TEXT SIZE

search for



CrossRef (0)
Bayesian baseline-category logit random effects models for longitudinal nominal data
Communications for Statistical Applications and Methods 2020;27:201-210
Published online March 31, 2020
© 2020 Korean Statistical Society.

Jiyeong Kima, Keunbaik Lee1,b

aLaboratory of Low Dose Risk Assessment, National Radiation Emergency Medical Center, Korea Institute of Radiological and Medical Sciences, Korea;
bDepartment of Statistics, Sungkyunkwan University, Korea
Correspondence to: 1Department of Statistics, Sungkyunkwan University, 25-2, Sungkyunkwan-ro, Jongno-gu, Seoul 03063, Korea. E-mail: keunabaik@skku.edu
This paper is based on part of Jiyeong Kim셲 PhD thesis.
Received September 19, 2019; Revised December 2, 2019; Accepted December 24, 2019.
 Abstract
Baseline-category logit random effects models have been used to analyze longitudinal nominal data. The models account for subject-specific variations using random effects. However, the random effects covariance matrix in the models needs to explain subject-specific variations as well as serial correlations for nominal outcomes. In order to satisfy them, the covariance matrix must be heterogeneous and high-dimensional. However, it is difficult to estimate the random effects covariance matrix due to its high dimensionality and positive-definiteness. In this paper, we exploit the modified Cholesky decomposition to estimate the high-dimensional heterogeneous random effects covariance matrix. Bayesian methodology is proposed to estimate parameters of interest. The proposed methods are illustrated with real data from the McKinney Homeless Research Project.
Keywords : covariance matrix, heterogeneous, high-dimensional, modified Cholesky decomposition, positive-definiteness
1. Introduction

Longitudinal data are collected over time from the same subjects. Therefore, the outcomes from the same subjects are correlated. Many models have been proposed to analyze data such as linear mixed models and generalized linear mixed models (GLMMs). Especially, the GLMMs are commonly used to analyze longitudinal categorical data (Breslow and Clayton, 1993), and the GLMMs specify the effects of covariates on response conditional for random effects.

Baseline-category logit random effects models are typically used to analyze longitudinal nominal data (Theil, 1969, 1970), and the models account for subject-specific variations using the random effects covariance matrix. However, the random effects covariance matrix in the models cannot explain the serial correlations of nominal outcomes. The random effects covariance matrix must be heterogeneous and high-dimensional to account for both the correlations and subject-specific variations. However, it is difficult to estimate the random effects covariance matrix due to high dimensionality and positive-definiteness (Lee et al., 2012). In this paper, we propose models to solve problems using modified Cholesky decomposition (MCD) (Pourahmadi, 1999).

The MCD approach uses the new unconstrained parametrization of an inverse of a covariance matrix. As a result, the parameters of the MCD are generalized autoregressive parameters (GARPs) and innovation variances (IVs). The GARPs are dependence parameters describing the serially dependence of the previous outcomes, and the IVs are prediction variances. The positive-definiteness restriction of the covariance matrix is that the IVs need to be positive (Pourahmadi, 1999, 2000). Pan and Mackenzie (2006) used the MCD to address joint mean-covariance estimation for linear models. Lee et al. (2012) and Lee (2013) also used the MCD to estimate the random effects covariance matrix in logistic random effects models for longitudinal binary data. In this paper, we also use the MCD to estimate the random effects covariance matrix in baseline-category logit random effects models for longitudinal nominal data.

There is much literature dealing with models for longitudinal nominal data. Multinomial logit models were developed by Theil (1969, 1970). Daniels and Gatsonis (1997) proposed a Bayesian two-level generalized logit model to accommodate clustered nominal data. Revelt and Train (1998) proposed discrete choice models with random coefficients that do not have the restrictive ‘independence from irrelevant alternatives’ property. Hartzel et al. (2001) studied logit random effects models of clustered ordinal and nominal data. For the analysis of clustered or longitudinal nominal data, mixed-effects multinomial logistic regression models were proposed by Hedeker (2003). Chen et al. (2009) developed a Markov model based on the likelihood approach to analyze longitudinal categorical data for the modeling of both marginal and conditional relationships. Lee and Mercante (2010) and Lee et al. (2011) proposed marginalized models using a Markovian dependence structure or random effects to analyze longitudinal nominal data, respectively.

This paper is organized as follows. In Section 2, we propose baseline-category logit random effects models for longitudinal nominal data using the MCD approach. In Section 3, we present Bayesian methodology for estimation of parameters. In Section 4, we illustrate real data and apply our proposed models to them. Finally, we summarize this paper in Section 5.

2. Bayesian baseline-category logit random effects models for longitudinal nominal data

In this section, we propose baseline-category logit random effects models with autoregressive random effects covariance matrix to analyze longitudinal nominal data.

2.1. Proposed models

Let Yit be a nominal response with K-categories on subject i (i = 1, . . . , N) at time t (t = 1, . . . , ni; niT) and let xit be the corresponding vector of covariates. We assume that each Yit is conditionally independent given random effects bit, that the responses for different subjects are independent, and that the regression model is given by log P ( Y i t = k | b i t , x i t ) P ( Y i t = K | b i t , x i t ) = x i t T β k + b i t k ,where k = 1, . . . , K − 1, xit is a p × 1 vector of covariates, βk is a p × 1 vector of regression coefficient, b i = ( b i 1 , , b i n i ) T indep ~ N ( 0 , Σ i ) ,with b i t T = ( b i t 1 , , b i t , K - 1 ), and ∑i is a (K − 1)ni × (K − 1)ni random effects covariance matrix.

Then conditional probabilities for Yit given the random effects bit are given by P ( Y i t = k | b i t , x i t ) = { exp ( x i t T β k + b i t k ) 1 + l = 1 K - 1 exp ( x i t T β k + b i t k ) , for  k = 1 , , K - 1 , 1 1 + l = 1 K - 1 exp ( x i t T β k + b i t k ) , for  k = K .

The random effects covariance matrix ∑i has subject variations and serial correlations. In addition, it is high-dimensional and should be positive definite. However, the estimation of the covariance matrix is not easy due to the constraints. Therefore, we consider the MCD to solve the problem of constraints.

2.2. Modeling of the random effects covariance matrix

In this section, we describe the random effects covariance matrix ∑i using the MCD. The random effects bit in equation (2.2) are assumed to be decomposed as follows b i 1 = e i 1 , b i t = j = 1 t - 1 Ψ i t j b i j + e i t ,  듼  듼  듼 for  t = 2 , , n i ,where e i t = ( e i t 1 e i t , K - 1 ) , Ψ i t j = ( φ i t j , 11 φ i t j , 12 φ i t j , 1 , K - 1 φ i t j , 21 φ i t j , 22 φ i t j , 2 , K - 1 φ i t j , K - 1 , 1 φ i t j , K - 1 , 2 φ i t j , K - 1 , K - 1 ) .

Note that the elements of the matrix Ψit j, which are called GARPs, present serial correlations of repeated nominal outcomes. The serial correlations are the correlation within categories over time and the cross-correlation between different categories at different times. A similar form to indicate dependence of multivariate longitudinal outcomes was presented in Lee et al. (2020). We refer to the matrix Ψit j as the generalized autoregressive matrices (GARMs).

We also assume that e i = ( e i 1 T , , e i n i T ) T indep ~ N ( 0 , A i ) ,where Ai = diag(Ai1, . . . , Aini) with A i t = diag ( σ i t 1 2 , , σ i t , K - 1 2 ). Note that diagonal matrix Ait presents the prediction variance matrix of bit. We refer to the new parameters Ait as the innovation variance matrices (IVMs).

Then we reexpress equations (2.3) and (2.4) in matrix form as T i b i = e i ,where T i = ( I 0 0 0 - [ Ψ i 21 ] I 0 0 - [ Ψ i 31 ] - [ Ψ i 32 ] I 0 - [ Ψ i n i 1 ] - [ Ψ i n i 2 ] - [ Ψ i n i 3 ] I ) .

From equation (2.5), we have T i Σ i T i T = Var ( e i ) T i Σ i T i T = A i Σ i = T i - 1 A i T i - T .

We note that the random effects covariance matrix is directly decomposed into the GARMs and IVMs. They can be modeled using time and/or subject-specific covariate vectors wit j and hit by setting φ i t j , l m = w i t j T α l m , log  σ i t k 2 = h i t T λ k ,where αlm is an a × 1 vector and λk is a b × 1 vector of unknown parameters, respectively.

Note that wit j’s are covariate design vectors controlling the time order of the model and the correlation of responses. As a result, the random effects covariance matrix can be nonstationary and heteroscedastic depending on covariates. We also note that the IVMs are positive definite using the loglinear model in equation (2.7). This result guarantees that ∑i is positive definite because the diagonal matrix of Ai in equation (2.6) are all positive. These results represent the advantages of the MCD.

3. Bayesian methodology

We now describe Bayesian approaches to estimate parameters in our proposed models. We derive the likelihood function for the model specified in Subsection 2.1. The parameters in model (2.7) are the regression coefficients which ranges on (−∞,∞). In this case, normal priors are commonly used for parameters and guarantee the propriety of posterior distributions. The normal priors with large prior variances remains relatively objective (Daniels and Zhao, 2003). The priors distributions for the parameters in the model with the AR structure of random effects covariance matrix are given by β k ~ N ( 0 , σ β 2 I ) , α l m ~ N ( 0 , σ α 2 I ) , λ k ~ N ( 0 , σ λ 2 I ) .

In general, σ β 2 , σ α 2, and σ λ 2 are in large to be noninformative such as 100. By combining likelihood function from (2.1) and prior distributions from (3.1)(3.3), we obtain the joint distribution given by P ( y , b , β , α , λ ) P ( y | b , β ) P ( b | α , λ ) P ( β ) P ( α ) P ( λ ) [ i = 1 N { t = 1 n i k = 1 K ( P i t k c ( b i t k ) ) y i t k } { t = 1 n i k = 1 K - 1 ( α i t k 2 ) - 1 2 } exp ( - 1 2 b i T Σ i - 1 b i ) ] × k = 1 K - 1 exp ( - 1 2 σ β 2 β k T β k ) × l = 1 K - 1 m = 1 K - 1 exp ( - 1 2 σ α 2 α l m T α l m ) × k = 1 K - 1 exp ( - 1 2 σ λ 2 λ k T λ k ) .

To generate parameters from the posterior distribution, MCMC methods are adapted to generate posterior samples for model estimation. The full condition posterior distribution are given below:

  • For bi (i = 1, . . . , N) P ( b i | y , β , β 0 , α , γ , λ ) { t = 1 n i k = 1 K ( P i t k c ( b i t ) ) y i t k } × exp ( - 1 2 i = 1 N b i T T i T A i - 1 T i T b i ) .

  • For βk P ( β k | y , b , β 0 , α , γ , λ ) [ i = 1 N { t = 1 n i k = 1 K ( P i t k c ( b i t ) ) y i t k } ] exp ( - 1 2 σ β k 2 β k T β k ) .

  • For αlm P ( α l m | y , b , β , β 0 , γ , λ ) exp ( - 1 2 i = 1 N b i T T i T A i - 1 T i T b i ) exp ( - 1 2 σ α 2 α T α ) .

  • For λk P ( λ k | y , b , β , β 0 , α , γ ) { i = 1 N t = 1 n i ( σ i t 2 ) - 1 2 } exp ( - 1 2 i = 1 N b i T T i T A i - 1 T i T b i ) exp ( - 1 2 σ λ k 2 λ k T λ k ) .

All full conditionals are not closed forms; therefore, we construct suitable proposals for a Metropolis-Hastings step. In practice, MCMC is implemented using JAGS (http://mcmc-jags.sourceforge.net/).

4. Analysis of McKinney homeless research project data

4.1. Data description

The McKinney homeless research project (MHRP), first described by Hurlburt et al. (1996), was used to demonstrate the use of our proposed models. The MHRP was designed to assess if the use of Section 8 housing certificates was effective in providing housing options to homeless individuals with severe mental illness. 361 clients were randomly assigned to one of two types of supportive case management (comprehensive and traditional) and to one of two levels of access for independent housing (Section 8 certificates: Group = 1 for yes; 0 for no). Nominal-level housing status outcomes were collected at baseline, and at 6, 12, and 24 months after post-randomization. When analyzing this data, time was set to 0, 1, 2, 3. Similar to Hedeker and Gibbons (2006), we focus on examining the effect of access to Section 8 certificates on repeated housing outcomes across time. There were three different housing outcomes of either street/shelter housing, community housing or independent housing for each time point.

Figure 1 presents marginal proportions of responses for MHRP data. The marginal proportion plots indicate the marginal probability for Section 8 certificate at each time, indicating the difference between the groups.

For more detailed explanations on categorical outcomes, housing categories by living arrangements are summarized in Table 1 according to the study by Hurlburt et al. (1996). About 25 % of the subjects dropped out of the study during the follow-up period resulting in some missing housing outcome status data. For our analysis, we assume missing at random missingness. The housing outcome status of street/shelter was chosen as the reference category; in addition, we focused on evaluating the effects of access to Section 8 certificates on housing outcomes across time.

4.2. Model fit

We fit five models which are a typical GLMM and four baseline-category logit random effects models with several structure of ∑i.

Table 2 presents the specification of the five models for ∑i using the various structures. The typical GLMM is a baseline-category logit random effects model with a homogeneous random effects variance. The other models considered only the time window difference using indicator function 1(I(|tj|=1)), which indicates the AR(1) structure. Note that ‘-C’ means a model with a constant random effects covariance matrix and ‘-A’ means a model with a random effect covariance matrix depending on group. Therefore, Model AR(1)-CC indicates a model with a homogeneous random effects covariance matrix with an AR(1) structure, and Model AR(1)-AA indicates a model with a group-dependent AR(1) random effects covariance matrix having IVs depending on a group with an AR(1) structure. Model AR(1)-CA is a model with an AR(1) random effects covariance matrix having an IVs depending on a group. Finally, Model AR(1)-AC is a model with a group-dependent AR(1) random effects covariance matrix that has constant IVs.

For the estimation of all parameters in the models, Gibbs sampler is implemented using JAGS in R. Posterior means were calculated with a sample size of 250,000, thin of 5 and burn-in period of 100,000. To use the Gelman and Rubin approach, we used multiple chains (chain of 2). We also checked the convergence of all parameters in the models using the trace plots of random numbers for the parameters. Using the plots, we observed that the lines of different chains were mixed and crossed; convergence was then satisfied.

Table 3 shows comparison of the four models using deviance information criterion (DIC) (Spiegelhalter et al., 2010). Since AR(1)-AA was the most complex model among the five models, pD was large. However, DIC was the smallest of all other models. This means that the Model AR(1)-AA was the best fit. Therefore, we now focus on Model AR(1)-AA for further analysis.

Table 4 is organized into three parts according to the nominal response categories to be compared (either community vs street/shelter or independent vs street/shelter) and GARPs. The top part of Table 4 presents the estimates of coefficients and associated 95% credible interval to compare the two nominal response categories of community housing and street/shelter housing. For these two categories, the estimated baseline-category logit is given by log P ^ ( Community ) P ^ ( Street / Shelter ) = - 0.27 + 3.01 Time 1 + 3.48 Time 2 + 2.48 Time 3 - 0.10 Section  8 - 1.34 Time 1 * Section  8 - 2.77 Time 2 * Section  8 - 0.54 Time 3 * Section  8 + b ^ i t K ,where Time1, Time2, and Time3 are indicators of month 6 (6 Month), month 12 (12 Month), and month 24 (14 Month), respectively. The posterior means of regression coefficients for 6 Month versus Baseline and 12 Month versus Baseline were not in 95% credible intervals. In Control (Groupi = 0), the odds ratio of the conditional posterior probability were e3.01 = 20.29, e3.48 = 32.46, and e2.48 = 11.94 in community housing as opposed to street/shelter housing at 6 Month, 12 Month and 24 Months. The posterior means of regression coefficients for the interaction between Section 8 certificates and 12 Month were not in the credible intervals. Combining the logit estimates for the main effects of Section 8 certificates (−0.10) and the association between 6 Month and Section 8 certificates (−1.34) yielded an estimated odds ratio of e−1.44 = 0.24, the association between 12 Month and Section 8 certificates (−2.77) yielded an estimated odds ratio of e−2.87 = 0.06, and the association between 24 Month and Section 8 certificates (−0.54) yielded an estimated odds ratio of e−0.64 = 0.53, suggesting that individuals with Section 8 certificates were less likely to be in community housing as opposed to street/shelter housing at 6 Month, 12 Month and 24 Months.

Second, to compare independent housing and street/shelter housing we consider the lower part of Table 4. The baseline-category logit is log P ^ ( Independent ) P ^ ( Street / Shelter ) = - 1.67 + 1.94 Time 1 + 3.36 Time 2 + 2.83 Time 3 - 0.75 Section  8 + 3.25 Time 10 * Section  8 + 2.70 Time 2 * Section  8 + 4.08 Time 3 * Section  8 + b ^ i t K .

The posterior means of regression coefficients for 6 Month versus Baseline, 12 Month versus Baseline and 24 Month versus Baseline were not in credible intervals. The odds ratios of the control group were e1.94 = 6.96, e3.36 = 28.79, and e2.83 = 16.95 in community housing as opposed to street/shelter housing at 6 Month, 12 Month, and 24 Months. Also the posterior means of regression coefficients of the interaction between Section 8 certificates and 6 Month, 12 Month, and 24 Month were not in credible intervals. Combining the logit estimates for the main effects of Section 8 certificates (−0.75) and the association between 6 Month and Section 8 certificates (3.25) yielded an estimated odds ratio of e2.5 = 12.18, the association between 12 Month and Section 8 certificates (2.70) yielded an estimated odds ratio of e1.95 = 7.03, and the association between 24 Month and Section 8 certificates (4.08) yielded an estimated odds ratio of e3.33 = 27.94, suggesting that individuals with Section 8 certificates were less likely to be in community housing as opposed to street/shelter housing at 6 Month, 12 Month and 24 Months.

Some may be interested in comparing community housing and independent housing. The baseline-category parameters of the K–1 equations in the model can be used to represent logit defined for pairs of different response categories as follows log P ( Community ) P ( Independent ) = log P ( Community ) P ( Street / Shelter ) - log P ( Independent ) P ( Street / Shelter ) .

Thus, log P ^ ( Community ) P ^ ( Independent ) = 1.40 + 1.07 Time 1 + 0.12 Time 2 - 0.35 Time 3 + 0.65 Section  8 - 4.59 Time 1 * Section  8 - 5.47 Time 2 * Section  8 - 4.62 Time 3 * Section  8 + b ^ i t K .

The posterior means for diagonal element matrices of Âi were given by A ^ i t = { diag  ( 29.96 , 0.85 ) , Control group , diag  ( 1.40 , 3.22 ) , Section  8.

In the GARMs, the posterior mean for Ψit j is given by Ψ ^ i t j = { ( 0.55 0.02 - 0.15 0.17 ) , for Control group , ( 0.56 3.16 2.00 - 0.78 ) , for Section  8 group .

Figure 2 compares fitted marginal probabilities for Section 8 group versus control group. In the street/shelter housing, two estimated marginal probabilities decreased as the month increased. In the community housing and independent housing, there were many difference between Section 8 and control groups.

5. Conclusion

We proposed Bayesian baseline-category logit random effects models for longitudinal nominal data. In the models, the modified Cholesky decomposition (MCD) was used to decompose the random effects covariance matrix to the generalized autoregressive matrices (GARMs) and innovation variance matrices (IVMs). The GARMs account for serial correlations of nominal outcomes, and the IVMs explain prediction error variances. The MCD represents a computationally attractive approach and provides a better fit than the competing random intercept model with a homogeneous covariance. The proposed models also were fitted using a Bayesian approach. McKinney homeless research project (MHRP) data were analyzed using our proposed models. We fitted five baseline-category logit models to compare. Among the models, the model with a heteroscedastic AR(1) random effects covariance matrix was the best fit to our data. The estimated conditional probabilities for three groups were different trends as months increased.

Figures
Fig. 1.

Response of marginal proportions under two group (Section 8 and control) for response 1 (street/shelter), 2 (community) and 3 (independent housings), respectively.


Fig. 2.

Model fit of marginal proportions under two group (Section 8 and control) for response 1 (street/shelter), 2 (community) and 3 (independent housings), respectively.


TABLES

Table 1

Nominal outcomes by living arrangements

Outcomes Living arrangement
Street/shelter housing Public or private shelter
Church/chapel
Indoor public place (bus station/theater)
Abandoned building
Car or other vehicle
Outside without shelter

Community housing Hotel
Family member’s home or room
Friend’s or acquaintance’s home or room
Boarding house/halfway house

Independent housing Private house or own apartment

Table 2

Models fit with wit j and hit for MHRP data

Model GARMs log(ICMs)
GLMM NA log  σ i t k 2 = λ 0 k
AR(1)-CC: it j,lm = αlm,0I(|tj|=1) log  σ i t k 2 = λ 0 k
AR(1)-CA: it j,lm = αlm,0I(|tj|=1) log  σ i t k 2 = λ 0 k + λ 1 k Group i
AR(1)-AC: it j,lm = αlm,0I(|tj|=1) + αlm,1Groupi I(|tj|=1) log  σ i t k 2 = λ 0 k
AR(1)-AA: it j,lm = αlm,0I(|tj|=1) + αlm,1Groupi I(|tj|=1) log  σ i t k 2 = λ 0 k + λ 1 k Group i

Table 3

DIC of Bayesian baseline-category logit random effects model for MHRP data

Model Dev ( θ ) ¯ D(θ) pD DIC
GLMM 2069.00 1666.34 402.66 2471.66
AR(1)-CC 1612.81 1227.92 384.89 1997.70
AR(1)-CA 1629.63 1261.44 368.19 1997.82
AR(1)-AC 1590.25 1208.74 381.51 1971.76
AR(1)-AA 1526.97 1090.70 436.27 1963.24

Table 4

Posterior means of Bayesian baseline-category logit random effects model using MCD for MHRP data (95% Bayesian confidence interval)

GLMM AR(1)-CC AR(1)-CA AR(1)-AC AR(1)-AA
C vs S
Intercept −0.09 (−0.50, 0.33) −0.24 (−0.58, 0.09) −0.24 (−0.60, 0.12) −0.21 (−0.55, 0.12) −0.27 (−1.34, 0.90)
6 Month vs Baseline 1.71* (0.74, 3.11) 1.62* (1.02, 2.30) 1.64* (1.04, 2.34) 1.61* (1.03, 2.26) 3.01* (1.40, 5.65)
12 Month vs Baseline 1.99* (0.68, 3.22) 2.60* (1.71, 3.62) 2.62* (1.71, 3.63) 2.73* (1.81, 3.79) 3.48* (1.37, 5.67)
24 Month vs Baseline 0.79 (−1.14, 1.96) 2.22* (0.80, 3.68) 2.26* (0.89, 3.68) 2.53* (1.02, 4.19) 2.48 (−0.30, 4.91)
Section 8 (YES, NO) −0.88 (−3.33, 0.29) 0.07 (−0.47, 0.55) 0.05 (−0.57, 0.57) 0.00 (−0.51, 0.48) −0.10 (−1.33, 0.88)
Section 8 by 6 Month −1.34 (−4.23, 0.49) −0.38 (−1.39, 0.59) −0.37 (−1.43, 0.64) −0.19 (−1.31, 0.93) −1.34 (−4.09, 0.61)
Section 8 by 12 Month −3.66 (−7.42, −1.20) −2.42* (−3.71, −1.21) −2.39* (−3.80, −1.09) −2.19* (−3.88, −0.62) −2.77* (−5.58, −0.31)
Section 8 by 24 Month −1.45 (−4.20, 0.36) −0.90 (−2.91, 0.81) −0.72 (−2.86, 1.19) −0.95 (−3.57, 1.44) −0.54 (−3.66, 2.63)
λ0 2.05 (−3.25, 4.62) −1.64 (−5.69, 1.11) −1.22 (−6.10, 1.46) −1.46 (−5.54, 0.44) 3.40* (0.61, 6.14)
λ1 (Group) −0.07 (−1.81, 1.40) −3.06* (−6.27, −0.27)

I vs S
Intercept −0.52* (−0.83,−0.21) −1.47* (−2.02,−0.98) −1.44* (−1.99, −0.96) −1.62* (−2.35, −1.06) −1.67* (−2.46, −1.08)
6 Month vs Baseline 0.98* (0.41,1.57) 1.52* (0.74,2.30) 1.58* (0.80, 2.34) 1.56* (0.74, 2.38) 1.94* (1.09, 2.89)
12 Month vs Baseline 1.97* (1.30,2.69) 2.52* (1.23,3.80) 2.64* (1.33, 3.87) 2.76* (1.38, 4.11) 3.36* (1.95, 5.02)
24 Month vs Baseline 1.66* (1.02,2.32) 2.07* (0.23,3.84) 2.24* (0.39, 3.95) 2.63* (0.61, 4.60) 2.83* (0.72, 5.15)
Section 8 (YES, NO) −0.70* (−1.27,−0.16) −0.02 (−0.70,0.64) −0.13 (−0.92, 0.60) 0.08 (−0.63, 0.79) −0.75 (−2.58, 0.62)
Section 8 by 6 Month 2.23* (1.29,3.23) 2.03* (0.89,3.26) 2.09* (0.88, 3.51) 2.25* (0.94, 3.81) 3.25* (1.15, 5.89)
Section 8 by 12 Month 0.86 (−0.11,1.84) 1.35 (−0.14,2.93) 1.39 (−0.18, 3.09) 1.73 (−0.05, 3.84) 2.70* (0.10, 5.70)
Section 8 by 24 Month 1.50* (0.52,2.52) 3.082* (0.91,5.41) 3.11* (0.77, 5.62) 2.95* (0.25, 5.82) 4.08* (0.88, 7.47)
λ0 −2.92* (−6.67,−0.21) −0.50 (−1.80,0.69) −0.74 (−2.40, 0.65) −0.26 (−1.91, 1.19) −0.16 (−2.11, 1.45)
λ1 (Group) 0.44 (−0.76, 1.46) 1.33 (−2.96, 3.73)

α11,0 (AR(1)) 1.91* (0.70,4.40) 1.72* (0.61, 4.91) 1.78* (0.80, 4.62) 0.55* (0.32, 0.97)
α12,0 (AR(1)) 0.05 (−0.18,0.25) 0.05 (−0.20, 0.26) −0.52 (−2.36, 0.73) 0.02 (−1.28, 1.22)
α21,0 (AR(1)) −0.69 (−5.27,0.90) −0.34 (−5.40, 0.98) 0.19 (−0.16, 0.52) −0.15 (−1.32, 0.57)
α22,0 (AR(1)) 2.11* (1.48,2.94) 2.11* (1.38, 3.09) −0.18 (−0.63, 0.23) 0.17 (−0.62, 1.35)
α11,1 (AR(1) by Group) −0.655 (−5.35, 0.88) 0.01 (−0.09, 0.21)
α12,1 (AR(1) by Group) 3.26 (−2.04, 8.48) 3.14* (0.18, 8.07)
α21,1 (AR(1) by Group) 1.94* (1.27, 2.93) 2.15* (1.33, 3.33)
α22,1 (AR(1) by Group) −0.266 (−1.97, 0.89) −0.95* (−2.34, 0.37)

*indicates the 95% credible interval does not include zero.

C vs S is Community vs Street/Shelter, I vs S is Independent vs Street/Shelter


References
  1. Breslow NE and Clayton DG (1993). Approximate inference in generalized linear mixed models, Journal of the American Statistical Association, 88, 125-134.
    CrossRef
  2. Chen B, Yi GY, and Cook RJ (2009). Likelihood analysis of joint marginal and conditional models for longitudinal categorical data, Canadian Journal of Statistics, 37, 182-205.
    CrossRef
  3. Daniels MJ and Gatsonis C (1997). Hierarchical polytomous regression models with applications to health services research, Statistics in Medicine, 16, 2311-2325.
    Pubmed CrossRef
  4. Daniels JM and Zhao YD (2003). Modeling the random effects covariance matrix in longitudinal data, Statistics in Medicine, 22, 1631-1647.
    CrossRef
  5. Hartzel J, Agresti A, and Caffo B (2001). Multinomial logit random effects models, Statistical Modelling, 1, 81-102.
    CrossRef
  6. Hedeker D (2003). A mixed-effects multinomial logistic regression model, Statistics in Medicine, 22, 1433-1446.
    Pubmed CrossRef
  7. Hedeker D and Gibbons RD (2006). Longitudinal Data Analysis, Wiley, Hoboken, New Jersey.
  8. Hurlburt MS, Wood PA, and Hough RL (1996). Providing independent housing for the homeless mentally ill: A novel approach to evaluating long-term longitudinal housing patterns, Journal of Community Psychology, 24, 291-310.
    CrossRef
  9. Lee K (2013). Bayesian modeling of random effects covariance matrix for generalized linear mixed models, Communication for Statistical Applications and Methods, 20, 235-240.
    CrossRef
  10. Lee K, Cho H, Kwak MS, and Jang EJ (2020). Estimation of covariance matrix of multivariate longitudinal data using modified Choleksky and hypersphere decompositions, Biometrics, 76, 75-86.
    CrossRef
  11. Lee K, Kang S, Liu X, and Seo D (2011). Likelihood-based approach for analysis of longitudinal nominal data using marginalized random effects models, Journal of Applied Statistics, 38, 1577-1590.
    CrossRef
  12. Lee K and Mercante D (2010). Longitudinal nominal data analysis using marginalized models, Computational Statistics & Data Analysis, 54, 208-218.
    CrossRef
  13. Lee K, Yoo JK, Lee J, and Hagan J (2012). Modeling the random effects covariance matrix for the generalized linear mixed models, Computational Statistics & Data Analysis, 56, 1545-1551.
    CrossRef
  14. Pan J and Mackenzie G (2006). Regression models for covariance structures in longitudinal studies, Statistical Modelling, 6, 43-57.
    CrossRef
  15. Pourahmadi M (1999). Joint mean-covariance models with applications to longitudinal data: unconstrained parameterisation, Biometrika, 86, 677-690.
    CrossRef
  16. Pourahmadi M (2000). Maximum likelihood estimation of generalized linear models for multivariate normal covariance matrix, Biometrika, 87, 425-435.
    CrossRef
  17. Revelt D and Train K (1998). Mixed logit with repeated choices: households’ choices of appliance efficiency level, Review of Economics and Statistics, 80, 647-657.
    CrossRef
  18. Theil H (1969). A multinomial extension of the linear logit model, International Economic Review, 10, 251-259.
    CrossRef
  19. Theil H (1970). On the estimation of relationships involving qualitative variables, American Journal of Sociology, 76, 103-154.
    CrossRef