MCMC maximum likelihood for latent state models

This paper develops a pure simulation-based approach for computing maximum likelihood estimates in latent state variable models using Markov Chain Monte Carlo methods (MCMC). Our MCMC algorithm simultaneously evaluates and optimizes the likelihood function without resorting to gradient methods. The approach relies on data augmentation, with insights similar to simulated annealing and evolutionary Monte Carlo algorithms. We prove a limit theorem in the degree of data augmentation and use this to provide standard errors and convergence diagnostics. The resulting estimator inherits the sampling asymptotic properties of maximum likelihood. We demonstrate the approach on two latent state models central to financial econometrics: a stochastic volatility and a multivariate jump-diffusion models. We find that convergence to the MLE is fast, requiring only a small degree of augmentation.     

An EM algorithm for the estimation of parametric and nonparametric hierarchical nonlinear models

It is shown how to implement an EM algorithm for maximum likelihood estimation of hierarchical nonlinear models for data sets consisting of more than two levels of nesting. This upward–downward algorithm makes use of the conditional independence assumptions implied by the hierarchical model. It cannot only be used for the estimation of models with a parametric specification of the random effects, but also to extend the two-level nonparametric approach – sometimes referred to as latent class regression – to three or more levels. The proposed approach is illustrated with an empirical application.     

Some remarks on measurement errors and the identification of panel data models*

We demonstrate that despite the common worry about the possible correlations between the unobserved individual effects and the explanatory variables in panel data models the likelihood approach can provide a unified framework towards the study of the identification of a panel data model subject to measurement errors. In fact, it can also serve as a basis for deriving efficient estimation methods.     

Bayesian modeling of joint and conditional distributions

In this paper, we study a Bayesian approach to flexible modeling of conditional distributions. The approach uses a flexible model for the joint distribution of the dependent and independent variables and then extracts the conditional distributions of interest from the estimated joint distribution. We use a finite mixture of multivariate normals (FMMN) to estimate the joint distribution. The conditional distributions can then be assessed analytically or through simulations. The discrete variables are handled through the use of latent variables. The estimation procedure employs an MCMC algorithm. We provide a characterization of the Kullback–Leibler closure of FMMN and show that the joint and conditional predictive densities implied by the FMMN model are consistent estimators for a large class of data generating processes with continuous and discrete observables. The method can be used as a robust regression model with discrete and continuous dependent and independent variables and as a Bayesian alternative to semi- and non-parametric models such as quantile and kernel regression. In experiments, the method compares favorably with classical nonparametric and alternative Bayesian methods.     

Estimating Transition Probabilities from a Time Series of Independent Cross Sections

This paper considers the implementation of a nonstationary, heterogeneous Markov model for the analysis of a binary dependent variable in a time series of independent cross sections. The model, previously considered by M offitt (1993), offers the opportunity to estimate entry and exit transition probabilities and to examine the effects of time-constant and time-varying covariates on the hazards. We show how ML estimates of the parameters can be obtained by Fisher's method-of-scoring and how to estimate both fixed and time-varying covariate effects. The model is exemplified with an analysis of the labor force participation decision of Dutch women using data from the Socio-economic Panel (SEP) study conducted in the Netherlands between 1986 and 1995. We treat the panel data as independent cross sections and compare the employment status sequences predicted by the model with the observed sequences in the panel. Some open problems concerning the application of the model are also discussed.     

Generalized extreme value discrete choice demand models : Existence and uniqueness of market equilibria

Generalized extreme value (GEV) random utility choice models have been suggested as a development of the multinomial logit models that allows the random components of various alternatives to be statistically dependent. This paper establishes the existence of and provides necessary and sufficient uniqueness conditions for the solutions to a set of equations that may be interpreted as an equilibrium of an economy, the demand side of which is described by a multiple-segment GEV random choice model. The same equations may alternatively be interpreted in a maximum likelihood estimation context. The method employed is based on optimization theory and may provide a useful computational approach. The uniqueness results suggest a way to introduce segregation/integration effects into logit type choice models. Generalization to non-GEV models are touched upon.     

Stationary-Increment Variance-Gamma and t Models: Simulation and Parameter Estimation

We detail a method of simulating data from long range dependent processes with variance-gamma or t distributed increments, test various estimation procedures [method of moments (MOM), product-density maximum likelihood (PMLE), non-standard minimum χ² and empirical characteristic function estimation] on the data, and assess the performance of each. The investigation is motivated by the apparent poor performance of the MOM technique using real data ( Tjetjep & Seneta, 2006 ); and the need to assess the performance of PMLE for our dependent data models. In the simulations considered the product-density method performs favourably.     

Estimating wage differentials without logarithms

Regression models of wage determination are typically estimated by ordinary least squares using the logarithm of the wage as the dependent variable. These models provide consistent estimates of the proportional impact of wage determinants only under the assumption that the distribution of the error term is independent of the regressors — an assumption that can be violated by the presence of heteroskedasticity, for example. Failure of this assumption is particularly relevant in the estimation of the impact of union status on wages. Alternative wage-equation estimators based on the use of quasi-maximum-likelihood methods are consistent under weaker assumptions about the dependence between the error term and the regressors. They also provide the ability to check the specification of the underlying wage model. Applying this approach to a standard data set, I find that the impact of unions on wages is overstated by a magnitude of 20-30 percent when estimates from log-wage regressions are used for inference.     

Mixed Hidden Markov Models for Longitudinal Data: An Overview

In this paper we review statistical methods which combine hidden Markov models (HMMs) and random effects models in a longitudinal setting, leading to the class of so‐called mixed HMMs. This class of models has several interesting features. It deals with the dependence of a response variable on covariates, serial dependence, and unobserved heterogeneity in an HMM framework. It exploits the properties of HMMs, such as the relatively simple dependence structure and the efficient computational procedure, and allows one to handle a variety of real‐world time‐dependent data. We give details of the Expectation‐Maximization algorithm for computing the maximum likelihood estimates of model parameters and we illustrate the method with two real applications describing the relationship between patent counts and research and development expenditures, and between stock and market returns via the Capital Asset Pricing Model.     

Censored regression models with unobserved,stochastic censoring thresholds

The ‘Tobit’ model is a useful tool for estimation of regression models with truncated or limited dependent variables, but it requires a threshold which is either a known constant or an observable and independent variable. The model presented here extends the Tobit model to the censored case where the threshold is an unobserved and not necessarily independent random variable. Maximum likelihood procedures can be employed for joint estimation of both the primary regression equation and the parameters of the distribution of that random threshold.     

Structural estimation of jump-diffusion processes in macroeconomics

This paper shows how to solve and estimate a continuous-time dynamic stochastic general equilibrium (DSGE) model with jumps. It also shows that a continuous-time formulation can make it simpler (relative to its discrete-time version) to compute and estimate the deep parameters using the likelihood function when non-linearities and/or non-normalities are considered. We illustrate our approach by solving and estimating the stochastic AK and the neoclassical growth models. Our Monte Carlo experiments demonstrate that non-normalities can be detected for this class of models. Moreover, we provide strong empirical evidence for jumps in aggregate US data.     

A unified approach to nonparametric trend tests for dependent and independent samples

We consider testing nonparametric hypotheses against ordered alternatives and propose a new unified approach for dependent and independent samples and factorial designs. The new approach allows for arbitrary underlying distributions, including quantitative and discrete ordinal (ordered categorical), or even binary data. It is compared to procedures available in the literature and applied to different data examples. The new method is not only invariant under monotone transformations of the response, but also under monotone transformations of the weights describing the alternative pattern.     

Simultaneous dependence between firm-level stock returns

We show that use of ordinary least-squares to explore relationships involving firm-level stock returns as the dependent variable in the face of structured dependence between individual firms leads to an endogeneity problem. This in turn leads to biased and inconsistent least-squares estimates. A maximum likelihood estimation procedure that will produce consistent estimates in these situations is illustrated. This is done using methods that have been developed to deal with spatial dependence between regional data observations, which can be applied to situations involving firm-level observations that exhibit a structure of dependence. In addition, we show how to correctly interpret maximum likelihood parameter estimates from these models in the context of firm-level dependence, and provide a Monte Carlo as well as applied illustration of the magnitude of bias that can arise.     

Analysis of treatment response data without the joint distribution of potential outcomes

In this paper we show how it is possible to develop a Bayesian framework for analyzing structural models for treatment response data without the joint distribution of the potential outcomes. That this is possible has not been noticed in the literature. We also discuss the computation of the model marginal likelihood and present recipes for finding relevant treatment effects, averaged over both parameters and covariates. As compared to an approach in which the counterfactuals are part of the prior-posterior analysis (as in the work to date), the approach we suggest is simpler in terms of the required prior inputs, computational burden and extensibility to more complex settings.     

Maximum likelihood estimation of limited and discrete dependent variable models with nested random effects

Gauss–Hermite quadrature is often used to evaluate and maximize the likelihood for random component probit models. Unfortunately, the estimates are biased for large cluster sizes and/or intraclass correlations. We show that adaptive quadrature largely overcomes these problems. We then extend the adaptive quadrature approach to general random coefficient models with limited and discrete dependent variables. The models can include several nested random effects (intercepts and coefficients) representing unobserved heterogeneity at different levels of a hierarchical dataset. The required multivariate integrals are evaluated efficiently using spherical quadrature rules. Simulations show that adaptive quadrature performs well in a wide range of situations.     

Bayesian inference for nonlinear structural time series models

We consider efficient methods for likelihood inference applied to structural models. In particular, we introduce a particle filter method which concentrates upon disturbances in the Markov state of the approximating solution to the structural model. A particular feature of such models is that the conditional distribution of interest for the disturbances is often multimodal. We provide a fast and effective method for approximating such distributions. We estimate a neoclassical growth model using this approach. An asset pricing model with persistent habits is also considered. The methodology we employ allows many fewer particles to be used than alternative procedures for a given precision.     

Particle filters for continuous likelihood evaluation and maximisation

In this paper, a method is introduced for approximating the likelihood for the unknown parameters of a state space model. The approximation converges to the true likelihood as the simulation size goes to infinity. In addition, the approximating likelihood is continuous as a function of the unknown parameters under rather general conditions. The approach advocated is fast and robust, and it avoids many of the pitfalls associated with current techniques based upon importance sampling. We assess the performance of the method by considering a linear state space model, comparing the results with the Kalman filter, which delivers the true likelihood. We also apply the method to a non-Gaussian state space model, the stochastic volatility model, finding that the approach is efficient and effective. Applications to continuous time finance models and latent panel data models are considered. Two different multivariate approaches are proposed. The neoclassical growth model is considered as an application.     

Bayesian median autoregression for robust time series forecasting

We develop a Bayesian median autoregressive (BayesMAR) model for time series forecasting. The proposed method utilizes time-varying quantile regression at the median, favorably inheriting the robustness of median regression in contrast to the widely used mean-based methods. Motivated by a working Laplace likelihood approach in Bayesian quantile regression, BayesMAR adopts a parametric model bearing the same structure as autoregressive models by altering the Gaussian error to Laplace, leading to a simple, robust, and interpretable modeling strategy for time series forecasting. We estimate model parameters by Markov chain Monte Carlo. Bayesian model averaging is used to account for model uncertainty, including the uncertainty in the autoregressive order, in addition to a Bayesian model selection approach. The proposed methods are illustrated using simulations and real data applications. An application to U.S. macroeconomic data forecasting shows that BayesMAR leads to favorable and often superior predictive performance compared to the selected mean-based alternatives under various loss functions that encompass both point and probabilistic forecasts. The proposed methods are generic and can be used to complement a rich class of methods that build on autoregressive models.     

Extension of the Wald statistic to models with dependent observations

A generalization of the Wald statistic for testing composite hypotheses is suggested for dependent data from exponential models which include Lévy processes and diffusion fields. The generalized statistic is proved to be asymptotically chi-squared distributed under regular composite hypotheses. It is simpler and more easily available than the generalized likelihood ratio statistic. Simulations in an example where the latter statistic is available show that the generalized Wald test achieves higher average power than the generalized likelihood ratio test. Received: February 29, 2000     

GENERALIZED AUTOREGRESSIVE SCORE MODELS WITH APPLICATIONS

We propose a class of observation‐driven time series models referred to as generalized autoregressive score (GAS) models. The mechanism to update the parameters over time is the scaled score of the likelihood function. This new approach provides a unified and consistent framework for introducing time‐varying parameters in a wide class of nonlinear models. The GAS model encompasses other well‐known models such as the generalized autoregressive conditional heteroskedasticity, autoregressive conditional duration, autoregressive conditional intensity, and Poisson count models with time‐varying mean. In addition, our approach can lead to new formulations of observation‐driven models. We illustrate our framework by introducing new model specifications for time‐varying copula functions and for multivariate point processes with time‐varying parameters. We study the models in detail and provide simulation and empirical evidence. Copyright © 2012 John Wiley & Sons, Ltd.