Les méthodes de Monte Carlo par chaines de Markov MCMC sont des outils mathématiques utilisés pour simuler des mesures de probabilités π définies sur des espaces de grandes dimensions. Une des questions les plus importantes dans ce contexte est de savoir à quelle vitesse converge la chaine de Markov P vers la mesure invariante π. Pour mesurer la vitesse de convergence de la chaine de Markov P vers sa mesure invariante π nous utilisons la distance de la variation totale. Il est bien connu que la vitesse de convergence d’une chaine de Markov réversible P dépend de la deuxième plus grande valeur propre en valeur absolue de la matrice P notée β!. Une partie importante dans l’estimation de β! consiste à estimer la deuxième plus grande valeur propre de la matrice P, qui est notée β1. Diaconis et Stroock (1991) ont introduit une méthode basée sur l’inégalité de Poincaré pour estimer β1 pour le cas général des chaines de Markov réversibles avec un nombre fini d'état. Dans cette thèse, nous utilisons la méthode de Shiu et Chen (2015) pour étudier le cas de l'algorithme de l'échantillonneur de Gibbs pour le modèle d'Ising unidimensionnel avec trois états ou plus appelé aussi modèle de Potts. Puis, nous généralisons le résultat de Shiu et Chen au cas du modèle d’Ising deux- dimensionnel avec deux états. Les résultats obtenus minorent ceux introduits par Ingrassia (1994). Puis nous avons pensé à perturber l'échantillonneur de Gibbs afin d’améliorer sa vitesse de convergence vers l'équilibre. / Monte Carlo Markov chain methods MCMC are mathematical tools used to simulate probability measures π defined on state spaces of high dimensions. The speed of convergence of this Markov chain X to its invariant state π is a natural question to study in this context.To measure the convergence rate of a Markov chain we use the total variation distance. It is well known that the convergence rate of a reversible Markov chain depends on its second largest eigenvalue in absolute value denoted by β!. An important part in the estimation of β! is the estimation of the second largest eigenvalue which is denoted by β1.Diaconis and Stroock (1991) introduced a method based on Poincaré inequality to obtain a bound for β1 for general finite state reversible Markov chains.In this thesis we use the Chen and Shiu approach to study the case of the Gibbs sampler for the 1−D Ising model with three and more states which is also called Potts model. Then, we generalize the result of Shiu and Chen (2015) to the case of the 2−D Ising model with two states.The results we obtain improve the ones obtained by Ingrassia (1994). Then, we introduce some method to disrupt the Gibbs sampler in order to improve its convergence rate to equilibrium.

Uncertainty is pervasive in risk assessment. In ecotoxicological risk assessments, it arises from such sources as a lack of data, the simplification and abstraction of complex situations, and ambiguities in assessment endpoints (Burgman 2005; Suter 1993). When evaluating and managing risks, uncertainty needs to be explicitly considered in order to avoid erroneous decisions and to be able to make statements about the confidence that we can place in risk estimates. Although informative, previous approaches to dealing with uncertainty in ecotoxicological modelling have been found to be limited, inconsistent and often based on assumptions that may be false (Ferson & Ginzburg 1996; Suter 1998; Suter et al. 2002; van der Hoeven 2004; van Straalen 2002a; Verdonck et al. 2003a). In this thesis a Generalised Linear Modelling approach is proposed as an alternative, congruous framework for the analysis and prediction of a wide range of ecotoxicological effects. This approach was used to investigate the results of toxicity experiments on the effect of 2,4-Dichlorophenoxyacetic Acid (2,4-D) formulations and 4-Chlorophenol (4-CP, an associated breakdown product) on Daphnia carinata. Differences between frequentist Maximum Likelihood (ML) and Bayesian Markov-Chain Monte-Carlo (MCMC) approaches to statistical reasoning and model estimation were also investigated. These approaches are inferentially disparate and place different emphasis on aleatory and epistemic uncertainty (O'Hagan 2004). Bayesian MCMC and Probability Bounds Analysis methods for propagating uncertainty in risk models are also compared for the first time. For simple models, Bayesian and frequentist approaches to Generalised Linear Model (GLM) estimation were found to produce very similar results when non-informative prior distributions were used for the Bayesian models. Potency estimates and regression parameters were found to be similar for identical models, signifying that Bayesian MCMC techniques are at least a suitable and objective replacement for frequentist ML for the analysis of exposureresponse data. Applications of these techniques demonstrated that Amicide formulations of 2,4-D are more toxic to Daphnia than their unformulated, Technical Acid parent. Different results were obtained from Bayesian MCMC and ML methods when more complex models and data structures were considered. In the analysis of 4-CP toxicity, the treatment of 2 different factors as fixed or random in standard and Mixed-Effect models was found to affect variance estimates to the degree that different conclusions would be drawn from the same model, fit to the same data. Associated discrepancies in the treatment of overdispersion between ML and Bayesian MCMC analyses were also found to affect results. Bayesian MCMC techniques were found to be superior to the ML ones employed for the analysis of complex models because they enabled the correct formulation of hierarchical (nested) datastructures within a binomial logistic GLM. Application of these techniques to the analysis of results from 4-CP toxicity testing on two strains of Daphnia carinata found that between-experiment variability was greater than that within-experiments or between-strains. Perhaps surprisingly, this indicated that long-term laboratory culture had not significantly affected the sensitivity of one strain when compared to cultures of another strain that had recently been established from field populations. The results from this analysis highlighted the need for repetition of experiments, proper model formulation in complex analyses and careful consideration of the effects of pooling data on characterising variability and uncertainty. The GLM framework was used to develop three dimensional surface models of the effects of different length pulse exposures, and subsequent delayed toxicity, of 4-CP on Daphnia. These models described the relationship between exposure duration and intensity (concentration) on toxicity, and were constructed for both pulse and delayed effects. Statistical analysis of these models found that significant delayed effects occurred following the full range of pulse exposure durations, and that both exposure duration and intensity interacted significantly and concurrently with the delayed effect. These results indicated that failure to consider delayed toxicity could lead to significant underestimation of the effects of pulse exposure, and therefore increase uncertainty in risk assessments. A number of new approaches to modelling ecotoxicological risk and to propagating uncertainty were also developed and applied in this thesis. In the first of these, a method for describing and propagating uncertainty in conventional Species Sensitivity Distribution (SSD) models was described. This utilised Probability Bounds Analysis to construct a nonparametric 'probability box' on an SSD based on EC05 estimates and their confidence intervals. Predictions from this uncertain SSD and the confidence interval extrapolation methods described by Aldenberg and colleagues (2000; 2002a) were compared. It was found that the extrapolation techniques underestimated the width of uncertainty (confidence) intervals by 63% and the upper bound by 65%, when compared to the Probability Bounds (P3 Bounds) approach, which was based on actual confidence estimates derived from the original data. An alternative approach to formulating ecotoxicological risk modelling was also proposed and was based on a Binomial GLM. In this formulation, the model is first fit to the available data in order to derive mean and uncertainty estimates for the parameters. This 'uncertain' GLM model is then used to predict the risk of effect from possible or observed exposure distributions. This risk is described as a whole distribution, with a central tendency and uncertainty bounds derived from the original data and the exposure distribution (if this is also 'uncertain'). Bayesian and P-Bounds approaches to propagating uncertainty in this model were compared using an example of the risk of exposure to a hypothetical (uncertain) distribution of 4-CP for the two Daphnia strains studied. This comparison found that the Bayesian and P-Bounds approaches produced very similar mean and uncertainty estimates, with the P-bounds intervals always being wider than the Bayesian ones. This difference is due to the different methods for dealing with dependencies between model parameters by the two approaches, and is confirmation that the P-bounds approach is better suited to situations where data and knowledge are scarce. The advantages of the Bayesian risk assessment and uncertainty propagation method developed are that it allows calculation of the likelihood of any effect occurring, not just the (probability)bounds, and that the same software (WinBugs) and model construction may be used to fit regression models and predict risks simultaneously. The GLM risk modelling approaches developed here are able to explain a wide range of response shapes (including hormesis) and underlying (non-normal) distributions, and do not involve expression of the exposure-response as a probability distribution, hence solving a number of problems found with previous formulations of ecotoxicological risk. The approaches developed can also be easily extended to describe communities, include modifying factors, mixed-effects, population growth, carrying capacity and a range of other variables of interest in ecotoxicological risk assessments. While the lack of data on the toxicological effects of chemicals is the most significant source of uncertainty in ecotoxicological risk assessments today, methods such as those described here can assist by quantifying that uncertainty so that it can be communicated to stakeholders and decision makers. As new information becomes available, these techniques can be used to develop more complex models that will help to bridge the gap between the bioassay and the ecosystem.

本論文主要討論貝氏漸近的比較，推導出參數的聯合後驗分配與利用圖形來診斷馬可夫鏈蒙地卡羅的收斂。Johnson (1970)利用泰勒展開式得到個別後驗分配的展開式，此展開式是根據概似函數與先驗分配。 Weng (2010b) 和 Weng and Hsu (2011) 利用 Stein’s 等式且由概似函數與先驗分配估計後驗動差；將這些後驗動差代入Edgeworth 展開式得到近似後驗分配，此近似分配的誤差可精確到大O的負3/2次方與Johnson’s 相同。另外Weng and Hsu (2011)發現Weng (2010b) 和Johnson (1970)的近似展開式各別項誤差到大O的負1次方不一致，由模擬結果得到Weng’s 在此項表現比Johnson’s 好。另外由Weng (2010b)得到一維參數 的Edgeworth 近似後驗分配延伸到二維參數的聯合後驗分配；並應用二維參數的聯合後驗分配於多階段資料。本論文我們提出利用圖形來診斷馬可夫鏈蒙地卡羅收斂的方法，並且應用一般化線性模型與混合常態模型做為模擬。 關鍵字: Edgeworth 展開式；馬可夫鏈蒙地卡羅；個別後驗分配；Stein’s 等式 / Johnson (1970) obtained expansions for marginal posterior distributions through Taylor expansions. The expansion in Johnson (1970) is expressed in terms of the likelihood and the prior. Weng (2010b) and Weng and Hsu (2011) showed that by using Stein's identity we can approximate the posterior moments in terms of the likelihood and the prior; then substituting these approximations into an Edgeworth series one can obtain an expansion which is correct to O(t{-3/2}), similar to Johnson's. Weng and Hsu (2011) found that the O(t{-1}) terms in Weng (2010b) and Johnson (1970) do not agree and further compared these two expansions by simulation study. The simulations confirmed this finding and revealed that our O(t{-1}) term gives better performance than Johnson's. In addition to the comparison of Bayesian asymptotics, we try to extend Weng (2010a)'s Edgeworth series for the distribution of a single parameter to the joint distribution of all parameters. Since the calculation is quite complicated, we only derive expansions for the two-parameter case and apply it to the experiment of multi-stage data. Markov Chain Monte Carlo (MCMC) is a popular method for making Bayesian inference. However, convergence of the chain is always an issue. Most of convergence diagnosis in the literature is based solely on the simulation output. In this dissertation, we proposed a graphical method for convergence diagnosis of the MCMC sequence. We used some generalized linear models and mixture normal models for simulation study. In summary, the goals of this dissertation are threefold: to compare some results in Bayesian asymptotics, to study the expansion for the joint distribution of the parameters and its applications, and to propose a method for convergence diagnosis of the MCMC sequence. Key words: Edgeworth expansion; Markov Chain Monte Carlo; marginal posterior distribution; Stein's identity.

On s'intéresse à l'ajout d'incertitudes supplémentaires dans les modèles de Markov cachés. L'inférence est réalisée dans un cadre bayésien à l'aide des méthodes de Monte Carlo. Dans un cadre multicapteur, on suppose que chaque capteur peut commuter entre plusieurs états de fonctionnement. Un modèle à saut original est développé et des algorithmes de Monte Carlo efficaces sont présentés pour différents types de situations, prenant en compte des données synchrones/asynchrones et le cas binaire capteur valide/défaillant. Le modèle/algorithme développé est appliqué à la localisation d'un véhicule terrestre équipé de trois capteurs, dont un récepteur GPS, potentiellement défaillant à cause de phénomènes de trajets multiples. <br />On s'intéresse ensuite à l'estimation de la densité de probabilité des bruits d'évolution et de mesure dans les modèles de Markov cachés, à l'aide des mélanges de processus de Dirichlet. Le cas de modèles linéaires est tout d'abord étudié, et des algorithmes MCMC et de filtrage particulaire sont développés. Ces algorithmes sont testés sur trois applications différentes. Puis le cas de l'estimation des densités de probabilité des bruits dans les modèles non linéaires est étudié. On définit pour cela des processus de Dirichlet variant temporellement, permettant l'estimation en ligne d'une densité de probabilité non stationnaire.

信用衍生性商品（credit derivatives）是用於移轉信用風險之契約，契約是由保護買方（protection buyer）與保護賣方（protection seller）所簽定，由保護買方支付保險金（可為躉繳或分期支付）以獲得信用的保護，而保護賣方則需在律定之信用事件發生時支付償金予保護買方做為補償。近年來頻傳金融事件，巴塞爾銀行監理委員會（Basel Committee on Banking Supervision）也不得不制定新版的巴塞爾協定以要求銀行強化信用風險控制與分散，而信用衍生性商品亦有助於信用風險的移轉與抵減的功能。 本篇針對利用conditional Markov chain來建構信用違約交換與第n次信用違約交換之評價模型，並利用模擬的方式來求算出各商品之利差。藉由現實中的資料取得參數的估計值放入模型內則可以模擬出各種不同的狀況，進而做出避險的策略。 此外，本篇亦探討如何利用Gibbs sampler來改良conditional Markov chain的模擬方法，以模擬當信用衍生性商品中的資產組合有傳染效果的情況。

本文採用 Chib， Nardari， 與 Shephard(2006) 的多變量動態因子隨機波動模型(MSV)， 來探討美、日、台三國的資訊、電腦類股股價報酬率波動的共同行為。 我們將模型中的因子解釋為產業的前景或信心，並藉由模擬的方式描繪出其樣貌，進而希望了解產業景氣循環在股價的波動行為中扮演什麼角色。 研究財務市場間的關聯性一值是一項重要的課題，也發展出各種的模型來描述既有的現象。 MSV 模型將看不到的解釋變量數量化，並將變數的波動行為切割為可由因子所解釋與不能解釋的部分。 且藉由將觀察值的誤差項以及單一因子的波動行為設定為隨機波動，放寬共變數變異數矩陣為定值的假設，讓每一時點都能依時變動，在同類的模型中對資料的設定是較少的。 在實證分析中我們有幾點發現：1. 因子能夠解釋資產間的波動行為，其反映在扣除因子波動之後的自有波動，其波動水準值的降低。 2. 在股價波動劇烈期間，因子解釋能力提高。 3. 因子的解釋能力在不同的國家中差異幅度很大，日本有超過一半的波動可以為因子的波動所解釋，而因子在台灣股價的波動行為只有兩成左右的解釋能力。

Dans ce mémoire, nous cherchons à modéliser des tables à deux entrées monotones en lignes et/ou en colonnes, pour une éventuelle application sur les tables de mortalité. Nous adoptons une approche bayésienne non paramétrique et représentons la forme fonctionnelle des données par splines bidimensionnelles. L’objectif consiste à condenser une table de mortalité, c’est-à-dire de réduire l’espace d’entreposage de la table en minimisant la perte d’information. De même, nous désirons étudier le temps nécessaire pour reconstituer la table. L’approximation doit conserver les mêmes propriétés que la table de référence, en particulier la monotonie des données. Nous travaillons avec une base de fonctions splines monotones afin d’imposer plus facilement la monotonie au modèle. En effet, la structure flexible des splines et leurs dérivées faciles à manipuler favorisent l’imposition de contraintes sur le modèle désiré. Après un rappel sur la modélisation unidimensionnelle de fonctions monotones, nous généralisons l’approche au cas bidimensionnel. Nous décrivons l’intégration des contraintes de monotonie dans le modèle a priori sous l’approche hiérarchique bayésienne. Ensuite, nous indiquons comment obtenir un estimateur a posteriori à l’aide des méthodes de Monte Carlo par chaînes de Markov. Finalement, nous étudions le comportement de notre estimateur en modélisant une table de la loi normale ainsi qu’une table t de distribution de Student. L’estimation de nos données d’intérêt, soit la table de mortalité, s’ensuit afin d’évaluer l’amélioration de leur accessibilité. / This master’s thesis is about the estimation of bivariate tables which are monotone within the rows and/or the columns, with a special interest in the approximation of life tables. This problem is approached through a nonparametric Bayesian regression model, in particular linear combinations of regression splines. By condensing a life table, our goal is to reduce its storage space without losing the entries’ accuracy. We will also study the reconstruction time of the table with our estimators. The properties of the reference table, specifically its monotonicity, must be preserved in the estimation. We are working with a monotone spline basis since splines are flexible and their derivatives can easily be manipulated. Those properties enable the imposition of constraints of monotonicity on our model. A brief review on univariate approximations of monotone functions is then extended to bivariate estimations. We use hierarchical Bayesian modeling to include the constraints in the prior distributions. We then explain the Markov chain Monte Carlo algorithm to obtain a posterior estimator. Finally, we study the estimator’s behaviour by applying our model on the Standard Normal table and the Student’s t table. We estimate our data of interest, the life table, to establish the improvement in data accessibility.

This dissertation is devoted to studying a fast and analytic approximation method, called the variational Bayesian (VB) method, and aims to give insight into its general applicability and usefulness, and explore its applications to various real-world problems. This work has three main foci: 1) The general applicability and properties; 2) Diagnostics for VB approximations; 3) Variational applications. Generally, the variational inference has been developed in the context of the exponential family, which is open to further development. First, it usually consider the cases in the context of the conjugate exponential family. Second, the variational inferences are developed only with respect to natural parameters, which are often not the parameters of immediate interest. Moreover, the full factorization, which assumes all terms to be independent of one another, is the most commonly used scheme in the most of the variational applications. We show that VB inferences can be extended to a more general situation. We propose a special parameterization for a parametric family, and also propose a factorization scheme with a more general dependency structure than is traditional in VB. Based on these new frameworks, we develop a variational formalism, in which VB has a fast implementation, and not be limited to the conjugate exponential setting. We also investigate its local convergence property, the effects of choosing different priors, and the effects of choosing different factorization scheme. The essence of the VB method relies on making simplifying assumptions about the posterior dependence of a problem. By definition, the general posterior dependence structure is distorted. In addition, in the various applications, we observe that the posterior variances are often underestimated. We aim to develop diagnostics test to assess VB approximations, and these methods are expected to be quick and easy to use, and to require no sophisticated tuning expertise. We propose three methods to compute the actual posterior covariance matrix by only using the knowledge obtained from VB approximations: 1) To look at the joint posterior distribution and attempt to find an optimal affine transformation that links the VB and true posteriors; 2) Based on a marginal posterior density approximation to work in specific low dimensional directions to estimate true posterior variances and correlations; 3) Based on a stepwise conditional approach, to construct and solve a set of system of equations that lead to estimates of the true posterior variances and correlations. A key computation in the above methods is to calculate a uni-variate marginal or conditional variance. We propose a novel way, called the VB Adjusted Independent Metropolis-Hastings (VBAIMH) method, to compute these quantities. It uses an independent Metropolis-Hastings (IMH) algorithm with proposal distributions configured by VB approximations. The variance of the target distribution is obtained by monitoring the acceptance rate of the generated chain. One major question associated with the VB method is how well the approximations can work. We particularly study the mean structure approximations, and show how it is possible using VB approximations to approach model selection tasks such as determining the dimensionality of a model, or variable selection. We also consider the variational application in Bayesian nonparametric modeling, especially for the Dirichlet process (DP). The posterior inference for DP has been extensively studied in the context of MCMC methods. This work presents a a full variational solution for DP with non-conjugate settings. Our solution uses a truncated stick-breaking representation. We propose an empirical method to determine the number of distinct components in a finite dimensional DP. The posterior predictive distribution for DP is often not available in a closed form. We show how to use the variational techniques to approximate this quantity. As a concrete application study, we work through the VB method on regime-switching lognormal models and present solutions to quantify both the uncertainty in the parameters and model specification. Through a series numerical comparison studies with likelihood based methods and MCMC methods on the simulated and real data sets, we show that the VB method can recover exactly the model structure, gives the reasonable point estimates, and is very computationally efficient.

The New Macroeconometrics may succinctly be described as the application of Bayesian analysis to the class of macroeconomic models called Dynamic Stochastic General Equilibrium (DSGE) models. A prominent local example from this research area is the development and estimation of the RAMSES model, the main macroeconomic model in use at Sveriges Riksbank.   Bayesian estimation of DSGE models is often computationally demanding. In this thesis fast algorithms for Bayesian inference are developed and tested in the context of the state space model framework implied by DSGE models. The algorithms discussed in the thesis deal with evaluation of the DSGE model likelihood function and sampling from the posterior distribution. Block Kalman filter algorithms are suggested for likelihood evaluation in large linearised DSGE models. Parallel particle filter algorithms are presented for likelihood evaluation in nonlinearly approximated DSGE models. Prefetching random walk Metropolis algorithms and adaptive hybrid sampling algorithms are suggested for posterior sampling. The generality of the algorithms, however, suggest that they should be of interest also outside the realm of macroeconometrics.

Disertacijoje nagrinėjami Markovo grandinės Monte-Karlo (MCMC) adaptavimo metodai, skirti efektyviems skaitiniams duomenų analizės sprendimų priėmimo su iš anksto nustatytu patikimumu algoritmams sudaryti. Suformuluoti ir išspręsti hierarchiniu būdu sudarytų daugiamačių skirstinių (asimetrinio t skirstinio, Puasono-Gauso modelio, stabiliojo simetrinio vektoriaus dėsnio) parametrų vertinimo uždaviniai. Adaptuotai MCMC procedūrai sukurti yra pritaikytas nuoseklaus Monte-Karlo imčių generavimo metodas, įvedant statistinį stabdymo kriterijų ir imties tūrio reguliavimą. Statistiniai uždaviniai išspręsti šiuo metodu leidžia atskleisti aktualias MCMC metodų skaitmeninimo problemų ypatybes. MCMC algoritmų efektyvumas tiriamas pasinaudojant disertacijoje sudarytu statistinio modeliavimo metodu. Atlikti eksperimentai su sportininkų duomenimis ir sveikatos industrijai priklausančių įmonių finansiniais duomenimis patvirtino, kad metodo skaitinės savybės atitinka teorinį modelį. Taip pat sukurti metodai ir algoritmai pritaikyti sociologinių duomenų analizės modeliui sudaryti. Atlikti tyrimai parodė, kad adaptuotas MCMC algoritmas leidžia gauti nagrinėjamų skirstinių parametrų įvertinius per mažesnį grandžių skaičių ir maždaug du kartus sumažinti skaičiavimų apimtį. Disertacijoje sukonstruoti algoritmai gali būti pritaikyti stochastinio pobūdžio sistemų tyrimui ir kitiems statistikos uždaviniams spręsti MCMC metodu. / Markov chain Monte Carlo adaptive methods by creating computationally effective algorithms for decision-making of data analysis with the given accuracy are analyzed in this dissertation. The tasks for estimation of parameters of the multivariate distributions which are constructed in hierarchical way (skew t distribution, Poisson-Gaussian model, stable symmetric vector law) are described and solved in this research. To create the adaptive MCMC procedure, the sequential generating method is applied for Monte Carlo samples, introducing rules for statistical termination and for sample size regulation of Markov chains. Statistical tasks, solved by this method, reveal characteristics of relevant computational problems including MCMC method. Effectiveness of the MCMC algorithms is analyzed by statistical modeling method, constructed in the dissertation. Tests made with sportsmen data and financial data of enterprises, belonging to health-care industry, confirmed that numerical properties of the method correspond to the theoretical model. The methods and algorithms created also are applied to construct the model for sociological data analysis. Tests of algorithms have shown that adaptive MCMC algorithm allows to obtain estimators of examined distribution parameters in lower number of chains, and reducing the volume of calculations approximately two times. The algorithms created in this dissertation can be used to test the systems of stochastic type and to solve other statistical... [to full text]