Abstract
In a finite/survey population setup, where categorical/multinomial responses are collected from individuals belonging to a cluster, in a recent study Skinner (International Statistical Review, 87, S64-S78 2019) has modeled the means of the clustered categorical responses as a function of regression parameters, and suggested a ‘working’ correlations based GEE (generalized estimating equations) approach for the estimation of the regression parameters. However, this mean model involving only regression parameters is not justified for clustered multinomial responses because of the fact that these responses share a common cluster effect which compels the clustered correlation parameter to enter into the mean function on top of the regression parameters. Consequently, the so-called GEE approach, which requires the means to be free of correlations, is not applicable for regression analysis in the clustered multinomial setup. As a remedy, in this paper we consider a multinomial mixed model which accommodates the clustered correlation parameter in the mean functions. For inferences in the present finite population setup, as the GQL (generalized quasi-likelihood) approach is known to produce consistent and more efficient estimate than the MM (method of moments) approach in an infinite population setup, we estimate the regression parameters of primary interest by using the first order response based survey weighted GQL (WGQL) approach. For the estimation of the random effects variance (also known as clustered correlation) parameter, as it is of secondary interest, we use the second order response based survey weighted MM (WMM) approach, which is simpler than the corresponding WGQL estimation approach. The estimation steps are presented clearly for the benefit to the practitioners. Also because, in practice, survey practitioners such as statistical agencies frequently deal with a large health or socio-economic data set at national or state levels, for example, we make sure for their benefit that our proposed WGQL and WMM estimators are consistent. Thus, the asymptotic properties such as asymptotic unbiasedness and consistency for both regression and clustered correlation parameters are studied in details. The asymptotic normality property, for the benefit of constructing confidence interval for the main regression parameters, is also studied.
Similar content being viewed by others
References
Amemiya, T. (1985). Advanced econometrics. Harvard University Press, Cambridge.
Breslow, N. E. and Clayton, D. G. (1993). Approximate inference in generalized linear mixed models. Journal of American Statistical Association 88, 9–25.
Cochran, W. G. (1977). Sampling techniques. Wiley, New York.
Godambe, V. P. and Thompson, M. E. (1986). Parameters of super-population and survey population: their relationships and estimation. International Statistical Review 54, 127–138.
Lee, Y. and Nelder, J. (1996). Hierarchical generalized linear models. Journal of Royal Statistical Society. B 58, 619–678.
Lee, S. E., Lee, P. R. and Shin, K. (2016). A composite estimator for stratified two stage cluster sampling. Communications for Statistical Applications and Methods 23, 47–55.
Laing, K. -Y. and Zeger, S. L. (1986). Longitudinal data analysis using generalized linear models. Biometrika 73, 13–22.
Mccullagh, P. (1983). Quasilikelihood functions. Annals of Statistics11, 59–67.
McCullagh, P. and Nelder, J. A (1989). Generalized linear models. Chapman and Hall, London.
McDonald, D. R. (2005). The local limit theorem: a historical perspective. J. Iran. Stat. Soc. 4, 73–86.
Molina, E. A. and Skinner, C. J. (1992). Pseudo-likelihood and quasi-likelihood estimation for complex sampling schemes. Computational Statistics and Data Analysis 13, 395–405.
Molina, E. A., Smith, T. M. F. and Sugden, R. A. (2001). Modelling overdispersion for complex survey data. Int. Stat. Rev. 69, 373–384.
Pfeffermann, D. (1993). The role of sampling weights when modelling survey data. Int. Statist. Rev. 61, 317–337.
Rao, R. P., Sutradhar, B. C. and Pandit, V. N. (2012). GMM And GQL inferences in linear dynamic panel data models. Brazilian Journal of Probability and Statistics 26, 167–177.
Roberts, G., Rao, J. N. K. and Kumar, S. (1987). Logistic regression analysis of sample survey data. Biometrika 74, 1–12.
Roberts, G., REN, Q. and RAO, J. N. K. (2009). Using marginal mean models for data from longitudinal surveys with a complex design: some advances in methods. Wiley, Chichester, Lynn, P. (ed.), p. 351–366.
Skinner, C. J. and Vieira, M. D. T. (2007). Variance estimation in the analysis of clustered longitudinal survey data. Survey Methodolgy 33, 3–12.
Skinner, C. (2019). Analysis of categorical data for complex surveys. International Statistical Review 87, S64–S78.
Sutradhar, B. C. and Kovacevic, M. (2000). Analyzing ordinal longitudinal survey data: Generalized estimating equations approach. Biomerika 87, 837–848.
Sutradhar, B. C. (2003). An overview on regression models for discrete longitudinal responses. Stat. Sci. 18, 377–393.
Sutradhar, B. C. (2004). On exact quasi-likelihood inference in generalized linear mixed models. Sankhya B: The Indian Journal of Statistics 66, 261–289.
Sutradhar, B. C. (2014). Longitudinal categorical data analysis. Springer, New York.
Sutradhar, B. C. (2018a). A parameter dimension-split based asymptotic regression estimation theory for a multinomial panel data model. Sankhya A: The Indian Journal of Statistics 80, 301–329.
Sutradhar, B. C. and Zheng, N. (2018b). Inferences in binary dynamic fixed models in a semi-parametric setup. Sankhya B 80, 263–291.
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.
Valliant, R. (1987). Generalized variance functions in stratified two-stage sampling. Journal of American Statistical Association 82, 499–508.
Wedderburn, R. (1974). Quasilikelihood functions, generalized linear models and the Gauss-Newton method. Biometrika 61, 439–447.
Weisstein, E. W. (2002). Newton’s method. http://mathworld.wolfram.com.
Zheng, N. and Sutradhar, B. C. (2018). Generalized quasi-likelihood inference in a semi-parametric binary dynamic mixed logit model. Australian and New Zealand Journal of Statistics 60, 343–373.
Acknowledgments
This research was supported partially by an NSERC grant. The author thanks two referees and the Associated Editor for their comments and suggestions that lead to the improvement of the paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A: Proof for the Design Unbiased Property of the Estimating Functions (4.1) of β, and Eq. 4.4 of \(\sigma ^{2}_{\gamma }\)
Appendix A1: Proof for the Estimating Function (4.1) of β
Here we check the design (D) unbiasedness of \(\hat {\tau }_{\beta |\sigma ^{2}_{\gamma }}(y;\cdot )\) (4.1) for \(\tau _{\beta |\sigma ^{2}_{\gamma }}(y;\cdot )\) in Eq. 3.9, by taking the design expectation (ED) as
where EC denotes the expectation over the clusters (C) and E|C is the expectation over the individuals within a given cluster. Hence,
which is the hypothetical estimating function (HEF) in the left hand side of the HEE in Eq. 3.9.
Appendix A2: Proof for the Estimating Function (4.4) of \(\sigma ^{2}_{\gamma }\)
Notice that by similar calculations as in (a.2) for β estimation, one can show that the first term in Eq. 4.4 is an unbiased estimating function for the first term in Eq. 3.24. That is,
However, the design expectation (ED) of the second term in Eq. 4.4 is given by
Hence, when eq. a.4 is combined with Eq. a.3, it follows that the naive estimating function \(\hat {\tau }_{NMM,\sigma ^{2}_{\gamma }|\beta }(y;\cdot )\) in (4.4) is not an unbiased estimating function for the HEF, in the left hand side of (3.24). However, by a bias correction adjustment, one can construct an unbiased function as done in (4.5).
Appendix B: Asymptotic Unbiasedness and Variance of the Survey Weighted GQL Estimator \(\hat {\beta }_{WGQL}\)
Asymptotic Unbiasedness:
By Eq. 5.2 we write
To compute the expectations involved in Eq. b.1, first we write from Eq. b.2 that
implying that
Next because
by Eq. 3.10, it follows from Eqs. b.2 and b.3 that
respectively.
Consequently, by applying (b.4) and (b.5), it follows from Eq. b.1 that
as \(n \rightarrow N \rightarrow \infty ,\) showing that \(\hat {\beta }_{WGQL}\) is asymptotically unbiased for β.
Computation of
\(\text {cov}\left ({\sum }^{k}_{c=1}{\sum }^{n_{c}}_{i=1}\tilde {w}_{cs^{*}}z_{cis^{*}} \right )\) Notice that as the sample of size \(n={\sum }^{k}_{c=1}n_{c}\) is selected from the FP of size \(N={\sum }^{K}_{c=1}N_{c}\) by using the TSCS, it then follows that
Now computing the within and between clustered expectations and covariances, one simplifies (b.7) and obtains \(V^{*}(\beta ,\sigma ^{2}_{\gamma }) = \text {cov}\left ({\sum }^{k}_{c=1}{\sum }^{n_{c}}_{i=1}\tilde {w}_{cs^{*}}z_{cis^{*}} \right )\) as in Eq. 5.8 under Section 5.1.
Asymptotic Variance of the Regression Estimator \(\hat {\beta }_{WGQL}:\)
Because \(\hat {\beta }_{WGQL},\) by Eq. b.6, is asymptotically unbiased for β, by applying (b.5) and (b.7), it follows from (5.2) (see also Eq. b.1) that
which, as expected, is the same as the covariance matrix of the normal distribution given by (5.26).
Appendix C: Proof for Consistency of the Survey Weighted MM Estimator \(\hat {\sigma }^{2}_{\gamma ,WMM}\)
Because \(\hat {\sigma }^{2}_{\gamma , WMM}\) is the solution of the WMM estimating equation (4.5), a first order Taylor series expansion of the estimating function in the left hand side of (4.5) about \(\sigma ^{2}_{\gamma }\) provides
where \(n={\sum }^{k}_{c=1}n_{c}\) (see also (4.6)). For convenience of further calculations, we re-express the equation in Eq. c.1 as
Notice that S1 is free of responses, whereas S2, y contains the responses {y} because \(p_{cs^{*}}\) and \(q_{cs^{*}}\) in Eq. c.1 are functions of second order responses. More specifically, by (4.4) (see also (3.20)), \(p_{cs^{*}}\) is constructed by using the second order responses from the matrix {yciyci′} for c = 1, … , k; i = 1, … , nc, where by (1.7), yci is the (J − 1)-dimensional multinomial response vector. Similarly, \(q_{cs^{*}}\) is constructed by using the elements of second order response matrix {yciycm′} given in (3.12), for i ≠ m; i, m = 1, … , nc. Consequently, one can show that
Suppose that this survey/finite population based quantity in Eq. c.3 is finite and bounded. More specifically, we assume that the following regularity condition holds:
C1.
For \(N={\sum }^{K}_{c=1}N_{c},\) let fN increases as N gets larger but it is a finite quantity. Also suppose that the scalar quantity in Eq. c.3 satisfies
By applying (c.4) to (c.3), one obtains the order of \(S^{-1}_{1}\) in Eq. c.2 as
Next, it is clear from Eq. c.1 that
in order of \([\text {var}(S_{2,y})]^{\frac {1}{2}}.\) To obtain this order we compute the variance var(S2, y) as follows:
where the variances and covariances can be computed as follows.
where si is a random indicator variable such that
[Cochran (1977, Section 2.9)]. It then follows that
and
Hence by putting (c.11) and (c.10) in (c.8), we obtain the variance
Notice that the computation of the remaining variance and the covariance is slightly more complicated because it will require the fourth order product moments of indicator random variables. Specifically, we write
where
by Eq. c.11. The covariance in Eq. c.13 is computed as
Using Eqs. c.15 and c.14 in Eq. c.13, we obtain the formula for Vc,2(⋅) as
Next we compute
Finally, by applying (c.12), (c.16) and (c.17), the desired variance follows from Eq. c.7 as
For simplicity, we assume that the sampling fractions are negligible. It then follows from Eqs. c.12, c.16 and c.17 that
respectively. Consequently, we can simplify the desired variance in Eq. c.18 as
Next we assume that the following regularity conditions hold with regard to the finite population:
C2
For \(N={\sum }^{K}_{c=1}N_{c},\) and for N dependent finite and bounded quantity rN, the summed quantities in Eq. c.22 satisfy
It then follows that the variance in Eq. c.22 satisfies the condition
Finally by applying (c.24), (c.6), and (c.5), the convergence result stated in (6.1)-(6.2) under Section 6 follows from Eq. c.2.
Rights and permissions
About this article
Cite this article
Sutradhar, B.C. Multinomial Logistic Mixed Models for Clustered Categorical Data in a Complex Survey Sampling Setup. Sankhya A 84, 743–789 (2022). https://doi.org/10.1007/s13171-020-00215-2
Received:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13171-020-00215-2
Keywords
- Clustered categorical data
- Clustered correlations through common random effect
- Complex survey design
- Consistency
- Correct mean specification
- Finite/survey population setup
- Generalized quasi-likelihood and method of moments estimation
- Large sample properties
- Normality
- Survey weighted estimation