Essays in Statistics

Abstract

This thesis is comprised of several contributions to the field of mathematical statistics, particularly with regards to computational issues of Bayesian statistics and functional data analysis.<br /> The first two chapters are concerned with computational Bayesian approaches that allow one to generate samples from an approximation to the posterior distribution in settings where the likelihood function of some statistical model of interest is unknown. This has led to a class of Approximate Bayesian Computation (ABC) methods whose performance depends on the ability to effectively summarize the information content of the data sample by a lower-dimensional vector of summary statistics. Ideally, these statistics are sufficient for the parameter of interest. However, it is difficult to establish sufficiency in a straightforward way if the likelihood of the model is unavailable. In Chapter 1 we propose an indirect approach to select sufficient summary statistics for ABC methods that borrows its intuition from the indirect estimation literature in econometrics. More precisely, we introduce an auxiliary statistical model that is large enough as to contain the structural model of interest. Summary statistics are then identified in this auxiliary model and mapped to the structural model of interest. We show sufficiency of these statistics for Indirect ABC methods based on parameter estimates (ABC-IP), likelihood functions (ABC-IL) and scores (ABC-IS) of the auxiliary model. A detailed simulation study investigates the performance of each proposal and compares it to a traditional, moment-based ABC approach. Particularly, the ABC-IL and ABC-IS algorithms are shown to perform better than both standard ABC and the ABC-IP methods. In Chapter 2 we extend the notion of Indirect ABC methods by proposing an efficient way of weighting the individual entries of the vector of summary statistics obtained from the score-based Indirect ABC approach (ABC-IS). In particular, the weighting matrix is given by the inverse of the asymptotic covariance matrix of the score vector of the auxiliary model and allows us to appropriately assess the distance between the true posterior distribution and the approximation based on the ABC-IS method. We illustrate the performance gain in a simulation study. An empirical application then implements the weighted ABC-IS method to the problem of estimating a continuous-time stochastic volatility model based on non-Gaussian Ornstein-Uhlenbeck processes. We show how a suitable auxiliary model can be constructed and confirm estimation results from concurring Bayesian estimation approaches suggested in the literature.<br /> In Chapter 3 we consider the problem of sampling from high-dimensional probability distributions that exhibit multiple, well-separated modes. Such distributions arise frequently, for instance, in the Bayesian estimation of macroeconomic DSGE models. Standard Markov Chain Monte Carlo (MCMC) methods, such as the Metropolis-Hastings algorithm, are prone to get trapped in local neighborhoods of the target distribution thus severely limiting the use of these methods in more complex models. We suggest the use of a Sequential Markov Chain Monte Carlo approach to overcome these difficulties and investigate its finite sample properties. The results show that Sequential MCMC methods clearly outperform standard MCMC approaches in a multimodal setting and can recover both the location as well as the mixture weights in a 12-dimensional mixture model. Moreover, we provide a detailed comparison of the effects different choices of tuning parameters have on the approximation to the true sampling distribution. These results can serve as valuable guidelines when applying this method to more complex economic models, such as the (Bayesian) estimation of Dynamic Stochastic General Equilibrium models.<br /> Chapters 4 and 5 study the statistical problem of prediction from a functional perspective. In many statistical applications, data is becoming available at ever increasing frequencies and it has thus become natural to think of discrete observations as realizations of a continuous function, say over the course of one day. However, as functions are generally speaking infinite-dimensional objects, the statistical analysis of such functional data is intrinsically different from standard multivariate techniques. In Chapter 4 we consider prediction in functional additive models of first-order autoregressive type for a time series of functional observations. This is a generalization of functional linear models that are commonly considered in the literature and has two advantages to be applied in a functional time series setting. First, it allows us to introduce a very general notion of time dependencies for functional data in this modeling framework. Particularly, it is rooted at the correlation structure of functional principal component scores and even allows for long memory behavior in the score series across the time dimension. Second, prediction in this modeling framework is straightforwardly implemented as it only concerns conditional means of scalar random variables and we suggest a k-nearest neighbors classification scheme. The theoretical contributions of this paper are twofold. In a first step, we verify the applicability of the functional principal components analysis under our notion of time dependence and obtain precise rates of convergence for the mean function and the covariance operator associated with the observed sample of functions. In a second step, we derive precise rates of convergence of the mean squared error for the proposed predictor, taking into account both the effect of truncating the infinite series expansion at some finite integer L as well as the effect of estimating the covariance operator and associated eigenelements based on a sample of N curves. In Chapter 5 we investigate the performance of functional models in a forecasting study of ground-level ozone-concentration surfaces over the geographical domain of Germany. Our perspective thus differs from the literature on spatially distributed functional processes (which are considered to be (univariate) functions of time that show spatial dependence) in that we consider smooth surfaces defined over some spatial domain that are sampled consecutively over time. In particular, we treat discrete observations that are sampled both over a spatial domain and over time as noisy realizations of some time series of smooth bivariate functions. In a first step we therefore discuss how smooth functions can be reconstructed from such noisy measurements through a finite element spline smoother that is defined over some triangulation of the spatial domain. In a second step we consider two forecasting approaches to functional time series. The first one is a functional linear model of first-order auto-regressive type, whereas the second considers the non-parametric extension to functional additive models discussed in Chapter 4. Both approaches are applied to predicting ground-level ozone concentration measured over the spatial domain of Germany and are shown to yield similar predictions.