This thesis consists of two parts: an exploration of new measures of backbench opinion in the UK House of Commons, and an exploration of variance stabilising transformations of count data.

We investigate two problems in modelling time series data that exhibit conditional heteroscedasticity. The first part deals with the local maximum likelihood estimation of volatility functions which are in the form of conditional variance functions. The existing estimation procedures yield plausible results. Yet, they often fail to take into account special features of the data at the cost of reduced accuracy of prediction. More precisely, many of the parametric and nonparametric conditional variance models ignore the fact that the error distribution departs significantly from gaussian distribution. We propose a novel nonparametric estimation procedure that replaces popular local least squares method with local maximum likelihood estimation. Intuitively, using information from the error distribution improves the estimators and therefore increases the accuracy in prediction. This conclusion is proved theoretically and illustrated by numerical examples. In addition, we show that the proposed estimator adapts asymptotically to the error distribution as well as to the mean regression function. Applications with real data examples demonstrate the potential use of the adaptive maximum likelihood estimator in financial risk management. The second part deals with the variable selection for a particular class of semipara-metric models known as the partial linear models. The existing selection methods are computationally demanding. The proposed selection procedure is computationally more efficient. In particular, if P and Q are the number of linear and nonparametric candidate regressors, respectively, then the proposed procedure reduces the order of the number of variable subsets to be investigated from 2 Q+P to 2Q + 2 P. At the same time, it maintains all the good properties of existing methods, such as consistency. The latter is proven theoretically and confirmed numerically by simulated examples. The results are presented for the mean regression function while the generalization to the conditional variance function is discussed separately.

Dans cette thèse, nous proposons une méthode d'analyse de variance pour des données discrètes issues du contrôle de qualité en microbiologie. Nous étudions tout d'abord la méthode d'analyse de variance actuellement utilisée, ses avantages, inconvénients, et limites. Nous proposons une première modélisation du problème par un modèle linéaire à deux facteurs fixes imbriqués. Nous utilisons la méthode d'analyse de déviance pour développer des tests de significativité des facteurs, qui s'avèrent efficaces sur des données d'essais interlaboratoires en microbiologie. Nous présentons ensuite une modélisation à facteurs aléatoires. Le caractère aléatoire des facteurs permet de caractériser la surdispersion des résultats de dénombrement en microbiologie, ce qui constitue l'un des objectifs principaux de ce travail. Le modèle développé correspond à un Modèle Linéaire Généralisé Hiérarchique Gamma-Poisson à trois facteurs aléatoires. Nous proposons alors une méthode d'estimation des effets fixes et aléatoires, ainsi que des paramètres de dispersion associés aux facteurs. Nous présentons des applications pratiques de cette méthode à des données d'essais interlaboratoires en microbiologie, qui prouvent l’ajustement du modèle aux données réelles. Nous proposons également une méthode de test de la significativité des facteurs, ainsi qu'une nouvelle méthode d'évaluation de la performance analytique des laboratoires participants à un essai. Nous présentons enfin une distribution presque-exacte du produit de variables aléatoires indépendantes de loi Gamma Généralisées, permettant d’effectuer des tests de détection de résultats de dénombrement aberrants. / In this thesis, we propose an analysis of variance method for discrete data from quality control in microbiology. To identify the issues of this work, we start by studying the analysis of variance method currently used in microbiology, its benefits, drawbacks, and limits. We propose a first model to respond the problem, corresponding to a linear model with two nested fixed factors. We use the analyse of deviance method to develop significance tests, that proved to be efficient on data sets of proficiency testings in microbiology. We then introduce a new model involving random factors. The randomness of the factors allow to assess and to caracterize the overdispersion observed in results of counts from proficiency testings in microbiology, that is one of the main objectives of this work. The new model corresponds to a Gamma-Poisson Hierarchical Generalized Linear Model with three random factors. We propose a method based on this model to estimate dispersion parameters, fixed, and random effects. We show practical applications of this method to data sets of proficiency testings in microbiology, that prove the goodness of fit of the model to real data. We also develop significance tests of the random factors from this new model, and a new method to assess the performance of the laboratories taking part in a proficiency testing. We finally introduce a near-exact distribution for the product of independent generalized Gamma random variables, in order to characterize the intensity of the Poisson distribution of the model. This approximation, developped from a factorization of the characteristic function, is very precise and can be used to detect outliers.

Η ανάλυση διασποράς μελετά τη σχέση που έχει μια εξαρτημένη μεταβλητή Υ, τις τιμές της οποίας μπορούμε να παρτηρήσουμε, με τις τιμές ενός παράγοντα Α ή πολλών παραγόντων που είναι και οι ανεξάρτητες μεταβλητές. Είναι κατάλληλη για δεδομένα που βασίζονται πάνω στην έρευνα ή σε πειράματα και μπορεί να χρησιμοποιηθεί είτε οι ανεξάρτητες μεταβλητές είναι ποσοτικές είτε ποιοτικές. Στην παρούσα εργασία, θα χρησιμοποιήσουμε ένα μοντέλο ανάλυσης διασποράς για να μελετήσουμε τις επιρροές των ανεξάρτητων μεταβλητών στην εξαρτημένη μεταβλητή χωρίς περιορισμούς στην φύση της στατιστικής σχέσης καθώς επίσης, θα γίνουν έλεγχοι για να μελετήσουμε αν οι πληθυσμοί μας έχουν ίσες μέσες τιμές ή ίσες διασπορές. Στο πρώτο κεφάλαιο της παρούσας εργασίας παρουσιάζονται τα γενικά χαρακτηριστικά μιας μονοπαραγοντικής ανάλυσης διασποράς καθώς επιίσης και οι έλεγχοι των μέσων τιμών για τις στάθμες του παράγοντα. Στο δεύτερο κεφάλαιο ακολουθεί η παρουσίαση της ανάλυσης των επιδράσεων του παράγοντα και οι διάφορες μέθοδοι για πολλαπλές συγκρίσεις. Επίσης, στο τρίτο κεφάλαιο αναλύεται ο σχεδιασμός των δειγματικών μεγεθών καθώς και οι έλεγχοι για ίσες διασπορές. Τέλος, στο τέταρτο κεφάλαιο αναφέρονται τα γενικά χαρακτηριστικά μιας πολυπαραγοντικής ανάλυσης διασποράς. / The analysis of variance studies the relation which has a dependent variable Y, whose the values we observe, with the values of a factor A or more factors that are the independent variables. It is suitable for data that are based on a research or on experiments and it can be used either the independent variables are quantitative or qualitative. In this paper, we are going to use a model of analysis of variance to study the influences of the independent variables to the dependent variable without restrictions to the nature of the statistical relation and they will be controls to study if our populations have equal means or variances, as well. In the first chapter of this paper, they are introduced the general characteristics of a single-factor analysis of variance and the controls of the means for the levels of the factor, as well. In the second chapter, it follows the presentation of the analysis of the factor effects and several methods for multiple comparisons. Furthermore, in the third chapter, it is analysed the planning of sample sizes and the controls for equal variances, as well. In the last chapter, the general charachteristics of a multiple analysis of variance are reported.

- / -

Dans cette thèse nous considérons une population branchante générale où les individus vivent et se reproduisent de manière i.i.d. La durée de vie de chaque individu est distribuée suivant une mesure de probabilité arbitraire et chacun d'eux donne naissance à taux exponentiel. L'arbre décrivant la dynamique de cette population est connu sous le nom de splitting tree. Le processus comptant le nombre d’individus vivant au temps t est connu sous le nom de processus de Crump-Mode-Jagers binaire homogène, et il est connu que ce processus, quand correctement renormalisé, converge presque sûrement en temps long vers une variable aléatoire. Grâce à l'étude du splitting tree sous-jacent à la population via les outils introduit par A. Lambert en 2010, nous montrons un théorème central limite pour cette convergence p.s. dans le cas surcritique. Nous supposons, de plus, que les individus subissent des mutations à taux exponentiel sous l'hypothèse d'infinité d'allèles. Nous nous intéressons alors au spectre de fréquence allélique de la population qui compte la fréquence des tailles de familles dans la population à un instant donnée. Grâce aux méthodes développées dans cette thèse, nous obtenons des résultats d’approximations du spectre de Fréquence. Enfin nous nous intéressons à des questions statistiques sur des arbres de Galton-Watson conditionnés par leurs tailles. Le but est d'estimer la variance de la loi de naissance rendue inaccessible par le conditionnement. On utilise le fait que le processus de contour d'un tel arbre converge vers une excursion Brownienne quand la taille de l'arbre grandit afin de construire des estimateurs de la variance à partir de forêts / In this thesis we consider a general branching population. The lifetimes of the individuals are supposed to be i.i.d. random variables distributed according to an arbitrary distribution. Moreover, each individual gives birth to new individuals at Poisson rate independently from the other individuals. The tree underlying the dynamics of this population is called a splitting tree. The process which count the number of alive individuals at given times is known as binary homogeneous Crump-Mode-Jagers processes. Such processes are known, when properly renormalized, to converge almost surely to some random variable. Thanks to the study of the underlying splitting tree through the tools introduced by A. Lambert in 2010, we show a central limit theorem associated to this a.s. convergence. Moreover, we suppose that individuals undergo mutation at Poisson rate under the infinitely many alleles assumption. We are mainly interested in the so called allelic frequency spectrum which describes the frequency of sizes of families (i.e. sets of individuals carrying the same type) at fixed times. Thanks to the methods developedin this thesis, we are able to get approximation results for the frequency spectrum. In a last part, we study some statistical problems for size constrained Galton-Watson trees. Our goal is to estimate the variance of the birth distribution. Using that the contour process of such tree converges to a Brownian excursion as the size of the tree growth, we construct estimators of the variance of the birth distribution

The aim of this thesis was to investigate the use of effect sizes to report the results of statistical credit rating models in a more practical way. Rating systems in the form of statistical probability models like logistic regression models are used to forecast the behaviour of clients and guide business in rating clients as “high” or “low” risk borrowers. Therefore, model results were reported in terms of statistical significance as well as business language (practical significance), which business experts can understand and interpret. In this thesis, statistical results were expressed as effect sizes like Cohen‟s d that puts the results into standardised and measurable units, which can be reported practically. These effect sizes indicated strength of correlations between variables, contribution of variables to the odds of defaulting, the overall goodness-of-fit of the models and the models‟ discriminating ability between high and low risk customers. / Statistics / M. Sc. (Statistics)

The aim of this thesis was to investigate the use of effect sizes to report the results of statistical credit rating models in a more practical way. Rating systems in the form of statistical probability models like logistic regression models are used to forecast the behaviour of clients and guide business in rating clients as “high” or “low” risk borrowers. Therefore, model results were reported in terms of statistical significance as well as business language (practical significance), which business experts can understand and interpret. In this thesis, statistical results were expressed as effect sizes like Cohen‟s d that puts the results into standardised and measurable units, which can be reported practically. These effect sizes indicated strength of correlations between variables, contribution of variables to the odds of defaulting, the overall goodness-of-fit of the models and the models‟ discriminating ability between high and low risk customers. / Statistics / M. Sc. (Statistics)

Classic parametric statistical tests, like the analysis of variance (ANOVA), are powerful tools used for comparing population means. These tests produce accurate results provided the data satisfies underlying assumptions such as homoscedasticity and balancedness, otherwise biased results are obtained. However, these assumptions are rarely satisfied in real-life. Alternative procedures must be explored. This thesis aims at investigating the impact of heteroscedasticity and unbalancedness on effect sizes in two-way fixed-effects ANOVA models. A real-life dataset, from which three different samples were simulated was used to investigate the changes in effect sizes under the influence of unequal variances and unbalancedness. The parametric bootstrap approach was proposed in case of unequal variances and non-normality. The results obtained indicated that heteroscedasticity significantly inflates effect sizes while unbalancedness has non-significant impact on effect sizes in two-way ANOVA models. However, the impact worsens when the data is both unbalanced and heteroscedastic. / Statistics / M. Sc. (Statistics)