparisons, and Reliability. Statistics for Industry and Technology, Birkhäuser Boston, pp 3–20. https://doi.org/10.1007/0-8176-4422-9_1
Roy A (2007) A note on testing of Kroneckar product covariance structures for doubly multivariate data. In: Proceedings of the American Statistical Association, Seattle, Washington, pp 2157–2162
Roy A, Leiva R, Žežula I, Klein D (2015) Testing of equality of mean vectors for paired doubly multivariate observations in blocked compound symmetric covariance matrix setup. J Multivar Anal 137:50–60
Roy A, Zmyślony R, Fonseca M, Leiva R (2016) Optimal estimation for doubly multivariate data in blocked compound symmetric covariance structure. J Multivar Anal 144:81–90
Roy A, Filipiak K, Klein D (2018) Testing a block exchangeable covariance matrix. Stat: A J Theor Appl Stat 52(2):393–408
Roy A, Leiva R (2011) Estimating and testing a structured covariance matrix for three-level multivariate data. Commun Stat Theory Methods 40(10):1945–1963
Sperling MR, Gur RC, Alavi A, Gur RE, Resnick S, O’Connor MJ, Reivich M (1990) Subcortical metabolic alterations in partial epilepsy. Epilepsia 31(2):145–155
Szatrowski TH (1982) Testing and estimation in the block compound symmetry problem. J Educ Stat 7(1):3–18
Tsukada S (2018) Hypothesis testing for independence under blocked compound symmetric covariance structure. Commun Math Stat 6:163–184
Wang WL (2015) Approximate methods for maximum likelihood estimation of multivariate nonlinear mixed-effects models. Entropy 17:5353–5381
Wilks S (1938) The large sample distribution of the likelihood ratio for testing composite hypotheses. Ann Math Stat 9:60–62
Žežula I, Klein D, Roy A (2018) Testing of multivariate repeated measures data with block exchangeable covariance structure. Test 27(2):360–378
Acknowledgements
The authors want to thank an anonymous reviewer for the careful reading and valuable suggestions that led to a quite improved version of the manuscript.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of Interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article is part of the topical collection “Celebrating the Centenary of Professor C. R. Rao” guest edited by, Ravi Khattree, Sreenivasa Rao Jammalamadaka , and M. B. Rao.
Appendices
A Appendix: Definitions and Results of Special Matrices
Many references exist for the following definitions and propositions. We direct the reader to but one, Magnus and Neudecker [19]. Let \({\varvec{N}}_m\) be the symmetric idempotent \(m^2 \times m^2\) matrix defined as \({\varvec{N}}_m=\frac{1}{2}({\varvec{I}}_{m^2}+ {\varvec{K}}_{m,m})\), where \({\varvec{I}}_{m^2}\) and \({\varvec{K}}_{m,m}\) represent the identity matrix and the commutation matrix, respectively. A unique \(m^2\times m(m+1)/2-\)dimensional matrix \({\varvec{D}}_m\) is called a duplication matrix if
Using the above definitions, we have the following proposition.
Proposition 1
The following equalities hold:
-
(i)
\({\varvec{K}}_{p,m}({\varvec{A}} \otimes {\varvec{y}}) = {\varvec{y}} \otimes {\varvec{A}}\), for any \(m \times n\) matrix \({\varvec{A}}\) and \(p \times 1\) vector \({\varvec{y}}\);
-
(ii)
\(\mathrm{vec}({\varvec{A}}\otimes {\varvec{B}})=({\varvec{I}}_l\otimes {\varvec{K}}_{n,k}\otimes {\varvec{I}}_m)(\mathrm{vec}{\varvec{A}}\otimes \mathrm{vec}{\varvec{B}})\) for any \(k\times l\) matrix \({\varvec{A}}\) and \(m\times n\) matrix \({\varvec{B}}\);
-
(iii)
\(\mathrm{vec}\left( {\varvec{A}}{\varvec{B}}{\varvec{C}}\right) = \left( {\varvec{C}}^{\prime } \otimes {\varvec{A}}\right) \mathrm{vec}\left( {\varvec{B}}\right) \);
-
(iv)
\({\varvec{N}}_m({\varvec{A}}\otimes {\varvec{A}}){\varvec{N}}_m={\varvec{N}}_m({\varvec{A}}\otimes {\varvec{A}})= ({\varvec{A}}\otimes {\varvec{A}}){\varvec{N}}_m\) for any \(m\times m\) symmetric matrix \({\varvec{A}}\);
-
(v)
\({\varvec{D}}_m{\varvec{D}}^+_m={\varvec{N}}_m\);
-
(vi)
\({\varvec{N}}_m \mathrm{vec}{\varvec{A}} = \mathrm{vec}{\varvec{A}}\) for any \(m\times m\) symmetric matrix \({\varvec{A}}\).
Lemma 1
The following lemma follows from Proposition 1(i):
-
(i)
\(({\varvec{I}}_n\otimes {\varvec{K}}_{p,n}\otimes {\varvec{I}}_p)(\mathrm{vec}({\varvec{P}}_n) \otimes {\varvec{I}}_p^2)= \frac{{\varvec{1}}_n}{n} \otimes {\varvec{I}}_p \otimes {\varvec{1}}_n \otimes {\varvec{I}}_p\);
-
(ii)
\(({\varvec{I}}_n\otimes {\varvec{K}}_{p,n}\otimes {\varvec{I}}_p)(\mathrm{vec}({\varvec{I}}_n) \otimes {\varvec{I}}_p^2)= \sum _{i=1}^n {\varvec{e}}_i \otimes {\varvec{I}}_p \otimes {\varvec{e}}_i \otimes {\varvec{I}}_p\);
Proof
Starting from the LHS of equation (i), we have
Likewise, starting from the LHS of equation (ii), we have
\(\square \)
B Appendix: Proofs of BCS Related Results
1.1 B.1 Intermediary Expectations
Using the results and notation in Sect. 3 with \(\varvec{\eta }=\varvec{\eta }^{\star },\) note that
Since \({\varvec{z}}_1, \ldots , {\varvec{z}}_N\) are independent, the proof of the expectations in Sect. 5.1 is
1.2 B.2 Fisher Information
Using the expectations developed in Appendix B.1 and Lemma 1(i) and (ii) in Appendix A, the blocks of the information matrix reported in Sect. 5.1 are proven as
1.3 B.3 MLEs under \(H_{0,int}: \varvec{\alpha }^{CP}={\varvec{0}}\)
The MLEs under \(H_{0,int}: \varvec{\alpha }^{CP} = {\varvec{0}}\) may be obtained using a similar profiling argument as in Sect. 5.2. First recall that \(\varvec{\alpha } = \varvec{\alpha }^{CP} - \sqrt{2/\pi } \left( 1 + n \varvec{\eta }^{\prime }\varvec{\varSigma }_{\small {{\varvec{z}}_1}} \varvec{\eta } \right) ^{-1/2} \varvec{\varSigma }_{\small {{\varvec{z}}_1}} \varvec{\eta }\). Let \({\varvec{z}}_{i,\alpha }\) denote \({\varvec{z}}_i\) evaluated at \(\varvec{\alpha }^{CP} = {\varvec{0}}\). From the previous transformation, we have
In addition let \({\varvec{Z}}_\alpha = \left( {\varvec{z}}_{1,\alpha }, \ldots , {\varvec{z}}_{N,\alpha }\right) \) and define
Since the transformation from \(\varvec{\alpha }\) to \(\varvec{\alpha }^{CP}\) does not introduce any terms including \(\varvec{\varSigma }_{\small {{\varvec{Z}}_2}}\), the MLE of \(\varvec{\varSigma }_{\small {{\varvec{Z}}_2}}\), given \(\varvec{\theta }^{TP} = \left( \mathrm{vec}(\varvec{\gamma }^{\prime }), \mathrm{vech}(\varvec{\varSigma }_{\small {{\varvec{z}}_1}}), \varvec{\eta } \right) \), is
However, since this transformation introduced new terms involving \(\varvec{\varSigma }_{\small {{\varvec{z}}_1}}\), it can be shown that no closed-form expression exists for the MLE of \(\varvec{\varSigma }_{\small {{\varvec{z}}_1}}\) given the other parameters are known.
Plugging in \(\widehat{\varvec{\varSigma }}_{\small {{\varvec{Z}}_2}, \varvec{\theta }^{TP}}\) into the likelihood given in Eq. (8), keeping the transformation of \({\varvec{Z}}\) to \({\varvec{Z}}_\alpha \) in mind, results in the profiled log likelihood (up to an additive constant)
and its first differential is
Using a quasi-Newton method, the profile likelihood can be maximized, and the process may be quickened by the inclusion of the gradient, formed by the above differential. Initial values for the optimization procedure follow from a similar argument discussed in Sect. 5.2.
Rights and permissions
About this article
Cite this article
Opheim, T., Roy, A. Score Tests for Intercept and Slope Parameters of Doubly Multivariate Linear Models with Skew-Normal Errors. J Stat Theory Pract 15, 30 (2021). https://doi.org/10.1007/s42519-020-00159-8
Accepted:
Published:
DOI: https://doi.org/10.1007/s42519-020-00159-8
Keywords
- Maximum likelihood estimates
- Model selection
- Hypotheses tests
- Repeated measures data
- Skew-normal distributions