Abstract
Capture-recapture is a widely used method for estimating population size and inferring demographic processes such as survival or migration rates. In recent years, the use of man-made tags to mark animals has been replaced by natural tags such as visual patterns or genetic fingerprints. Natural marks have the advantage of enabling the individual to be 'captured' non-invasively, using photographic traps or by collecting feces, hairs, or feathers for instance. It is therefore not necessary to see the individuals in order to capture them.Although non-invasive sampling is increasingly used in capture-recapture experiments, it carries a risk of individual misidentification that cannot be ignored. In most studies, data susceptible to individual misidentification are simply discarded. As a result, the proportion of discarded data may be significant. To overcome this problem, models have been proposed that can deal with individual identification errors in order to use a larger amount of data. However, these models do not take into account several characteristics common to non-invasive sampling data. First, identification quality is only modelled globally, although a measure of identification quality is often available at the sample level.Second, the models do not take into account repeated observations, i.e. different samples belonging to the same individual and collected on the same occasion.Third, most models have only been proposed for closed populations, which only concerns a limited number of studies.By going through the different models available, I selected one that had the potential to address all these limitations. I implemented the selected model in the R language, specifically the NIMBLE package, in a Bayesian approach, and extended it to overcome the identified limitations. I used simulations to test the performances of the models I had developed and compared them with appropriate pre-existing models.My work has allowed the development of a complete framework for all basic cases of capture-recapture in the presence of individual misidentification. It covers closed or open populations, in single or multiple states, and with or without an identification quality covariate. This work also provides an example of the potential of modelling misidentifications through a simulation study of capture-recapture on mosquito larvae, where discarding the poor quality samples would likely result in almost no sample being retained. Finally, the implementation of the model will make it usable by modellers and should contribute to the dissemination of these new models in a wider context.