Model selection and error estimation

We study model selection strategies based on penalized empirical loss minimization. We point out a tight relationship between error estimation and data-based complexity penalization: any good error estimate may be converted into a data-based penalty function and the performance of the estimate is governed by the quality of the error estimate. We consider several penalty functions, involving error estimates on independent test data, empirical {\sc vc} dimension, empirical {\sc vc} entropy, andmargin-based quantities. We also consider the maximal difference between the error on the first half of the training data and the second half, and the expected maximal discrepancy, a closely related capacity estimate that can be calculated by Monte Carlo integration. Maximal discrepancy penalty functions are appealing for pattern classification problems, since their computation is equivalent to empirical risk minimization over the training data with some labels flipped.

Fixed and random effects in Classical and Bayesian regression

This paper proposes a common and tractable framework for analyzingdifferent definitions of fixed and random effects in a contant-slopevariable-intercept model. It is shown that, regardless of whethereffects (i) are treated as parameters or as an error term, (ii) areestimated in different stages of a hierarchical model, or whether (iii)correlation between effects and regressors is allowed, when the sameinformation on effects is introduced into all estimation methods, theresulting slope estimator is also the same across methods. If differentmethods produce different results, it is ultimately because differentinformation is being used for each methods.

Random measurement error and regression dilution bias.

Summary points: - The bias introduced by random measurement error will be different depending on whether the error is in an exposure variable (risk factor) or outcome variable (disease) - Random measurement error in an exposure variable will bias the estimates of regression slope coefficients towards the null - Random measurement error in an outcome variable will instead increase the standard error of the estimates and widen the corresponding confidence intervals, making results less likely to be statistically significant - Increasing sample size will help minimise the impact of measurement error in an outcome variable but will only make estimates more precisely wrong when the error is in an exposure variable

Unfolding nature "in silico" : overcoming practical and metodological limitations of species distribution models

Abstract : The existence of a causal relationship between the spatial distribution of living organisms and their environment, in particular climate, has been long recognized and is the central principle of biogeography. In turn, this recognition has led scientists to the idea of using the climatic, topographic, edaphic and biotic characteristics of the environment to predict its potential suitability for a given species or biological community. In this thesis, my objective is to contribute to the development of methodological improvements in the field of species distribution modeling. More precisely, the objectives are to propose solutions to overcome limitations of species distribution models when applied to conservation biology issues, or when .used as an assessment tool of the potential impacts of global change. The first objective of my thesis is to contribute to evidence the potential of species distribution models for conservation-related applications. I present a methodology to generate pseudo-absences in order to overcome the frequent lack of reliable absence data. I also demonstrate, both theoretically (simulation-based) and practically (field-based), how species distribution models can be successfully used to model and sample rare species. Overall, the results of this first part of the thesis demonstrate the strong potential of species distribution models as a tool for practical applications in conservation biology. The second objective this thesis is to contribute to improve .projections of potential climate change impacts on species distributions, and in particular for mountain flora. I develop and a dynamic model, MIGCLIM, that allows the implementation of dispersal limitations into classic species distribution models and present an application of this model to two virtual species. Given that accounting for dispersal limitations requires information on seed dispersal, distances, a general methodology to classify species into broad dispersal types is also developed. Finally, the M~GCLIM model is applied to a large number of species in a study area of the western Swiss Alps. Overall, the results indicate that while dispersal limitations can have an important impact on the outcome of future projections of species distributions under climate change scenarios, estimating species threat levels (e.g. species extinction rates) for a mountainous areas of limited size (i.e. regional scale) can also be successfully achieved when considering dispersal as unlimited (i.e. ignoring dispersal limitations, which is easier from a practical point of view). Finally, I present the largest fine scale assessment of potential climate change impacts on mountain vegetation that has been carried-out to date. This assessment involves vegetation from 12 study areas distributed across all major western and central European mountain ranges. The results highlight that some mountain ranges (the Pyrenees and the Austrian Alps) are expected to be more affected by climate change than others (Norway and the Scottish Highlands). The results I obtain in this study also indicate that the threat levels projected by fine scale models are less severe than those derived from coarse scale models. This result suggests that some species could persist in small refugias that are not detected by coarse scale models. Résumé : L'existence d'une relation causale entre la répartition des espèces animales et végétales et leur environnement, en particulier le climat, a été mis en évidence depuis longtemps et est un des principes centraux en biogéographie. Ce lien a naturellement conduit à l'idée d'utiliser les caractéristiques climatiques, topographiques, édaphiques et biotiques de l'environnement afin d'en prédire la qualité pour une espèce ou une communauté. Dans ce travail de thèse, mon objectif est de contribuer au développement d'améliorations méthodologiques dans le domaine de la modélisation de la distribution d'espèces dans le paysage. Plus précisément, les objectifs sont de proposer des solutions afin de surmonter certaines limitations des modèles de distribution d'espèces dans des applications pratiques de biologie de la conservation ou dans leur utilisation pour évaluer l'impact potentiel des changements climatiques sur l'environnement. Le premier objectif majeur de mon travail est de contribuer à démontrer le potentiel des modèles de distribution d'espèces pour des applications pratiques en biologie de la conservation. Je propose une méthode pour générer des pseudo-absences qui permet de surmonter le problème récurent du manque de données d'absences fiables. Je démontre aussi, de manière théorique (par simulation) et pratique (par échantillonnage de terrain), comment les modèles de distribution d'espèces peuvent être utilisés pour modéliser et améliorer l'échantillonnage des espèces rares. Ces résultats démontrent le potentiel des modèles de distribution d'espèces comme outils pour des applications de biologie de la conservation. Le deuxième objectif majeur de ce travail est de contribuer à améliorer les projections d'impacts potentiels des changements climatiques sur la flore, en particulier dans les zones de montagnes. Je développe un modèle dynamique de distribution appelé MigClim qui permet de tenir compte des limitations de dispersion dans les projections futures de distribution potentielle d'espèces, et teste son application sur deux espèces virtuelles. Vu que le fait de prendre en compte les limitations dues à la dispersion demande des données supplémentaires importantes (p.ex. la distance de dispersion des graines), ce travail propose aussi une méthode de classification simplifiée des espèces végétales dans de grands "types de disperseurs", ce qui permet ainsi de d'obtenir de bonnes approximations de distances de dispersions pour un grand nombre d'espèces. Finalement, j'applique aussi le modèle MIGCLIM à un grand nombre d'espèces de plantes dans une zone d'études des pré-Alpes vaudoises. Les résultats montrent que les limitations de dispersion peuvent avoir un impact considérable sur la distribution potentielle d'espèces prédites sous des scénarios de changements climatiques. Cependant, quand les modèles sont utilisés pour évaluer les taux d'extinction d'espèces dans des zones de montages de taille limitée (évaluation régionale), il est aussi possible d'obtenir de bonnes approximations en considérant la dispersion des espèces comme illimitée, ce qui est nettement plus simple d'un point dé vue pratique. Pour terminer je présente la plus grande évaluation à fine échelle d'impact potentiel des changements climatiques sur la flore des montagnes conduite à ce jour. Cette évaluation englobe 12 zones d'études réparties sur toutes les chaines de montages principales d'Europe occidentale et centrale. Les résultats montrent que certaines chaines de montagnes (les Pyrénées et les Alpes Autrichiennes) sont projetées comme plus sensibles aux changements climatiques que d'autres (les Alpes Scandinaves et les Highlands d'Ecosse). Les résultats obtenus montrent aussi que les modèles à échelle fine projettent des impacts de changement climatiques (p. ex. taux d'extinction d'espèces) moins sévères que les modèles à échelle large. Cela laisse supposer que les modèles a échelle fine sont capables de modéliser des micro-niches climatiques non-détectées par les modèles à échelle large.

Interactions between respiration and systemic hemodynamics. Part II: practical implications in critical care.

In Part I of this review, we have covered basic concepts regarding cardiorespiratory interactions. Here, we put this theoretical framework to practical use. We describe mechanisms underlying Kussmaul's sign and pulsus paradoxus. We review the literature on the use of respiratory variations of blood pressure to evaluate volume status. We show the possibilities of attaining the latter aim by investigating with ultrasonography how the geometry of great veins fluctuates with respiration. We provide a Guytonian analysis of the effects of PEEP on cardiac output. We terminate with some remarks on the potential of positive pressure breathing to induce acute cor pulmonale, and on the cardiovascular mechanisms that at times may underly the failure to wean a patient from the ventilator.

Décubitus ventral et syndrome de détresse respiratoire aiguë de l'adulte: de la théorie à la pratique [Prone position mechanical ventilation in the Acute Respiratory Distress syndrome: theoretical and practical considerations].

Mortality of the acute respiratory distress syndrome (ARDS) remains extremely high and only few evidence-based specific treatments are currently available. Protective mechanical ventilation has emerged as the comer stone of the management of ARDS to avoid the occurrence of ventilation-induced lung injuries (VILI). Mechanical ventilation in the prone position has often been considered as a rescue therapy reserved to refractory hypoxemia. Since the publication of the PROSEVA study in 2013, early prone positioning for mechanical ventilation should be recommended to improve survival of patients with severe ARDS. In this article, both the theoretical and practical aspects of mechanical ventilation in prone position are reviewed.

Hallux valgus: résumé pratique [Hallux valgus: practical summary].

Hallux valgus arises from static and dynamic unbalance of the forefoot. Possibilities of conservative treatment are limited. This paper recalls the essentials of physiopathology and exposes briefly to the non-operators the most current amongst 150 operative techniques described. While becoming more complex, the surgery of hallux valgus has proven to gain efficiency. Surgical indication, based on biomechanics, should be very careful, which gives a greater interest to this discipline.

Validation of the Mini Nutritional Assessment short-form (MNA-SF): a practical tool for identification of nutritional status.

OBJECTIVE: To validate a revision of the Mini Nutritional Assessment short-form (MNA(R)-SF) against the full MNA, a standard tool for nutritional evaluation. METHODS: A literature search identified studies that used the MNA for nutritional screening in geriatric patients. The contacted authors submitted original datasets that were merged into a single database. Various combinations of the questions on the current MNA-SF were tested using this database through combination analysis and ROC based derivation of classification thresholds. RESULTS: Twenty-seven datasets (n=6257 participants) were initially processed from which twelve were used in the current analysis on a sample of 2032 study participants (mean age 82.3y) with complete information on all MNA items. The original MNA-SF was a combination of six questions from the full MNA. A revised MNA-SF included calf circumference (CC) substituted for BMI performed equally well. A revised three-category scoring classification for this revised MNA-SF, using BMI and/or CC, had good sensitivity compared to the full MNA. CONCLUSION: The newly revised MNA-SF is a valid nutritional screening tool applicable to geriatric health care professionals with the option of using CC when BMI cannot be calculated. This revised MNA-SF increases the applicability of this rapid screening tool in clinical practice through the inclusion of a "malnourished" category.

Information security and the enforcement of secrecy : the practical application of quantum key distribution

There is no doubt about the necessity of protecting digital communication: Citizens are entrusting their most confidential and sensitive data to digital processing and communication, and so do governments, corporations, and armed forces. Digital communication networks are also an integral component of many critical infrastructures we are seriously depending on in our daily lives. Transportation services, financial services, energy grids, food production and distribution networks are only a few examples of such infrastructures. Protecting digital communication means protecting confidentiality and integrity by encrypting and authenticating its contents. But most digital communication is not secure today. Nevertheless, some of the most ardent problems could be solved with a more stringent use of current cryptographic technologies. Quite surprisingly, a new cryptographic primitive emerges from the ap-plication of quantum mechanics to information and communication theory: Quantum Key Distribution. QKD is difficult to understand, it is complex, technically challenging, and costly-yet it enables two parties to share a secret key for use in any subsequent cryptographic task, with an unprecedented long-term security. It is disputed, whether technically and economically fea-sible applications can be found. Our vision is, that despite technical difficulty and inherent limitations, Quantum Key Distribution has a great potential and fits well with other cryptographic primitives, enabling the development of highly secure new applications and services. In this thesis we take a structured approach to analyze the practical applicability of QKD and display several use cases of different complexity, for which it can be a technology of choice, either because of its unique forward security features, or because of its practicability.

Développement, implémentation et utilisation pratique d'un score diagnostique [Development, implementation and practical use of a diagnostic score].

In recent years many clinical prediction rules (CPR) have been developed. Before a CPR can be used in clinical practice, different methodical steps are necessary, from the development of the score, the internal and external validation to the impact study. Before using a CPR in daily practice family doctors have to verify how the rules have been developed and whether this has been done in a population similar to the population in which they would use them. The aim of this paper is to describe the development of a CPR, and to discuss advantages and risks related to the use of CPR in order to help family doctors in their choice of scores for use in their daily practice.

Memoria de la elaboración de subtítulos para "Babel", primer capítulo de la serie documental Fry's Planet Word

This work tries to identify some of the skills an audio visual translator must develop, from a practical point of view, in order to pursue a career in this field, putting the stress on mastering subtitling-specific software. This report describes trial and error process during the making of the subtitles for a documentary and identifies some of the difficulties we might encounter while working on an assignment of this kind if we work with free licensing software. Moreover, it tries to contribute with some answers to these issues.

The effects of dominance, regular inbreeding and sampling design on Q(ST), an estimator of population differentiation for quantitative traits.

To test whether quantitative traits are under directional or homogenizing selection, it is common practice to compare population differentiation estimates at molecular markers (F(ST)) and quantitative traits (Q(ST)). If the trait is neutral and its determinism is additive, then theory predicts that Q(ST) = F(ST), while Q(ST) > F(ST) is predicted under directional selection for different local optima, and Q(ST) < F(ST) is predicted under homogenizing selection. However, nonadditive effects can alter these predictions. Here, we investigate the influence of dominance on the relation between Q(ST) and F(ST) for neutral traits. Using analytical results and computer simulations, we show that dominance generally deflates Q(ST) relative to F(ST). Under inbreeding, the effect of dominance vanishes, and we show that for selfing species, a better estimate of Q(ST) is obtained from selfed families than from half-sib families. We also compare several sampling designs and find that it is always best to sample many populations (>20) with few families (five) rather than few populations with many families. Provided that estimates of Q(ST) are derived from individuals originating from many populations, we conclude that the pattern Q(ST) > F(ST), and hence the inference of directional selection for different local optima, is robust to the effect of nonadditive gene actions.

Biases of the ordinary least squares and instrumental variables estimators of the intergenerational earnings elasticity : revisited in the light of panel data

The OLS estimator of the intergenerational earnings correlation is biased towards zero, while the instrumental variables estimator is biased upwards. The first of these results arises because of measurement error, while the latter rests on the presumption that the education of the parent family is an invalid instrument. We propose a panel data framework for quantifying the asymptotic biases of these estimators, as well as a mis-specification test for the IV estimator. [Author]