Abstract
There exist many studies on regression analysis for spatial binary data, espsecially in ecological, environmental and socio-economic setups, where spatial responses from neighboring locations within a given threshold distance are correlated. However, in some of these studies, it could be more natural to consider a spatial regression analysis for categorical response data with more than two categories, as an improvement over the spatial binary analysis. But, this type of regression analysis for spatial categorical/multinomial data is not adequately addressed in the literature. One of the main reasons is the difficulty of modeling the spatial familial correlations for categorical data, where a spatial family is generated within the threshold distance for each of the two selected neighboring locations. Also, some of the locations from two families may be pair-wise correlated. Unlike the existing studies, in this paper we propose a familial random effects based multinomial logits mixed (MLM) effects model which accommodates both within and between familial correlations for spatial multinomial data. In this context, the proposed spatial multinomial correlations are contrasted with existing longitudinal multinomial correlations so that the longitudinal correlation models are avoided for spatial multinomial data. Both regression effects and the random effects influence parameters are estimated using the generalized quasi-likelihood approach, whereas the random effects variance and correlation parameters are estimated by the well known method of moments. The large sample properties such as consistency of the proposed estimators are studied analytically. The asymptotic normality of the regression estimators is also studied for the convenience of constructing the confidence intervals when needed. The devirations and proofs are given in details, as opposed to conducting a limited simulation study, to justify the validity and convergence properties of the proposed estimators. The estimating equations those produced consistent estimates are clearly formulated for the computational benefit to the practitioners.
Similar content being viewed by others
References
Ainsworth, L. M., Dean, C. B. and Joy, R. (2016). Zero-inflated spatial models: Applications and interpretation, 218. Springer, Switzerland, Sutradhar, B. C. (ed.), p. 75–96.
Amemiya, T. (1985). Advanced econometrics. Harvard University Press, Cambridge.
Asif, A. and Moura, J. M. F. (2005). Block matrices with L-block-banded inverse: Inversion algorithms. IEEE Transactions on Signal Processing 53, 630–642.
Bartlett, M. S. (1966). An introduction to stochastic processes. University Press, Cambridge.
Bartlett, M. S. and Besag, J. (1969). Correlation properties of some nearest-neighbour model. Bulletin of International Statistical Institute 43, 191–193.
Basu, S. and Reinsel, G. C. (1994). Regression models with spatially correlated errors. Journal of the American Statistical Association 89, 88–99.
Bevilacqua, M., Crudu, F and Porcu, E (2015). Combining Euclidean and composite likelihood for binary spatial data estimation. Stochastic Environmental Research and Risk Assessment 29, 335–346.
Besag, J. (1972a). Nearest-neighbour systems and the auto-logistic model for binary data. Journal of the Royal Statistical Society Series B 34, 75–83.
Besag, J. (1972b). On the correlation structure of some two-dimensional stationary processes. Biometrika 59, 43–48.
Besag, J. (1974). Spatial interaction and the statistical analysis of lattice systems. Journal of the Royal Statistical Society. Series B 36, 192–236.
Boots, B. (2003). Developing local measures of spatial association for categorical data. Journal of Geographical System 5, 139–160.
Brook, D. (1964). On the distinction between the conditional probability and joint probability approaches in the specification of nearest-neighbour Systems. Biometrika 51, 481–483.
Cao, G., Kyriakidid, P. C. and Goodchild, M. F. (2011). A multinomial logistic mixed model for the prediction of categorical spatial data. International Journal of geographical Information Science 25, 2071–2086.
Gayawan, E., Arogundade, E. D. and Adebato, S. B. (2014). A Bayesisn multinomial modeling of spatial pattern of co-morbidity pf malaria and non-malaria febrile illness among young children in Nigeria. Transactions of the Royal Society of Tropical Medicine and Hygiene 108, 415–424.
Heagerty, P. J. and Lele, S. (1998). A Composite likelihood approach to binary spatial data. Journal of American Statistical Association 93, 1099–1111.
Jain, J., Li, H., Cauley, S., Koh, C. -K. and Balakrishnan, V. (2007). Numerically stable algorithms for inversion of block tridiagonal and banded matrices. Electrical and Computer Engineering Technical Reports. Purdue University, Purdue e-Pubs.
Kazembe, L. N. and Namangale, J. J. (2007). A Bayesian multinomial model to analyse spatial patterns of childhood co-morbidity in Malawi. European Journal of Epidemiology 22, 545–556.
Lin, P.-S. and Clayton, M. K. (2005). Analysis of binary spatial data by quasi-likelihood estimating equations. The Annals of Statistics 33, 542–555.
Mariathas, H. H. and Sutradhar, B. C. (2016). Variable family size based spatial moving correlations model. Sankhya B 78, 1–38.
McCulloch, C. (1997). Maximum likelihood algorithms for generalized linear mixed models. Journal of American Statistical Association 92, 162–170.
Muniz, D. G., Santos, E. S., Guimaraes, Jr., P. R., Nakagawa, S. and Machado, G. (2017). A multinomial network method for the analysis of mate choice and assortative mating in spatially structured populations. Methods in Ecology and Evolution 8, 1321–1331.
Rathbun, S. and Cressie, N. A. (1994). A space-time survival point process for a longleaf pine forest in Southern Georgia. Journal of the American Statistical Association 89, 1164–1174.
Sutradhar, B. C. (2003). An overview on regression models for discrete longitudinal responses. Statistical Science 18, 377–393.
Sutradhar, B. C. (2004). On exact quasi-likelihood inference in generalized linear mixed models. Sankhya B 66, 261–289.
Sutradhar, B. C. (2014). Longitudinal categorical data analysis. Springer, New York.
Sutradhar, B. C. (2021). Block-band behavior of spatial correlations: An analytical asymptotic study in a spatial exponential family data setup. Journal of Multivariate Analysis, 186.
Sutradhar, B. C. and Oyet, A. (2020). An unbalanced multidimensional latent effects-based logistic mixed model and GQL estimation for spatial binary data. Journal of Statistical Computation and Simulation 90, 3338–3359.
Sutradhar, B. C., Rao, R. P. and Pandit, V. (2008). Generalized method of moments versus generalized quasi-likelihood inferences in binary panel data models. Sankhya B 70, 34–62.
Ten Have, T. R. and Morabia, A. (1999). Mixed effects models with bivariate and univariate association parameters for longitudinal bivariate binary response data. Biometrics 55, 85–93.
Waller, L. A. (2005). Spatial models for categorical data. In: Encyclopedia of Biostatistics.
Whittle, P. (1963). Stochastic processes in several dimensions. Bulletin of international Statistical Institute 40, 974–994.
Wijekoon, P., OYET, A. and SUTRADHAR, B. C. (2019). Pair-wise family-based correlation model for spatial count data. Sankhya B 81, 133–184.
Yoo, E. -H., Hoagland, B. W., Cao, G. and Fagin, T. (2013). Spatial distribution of trees and landscapes of the past: a mixed spatially correlated multinomial logit model approach for the analysis of the public land survey data. Geographical Analysis 45, 419–440.
Acknowledgements
The authors would like to thank two reviewers for their valuable comments and suggestions that led to the improvement of the paper. Thanks are also due to the Editor-in-Chief, the Editor, and the Associate Editor for their helpful suggestions.
Funding
This research was partially supported by a grant (RGPIN-04503-2015) from the Natural Sciences and the Engineering Research Council of Canada, awarded to the first author.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of Interests
There is no conflict of interest.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix: Unconditional Mean (2.8) and Variance (2.13) Computation Using Binomial Approximation to Normal Integration
Mean Computation:
By Eq. 2.10 (see also Assumption 1 and Lemma 2.1) γi ≡ γi(ni) has the multivariate normal distribution as
Consequently, by using a non-singular transformation \({g}_{i}={\Phi }^{-\frac {1}{2}}_{i}(\sigma ^{2}_{\gamma },\phi ){\gamma }_{i}\) so that \({g}_{i} \sim N_{n_{i}}(0,I),\) we first replace γi in Eq. 2.8 with \({\gamma }_{i}={\Phi }^{\frac {1}{2}}_{i}{g}_{i},\) and re-express the unconditional multinomial mean as
where \({{{a}}}^{*'}_{i}={a}'_{i}{\Phi }^{\frac {1}{2}}_{i}=({a^{*}}_{i1},\ldots , {a^{*}}_{iu_{i}},\ldots ,{a^{*}}_{in_{i}})',\) and \(g^{*}_{N}(g_{iu_{i}}) \equiv N(0,1)\) for all ui = 1,…,ni.
Next we use a Binomial approximation (Ten Have and Morabia (1999, Eqn. (7)), Sutradhar (2014, Eqns. (5.48)-(5.50))) to perform the standard normal integration in Eq. a.2. More specifically, suppose that \(v_{iu_{i}}\) is binomial variable which ranges from 0 to V, V being a positive integer. Now replace the standard normal variable \(g_{iu_{i}}\) in Eq. a.2 with a standardized binomial variable
with large V such as V = 10, and re-express the functions involved under the integration as
and
It then follows that the unconditional spatial multinomial mean in Eq. a.2 may be computed using the above binomial approximation (a.4)–(a.6), as
Variance Computation by Eqs. 2.13 and 2.18:
Recall from Section 2.4.1 that
where \({{C}}^{\dag }_{ij}(\phi ) ): (n_{i}+n^{*}_{j}) \times (n_{i}+n^{*}_{j})\) is the correlation matrix as in Eq. 2.26, where \(n^{*}_{j}=n_{j}-n_{ij}\). This is equivalent to write
It then follows that
Next write
Using Eq. a.12, write
Thus by Eqs. a.11 and a.13, we write
Use them in Eq. 2.13 and follow the notations from Eq. a.2 to compute the joint probability as follows:
where
and \(g^{*}_{N}(g^{\dagger }_{ij,k}) \equiv N(0,1)\) for all \(k=1,\ldots ,n_{i}+n^{*}_{j}.\)
Next, these four multiple integrations in Eq, a.15 may be computed by using the Binomial approximation in the same as done through Eqs. a.3–a.7, with a difference that instead of Eq. a.3, we now use
for \(k=1,\ldots ,n_{i}+n^{*}_{j}.\)
Appendix B. Proof for Asymptotic Normality of the GQL Regression Estimator
Asymptotic Normality:
Recall from Eq. 4.2 that \(\hat {\beta }_{GQL}\) is obtained by solving \(\frac {\partial Q(\beta |\tau _{\gamma },\sigma _{\gamma },\phi )}{\partial \beta }=0.\) More specifically, by using a first order Taylor series approximation, it can be shown that \(\hat {{\beta }}_{GQL}\) from Eq. 4.2 satisfies
where \({y}=({y}'_{s_{1}},\ldots ,{y}'_{s_{i}},\ldots ,{y}'_{s_{K}})'\) is the complete spatial multinomial vector response of dimension K(C − 1) × 1, as defined in Eq. 3.1. Use
and for large K, using Eq. 4.7, re-express (b.1) as
where \({M}({\beta },\tau _{\gamma },\sigma _{\gamma },\phi ) =\frac {\partial {\pi }'({\beta }|\cdot )}{\partial {\beta }}{\Sigma }^{-\frac {1}{2}}({\beta },\tau _{\gamma },\sigma _{\gamma },\phi ).\) Notice that
justifying that z1,…,zi,…,zK, are uncorrelated (C − 1)-dimensional vectors. Further, re-express (b.3) as
Notice that
because cov[zi,zj′] = 0 by Eq. b.4. Thus, \(\bar {{h}}({\beta }|\cdot )\) in Eq. b.5 has its mean and covariance as
We now assume that the following regularity condition holds.
ASSUMPTION 3
Suppose that hi(β|⋅) with moment properties as in Eq. b.7 satisfy the Lindeberg regularity condition that
for all 𝜖 > 0, g(⋅) being the p-dimensional probability distribution of hi(⋅).
One may then exploit the Lindeberg-Feller central limit theorem (Amemiya 1985, Theorem 3.3.6) and obtain the limiting distribution of
as
Now because by Eq. b.9 it follows from Eq. b.5 that
Hence by Eq. b.10,
That is
justifying the limiting distributional result in Eq. 4.3 under the Theorem 4.1.
Appendix C: Formulas for the Derivatives in Estimating Equations
Computation of the Derivatives \(\frac {\partial {\pi _{i,c}}({\beta }|\cdot )}{\partial {\beta }}\) in Eq. 4.6:
Recall from Eq. 2.2 that β = (β1′,…,βc′,…,βC− 1′)′ is a (C − 1)(p + 1)-dimensional vector of regression parameters,βc being the (p + 1)-dimensional vector corresponding to the c-th category. Because this β parameter vector is involved in πi,c (marginal probability for i-th spatial location (si) response to be in c-th category) in the way as shown by Eq. 2.8, we can compute the derivatives in Eq. 4.6 as follows.
where δc = (0 ⊗ 1c− 1′,1,0 ⊗ 1C− 1−c′)′, \({\pi }^{*}_{i,c}(\cdot ) \equiv {\pi }^{*}_{i,c}(\beta ,\tau _{\gamma },\gamma _{i})\) is given by Eq. 2.2, and \({\pi }^{*}_{i}(\cdot )=({\pi }^{*}_{i,1}(\cdot ),\ldots ,{\pi }^{*}_{i,c}(\cdot ), \ldots ,{\pi }^{*}_{i,C-1}(\cdot ))'.\)
Next, by using the non-singular transformation from γi to gi as in Eq. a.2 and the standardized form of the binomial variable \(v_{iu_{i}}\) as in Eq. a.3, this integration in Eq. c.2 may be approximated as
Computation of the Derivatives \(\frac {\partial {\pi _{i,c}}({\tau _{\gamma }}|\cdot )}{\partial {\tau _{\gamma }}}\) in Eq. 4.16:
The formula for the probability function πi,c(τγ|⋅) for c = 1,…,C − 1 is defined in Eq. 2.8, where τγ,1 = 1. Thus, the estimating equation given by Eq. 4.15 is written for τγ = (τγ,2,…,τγ,c,…,τγ,C− 1)′. Note that as τγ,1 = 1, the pattern for the derivatives \(\frac {\partial {\pi _{i,c}}({\tau _{\gamma }}|\cdot )}{\partial {\tau _{\gamma }}}\) for c = 2,…,C − 1, will be similar and they will be different than that of \(\frac {\partial {\pi _{i,1}}({\tau _{\gamma }}|\cdot )}{\partial {\tau _{\gamma }}}.\) Following Eqs. c.1–c.3, these derivatives may be obtained as
and
where \(\delta ^{*}_{c}=(0\otimes 1'_{c-2},1,0 \otimes 1'_{C-2-(c-1)})',\)\({\pi }^{*}_{i,c}(\cdot ) \equiv {\pi }^{*}_{i,c}(\beta ,\tau _{\gamma },\gamma _{i})\) is given by Eq. 2.2, and \({\pi }^{**}_{i}(\cdot )=({\pi }^{*}_{i,2}(\cdot ),\ldots ,{\pi }^{*}_{i,c}(\cdot ), \ldots ,{\pi }^{*}_{i,C-1}(\cdot ))'.\)
Next by similar algebras as in Eqs. c.2 and c.3, the above two integrals in Eqs. c.4 and c.5 may be computed as follows by using the Binomial approximation:
where \(g_{i}=(g_{i1},\ldots ,g_{iu_{i}},\ldots ,g_{in_{i}})'\) with \(g_{iu_{i}}(v_{iu_{i}})\) as given in Eq. a.3.
Computation of the Derivatives \(\frac {\partial \pi _{i,c}(\cdot )} {\partial \sigma ^{2}_{\gamma }}\) and \(\frac {\partial \lambda _{ij,cr}(\cdot )} {\partial \sigma ^{2}_{\gamma }}\) in Eq. 4.27:
Similar to the derivatives in Eq. c.1 with respect to β, we write the derivative of the same function \(\pi _{i,c}({\sigma ^{2}_{\gamma }}|\cdot )\) with respect to \(\sigma ^{2}_{\gamma },\) as
using \({\Phi }_{i}(\sigma ^{2}_{\gamma },\phi )=\sigma ^{2}_{\gamma } {C}_{ii}(\phi )\) from Eqs. a.1 into a.2.
Because σγ is unplugged in Eq. c.9, following the notation from Eq. a.6, the derivative and integration gives the formula as
where gi in terms of vi is defined in Eq. a.3.
Next to compute \(\frac {\partial \lambda _{ij,cr}(\cdot )} {\partial \sigma ^{2}_{\gamma }}\) for c,r = 1,…,C − 1, we follow the formula for λij,cr(⋅) from (a.15), and compute the desired derivative as
Further because \(\tilde {\Phi }_{ij}(\sigma ^{2}_{\gamma },\phi )=\sigma ^{2}_{\gamma } {{C}}^{\dag }_{ij}(\phi )\) as in Eq. a.9, we can unplug \(\sigma ^{2}_{\gamma }\) and re-express (a.14) as
It then follows that
where \(g^{\dagger }_{ij,k}\) in terms of vij,k is defined in Eq. a.16.
Computation of the Derivative \(\frac {\partial \lambda _{i(i-1),cr}(\cdot )} {\partial \phi }\) in Eq. 4.36:
Notice that the structure of the scale matrices \(\tilde {A}(\phi )\) and \({\tilde {B}}^{*}(\phi )\) as a function of ϕ is known from the relationships (a.9), and (a.11)–(a.13). More specifically, one writes from Eqs. a.9 and a.11 that
and from Eq. a.13
Hence by similar calculations as in Eq. c.13, we compute the desired derive as
Rights and permissions
About this article
Cite this article
Sutradhar, B.C., Rao, R.P. Asymptotic Inferences in a Multinomial Logit Mixed Model for Spatial Categorical Data. Sankhya A 85, 885–930 (2023). https://doi.org/10.1007/s13171-022-00282-7
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13171-022-00282-7
Keywords
- Categorical/multinomial responses in a spatial setup
- Moving correlations
- Multinomial mixed logits
- Normality and consistency of the estimators
- Spatial correlations
- Spatial statistics