Résumé : Les méthodes de capture-recapture servent à décrire l’état d’une population et les processus qui en sous-tendent la dynamique, lorsque les méthodes d’observation et de détection de cette population sont imparfaites. En surveillance des maladies infectieuses, elles peuvent simplement être utilisées pour estimer la taille totale de la population infectée par un pathogène et estimer quantitativement la sensibilité du système de surveillance de ce pathogène. Bien qu'exploitées très largement dans le domaine de la santé publique pour répondre à cet objectif, l'utilisation des méthodes de capture-recapture (CR) en surveillance des maladies infectieuses animales peut être considérée comme récente. Parce que le contexte de surveillance des maladies animales est nettement différent de celui de la surveillance des maladies humaines, des questions demeurent quant à l'intérêt et aux limites de ces méthodes pour estimer la sensibilité des systèmes de surveillance des maladies animales. Pour tenter de répondre à ces questions, nous avons identifié quatre systèmes de surveillance différents par leur complexité, leur efficacité et leur maladie d'intérêt : nous avons retenu les systèmes de surveillance de la fièvre aphteuse au Cambodge, de l’influenza aviaire hautement pathogène (IAHP) H5N1 en Egypte, de la tremblante ovine en France et de l’IAHP H5N1 en Thaïlande. Pour chacun de ces systèmes de surveillance, nous avons déterminé le modèle de CR le plus adapté compte tenu des données générées (respectivement le modèle de CR à deux sources, le modèle log-linéaire à 4 sources, le modèle de comptage tronqué en zéro et le modèle de comptage enflé en zéro). Pour chaque application, nous avons donc estimé le nombre total d'unités épidémiologiques non détectées par les systèmes de surveillance considérés ce qui nous a permis d’apprécier la sensibilité de chaque système de surveillance considéré. Il est ressorti de ces applications que les méthodes de capture-recapture sont relativement faciles à conduire et qu’elles permettent à faible coût d’estimer l’importance réelle d’une maladie sur un territoire quand celle-ci est surveillée de manière imparfaite. Il semble cependant que les pratiques de surveillance et de contrôle des maladies animales limitent les applications à l’échelle de l’animal, et nécessitent d’élargir l’unité épidémiologique à une échelle supérieure (troupeau, commune, etc…). Cet élargissement introduit de nouvelles contraintes (notamment l’hétérogénéité d’abondance) qu’il est nécessaire de prendre en compte pour ne pas biaiser les estimations finales. Ce travail propose des perspectives d’application en épidémiologie descriptive, ainsi que des perspectives méthodologiques de recherche en statistique et en modélisation. / Abstract: Capture-recapture methods are generally used to describe populations when observation processes are imperfect. In the context of disease surveillance, they can be used simply for estimating the total size of the populations infected by a given pathogen, and hence, estimating quantitatively the sensitivity of the surveillance of this pathogen. Although they are widely used in public health, capture-recapture methods have been barely applied to the surveillance of animal diseases. Because the context of animal health is quite different from the context of public health, some questions remain concerning the benefits and the limitations of such methods for estimating the sensitivity of surveillance systems in animal health. For answering this research question, we identified four animal disease surveillance systems that differ by their complexity, their efficiency and their disease of interest. We selected the surveillance of foot-and-mouth disease in Cambodia, of highly pathogenic avian influenza (H5N1) in Egypt and Thailand, and of classical scrapie in France. For each surveillance system, we identified the most appropriate capture-recapture approach (respectively the two-source approach, the three-source approach, the zero-inflated approach and the zero-truncated approach). For each application, we estimated the total number of infected epidemiological units that remained undetected, and accessed an estimation of the sensitivity of each surveillance system. From these applications, we highlighted that these models are relatively easy to implement, and that they allow with little additional income to get an unbiased representation of the disease burden in a population when it is monitored with imperfect surveillance processes. However, it seems that practices used for the monitoring and controlling animal diseases tend to limit the applicability of these methods at the scale of the monitored unit. As a consequence, it is often necessary to enlarge the epidemiological unit (holding, commune, etc…) so that it comprises several monitored units. This enlargement introduces new constraints (abundance induced heterogeneity), that need to be taken into account in order not to bias final estimates. Finally, this work proposes surveillance perspectives for descriptive epidemiology, and methodological perspectives in statistics and modeling as well.

The study showed that density estimation of Eurasian lynx corresponds to 1.3 and 1.2 independent individuals per 100 km2 in the Jvorniky study area for the two time periods and 0.8 independent individuals per 100 km2 for Beskydy study area. The study failed to identify other large carnivores such a wolf (Canis lupus) and bear (Ursus arctos) from both Beskydy and Javorniky study areas. The estimated density of Lynx is low and their numbers should increase in future. There is enough prey base to support the existing population in the two study areas. The main factors effecting Lynx distribution are habitat fragmentation, poaching by humans, depleting the prey base by over hunting leading to starvation, vehicle collisions. Strict measures should be implemented to protect the species and long term study programmes must be started to get a comprehensive knowledge about the biology of species. Reintroductions must be carried over where there are suitable habitat for the survival and propagation of Lynx. The reason for not detecting wolf or bear might be due to the fact that the study areas are wide and the few migrating wolf or bear might be present outside my study area. Also there is lot of possibility to reintroduce wolf in my study area and I hope this will be done in future to ensure better biodiversity and to ensure wildlife conservation.

Data sources and data quality are indispensable in economical, medical, pharmaceutical or other studies and provide the basis for reliable study results in numerous research questions. Depending on the purpose of use, a high quality of data is a prerequisite. However, with increasing registry quality, costs also increase accordingly. Considering these time and cost consuming factors, this work is an attempt to estimate the cost advantages when applying statistical tools to existing registry data. This includes methodological considerations and suggestions regarding the evaluation of data quality including factors such as bias and reliability after dealing properly (or not) with missing data (MD), and possible consequences when ignoring the incompleteness of data. Results for the quality analysis of the gastric cancer patients’ data example showed that millions of Euros in study costs can be saved by reducing the time horizon. On average, €523,126.70 can be saved for every year that the study duration is shortened. Replacing additionally the over 25% of MD in some variables, data quality was immensely improved, but still showed quality difficulties, which – beside MD in variables – could be an indication for completely missing entries of patients in the registry. Capturerecapture methods were therefore discussed to demonstrate how the total completeness in a registry can be estimated. Since it was not possible to illustrate the CARE method with the example of the gastric cancer patients due to the given data structure (no access to required variables), other data sets had to be chosen – the publicly accessible data of the amyotrophic lateral sclerosis (ALS) and data of towed vehicles in the City of Chicago. The consequence of ignoring MD was further analyzed using bankruptcy prediction data sets of agribusiness companies and confirmed the assumption that MD have a negative impact on the data quality, in this case also regarding the misclassifications of predictions of bankrupted companies. Using the decision tree method (known as one of the most suitable methods in predicting financial distress), the percentage of correctly bankruptcy-predicted of bankrupted companies (one year to bankruptcy) with MD imputation was 87.5%, whereas it was only 60% when completely omitting MD. Overall, my findings showed dearly the importance of statistical methods to improve data quality which in turn helps to avoid drawing biased conclusions due to incomplete data.

L’estimation de fréquence des maladies est nécessaire pour générer des hypothèses étiologiques et évaluer leur l’impact médico-économique. Cependant, il est possible que ces estimations ne reflètent pas uniquement les caractéristiques de fréquence intrinsèque d’une maladie mais aussi de variations dues à des choix méthodologiques qui diffèrent entre études. Cette hypothèse d’une variabilité des fréquences liée à des facteurs méthodologiques a été initialement soulevée par notre premier travail qui visait à estimer la prévalence du syndrome de Gougerot-Sjögren (SGS) primitif au sein de la population française de la Seine-Saint-Denis pendant la période 2007. Malgré une bonne exhaustivité d’identification des cas de 90 % calculée par la méthode de capture-recapture, notre estimation de prévalence de 1,02 cas pour 10 000 habitants est la plus faible estimation de fréquence du SGS primitif publiée dans la littérature à ce jour. En particulier, ce travail a souligné le possible impact des méthodologies de type « recensement » ou « échantillonnage » sur les estimations établies par des études de fréquence. Cette étude a montré aussi pour la première fois une prévalence 2 fois supérieure chez les patients non-européens par rapport aux européens avec des possibles phénotypes distincts. Pour explorer plus en détail l’hypothèse d’une variabilité entre les estimations générées par les approches de « recensement » ou de « échantillon », nous avons réalisé un deuxième travail qui consistait en une méta-analyse de la prévalence de la maladie de Behçet (MB) rapportée par 45 études internationales publiées dans la littérature. Des analyses en sous-groupes et par méta-régression ont montré des variations notables de la prévalence de MB entre zones géographiques mais aussi selon le type d’étude (recensement vs échantillonnage) utilisé. En analyse par méta-régression multivariée, seule la variable « type d’étude » était significativement associée aux valeurs de prévalence de la MB. En conclusion, ces travaux soulignent l’impact de la méthodologie utilisée pour conduire les études de fréquence des maladies. Les différences conceptuelles entre les études de recensement et les études échantillonnage soulevant la question sur la comparabilité des estimations obtenues par ces deux approches. / Estimating the frequency of diseases is of major importance for generating etiologic hypotheses and for assessing their global burden. Although such estimates may reflect the intrinsic frequency characteristics of a disease, they may also reflect variations due to methodological differences between studies. In our first study, we raised this assumption that the variability of disease frequency estimates is linked to methodological factors. This population-based survey aimed at estimating the prevalence of primary Sjögren's syndrome (pSS) in Seine-Saint-Denis, France, during 2007. Despite a high completeness of case-finding, 90%, assessed by capture-recapture analysis, the estimated prevalence of pSS of 1.02 cases per 10,000 adults was the lowest prevalence estimate published in the literature. Also, this study was the first to show a two-fold higher prevalence of pSS in people of non-European than European background and possible ethnicity-related differences in disease phenotypes. In addition, this study highlighted the possible effect of "census" or "sampling" designs on frequency estimates reported from population-based surveys. To explore in more detail the assumption that "census" or "sampling" approaches generate variability in frequency estimates of diseases, we performed a second study, involving a meta-analysis of the prevalence of Behçet's disease (BD) in published reports of 45 international studies. Subgroup and meta-regression analyses showed notable variations in BD prevalence estimates geographically but also by study design (census vs sampling). On multivariate meta-regression analysis, only study design significantly predicted BD prevalence estimates. In conclusion, this thesis highlights the importance of study design in population-based estimates of disease frequency. Conceptual differences between census and sample studies raise questions about the extent to which estimates obtained by either of these two approaches are comparable.

Noninvasive methods for monitoring wildlife species have revolutionized the way population parameters, such as population density and survival and recruitment rates, are estimated while accounting for imperfect detection using capture-recapture models. Reliable estimates of these parameters are vital information required for making sound conservation decisions; however to date, noninvasive sampling methods have been of limited use for a vast number of species which are difficult to identify to the individual level–a general requirement of capture-recapture models. Capture-recapture models that utilize partial identity information have only recently been introduced and have not been extended to most types of noninvasive sampling scenarios in a manner that uses the spatial location where noninvasive samples were collected to further inform complete identity (i.e. spatial partial identity models). Herein, I extend the recently introduced spatial partial identity models to the noninvasive methods of remote cameras for species that are difficult to identify from photographs and DNA from hair or scat samples. The ability of these novel models to improve parameter estimation and extend study design options are investigated and the methods are made accessible to applied ecologists via statistical software. This research has the potential to greatly improve wildlife conservation decisions by improving our knowledge of parameters related to population structure and dynamics that inform those decisions. Unfortunately, many species of conservation concern (e.g., Florida panthers, Andean bears) are managed without having the necessary information on population status or trends, largely a result of the cost and difficulty of studying species in decline and because of the difficulty of applying statistical models to sparse data, which can produce imprecise and biased estimates of population parameters. By leveraging partial identity information in noninvasive samples, the models I developed will improve these parameter estimates and allow noninvasive methods to be used for more species, leading to more informed conservation decisions, and a more efficient allocation of conservation resources across species and populations. / Ph. D.

Rheumatoid arthritis (RA) is a chronic disease characterized by an overactive immune system and joint inflammation. Population-based administrative health data (AHD) are widely used for RA outcomes research and surveillance. However, AHD may not completely capture all cases of RA in the population. Capture-recapture (CR) methods have been proposed to describe the completeness of AHD for estimating disease population size, but AHD may not conform to the assumptions that underlie CR models. A Monte Carlo simulation study was used to investigate the effects of violations of the assumptions for two-source CR methods: dependence between data sources and heterogeneity of capture probabilities. We compared the Chapman estimator and an estimator based on the multinomial logistic regression model (MLRM) to study relative bias (RB), coverage probability (CP) of 95% confidence intervals, width of 95% confidence intervals (WCI), and root-mean-square-error (RMSE) in prevalence estimates. The effects of misspecification of the MLRM were also investigated. In addition, the Chapman and MLRM estimators were used to estimate RA prevalence using AHD data from Saskatchewan, Canada. Population sizes were consistently underestimated for CR methods when the assumptions were violated. The estimated population size for both of the estimators did not differ substantially except for the RMSE values. Parameter estimates became biased when the MLRM model was misspecified, but there was little impact on population size estimates. In conclusion, CR methods are recommended to reduce bias in prevalence estimates based on AHDS. Because these methods may be sensitive to assumption violations, researchers should consider potential dependence between data sources. As well, sufficient overlap in the cases captured by each data source (e.g., 50% of the cases are captured by both data sources) or balanced capture probability in each data source is needed to effectively implement these methods. Researchers who estimate population size using CR methods in AHDs should favour the MLRM estimator over the Chapman estimator.

Capture-recapture (CRC) models use two or more samples, or lists, to estimate the size of a population. In the canonical example, a researcher captures, marks, and releases several samples of fish in a lake. When the fish that are captured more than once are few compared to the total number that are captured, one suspects that the lake contains many more uncaptured fish. This basic intuition motivates CRC models in fields as diverse as epidemiology, entomology, and computer science. We use simulations to study the performance of conventional log-linear models for CRC. Specifically we evaluate model selection criteria, model averaging, an asymptotic variance formula, and several small-sample data adjustments. Next, we argue that interpretable models are essential for credible inference, since sets of models that fit the data equally well can imply vastly different estimates of the population size. A secondary analysis of data on survivors of the World Trade Center attacks illustrates this issue. Our main chapter develops local log-linear models. Heterogeneous populations tend to bias conventional log-linear models. Post-stratification can reduce the effects of heterogeneity by using covariates, such as the age or size of each observed unit, to partition the data into relatively homogeneous post-strata. One can fit a model to each post-stratum and aggregate the resulting estimates across post-strata. We extend post-stratification to its logical extreme by selecting a local log-linear model for each observed point in the covariate space, while smoothing to achieve stability. Local log-linear models serve a dual purpose. Besides estimating the population size, they estimate the rate of missingness as a function of covariates. Simulations demonstrate the superiority of local log-linear models for estimating local rates of missingness for special cases in which the generating model varies over the covariate space. We apply the method to estimate bird species richness in continental North America and to estimate the prevalence of multiple sclerosis in a region of France.

The study focuses in the ecology and conservation of the ocelot (Leopardus pardalis) in the conservation and tourism concession owned by the ARCC. The study site is 11 000 hectares and it is located in the Las Piedras Region, north of Tambopata province, Madre de Dios. Camera traps were used to monitor the ocelot population during a 7-month period (from August 2012 to February 2013), divided in 9 rounds were 73 cameras were installed. The camera traps found 8 independent ocelots, from which only 3 (A1, A3 and A6) were recaptured in more than one occasion. The study did a capture-recapture analysis. The distance traveled by the ocelots from a capture to a recapture site was used to estimate the effective sampled area using the Mean Maximum Distance Moved - MMDM and Half MMDM. The methods yielded a density of 70 individuals/100km2 and 180 individuals/100km2, with full MMDM and Half MMDM, respectively. The study also analyzed the camera trap capture probability with PRESENCE software. Using a closed CR analysis followed by a model of constant capture probability, it yielded a capture probability rate of 0,3 (SE 0,0567). Finally, the ocelot’s habitat preference was also studied using a combination of satellite imagery and GIS software. It was found that these animals frequently use transects aimed for tourists, prefer sites near water and that they avoid bamboo forests. / Este estudio trata sobre la ecología y conservación del ocelote (Leopardus pardalis), en la concesión de conservación y ecoturismo del albergue Amazon Research and Conservation Center - ARCC. El área de estudio, de 11 000 hectáreas, se encuentra ubicada en el distrito de Las Piedras, norte de la provincia de Tambopata, departamento de Madre de Dios, Perú. Se utilizaron cámaras trampa para monitorear la población del felino en un período de siete meses (de agosto de 2012 a febrero de 2013), dividido en nueve rondas donde se instalaron 73 cámaras en total. Fueron ocho ocelotes independientes los identificados, de los cuales solo tres (A1, A3 y A6) fueron recapturados visualmente en más de una ocasión. Se realizó un análisis de captura-recaptura. Las distancias recorridas por los ocelotes entre captura y recaptura se utilizaron para estimar el área efectiva de muestreo usando el método del Promedio de la Máxima Distancia Recorrida - MMDM y Mitad del MMDM. Los métodos dieron como resultado una densidad poblacional de 700 ocelotes/100 km2 y 180 ocelotes/100 km2 con MMDM y Mitad del MMDM, respectivamente. Por otro lado, se analizó la probabilidad de captura de las cámaras trampa con el software PRESENCE. Utilizando un análisis poblacional cerrado y un modelo constante, se halló una detección por ronda de 0,3 (SE 0,0567). Finalmente, también se evaluó la preferencia de hábitat de los ocelotes a través de imágenes satélite. Se halló que la mayoría de los felinos usan transectos turísticos, que prefieren las llanuras aluviales cercanas a las riberas de los ríos y cochas, y que evitan los pantanos.

We propose a ratio estimator to determine population estimates using capture-recapture sampling. It's different than traditional approaches in the following ways: (1) Ordering of recaptures: Currently data sets do not take into account the "ordering" of the recaptures, although this crucial information is available to them at no cost. (2) Dependence of trials and cluster sampling: Our model explicitly considers trials to be dependent and improves existing literature which assumes independence. (3) Rate of convergence: The percentage sampled has an inverse relationship with population size, for a chosen degree of accuracy. (4) Asymptotic Attainment of Minimum Variance (Open Systems: (=population variance). (5) Full use of data and model applicability (6) Non-parametric (7) Heterogeneity: When units being sampled are hard to identify. (8) Open and closed systems: Simpler results are presented separately for closed systems. (9) Robustness to assumptions in open systems

Introduction: Les études populationnelles sur le cancer nécessitent un recensement de référence fiable et exhaustif, en théorie possible à partir d'un recueil histologique. Méthode: Depuis 2005, toutes les structures d'anatomopathologie des Alpes-Maritimes adressent les codes ADICAP des tumeurs malignes et invasives et identifiants patients. L'exhaustivité pour les cancers du sein et colorectaux des 50-75 ans a été évalué par méthode de capture recapture en modélisation log-linéaire et en Bayesien à partir des cas communs ou non dépistés et vus en Réunion de Concertation Pluridisciplinaire. RésultatUn programme d'assurance qualité a permis de s'assurer de la fiabilité des données recueillies.L'estimation de l'exhaustivité était de plus de 90 % pour les cancers du sein et colorectaux des 50-75 ans. Les taux observés sur le département des Alpes-Maritimes, comparés aux taux estimés en France, se sont révélés cohérents.Enfin, la base a été utilisée pour déterminer l'existant les lésions prénéoplasiques du col de l'utérus avant la vaccination anti-HPV. ConclusionCe travail conclut à l'intérêt d'un recueil histologique des cas de cancers incidents. / Introduction Cancer population studies require reliable and complete baseline data, which should theoretically be available by collecting histopathology records.Method Since 2005, all histopathology laboratories from Alpes-Maritimes address ADICAP codes for invasive cancer and patient identifiers. The completeness of such a collection was evaluated using capture-recapture analysis based on three data sources concerning breast and colorectal cancers with the number of cases which were common or not between sources recording screened, diagnosed and treated cancers in the French Alpes Maritimes districtResult Data quality for the ADICAP code database may be considered satisfactoryThe estimated completeness of cancer records collected from histopathology laboratories was higher than 90%.Rates observed in the Alpes-Maritimes, compared with estimated rates in France have proven consistent. Rates of CIN for the entire female population of the Alpes-Maritimes in 2006 has been established.Conclusion A verified and validated histopathology data collection may be useful for cancer population studies.