From the individual to species level, it is common for animals to have connections with one another. These connections can exist in a variety of forms; from the social relationships within an animal society, to hybridisation between species. The structure of these connections in animal systems can be depicted using networks, often revealing non-trivial structure which can be biologically informative. Understanding the factors which drive the structure of animal networks can help us understand the costs and benefits of forming and maintaining relationships. Multivariate modelling provides a means to evaluate the relative contributions of a set of explanatory factors to a response variable. However, conventional modelling approaches use statistical tests which are unsuitable for the dependencies inherent in network and relational data. A solution to this problem is to use specialised models developed in the social sciences, which have a long history in modelling human social networks. Taking predictive multivariate models from the social sciences and applying them to animal networks is attractive given that current analytical approaches are predominantly descriptive. However, these models were developed for human social networks, where participants can self-identify relationships. In contrast, relationships between animals have to be inferred through observations of associations or interactions, which can introduce sampling bias and uncertainty to the data. Without appropriate care, these issues could lead us to make incorrect or overconfident conclusions about our data. In this thesis, we use an established network model, the multiple regression quadratic assignment procedure (MRQAP), and propose approaches to facilitate the application of this model in animal network studies. Through demonstrating these approaches on three animal systems, we make new biological findings and highlight the importance of considering data-sampling issues when analysing networks. Additionally, our approaches have wider applications to animal network studies where relationships are inferred through observing dyadic interactions.

I have studied persistent environmental pollutants in herring (Clupea harengus), in adult guillemot (Uria aalge) and in guillemot eggs from the Baltic Sea. The studied contaminants were organochlorines (OCs); dichlorodiphenyltrichloroethanes (DDTs), polychlorinated biphenyls (PCBs), hexachlorobenzene (HCB), hexachlorocyclohexanes (HCHs), and brominated flame retardants (BFRs); polybrominated diphenylethers (PBDEs) and hexabromocyclododecane (HBCD). The highest concentration in both species was shown by p,p′DDE with a concentration in guillemot egg (geometric mean (GM) with 95% confidence interval) of 18200 (17000 – 19600) ng/g lipid weight. The BFR with the highest concentration in guillemot egg was HBCD with a GM concentration of 140 (120 – 160) ng/g lw. To extract additional and essential information from the data, not possible to obtain using only univariate or bivariate statistics, I used multivariate data analysis techniques; principal components analysis (PCA), partial least squares regression (PLS), soft independent modelling of class analogy (SIMCA), and PLS discriminant analysis (PLS-DA). I found e.g.; that there are significant negative correlations between egg weight and the concentrations of HCB and p,p'DDE; that concentrations of OCs and BFRs in the organisms co-varied so that concentrations of OCs can be used to calculate concentrations of BFRs; and, that several contaminants (e.g. HBCD) had higher concentration in guillemot egg than in guillemot muscle, that several (e.g. BDE47) showed no concentration difference between muscle and egg and that one contaminant (BDE154) showed higher concentration in the guillemot muscles than in egg. In this thesis I developed a new method, “randomly sampled ratios” (RSR), to calculate biomagnification factors (BMFs) i.e. the ratio between the concentration of a contaminant in an organism and the concentration of the same contaminant in its food. With this new method BMFs are denoted with an estimate of variation. Contaminants that biomagnify are e.g., p,p′DDE, CB118, HCB, βHCH and all of the BFRs. Those that do not biomagnify are e.g., p,p′DDT, αHCH and CB101. Lastly, to investigate which of the contaminants descriptors (physical-chemical/other properties and characteristics) that correlates to the biomagnification of the contaminants, I modeled the contaminants’ respective BMFRSR versus ~100 descriptors and showed that ~20 descriptors in combination were important, either favoring or counteracting biomagnification between herring and guillemot.

A systematic methodology for quantitative structure-activity relationship (QSAR) development in environmental toxicology is provided. The methodology is summarized in a strategy with six sequential steps. The strategy rests on two cornerstones, namely (1) the use of statistical design to select a series of representative compounds (the so-called training set) on which to base a QSAR, and (2) the multivariate modelling of the relationship between physicochemical and biological properties of the training set compounds. The first step of the strategy is the division of chemicals into classes of structurally similar compounds. Briefly, steps 2 to 6 are: (2) physico-chemical and structural characterization of the compounds in a class, (3) selection of a training set of representative compounds, (4) biological testing of the selected training set, (5) QSAR model development, and (6) experimental validation of the QSAR and predictions for non-tested compounds. The thesis summarizes the results obtained from the application of the strategy to the class of halogenated aliphatic compounds. Biological measurements were made in four biological test systems, reflecting acute toxicity, mutagenicity, relative cytotoxicity and genotoxicity. QSARs were developed relating each biological endpoint to the structural descriptors of the compounds. Multivariate PLS modelling was used in the data analysis. The developed QSARs were used for predicting the biological activity pattern of the non-tested compounds in the class. These predictions may be used as a starting point for a priority ranking for further biological testing of these compounds. The strategy has not been developed solely for establishing QSARs for the halogenated aliphatics class. On the contrary, this work is intended to demonstrate a generally applicable QSAR methodology. / <p>Diss. (sammanfattning) Umeå : Umeå universitet, 1991</p> / digitalisering@umu

This thesis presents a number of studies, where a multivariate approach was taken to construct models that predict wood species and thermo mechanical pulp brightness from roundwood of Norway spruce and Scots pine. The first and second studies produced multivariate prediction models for wood species from the bark of spruce and pine. These models can be used for wood species classification and would replace the manual log assessment that takes place today. Principal Component Analysis, PCA, and Partial least squares projections to Latent Structures, PLS, were used to predict the wood species from multivariate measurements recorded from the bark of spruce and pine. Two different kinds of measurements were employed, near-infrared spectroscopy and digital imaging. Both methods showed that it was possible to predict the wood species with a high accuracy. The third and fourth studies of the thesis are related to the wood storage of roundwood and the deterioration of wood that occurs during the storage. The third study used an experimental design with five storage factors that provided different conditions for the analysed wood. The experimental design made it possible to identify the factors and the interaction between factors, which were important for the ISO brightness of peroxide and dithionite bleached thermo mechanical pulp, TMP. The final study of the thesis used NIR spectroscopy for predicting the ISO brightness of bleached TMP. Spectra recorded from stored wood were used to construct PLS prediction models.

MSc (Statistics) / Department of Statistics / See the attached abstract below

Cette thèse de doctorat porte sur l’analyse des dommages et sur l’évaluation des coûts induits sur les habitations par les submersions marines. L’étude se base sur les données d’assurance de deux évènements récents ayant touché la France et causé des submersions sur les côtes bretonnes et atlantiques : les tempêtes Johanna (mars 2008) et Xynthia (février 2010).Dans un premier temps, l’analyse des données d’expertise et d’indemnisation d’assurance, en lien avec celle des paramètres de l’aléa et des enjeux exposés, a eu pour but la meilleure compréhension des différents types de dommages, et l’explication des coûts observés. En parallèle, un travail de modélisation de l’aléa a été réalisé à une échelle régionale, afin de déterminer des indicateurs des forçages météo-marins, et à une échelle locale, afin de préciser les processus d’endommagement sur les sites étudiés pour les deux tempêtes. La caractérisation de la vulnérabilité et de la valeur des enjeux (coûts de construction) a été menée à l’aide de différents paramètres issus de bases de données nationales (INSEE et IGN) et de campagnes de terrain.Dans un second temps, les informations recueillies ont permis la construction de modèles empiriques de prédiction du coût des dommages aux habitations spécifiques à l’aléa submersion marine, outils aujourd’hui inexistants en France. Les différents types de modèles testés sont basés sur des approches statistiques univariées (fonctions d’endommagement) et multivariées. L’apport des données d’assurance à la réalisation de tels modèles est discuté, et des recommandations ainsi que des perspectives de recherche sont évoquées, afin de rendre ces modèles opérationnels et d’augmenter leur capacité de prédiction des coûts d’évènements catastrophiques futurs. / This PhD work aims at analysing damage and evaluating costs related to coastal flooding on residential buildings. The study is based on insurance data from two recent storm events, which caused coastal flooding in the Brittany and Atlantic regions in France: the storms Johanna (March 2008) and Xynthia (February 2010).At first, the analysis of insurance indemnities and loss adjustment data, in connection with hazard parameters, and exposed assets characteristics, allowed a better understanding of the different types of damage and costs observed. At the same time, hazard models were carried out at a regional level, in order to identify meteorological forcing indicators, and at a local level, in order to link damages to the associated physical flooding processes on the studied sites. The characterization of the asset’s vulnerability and values (construction costs) was conducted using different parameters from national databases (INSEE and IGN) and field survey.In a second step, empirical cost-assessment models were built on the basis of the data analysed, using univariate (damage functions) and multivariate statistical approaches. This study is the first attempt in France to elaborate models for the prediction of damage costs linked to coastal flooding on housing. The contribution of insurance data to the implementation of such models is discussed, and recommendations and research perspectives are expressed, in order to make the models operational and to increase their capacity to predict future catastrophic events costs.