Small area estimation for unemployment using latent Markov models
Section 4. The proposed model

• Previous
• Next

In this section, the proposed SAE model based on LMMs is illustrated. It can be considered as a compromise between the YRG model based on (3.1), which leads to possible oversmoothing, and the computationally demanding alternative proposed in Datta et al. (1999), based on (3.2). We first outline a general description on LMMs and then move to the specification of the area level model and to its estimation.

4.1  Preliminaries

In LMMs, the existence of two types of process is assumed: an unobservable finite-state first-order Markov chain $U_{it}$ with state space $\left\{\mathrm{1,}\dots \mathrm{,}k\right\}$ and an observed process, which in our case corresponds to ${\theta }_{it},$ with $i\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}m$ and $t\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}T.$ It is assumed that the distribution of ${\theta }_{it}$ depends only on $U_{it}\text{};$ specifically, the ${\theta }_{it}$ are conditionally independent given the $U_{it}.$ In addition, the latent state to which a small area belongs at a certain time point only depends on the latent state at the previous occasion.

The state-dependent distribution, namely the distribution of ${\theta }_{it}$ given $U_{it},$ can be a continuous or discrete. Such a distribution is typically taken from the exponential family. Thus, the overall vector of parameters of LMM, denoted by $\varphi ,$ includes parameters of the Markov chain, denoted by ${\varphi }_{\text{lat}}\text{},$ and the vector of parameters ${\varphi }_{\text{obs}}$ of the state-dependent distribution. In fact, the model consists of two components, the measurement model and the latent model, which concern the conditional distribution of the response variables given the latent variables and the distribution of the latent variables, respectively. By jointly considering these components, the so-called manifest distribution is obtained: it is the marginal distribution of the response variables, once the latent variables have been integrated out.

The measurement model, based on parameters ${\varphi }_{\text{obs}}\text{},$ can be written as

$${\theta }_{it}|U_{it}\mathrm{<mo>=</mo>}u\sim p\left({\theta }_{it}|u,{\varphi }_{\text{obs}}\right)\mathrm{.}$$

Moreover, the parameters ${\varphi }_{\text{lat}}$ of the Markov chain are:

• the vector of initial probabilities $\pi \mathrm{<mo>=</mo>}{\left({\pi }_1\mathrm{,}\dots \mathrm{,}{\pi }_k\right)}^{\prime }$ where

$${\pi }_u\mathrm{<mo>=</mo>}P\left(U_{i1}\mathrm{<mo>=</mo>}u\right)\mathrm{,}u\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k\mathrm{<mo>;</mo>}$$

• the transition probability matrix

$$\Pi =\left(\begin{array}{ccc}{\pi }_{1|1}& \dots & {\pi }_{1|k}\\ ⋮& \ddots & ⋮\\ {\pi }_{k|1}& \dots & {\pi }_{k|k}\end{array}\right)\mathrm{,}$$

• where

$${\pi }_{u|\overline{u}}=P\left(U_{it}=u|U_{i\text{}\mathrm{,}t-1}=\overline{u}\right)\mathrm{,}\overline{u}\mathrm{,}u\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k\mathrm{,}$$

• is the probability that area $i$ visits state $u$ at time $t$ given that at time $t-1$ it was in state $\overline{u}.$

In this work we consider homogeneous LMMs, namely LMMs where, in agreement with the previous definition, the transition probability matrix is constant in time. Generalizations to non-homogeneous hidden Markov chains and time-varying transition probabilities could also be considered (Bartolucci and Farcomeni, 2009). Individual covariates could be included in the measurement or in the latent model. When the covariates are included in the measurement model (Bartolucci and Farcomeni, 2009), they affect the response variables directly and the latent process is conceived as a way to account for the unobserved heterogeneity between areas. Differently, when the covariates are in the latent model (Vermunt and Magidson, 2002; Bartolucci, Pennoni and Francis, 2007) they influence initial and transition probabilities of the latent process. In a SAE context, we will consider the former approach, so that auxiliary information can be used to improve predictions. Bayesian inference approaches to LMMs are already available in the literature (e.g., in Marin, Mengersen and Robert, 2005; Spezia, 2010). In the following section we illustrate how to incorporate an LMM into an area level SAE model.

4.2  Proposed approach to area level SAE

The proposed model is based on two levels in an HB framework: at the first level, a sampling error model is assumed, then an LMM is used as linking model. The latter is based on two equations, corresponding to the measurement model and to the latent component. In particular, we adopt the following structure:

• Sampling Model:

$${\widehat{\theta }}_i|{\theta }_i\sim N_T\left({\theta }_i\mathrm{,}{\Psi }_i\right)\mathrm{,}i\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}m\mathrm{<mo>;</mo>}$$

• Linking Model:
  • Measurement Model:

$${\theta }_{it}|U_{it}\mathrm{<mo>=</mo>}u\mathrm{,}{x}_{it}\sim N\left(x_{it}^{\prime }{\beta }_u\mathrm{,}{\sigma }_u^2\right)i\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}m\mathrm{<mo>;</mo>}t\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}T\mathrm{<mo>;</mo>}$$

• Latent Model, based on the initial probabilities ${\pi }_u,$ $u\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k,$ and on the transition probabilities ${\pi }_{u|\overline{u}}\text{},$ $t\mathrm{<mo>=</mo>\; 2,}\dots \mathrm{,}T\text{}\mathrm{,}\overline{u}\mathrm{,}u\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k,$ already defined.

Here ${\beta }_u$ is the $p\times 1$ vector of the regression coefficients for the latent state to which area $i$ at time $t$ belongs, ${\sigma }_u^2$ is the corresponding error variance, and ${\Psi }_i$ is the matrix of sampling variances, which is assumed to be known.

It must be noticed that, while in the classical area level SAE models heterogeneity is modeled using continuous (usually Normally distributed) random variables, here it is modeled with a discrete dynamic variable. As we can deduce from Figure 4.1, our data have a skewed distribution. However, the empirical distribution is not far from a Normal distribution. D’Alò, Di Consiglio, Falorsi, Ranalli and Solari (2012) show that the differences in estimates between adopting a Normal or a Binomial model are not as relevant as expected and Normal models are often used for estimation of unemployment rates (You et al., 2003; Boonstra, 2014). Finally adopting the Normal distribution has computational advantages which are clarified later in this section.

The model parameters of interest can be divided into three groups:

• the matrix of small area parameters:

$$\Theta \mathrm{<mo>=</mo>}\left(\begin{array}{ccc}{\theta }_{11}& \cdots & {\theta }_{1T}\\ ⋮& \ddots & ⋮\\ {\theta }_{m1}& \cdots & {\theta }_{mT}\end{array}\right)\mathrm{<mo>;</mo>}(4.1)$$

• the vector of the measurement parameters:

$${\varphi }_{\text{obs}}\mathrm{<mo>=</mo>}{\left({\beta }_1^{\prime }\mathrm{,}\dots \mathrm{,}{\beta }_k^{\prime }\mathrm{,}{\sigma }_1^2\mathrm{,}\dots \mathrm{,}{\sigma }_k^2\right)}^{\prime }\text{}\mathrm{<mo>;</mo>}$$

• the set of latent parameters:

$${\varphi }_{\text{lat}}\mathrm{<mo>=</mo>}\left\{\pi \mathrm{,}\Pi \right\}\mathrm{.}$$

To complete the Bayesian formulation of the proposed model, it is necessary to choose priors for the model parameters. Small area parameters do not need a specific prior because direct estimates based on observed data are available; therefore, a set of priors is chosen for the measurement and the latent parameters. Regarding ${\varphi }_{\text{obs}},$ diffuse normal priors are assumed for the regression coefficients. These priors are conjugate and computationally more convenient than the usually flat priors over the real line (see Rao, 2003, Chapter 10). In particular, we assume

$${\beta }_u\sim N_p\left({\eta }_0\mathrm{,}{\Sigma }_0\right)\mathrm{,}u\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k\mathrm{,}$$

with ${\Sigma }_0\mathrm{<mo>=</mo>}{\sigma }_u^2{\Lambda }_0^{-1}$ and ${\Lambda }_0$ is a known diagonal matrix.

Variances ${\sigma }_u^2,$ $u\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k,$ are unknown and, therefore, it is necessary to set a prior also on these parameters. The choice of the prior distribution for the variance components is critical as in Bayesian mixed models the posterior distributions of these parameters are known to be sensitive to this specification. The inverse Gamma distribution is a popular choice, see e.g., You et al. (2003) and Datta, Lahiri, Maiti and Lu (1999) among others. Gelman (2006), Gelman, Jakulin, Pittau and Su (2008), and Polson and Scott (2012) propose to assume a half-Cauchy distribution for the variance of the random effect. Alternatively, a Uniform distribution can also be considered. Fabrizi et al. (2016) conduct an exhaustive sensitivity analysis when using a latent class model in a multivariate setting and find no significant difference among these different alternatives. For this reason, we choose the same prior distribution considered in You et al. (2003) and use an inverse Gamma distribution with shape parameter $a_0$ and scale parameter $b_0;$ then ${\sigma }_u^2\sim \text{IG}\left(a_0\mathrm{,}{b}_0\right),$ $u\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k\mathrm{,}$ where $a_0\mathrm{,}{b}_0\mathrm{<mo>></mo>\; 0}$ are set to very small values. This choice makes it also easier to derive the full conditional distributions for the Gibbs sampler.

For ${\varphi }_{\text{lat}}\text{},$ a system of Dirichlet priors is set on the initial probabilities and on the transition probabilities. The Dirichlet distribution is a conjugate prior for the multinomial distribution. This means that if the prior distribution of the multinomial parameters is Dirichlet then the posterior distribution belongs to the same family. The benefit of this choice is that the posterior distribution is easy to compute and, in some sense, it is possible to quantify how much our beliefs have changed after collecting the data. Then, we assume

$$\begin{array}{ll}\pi \hfill & \sim \text{Dirichlet}\left(1_k\right)\mathrm{,}\hfill \\ {\pi }_{\overline{u}}={\left({\pi }_{1|\overline{u}}\mathrm{,}\dots \mathrm{,}{\pi }_{k|\overline{u}}\right)}^{\prime }\hfill & \sim \text{Dirichlet}\left(1_k\right)\mathrm{,}\overline{u}\mathrm{<mo>=</mo>\; 1,}\dots \mathrm{,}k\mathrm{.}\hfill \end{array}$$

4.3  Estimation and model selection

In this work we make use of a data augmentation Markov Chain Monte Carlo (MCMC) method (Tanner and Wong, 1987; Liu, Wong and Kong, 1994; Van Dyk and Meng, 2001) based on the Gibbs sampler, in which the latent variables are treated as missing data (Marin et al., 2005; Germain, 2010). There are two main reasons for this choice. First of all, there is evidence that data augmentation has a better performance than other methods, as the marginal updating scheme (Boys and Henderson, 2003). Moreover, it simplifies the process of sampling from the posterior distribution. Details on this method and the full conditionals employed in the Gibbs sampler are given in Appendix A.1.

The choice of the number of latent states is a crucial step in applications. In the framework of LMMs, this requires a model selection procedure. From a Bayesian perspective, a fundamental goal is the computation of the marginal likelihood of the data for a given model. In this paper we use a model selection method based on the marginal likelihood and to estimate this quantity we use the method proposed by Carlin and Chib (1995), applied for each available model on the basis of the output of the MCMC algorithm. Technical details are provided in Appendix A.2.

A well-known problem occurring in Bayesian latent class and LMMs is the label switching. This implies that the component parameters are not identifiable as they are exchangeable. In a Bayesian context, if the prior distribution does not distinguish the component parameters between each other, then the resulting posterior distribution will be invariant with respect to permutations of the labels. Several solutions have been proposed; for a general review see Jasra, Holmes and Stephens (2005). The easiest approach is to use relabeling techniques retrospectively, by post-processing the MCMC output (Marin et al., 2005). However, in our case, we are interested in the prediction of the small area parameters, whose distribution depends on the number of areas in each latent state. Therefore, we do not use the post-processing approach and the MCMC output is permuted at every iteration according to the ordering of the mean of the response variables in each class.

• Previous
• Article main page
• Next