Les logiciels utilisés sont Splus et R. / Résumé La technique connue comme l'imputation multiple semble être la technique la plus appropriée pour résoudre le problème de non-réponse. La littérature mentionne des méthodes qui modélisent la nature et la structure des valeurs manquantes. Une des méthodes les plus populaires est l'algorithme « Pan » de (Schafer & Yucel, 2002). Les imputations rapportées par cette méthode sont basées sur un modèle linéaire multivarié à effets mixtes pour la variable réponse. La méthode « BHLC » de (Murua et al, 2005) est une extension de « Pan » dont le modèle est bayésien hiérarchique avec groupes. Le but principal de ce travail est d'étudier le problème de sélection du modèle pour l'imputation multiple en termes d'efficacité et d'exactitude des prédictions des valeurs manquantes. Nous proposons une mesure de performance liée à la prédiction des valeurs manquantes. La mesure est une erreur quadratique moyenne reflétant la variance associée aux imputations multiples et le biais de prédiction. Nous montrons que cette mesure est plus objective que la mesure de variance de Rubin. Notre mesure est calculée en augmentant par une faible proportion le nombre de valeurs manquantes dans les données. La performance du modèle d'imputation est alors évaluée par l'erreur de prédiction associée aux valeurs manquantes. Pour étudier le problème objectivement, nous avons effectué plusieurs simulations. Les données ont été produites selon des modèles explicites différents avec des hypothèses particulières sur la structure des erreurs et la distribution a priori des valeurs manquantes. Notre étude examine si la vraie structure d'erreur des données a un effet sur la performance du choix des différentes hypothèses formulées pour le modèle d'imputation. Nous avons conclu que la réponse est oui. De plus, le choix de la distribution des valeurs manquantes semble être le facteur le plus important pour l'exactitude des prédictions. En général, les choix les plus efficaces pour de bonnes imputations sont une distribution de student avec inégalité des variances dans les groupes pour la structure des erreurs et une loi a priori choisie pour les valeurs manquantes est la loi normale avec moyenne et variance empirique des données observées, ou celle régularisé avec grande variabilité. Finalement, nous avons appliqué nos idées à un cas réel traitant un problème de santé. Mots clés : valeurs manquantes, imputations multiples, modèle linéaire bayésien hiérarchique, modèle à effets mixtes. / Abstract The technique known as multiple imputation seems to be the most suitable technique for solving the problem of non-response. The literature mentions methods that models the nature and structure of missing values. One of the most popular methods is the PAN algorithm of Schafer and Yucel (2002). The imputations yielded by this method are based on a multivariate linear mixed-effects model for the response variable. A Bayesian hierarchical clustered and more flexible extension of PAN is given by the BHLC model of Murua et al. (2005). The main goal of this work is to study the problem of model selection for multiple imputation in terms of efficiency and accuracy of missing-value predictions. We propose a measure of performance linked to the prediction of missing values. The measure is a mean squared error, and hence in addition to the variance associated to the multiple imputations, it includes a measure of bias in the prediction. We show that this measure is more objective than the most common variance measure of Rubin. Our measure is computed by incrementing by a small proportion the number of missing values in the data and supposing that those values are also missing. The performance of the imputation model is then assessed through the prediction error associated to these pseudo missing values. In order to study the problem objectively, we have devised several simulations. Data were generated according to different explicit models that assumed particular error structures. Several missing-value prior distributions as well as error-term distributions are then hypothesized. Our study investigates if the true error structure of the data has an effect on the performance of the different hypothesized choices for the imputation model. We concluded that the answer is yes. Moreover, the choice of missing-value prior distribution seems to be the most important factor for accuracy of predictions. In general, the most effective choices for good imputations are a t-Student distribution with different cluster variances for the error-term, and a missing-value Normal prior with data-driven mean and variance, or a missing-value regularizing Normal prior with large variance (a ridge-regression-like prior). Finally, we have applied our ideas to a real problem dealing with health outcome observations associated to a large number of countries around the world. Keywords: Missing values, multiple imputation, Bayesian hierarchical linear model, mixed effects model.

Treatment of patients with multiple sclerosis (MS) and development of new therapies have been challenging due to the disease complexity and slow progression, and the limited sensitivity of available clinical outcomes. Modeling and simulation has become an increasingly important component in drug development and in post-marketing optimization of use of medication. This thesis focuses on development of pharmacometric models for characterization and quantification of the relationships between drug exposure, biomarkers and clinical endpoints in relapse-remitting MS (RRMS) following cladribine treatment. A population pharmacokinetic model of cladribine and its main metabolite, 2-chloroadenine, was developed using plasma and urine data. The renal clearance of cladribine was close to half of total elimination, and was found to be a linear function of creatinine clearance (CRCL). Exposure-response models could quantify a clear effect of cladribine tablets on absolute lymphocyte count (ALC), burden of disease (BoD), expanded disability status scale (EDSS) and relapse rate (RR) endpoints. Moreover, they gave insight into disease progression of RRMS. This thesis further demonstrates how integrated modeling framework allows an understanding of the interplay between ALC and clinical efficacy endpoints. ALC was found to be a promising predictor of RR. Moreover, ALC and BoD were identified as predictors of EDSS time-course. This enables the understanding of the behavior of the key outcomes necessary for the successful development of long-awaited MS therapies, as well as how these outcomes correlate with each other. The item response theory (IRT) methodology, an alternative approach for analysing composite scores, enabled to quantify the information content of the individual EDSS components, which could help improve this scale. In addition, IRT also proved capable of increasing the detection power of potential drug effects in clinical trials, which may enhance drug development efficiency. The developed nonlinear mixed-effects models offer a platform for the quantitative understanding of the biomarker(s)/clinical endpoint relationship, disease progression and therapeutic response in RRMS by integrating a significant amount of knowledge and data.

Les traitements existants des métastases hépatiques du cancer colorectal montrent une efficacité insuffisante. Le projet GR5FU visait à améliorer cette efficacité et consistait à délivrer le 5-fluorouracile (5FU) dans le foie via son encapsulation dans des globules rouges (GR). Dans ce contexte, la modélisation visait à prédire la quantité de 5FU à encapsuler dans les GR pour atteindre une efficacité équivalente à celle du 5FU standard. Dans cette thèse, nous avons construit et implémenté un modèle mathématique multi-échelle qui relie l’injection du 5FU à son efficacité sur la croissance tumorale en intégrant sa pharmacocinétique et son mécanisme d’action intracellulaire. Des simulations de population de ce modèle, s’appuyant sur des paramètres de la littérature, nous ont permis (i) de reproduire des résultats cliniques montrant le pouvoir prédictif de l’enzyme Thymidylate Synthase (TS) et (ii) d’identifier deux prédicteurs potentiels de la réponse au 5FU à l’échelle d’une population virtuelle, en complément du niveau de TS : la vitesse de croissance tumorale et le métabolisme intracellulaire des pyrimidines. Nous avons également analysé, à l’aide de modèles à effets mixtes, (i) la croissance in vivo de la tumeur intra-hépatique VX2 sans traitement, tenant lieu de modèle animal de métastase hépatique, et (ii) la distribution plasmatique et hépatique du 5FU chez l’animal. Cette modélisation statistique nous a permis d’identifier les modèles décrivant des données expérimentales, d’estimer les paramètres de ces modèles et leur variabilité, et de générer une meilleure connaissance de la croissance de la tumeur VX2 et de la pharmacocinétique animale du 5FU, en particulier hépatique. Dans cette thèse, nous avons illustré comment l’intégration du métabolisme d’un médicament et de son mécanisme d’action dans un modèle global et la simulation de ce modèle à l’échelle d’une population virtuelle, constituent une approche prometteuse pour optimiser le développement d’hypothèses thérapeutiques innovantes en collaboration avec des expérimentateurs. / Existing treatments for liver metastases of colorectal cancer show a lack of efficacy. In order to improve the prognosis of patients, the GR5FU project has been implemented. It consisted in delivering the drug 5-fluorouracil (5FU) in the liver via its encapsulation in red blood cells (RBC) to increase its efficacy / toxicity ratio. In this context, the modeling aimed at predicting the amount of 5FU to encapsulate in RBC to achieve an efficacy equivalent to standard 5FU. In this thesis, we have created and implemented a multiscale mathematical model that links the injection of 5FU to its efficacy on tumor growth by integrating its pharmacokinetics and mechanism of intracellular action. Population simulations of this model, using parameters from the literature, allowed us (i) to reproduce clinical results showing the predictive power of TS enzyme level and (ii) to identify two potential predictors of response to 5FU at the level of a population of virtual patients, in addition to TS level. We also analyzed, using mixed effects models, (i) the in vivo growth of intrahepatic VX2 tumor without treatment, serving as an animal model of liver metastasis, and (ii) the distribution of 5FU in the animal’s organism. This statistical modelization enabled us to identify the models describing experimental data, to estimate the parameters of these models and their variability, and generate a better knowledge of VX2 tumor growth and animal 5FU pharmacokinetics. In this thesis, we illustrated how the integration of drug metabolism and its mechanism of action in a global model and the simulation of this model at the scale of a virtual population, form a promising approach to optimize the development of innovative therapeutic hypotheses in collaboration with experimentalists.

La première partie de cette thèse est consacrée a l'estimation par maximum de vraisemblance dans les modèles mixtes a dynamique markovienne. Nous considérons plus précisément des modèles de Markov cachés a effets mixtes et des modèles de diffusion à effets mixtes. Dans le Chapitre 2, nous combinons l'algorithme de Baum-Welch a l'algorithme SAEM pour estimer les paramètres de population dans les modèles de Markov cachés à effets mixtes. Nous proposons également des procédures spéciques pour estimer les paramètres individuels et les séquences d'états cachés. Nous étudions les propriétés de cette nouvelle méthodologie sur des données simulées et l'appliquons sur des données réelles de nombres de crises d'épilepsie. Dans le Chapitre 3, nous proposons d'abord des modèles de diffusion à effets mixtes pour la pharmacocinétique de population. Nous en estimons les paramètres en combinant l'algorithme SAEM a un filtre de Kalman étendu. Nous étudions ensuite les propriétés asymptotiques de l'estimateur du maximum de vraisemblance dans des modèles de diffusion observés sans bruit de mesure continûment sur un intervalle de temps fixé lorsque le nombre de sujets tend vers l'infini. Le Chapitre 4 est consacré à la sélection de covariables dans des modèles mixtes généraux. Nous proposons une version du BIC adaptée au contexte de double asymptotique ou le nombre de sujets et le nombre d'observations par sujet tendent vers l'infini. Nous présentons quelques simulations pour illustrer cette procédure. / The first part of this thesis deals with maximum likelihood estimation in Markovianmixed-effects models. More precisely, we consider mixed-effects hidden Markov models and mixed-effects diffusion models. In Chapter 2, we combine the Baum-Welch algorithm and the SAEM algorithm to estimate the population parameters in mixed-effects hidden Markov models. We also propose some specific procedures to estimate the individual parameters and the sequences of hidden states. We study the properties of the proposed methodologies on simulated datasets and we present an application to real daily seizure count data. In Chapter 3, we first suggest mixed-effects diffusion models for population pharmacokinetics. We estimate the parameters of these models by combining the SAEM algorithm with the extended Kalman filter. Then, we study the asymptotic properties of the maximum likelihood estimatein some mixed-effects diffusion models continuously observed on a fixed time interval when the number of subjects tends to infinity. Chapter 4 is dedicated to variable selection in general mixed-effects models. We propose a BIC adapted to the asymptotic context where both of the number of subjects and the number of observations per subject tend to infinity. We illustrate this procedure with some simulations.

Cette thèse est consacrée au développement de nouvelles méthodologies pour l'analyse des modèles non-linéaires à effets mixtes, à leur implémentation dans un logiciel accessible et leur application à des problèmes réels. Nous considérons particulièrement des extensions des modèles non-linéaires à effets mixtes aux modèles de mélange et aux modèles conjoints. Dans la première partie, nous proposons, dans le but d'avoir une meilleure maîtrise de l'hétérogénéité liée aux données sur des patients issus de plusieurs clusters, des extensions des MNLEM aux modèles de mélange. Nous proposons ensuite de combiner l'algorithme EM, utilisé traditionnellement pour les modèles de mélanges lorsque les variables étudiées sont observées, et l'algorithme SAEM, utilisé pour l'estimation de paramètres par maximum de vraisemblance lorsque ces variables ne sont pas observées. La procédure résultante, dénommée MSAEM, permet ainsi d'éviter l'introduction d'une étape de simulation des covariables catégorielles latentes dans l'algorithme d'estimation. Cet algorithme est extrêmement rapide, très peu sensible à l'initialisation des paramètres, converge vers un maximum (local) de la vraisemblance et est implémenté dans le logiciel Monolix.La seconde partie de cette Thèse traite de la modélisation conjointe de l'évolution d'un marqueur biologique au cours du temps et les délais entre les apparitions successives censurées d'un évènement d'intérêt. Nous considérons entre autres, les censures à droite, les multiples censures par intervalle d'évènements répétés. Les paramètres du modèle conjoint résultant sont estimés en maximisant la vraisemblance jointe exacte par un algorithme de type MCMC-SAEM. Cette méthodologie est désormais disponible sous Monolix / The main goal of this thesis is to develop new methodologies for the analysis of non linear mixed-effects models, along with their implementation in accessible software and their application to real problems. We consider particularly extensions of non-linear mixed effects model to mixture models and joint models. The study of these two extensions is the essence of the work done in this document, which can be divided into two major parts. In the first part, we propose, in order to have a better control of heterogeneity linked to data of patient issued from several clusters, extensions of NLMEM to mixture models. We suggest in this Thesis to combine the EM algorithm, traditionally used for mixtures models when the variables studied are observed, and the SAEM algorithm, used to estimate the maximum likelihood parameters when these variables are not observed. The resulting procedure, referred MSAEM, allows avoiding the introduction of a simulation step of the latent categorical covariates in the estimation algorithm. This algorithm appears to be extremely fast, very little sensitive to parameters initialization and converges to a (local) maximum of the likelihood. This methodology is now available under the Monolix software. The second part of this thesis deals with the joint modeling of the evolution of a biomarker over time and the time between successive appearances of a possibly censored event of interest. We consider among other, the right censoring and interval censorship of multiple events. The parameters of the resulting joint model are estimated by maximizing the exact joint likelihood by using a MCMC-SAEM algorithm. The proposed methodology is now available under Monolix.

In the post-harvest area, a common approach to quantify the average color of fruits peel over time is the sampling of small number of points generally on its equatorial region using a colorimeter. However, when we use a colorimeter to classify an uneven-colored fruit misclassification may occur because points in the peel region may not be representative of average color of fruit. The main problem when we use this method is to determine the number of points to be sampled as well as the location of these points on the fruit\'s surface. An alternative method to evaluate measure of color is digital image analysis because it covers whole of the object surface, by using a sample of pixels taken from the image. As the colorimeter approach is faster and easier than image analysis, it may not be suitable for assessing the overall mean color of the papaya\'s peel and its performance will depend on the number of measured points and choice of sampled region. In this sense, the comparison between these approach is still necessary because we need to know if a sample on the equatorial region can reproduce a sample over the whole region, and if the colorimeter can compete with a scanner or digital camera in measuring the mean hue of papaya peel over time. Thus, we proposed a longitudinal concordance correlation (LCC) based on polynomial mixed-effects regression model to evaluate the extent of agreement among methods. The results show that ideally image analysis of whole fruit\'s region should be used to compute the mean hue and that the topography and curved surface of papaya fruit did not affect the mean hue obtained by the scanner. Since there are still no packages available to estimate the LCC in the free software environment R, we are developing a package called lcc, which provides functions for estimating the longitudinal concordance correlation (LCC) among methods based on variance components and fixed effects of polynomial mixed-effects model. Additionally, we implemented arguments in this function to estimating the longitudinal Pearson correlation (LPC), as precision measure, and longitudinal bias corrector factor (LA), as accuracy measure. Moreover, these components can be estimated using different structures for variance- covariance matrices of random effects and variance functions to model heteroscedasticity among within-group errors using or not the time as variance covariate. / No setor de pós-colheita é muito comum a utilização de colorímetros para avaliar a cor média da casca de frutos ao longo do tempo. No entanto, muitas vezes as técnicas de amostragem utilizando esse equipamento podem levar a medidas tendenciosas da média amostral. Alternativamente, a utilização de imagens digitais pode levar a um menor viés, uma vez que toda a região da casca do fruto é amostrada de forma sistemática. No entanto, ainda é necessária a comparação de ambas abordagens, pois o colorímetro tem vantagens em relação a facilidade de utilização e menor tempo para realizar a amostragem em cada fruto quando comparado a um scanner de mesa. Assim, no caso de variáveis respostas medidas em uma escala contínua, a reprodutibilidade das medidas tomadas por ambos equipamentos pode ser avaliada por meio do coeficiente de correlação de concordância. Dessa forma, para avaliar o perfil da concordância entre métodos, nós propomos uma correlação de concordância longitudinal (LCC), baseada em um modelo de regressão polinomial com efeitos mistos. Os resultados sugeriram que as técnicas por meio de imagens digitais devem ser utilizadas para a quantificação da tonalidade média de frutos. Adicionalmente, a partir do perfil de concordância estimado notamos que existe um período em que ambos os equipamentos podem ser utilizados. A performance do coeficiente de concordância longitudinal foi avaliada por meio de um estudo de simulação, o qual sugeriu que nossa metodologia é robusta a dados desbalanceados (\"dropout\") e que a probabilidade de convergência é aceitavel para uma amostra de 20 frutos e ideal para amostras a partir de 100 frutos. Uma vez que ainda não existem pacotes disponibilizados no ambiente computacional R para a estimação da correlação de concordância longitudinal, nós estamos desenvolvendo um pacote intitulado lcc, o qual será submetido ao \"Comprehensive R Archive Network\" (CRAN). Nesse pacote nós implementamos procedimentos para estimação da correlação de concordância longitudinal, da correlação de Person longitudinal e de uma medida de acurácia longitudinal. Além disso, nosso pacote foi desenvolvido para dados balanceados e desbalanceados, permitindo modelar a heteroscedasticidade entre erros dentro do grupo usando ou não o tempo como covariável, e, também, permitindo a inclusão de covariáveis no preditor linear para controlar variações sistemáticas na variável resposta.

Résumé La technique connue comme l'imputation multiple semble être la technique la plus appropriée pour résoudre le problème de non-réponse. La littérature mentionne des méthodes qui modélisent la nature et la structure des valeurs manquantes. Une des méthodes les plus populaires est l'algorithme « Pan » de (Schafer & Yucel, 2002). Les imputations rapportées par cette méthode sont basées sur un modèle linéaire multivarié à effets mixtes pour la variable réponse. La méthode « BHLC » de (Murua et al, 2005) est une extension de « Pan » dont le modèle est bayésien hiérarchique avec groupes. Le but principal de ce travail est d'étudier le problème de sélection du modèle pour l'imputation multiple en termes d'efficacité et d'exactitude des prédictions des valeurs manquantes. Nous proposons une mesure de performance liée à la prédiction des valeurs manquantes. La mesure est une erreur quadratique moyenne reflétant la variance associée aux imputations multiples et le biais de prédiction. Nous montrons que cette mesure est plus objective que la mesure de variance de Rubin. Notre mesure est calculée en augmentant par une faible proportion le nombre de valeurs manquantes dans les données. La performance du modèle d'imputation est alors évaluée par l'erreur de prédiction associée aux valeurs manquantes. Pour étudier le problème objectivement, nous avons effectué plusieurs simulations. Les données ont été produites selon des modèles explicites différents avec des hypothèses particulières sur la structure des erreurs et la distribution a priori des valeurs manquantes. Notre étude examine si la vraie structure d'erreur des données a un effet sur la performance du choix des différentes hypothèses formulées pour le modèle d'imputation. Nous avons conclu que la réponse est oui. De plus, le choix de la distribution des valeurs manquantes semble être le facteur le plus important pour l'exactitude des prédictions. En général, les choix les plus efficaces pour de bonnes imputations sont une distribution de student avec inégalité des variances dans les groupes pour la structure des erreurs et une loi a priori choisie pour les valeurs manquantes est la loi normale avec moyenne et variance empirique des données observées, ou celle régularisé avec grande variabilité. Finalement, nous avons appliqué nos idées à un cas réel traitant un problème de santé. Mots clés : valeurs manquantes, imputations multiples, modèle linéaire bayésien hiérarchique, modèle à effets mixtes. / Abstract The technique known as multiple imputation seems to be the most suitable technique for solving the problem of non-response. The literature mentions methods that models the nature and structure of missing values. One of the most popular methods is the PAN algorithm of Schafer and Yucel (2002). The imputations yielded by this method are based on a multivariate linear mixed-effects model for the response variable. A Bayesian hierarchical clustered and more flexible extension of PAN is given by the BHLC model of Murua et al. (2005). The main goal of this work is to study the problem of model selection for multiple imputation in terms of efficiency and accuracy of missing-value predictions. We propose a measure of performance linked to the prediction of missing values. The measure is a mean squared error, and hence in addition to the variance associated to the multiple imputations, it includes a measure of bias in the prediction. We show that this measure is more objective than the most common variance measure of Rubin. Our measure is computed by incrementing by a small proportion the number of missing values in the data and supposing that those values are also missing. The performance of the imputation model is then assessed through the prediction error associated to these pseudo missing values. In order to study the problem objectively, we have devised several simulations. Data were generated according to different explicit models that assumed particular error structures. Several missing-value prior distributions as well as error-term distributions are then hypothesized. Our study investigates if the true error structure of the data has an effect on the performance of the different hypothesized choices for the imputation model. We concluded that the answer is yes. Moreover, the choice of missing-value prior distribution seems to be the most important factor for accuracy of predictions. In general, the most effective choices for good imputations are a t-Student distribution with different cluster variances for the error-term, and a missing-value Normal prior with data-driven mean and variance, or a missing-value regularizing Normal prior with large variance (a ridge-regression-like prior). Finally, we have applied our ideas to a real problem dealing with health outcome observations associated to a large number of countries around the world. Keywords: Missing values, multiple imputation, Bayesian hierarchical linear model, mixed effects model. / Les logiciels utilisés sont Splus et R.

General attrition rates in drug development pipeline have been recognized as a necessity to shift gears towards new methodologies that allow earlier and correct decisions, and the optimal use of all information accrued throughout the process. The quantitative science of pharmacometrics using pharmacokinetic-pharmacodynamic models was identified as one of the strategies core to this renaissance. Coupled with Optimal Design (OD), they constitute together an attractive toolkit to usher more rapidly and successfully new agents to marketing approval. The general aim of this thesis was to investigate how the use of novel pharmacometric methodologies can improve the design and analysis of clinical trials within drug development. The implementation of a Monte-Carlo Mapped power method permitted to rapidly generate multiple hypotheses and to adequately compute the corresponding sample size within 1% of the time usually necessary in more traditional model-based power assessment. Allowing statistical inference across all data available and the integration of mechanistic interpretation of the models, the performance of this new methodology in proof-of-concept and dose-finding trials highlighted the possibility to reduce drastically the number of healthy volunteers and patients exposed to experimental drugs. This thesis furthermore addressed the benefits of OD in planning trials with bio analytical limits and toxicity constraints, through the development of novel optimality criteria that foremost pinpoint information and safety aspects. The use of these methodologies showed better estimation properties and robustness for the ensuing data analysis and reduced the number of patients exposed to severe toxicity by 7-fold.  Finally, predictive tools for maximum tolerated dose selection in Phase I oncology trials were explored for a combination therapy characterized by main dose-limiting hematological toxicity. In this example, Bayesian and model-based approaches provided the incentive to a paradigm change away from the traditional rule-based “3+3” design algorithm. Throughout this thesis several examples have shown the possibility of streamlining clinical trials with more model-based design and analysis supports. Ultimately, efficient use of the data can elevate the probability of a successful trial and increase paramount ethical conduct.

Biological invasions are reportedly one of the major contributory factors to biodiversity loss worldwide. The impacts of invasive alien plant (IAP) species on native communities are widely documented in the scientific literature, however, there is still a lack of detailed information on their impacts within the most vulnerable habitats. Riparian habitats are highly dynamic systems and naturally disturbed, making them particularly vulnerable to invasion. Climate change, directly or indirectly, is also predicted to adversely impact river systems, which may subsequently alter invasion rates and the impacts of IAPs. However, the interactions between climate and IAPs and their combined effects on vegetation have rarely been examined. To address these knowledge gaps, this thesis investigates: (1) the role of environmental variables, such as sediment loading or climate-related changes to river flow regime, on the abundance of IAPs within riparian zones; (2) how variation in IAP abundance impacts native vegetation, relative to the effects of native dominant plant species and (3) some of the mechanisms underlying the effects of IAPs in riparian habitats. Historic and recent field survey data were used to investigate changes in riparian vegetation on British rivers during the last 20 years. Analyses indicate that IAPs had a negative but small effect on native plant diversity. Overall, changes in land use and differences in flow regime between recording periods were the most important predictors of plant community change. Specifically, IAPs had a greater probability of being present along lowland rivers that experienced increased frequency of high flow events. On a local scale across rivers in Scotland, the abundance of IAPs was constrained by greater soil moisture in summer, whilst greater abundance was associated with tree-lined banks. Both native dominant species and IAPs negatively affected subordinate species abundance to a greater extent than species richness, although this effect varied spatially with bank elevation. Artificial turf mats were used to quantify viable propagules within riverine sediment deposited over-winter along invaded riverbanks. The data indicate that there is a legacy effect of IAP abundance, with the most invaded sites being associated with higher sediment loading the following year, though, contrary to the general pattern, 12 sediment associated propagules were scarcer at invaded sites. Moreover, lower above-ground native diversity was associated with sites which had been previously invaded. Plant species composition in the propagule bank and above-ground vegetation were highly dissimilar, particularly closest to the water’s edge at highly invaded sites. This suggests that mono-specific stands of IAPs proliferate best under less disturbed environmental conditions, although fluvial disturbance events may be required to create opportunities for initial establishment. The propagule bank contributed very little to the above-ground vegetation, nor did it limit invasion, suggesting that above-ground plant composition is largely dictated by competitive interactions. The findings presented in this thesis suggest that invasion by IAPs is an additional stressor for native vegetation within riparian habitats, modifying above-ground plant communities via competition and suppressing recruitment from the propagule bank. However, native dominant species common in riparian habitats also negatively impact, subordinate species via competition, in some cases equalling the effect of IAPs. Native dominant and IAP species are differently affected by environmental factors operating in the riparian zone, which may provide future opportunities for reducing and managing invasions.

En agriculture, la météo est le principal facteur de variabilité d’une année sur l’autre. Cette thèse vise à construire des modèles statistiques à grande échelle qui estiment l’impact des conditions météorologiques sur les rendements agricoles. Le peu de données agricoles disponibles impose de construire des modèles simples avec peu de prédicteurs, et d’adapter les méthodes de sélection de modèles pour éviter le sur-apprentissage. Une grande attention a été portée sur la validation des modèles statistiques. Des réseaux de neurones et modèles à effets mixtes (montrant l’importance des spécificités locales) ont été comparés. Les estimations du rendement de maïs aux États-Unis en fin d’année ont montré que les informations de températures et de précipitations expliquent en moyenne 28% de la variabilité du rendement. Dans plusieurs états davantage météo-sensibles, ce score passe à près de 70%. Ces résultats sont cohérents avec de récentes études sur le sujet. Les prévisions du rendement au milieu de la saison de croissance du maïs sont possibles à partir de juillet : dès juillet, les informations météorologiques utilisées expliquent en moyenne 25% de la variabilité du rendement final aux États-Unis et près de 60% dans les états plus météo-sensibles comme la Virginie. Les régions du nord et du sud-est des États-Unis sont les moins bien prédites. Le rendements extrêmement faibles ont nécessité une méthode particulière de classification : avec seulement 4 prédicteurs météorologiques, 71% des rendements très faibles sont bien détectés en moyenne. L’impact du changement climatique sur les rendements jusqu’en 2060 a aussi été étudié : le modèle construit nous informe sur la rapidité d’évolution des rendements dans les différents cantons des États-Unis et localisent ceux qui seront le plus impactés. Pour les états les plus touchés (au sud et sur la côte Est), et à pratique agricole constante, le modèle prévoit des rendements près de deux fois plus faibles que ceux habituels, en 2060 sous le scénario RCP 4.5 du GIEC. Les états du nord seraient peu touchés. Les modèles statistiques construits peuvent aider à la gestion sur le cours terme (prévisions saisonnières) ou servent à quantifier la qualité des récoltes avant que ne soient faits les sondages post-récolte comme une aide à la surveillance (estimation en fin d’année). Les estimations pour les 50 prochaines années participent à anticiper les conséquences du changement climatique sur les rendements agricoles, pour définir des stratégies d’adaptation ou d’atténuation. La méthodologie utilisée dans cette thèse se généralise aisément à d’autres cultures et à d’autres régions du monde. / In agriculture, weather is the main factor of variability between two consecutive years. This thesis aims to build large-scale statistical models that estimate the impact of weather conditions on agricultural yields. The scarcity of available agricultural data makes it necessary to construct simple models with few predictors, and to adapt model selection methods to avoid overfitting. Careful validation of statistical models is a major concern of this thesis. Neural networks and mixed effects models are compared, showing the importance of local specificities. Estimates of US corn yield at the end of the year show that temperature and precipitation information account for an average of 28% of yield variability. In several more weather-sensitive states, this score increases to nearly 70%. These results are consistent with recent studies on the subject. Mid-season maize crop yield forecasts are possible from July: as of July, the meteorological information available accounts for an average of 25% of the variability in final yield in the United States and close to 60% in more weather-sensitive states like Virginia. The northern and southeastern regions of the United States are the least well predicted. Predicting years for which extremely low yields are encountered is an important task. We use a specific method of classification, and show that with only 4 weather predictors, 71% of the very low yields are well detected on average. The impact of climate change on yields up to 2060 is also studied: the model we build provides information on the speed of evolution of yields in different counties of the United States. This highlights areas that will be most affected. For the most affected states (south and east coast), and with constant agricultural practice, the model predicts yields nearly divided by two in 2060, under the IPCC RCP 4.5 scenario. The northern states would be less affected. The statistical models we build can help for management on the short-term (seasonal forecasts) or to quantify the quality of the harvests before post-harvest surveys, as an aid to the monitoring (estimate at the end of the year). Estimations for the next 50 years help to anticipate the consequences of climate change on agricultural yields, and to define adaptation or mitigation strategies. The methodology used in this thesis is easily generalized to other cultures and other regions of the world.