Abstract
Dynamic linear models are typically developed assuming that both the observational and system distributions are normal. In this work, we relax this assumption by considering a skew-normal distribution for the observational random errors, providing thus an extension of the standard normal dynamic linear model. Full Bayesian inference is carried out using the hierarchical representation of the model. The inference scheme is led by means of the adaptation of the Forward Filtering Backward sampling and the usual MCMC algorithms to perform the inference. The proposed methodology is illustrated by a simulation study and applied to the condition factor index of male and female anchovies off northern Chile. These indexes have not been studied in a dynamic linear model framework.
Similar content being viewed by others
References
Akaike H (1974) A new look at the statistical model identification. IEEE Trans Autom Control 19(6):716–723
Arellano-Valle RB, Azzalini A (2006) On the unification of families of skew-normal distributions. Scand J Stat 33(3):561–574
Arellano-Valle RB, Genton MG (2005) On fundamental skew distributions. J Multivar Anal 96(1):93–116
Arellano-Valle RB, Genton MG (2010) Multivariate unified skew-elliptical distributions. Chil J Stat 1(1):17–33
Arellano-Valle RB, Castro LM, Genton MG, Gómez HW (2008) Bayesian inference for shape mixtures of skewed distributions, with application to regression analysis. Bayesian Anal 3(3):513–539
Arellano-Valle RB, Contreras-Reyes JE, Stehlík M (2017) Generalized skew-normal negentropy and its application to fish condition factor time series. Entropy 19(10):528
Azzalini A, Capitanio A (1999) Statistical applications of the multivariate skew normal distribution. J R Stat Soc B 61(3):579–602
Azzalini A, Dalla Valle A (1996) The multivariate skew-normal distribution. Biometrika 83(4):715–726
Cabral CRB, Da-Silva CQ, Migon HS (2014) A dynamic linear model with extended skew-normal for the initial distribution of the state parameter. Comput Stat Data Anal 74:64–80
Carter CK, Kohn R (1994) On Gibbs sampling for state space models. Biometrika 81(3):541–553
Contreras-Reyes JE (2016) Analyzing fish condition factor index through skew-gaussian information theory quantifiers. Fluct Noise Lett 15(2):1650013
Contreras-Reyes JE, Canales TM, Rojas PM (2016) Influence of climate variability on anchovy reproductive timing off northern Chile. J Mar Syst 164:67–75
Fahrmeir L (1992) Posterior mode estimation by extended Kalman filtering for multivariate dynamic generalized linear models. J Am Stat Assoc 87(418):501–509
Frühwirth-Schnatter S (1994) Data augmentation and dynamic linear models. J Time Ser Anal 15(2):183–202
Gamerman D, Lopes H (2006) Monte Carlo Markov chain: stochastic simulation for Bayesian inference. Chapman & Hall CRC Londres, London
Gamerman D, Moreira AR (2004) Multivariate spatial regression models. J Multivar Anal 91(2):262–281
Gelfand AE, Ghosh SK (1998) Model choice: a minimum posterior predictive loss approach. Biometrika 85(1):1–11
González-Farías G, Domínguez-Molina J, Gupta A (2004) The closed skew-normal distribution. In Skew-elliptical distributions and their applications: a journey beyond normality, pp 25–42. Chapman & Hall/CRC, Boca Raton, FL
Gupta AK, Aziz MA, Ning W (2013) On some properties of the unified skew normal distribution. J Stat Theor Pract 7(3):480–495
Harrison PJ, Stevens CF (1976) Bayesian forecasting. J R Stat Soc B 38(3):205–247
Kalman RE (1960) A new approach to linear filtering and prediction problems. J Basic Eng 82(1):35–45
Kalman RE, Bucy RS (1961) New results in linear filtering and prediction theory. J Basic Eng 83(1):95–108
Kim H, Ryu D, Mallick B, Genton M (2014) Mixtures of skewed kalman filters. J Multivar Anal 123:228–251
Kitagawa G (1987) Non-gaussian state-space modeling of nonstationary time series. J Am Stat Assoc 82(400):1032–1041
Le Cren E (1951) The length–weight relationship and seasonal cycle in gonad weight and condition in the perch (Perca fluviatilis). J. Anim Ecol 20:201–219
Naveau P, Genton MG, Shen X (2005) A skewed Kalman filter. J Multivar Anal 94(2):382–400
Plummer M (2003) JAGS: a program for analysis of Bayesian graphical models using Gibbs sampling. In: Proceedings of the 3rd international workshop on distributed statistical computing, vol 124, p 125. Vienna
Plummer M (2016) rjags: Bayesian graphical models using MCMC. R package version 4-6
Plummer M, Best N, Cowles K, Vines K (2006) CODA: convergence diagnosis and output analysis for MCMC. R News 6(1):7–11
Rezaie J, Eidsvik J (2014) Kalman filter variants in the closed skew normal setting. Comput Stat Data Anal 75:1–14
Rezaie J, Eidsvik J (2016) A skewed unscented Kalman filter. Int J Control 89:2572–2583
Schwarz G (1978) Estimating the dimension of a model. Ann Stat 6(2):461–464
Spiegelhalter DJ, Best NG, Carlin BP, Van Der Linde A (2002) Bayesian measures of model complexity and fit. J R Stat Soc B 64(4):583–639
Tong H, Lim KS (1980) Threshold autoregression, limit cycles and cyclical data. J R Stat Soc B 42:245–292
West M, Harrison J (1999) Bayesian forecasting & dynamic models. Springer, Berlin
West M, Harrison PJ, Migon HS (1985) Dynamic generalized linear models and Bayesian forecasting. J Am Stat Assoc 80(389):73–83
Acknowledgements
We thank the Instituto de Fomento Pesquero (IFOP, Valparaíso, Chile), for providing the biological information used in this work; and to Francisco Torres-Avilés (R.I.P) and José S. Romeo for all their useful comments on an early version of this paper. Arellano-Valle’s research was supported by Grant FONDECYT (Chile) Nro. 1120121 and 1150325. We are sincerely grateful to the two anonymous reviewers for their comments and suggestions that greatly improved an early version of this manuscript. All R codes used in this paper are available in bitbucket repository: https://bitbucket.org/vate01/sndlm_v2.
Author information
Authors and Affiliations
Corresponding author
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendix: Proofs
Appendix: Proofs
Theorem 1
Proof
The proof is by induction and using some well-known multivariate normal distributional properties, as the following marginal–conditional relation:
where \({\varvec{\Omega }}={\varvec{A}}{\varvec{\Sigma }}{\varvec{A}}' + {\varvec{\Psi }}\). In fact, suppose that the first step is true. Then, by the conditional independence assumptions considered in the Step 2, the evolutive prior distribution is
which is the kernel of the pdf of \({\varvec{\theta }}_t \mid {\varvec{D}}_{t-1} \sim SUN_{p,t-1}\big ( {\varvec{m}}_{t|t-1}, {\varvec{C}}_{t|t-1}, {\varvec{\Lambda }}_{t|t-1},{\varvec{\tau }}_{t|t-1},{\varvec{\Gamma }}_{t|t-1}\big )\). To proof the Step 3 in the sequential inference scheme, we have by Eq. (32) and the relation \({\varvec{m}}_{t|t}-{\varvec{m}}_{t|t-1}= {\varvec{C}}_{t|t-1}{\varvec{F}}_t {\varvec{Q}}_t^{-1}({\varvec{y}}_t - {\varvec{f}}_t)\) that
where
and \({\varvec{\Gamma }}_{t\mid t} = \mathrm{diag}(1,{\varvec{\Gamma }}_{t|t-1})\). Thus, by Eq. (7) we have
which corresponds to the kernel of the pdf of \( {\varvec{Y}}_t | {\varvec{D}}_{t-1} \sim SUN_{r,t}({\varvec{f}}_t , {\varvec{Q}}_t , \tilde{{\varvec{\Lambda }}}_t, \tilde{{\varvec{\tau }}}_t,\tilde{{\varvec{\Gamma }}}_t )\). To prove the Step 4 in the sequential inference scheme, we proceed as follows
where the relations \({\varvec{\Lambda }}_t ({\varvec{y}}_t -{\varvec{F}}'_t {\varvec{\theta }}_t + \sqrt{\frac{2}{\pi }}{\varvec{\Delta }}_t ) = - {\varvec{\Lambda }}_t {\varvec{F}}'_t ( {\varvec{\theta }}_t -{\varvec{m}}_{t|t}) + {\varvec{\Lambda }}_t {\varvec{V}}_t {\varvec{Q}}^{-1}_t ({\varvec{y}}_t - {\varvec{f}}_t)\) and \({\varvec{m}}_{t|t}-{\varvec{m}}_{t|t-1}= {\varvec{C}}_{t|t-1}{\varvec{F}}_t {\varvec{Q}}_t^{-1}({\varvec{y}}_t - {\varvec{f}}_t)\) are used. This last expression is the density kernel of \({\varvec{\theta }}_{t}\mid {\varvec{D}}_{t} \sim SUN_{p,t}({\varvec{m}}_{t|t} , {\varvec{C}}_{t|t} , {\varvec{\Lambda }}_{t|t}, {\varvec{\tau }}_{t|t},{\varvec{\Gamma }}_{t|t} )\), concluding thus the proof. \(\square \)
Theorem 2
Proof
Since \({\varvec{\theta }}_{t-k} \mid {\varvec{D}}_{t-k} \sim SUN_{p,t-k} ({\varvec{m}}_{t-k|t-k},{\varvec{C}}_{t-k|t-k},{\varvec{\Lambda }}_{t-k|t-k},{\varvec{\tau }}_{t-k|t-k},{\varvec{\Gamma }}_{t-k|t-k})\) and \({\varvec{\theta }}_{t-k+1} \mid {\varvec{\theta }}_{t-k}, {\varvec{D}}_{t-k} \sim N_p ({\varvec{G}}_{t-k+1} {\varvec{\theta }}_{t-k} ,{\varvec{W}}_{t-k+1})\), we have by Eq. (32)
where \({\varvec{h}}_{t}(k) - {\varvec{m}}_{t-k\mid t-k}= {\varvec{B}}_{t-k} ({\varvec{\theta }}_{t-k+1} - {\varvec{m}}_{t-k+1\mid t-k})\). Thus, this last result corresponds to the density kernel of the distribution given in Eq. (20). \(\square \)
Theorem 3
Proof
We need to solve the following integral
For this, note first that \(p({\varvec{\theta }}_{t-k} \mid {\varvec{\theta }}_{t-k+1} , {\varvec{D}}_t) = p({\varvec{\theta }}_{t-k} \mid {\varvec{\theta }}_{t-k+1} , {\varvec{D}}_{t-k})\), where by Theorem 2 \({\varvec{\theta }}_{t-k} \mid {\varvec{\theta }}_{t-k+1}, {\varvec{D}}_{t-k} \sim SUN_{p,t-k} ({\varvec{h}}_{t}(k),{\varvec{H}}_{t}(k),{\varvec{\Lambda }}_t(k),{\varvec{\tau }}_t(k), {\varvec{\Gamma }}_t(k))\) and by induction \({\varvec{\theta }}_{t-k+1}\mid {\varvec{D}}_t \sim \)\(SUN_{p,t} ({\varvec{m}}_{t-k+1|t},{\varvec{C}}_{t-k+1|t},{\varvec{\Lambda }}_{t-k+1|t},{\varvec{\tau }}_{t-k+1|t}, {\varvec{\Gamma }}_{t-k+1|t})\).
Hence, from the definitions of \({\varvec{h}}_{t}(k),\)\({\varvec{H}}_{t}(k)\), \(\Lambda _t(k)\), \({\varvec{\tau }}_t(k)\) and \({\varvec{\Gamma }}_t(k)\), the relations \({\varvec{\Lambda }}_{t-k+1|t-k}={\varvec{\Lambda }}_{t-k|t-k}{\varvec{B}}_{t-k}\), \({\varvec{\tau }}_{t-k+1|t-k}={\varvec{\tau }}_{t-k|t-k}\), \({\varvec{\Gamma }}_{t-k+1|t-k}={\varvec{\Gamma }}_t(k)+{\varvec{\Lambda }}_t(k){\varvec{H}}_t(k){\varvec{\Lambda }}_t(k)'\), by applying the Eqs. (32) and (7), we have
Thus, we have the kernel of the \(SUN_{p,t}({\varvec{m}}_{t-k|t} , {\varvec{C}}_{t-k|t}, {\varvec{\Lambda }}_{t-k|t}, {\varvec{\tau }}_{t-k|t},{\varvec{\Gamma }}_{t-k|t} )\) pdf, which concludes the proof. \(\square \)
Rights and permissions
About this article
Cite this article
Arellano-Valle, R.B., Contreras-Reyes, J.E., Quintero, F.O.L. et al. A skew-normal dynamic linear model and Bayesian forecasting. Comput Stat 34, 1055–1085 (2019). https://doi.org/10.1007/s00180-018-0848-1
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00180-018-0848-1