Abstract
A general class of mixed Poisson regression models is introduced. This class is based on a mixing between the Poisson distribution and a distribution belonging to the exponential family. With this, we unified some overdispersed models which have been studied separately, such as negative binomial and Poisson inverse gaussian models. We consider a regression structure for both the mean and dispersion parameters of the mixed Poisson models, thus extending, and in some cases correcting, some previous models considered in the literature. An expectation–maximization (EM) algorithm is proposed for estimation of the parameters and some diagnostic measures, based on the EM algorithm, are considered. We also obtain an explicit expression for the observed information matrix. An empirical illustration is presented in order to show the performance of our class of mixed Poisson models. This paper contains a Supplementary Material.
Similar content being viewed by others
References
Atkinson, A.C.: Plots, Transformations and Regression. Oxford University Press, Oxford (1985)
Breslow, N.E.: Extra-Poisson variation in log-linear models. J. R. Stat. Soc. Ser. C 33, 38–44 (1984)
Brillinger, D.R.: The natural variability of vital rates and associated statistics (with discussion). Biometrics 42, 693–734 (1986)
Cameron, A.C., Trivedi, P.K.: Regression Analysis of Count Data. Cambridge University Press, New York (1998)
Cook, R.D.: Detection of influential observations in linear regression. Technometrics 19, 15–18 (1977)
Cook, R.D.: Assessment of local influence. J. R. Stat. Soc. Se. B 48, 133–169 (1986)
Dean, C.B., Nielsen, J.D.: Generalized linear mixed models: a review and some extensions. Lifetime Data Anal. 13, 497–512 (2007)
Dean, C.B., Lawless, J., Willmot, G.E.: A mixed Poisson inverse Gaussian regression model. Can. J. Stat. 17, 171–182 (1989)
Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the EM algorithm (with discussion). J. R. Stat. Soc. Se. B 39, 1–38 (1977)
Dunn, P.K., Smyth, G.K.: dglm: Double generalized linear models. R package version 1.6.2. http://CRAN.R-project.org/package=dglm (2012)
Engel, J.: Models for response data showing extra-Poisson variation. Stat. Neerlandica 38, 159–167 (1984)
Ghitany, M.E., Karlis, D., Al-Mutairi, D.K., Al-Awadhi, F.A.: An EM algorithm for multivariate mixed poisson regression models and its application. Appl. Math. Sci. 6, 6843–6856 (2012)
Hall, D.: Zero-inflated Poisson and binomial regression with random effects: a case study. Biometrics 56, 1030–1039 (2000)
Hilbe, J.M.: Negative Binomial Regression. Cambridge University Press, New York (2008)
Hinde, J., Demétrio, C.G.B.: Overdispersion: models and estimation. Comput. Stat. Data Anal. 27, 151–170 (1998)
Holla, M.S.: On a Poisson-inverse Gaussian distribution. Metrika 11, 115–121 (1966)
Karlis, D.: A general EM approach for maximum likelihood estimation in mixed Poisson regression models. Stat. Model. 1, 305–318 (2001)
Karlis, D., Xekalaki, E.: Mixed Poisson distributions. Int. Stat. Rev. 73, 35–58 (2005)
Kassahun, W., Neyens, T., Faes, C., Molenberghs, G., Verbeke, G.: A zero-inflated overdispersed hierarchical Poisson model. Stat. Model. 14, 439–456 (2014)
Lambert, D.: Zero-inflated poisson regression, with an application to defects in manufacturing. Technometrics 34, 1–14 (1992)
Lawless, J.F.: Negative binomial and mixed Poisson regression. Can. J. Stat. 15, 209–225 (1987)
Lesaffre, E., Verbeke, G.: Local influence in linear mixed model. Biometrics 54, 570–583 (1998)
Lim, H.K., Li, W.K., Yu, P.L.H.: Zero-inflated Poisson regression mixture model. Comput. Stat. Data Anal. 71, 151–158 (2014)
Louis, T.A.: Finding the observed information matrix when using the EM algorithm. J. R. Stat. Soc. 44, 226–233 (1982)
Ord, J.K., Whitmore, G.A.: The Poisson-inverse gaussian distribution as a model for species abundance. Commun. Stat. Theory Methods 15, 853–871 (1986)
Poon, W., Poon, Y.: Conformal normal curvature and assessment of local influence. J. R. Stat. Soc. Ser. B 61, 51–61 (1999)
R Core Team.: R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. url: http://www.R-project.org (2014)
Ridout, M., Hinde, J., Demétrio, C.G.B.: A Score test for testing a zero-inflated Poisson regression model against zero-inflated negative binomial alternatives. Biometrics 57, 219–223 (2001)
Rigby, R.A., Stasinopoulos, D.M.: Generalized additive models for location, scale and shape (with discussion). Appl. Stat. 54, 507–554 (2005)
Sankaran, M.: Mixtures by the inverse Gaussian distribution. Sankhya B 30, 455–458 (1968)
Sellers, K.F., Shmueli, G.: A flexible regression model for count data. Ann. Appl. Stat. 4, 943–961 (2010)
Sichel, H.S.: On a family of discrete distributions particularly suited to represent long-tailed frequency data. In:Proceedings of the Third Symposium on Mathematical Statistics, Pretoria, CSIR, pp. 51–97 (1971)
Willmot, G.E.: The Poisson-inverse Gaussian distribution as an alternative to the negative binomial. Scand. Actuar. J. 87, 113–127 (1989)
Wu, C.F.J.: On the convergence properties of the EM algorithm. Ann. Stat. 11, 95–103 (1983)
Xie, F.C., Wei, B.C.: Influence analysis for Poisson inverse Gaussian regression models based on the EM algorithm. Metrika 67, 49–62 (2008)
Xie, F.C., Lin, J.G., Wei, B.C.: Bayesian zero-inflated generalized Poisson regression model: estimation and case influence diagnostics. J. Appl. Stat. 41, 1383–1392 (2014)
Yau, K.K.W., Wang, K., Lee, A.H.: Zero-inflated negative binomial mixed regression modeling of over-dispersed count data with extra zeros. Biom. J. 45, 437–452 (2003)
Zhu, H.T., Lee, S.Y.: Local influence for incomplete-data models. J. R. Stat. Soc. Ser. B 63, 111–126 (2001)
Zhu, H.T., Lee, S.Y., Wei, B.C., Zhu, J.: Case-deletion measures for models with incomplete data. Biometrika 88, 727–737 (2001)
Acknowledgments
We would like to thank two referees for their careful reading and suggestions which improved the paper. The authors thank the financial support from FAPEMIG (W. Barreto-Souza) and CNPq-Brazil (A.B. Simas).
Author information
Authors and Affiliations
Corresponding author
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendix
Appendix
We begin this section by providing correction of some results presented in Xie and Wei (2008).
In Xie and Wei (2008, p. 52), the correct expression for \(E(\nu _i^{-1}|Y_o,\theta ^{(r)})\) in the case \(y_i=0\) is \(\lambda ^{(r)}\big [ 1+\sqrt{(\lambda ^{(r)})^{-1}(2\mu _i+(\lambda ^{(r)})^{-1})}\big ]\) instead of \(\lambda ^{(r)}\big [ 1+\sqrt{(\lambda ^{(r)})^{-1}}(2\mu _i+(\lambda ^{(r)})^{-1})\big ]\) .
In Xie and Wei (2008, p. 58), where one finds “\(-T\varvec{1}_n\varvec{1}_m^TS\) \(- \beta ^TS\varvec{1}_mGX\)”, the correct expression is “\(-T\varvec{1}_n\varvec{1}_m^TS\) \(-GX\beta \varvec{1}_mS\)”.
Finally, the perturbation of responses (case 3) found in Xie and Wei (2008, p. 58) does not make sense since the random variable \(Y_i\) is discrete and, therefore, does not admit infinitesimal perturbations (all perturbations of \(Y_i\) must be integer-valued). Therefore, this perturbation scheme is wrong and should not be considered. What does make sense is to do an infinitesimal perturbation on the latent variable Z, since it is continuous. We provide such a perturbation scheme in the present article (see the perturbation of the hidden variable scheme).
In what follows we present the proofs of two propositions given in the main text.
Proof of Proposition 1
We have that
To obtain the second conditional expectation, we first get the conditional moment generating function of g(Z) given \(Y=y\), that is
with \(\mu _t^*\) as defined in the proposition. Hence, computing the derivative of \(E\left( \exp \{t\,g(Z)\}|Y=y\right) \) with respect to t at \(t=0\), after some simplifications we get the desired result. \(\square \)
Proof of Proposition 2
A simple computation shows that
and
Furthermore,
and
Thus
\(\square \)
The elements of the observed information matrix (6) can be obtained by using the following quantities:
for \(j,l=1,\ldots ,p\),
for \(j,l=1,\ldots ,q\),
for \(j=1,\ldots ,p\) and \(l=1,\ldots ,q\),
for \(j,l=1,\ldots ,p\),
for \(j,1,\ldots ,p\) and \(l=1,\ldots ,q\),
for \(j,l=1,\ldots ,q\), where \(\lambda _i\) and \(\kappa _i\) are defined as before and here we have defined \(\gamma _i=E(Z_i^2|\mathbf{Y})\), \(\rho _i=E(Z_i g(Z_i)|\mathbf{Y})\) and \(\nu _i=E(g(Z_i)^2|\mathbf{Y})\), with \(i=1,\ldots ,n\).
The conditional expectations that appears above are given explicitly in the following proposition.
Proposition 3
Let \(Y\sim \hbox {MP}(\mu ,\phi )\) with latent random effect Z belonging the exponential family as defined previously. Then, we have that
and
Corollary 1
For the PIG case, we have that
In the NB case, we have that
where \(\Psi '(x)=d\Psi (x)/dx\).
Rights and permissions
About this article
Cite this article
Barreto-Souza, W., Simas, A.B. General mixed Poisson regression models with varying dispersion. Stat Comput 26, 1263–1280 (2016). https://doi.org/10.1007/s11222-015-9601-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11222-015-9601-6