Biais d'autosélection | EPFL Graph Search

En statistiques, un biais d'autosélection peut se produire lorsque les individus formant les groupes que l'on cherche à comparer n'ont pas été assignés aléatoirement mais ont eux-mêmes choisi à quel groupe ils appartenaient. Par exemple, dans un essai clinique concernant l'efficacité de deux traitements (A vs. B), on prendra soin d'éviter que les individus du traité par A n'avaient pas au préalable des caractéristiques qui les ont fait opter pour A plutôt que B, par exemple, si leur état était moins sévère, on pourrait à tort conclure que A est plus efficace que B. Cette source d'erreur est particulièrement critique dans les situations où l'on peut supposer que certaines caractéristiques des individus peuvent les avoir incités à choisir l'un des groupes plutôt que les autres, comme cela peut être le cas en sociologie, en psychologie ou en économie. Le biais d'autosélection aboutit donc à un échantillon biaisé de la population étudiée. En sociologie, ces termes désignent aussi le pro

Getting cited: does open access help?

Patrick Gaulé

We reexamine the widely held belief that free availability of scientific articles increases the number of citations they receive. Since open access is relatively more attractive to authors of higher quality papers, regressing citations on open access and other controls yields upward-biased estimates. Using an instrumental variable approach, we find no significant effect of open access. Instead, self-selection of higher quality articles into open access explains at least part of the observed open access citation advantage.

2008

Three essays on access to knowledge

Patrick Gaulé

The first chapter of this thesis, coauthored with Annamaria Conti, seeks to add an economic contribution to the current debate on using university licensing contracts to improve access to medicines in developing countries. We build a simple model in which we have a university licensing out an academic invention to a profit-maximizing pharmaceutical company. We compare three different types of licensing contracts that the university might use to enhance access to pharmaceuticals in the South: (1) an exclusive license limited to the North; (2) an exclusive license worldwide with a price cap in the South; and (3) an exclusive license worldwide with a price cap in the South and a clause specifying that the licensee would lose its exclusivity in the South if it does not supply the Southern market. We show that in a simple model with asymmetric information on production costs the latter type of contract dominates the two others. In the second chapter, coauthored with Nicolas Maystre, we reexamine the widely held belief that free availability of scientific articles increases the number of citations they receive. Since open access is relatively more attractive to authors of higher quality papers, regressing citations on open access and other controls yields upward-biased estimates. Using an instrumental variable approach, we find no significant effect of open access. Instead, self-selection of higher quality articles into open access explains at least part of the observed open access citation advantage. The third chapter uses an evidence-based approach to assess the difficulties faced by developing country scientists in accessing the scientific literature. I compare backward citations patterns of Swiss and Indian scientists in a database of 43'150 scientific papers published by scientists from either country in 2007. Controlling for fields and quality with citing journal fixed effects, I find that Indian scientists (1) have shorter references lists (2) are more likely to cite articles from open access journals and (3) are less likely to cite articles from expensive journals. The magnitude of the effects is small which can be explained by informal file sharing practices among scientists.

EPFL2009

Transforming cumulative hazard estimates

Pal Christie Ryalen, Mats Julius Stensrud

Time-to-event outcomes are often evaluated on the hazard scale, but interpreting hazards may be difficult. Recently in the causal inference literature concerns have been raised that hazards actually have a built-in selection bias that prevents simple causal interpretations. This is a problem even in randomized controlled trials, where hazard ratios have become a standard measure of treatment effects. Modelling on the hazard scale is nevertheless convenient, for example to adjust for covariates; using hazards for intermediate calculations may therefore be desirable. In this paper we present a generic method for transforming hazard estimates consistently to other scales at which these built-in selection biases are avoided. The method is based on differential equations and generalizes a well-known relation between the Nelson–Aalen and Kaplan–Meier estimators. Using the martingale central limit theorem, we show that covariances can be estimated consistently for a large class of estimators, thus allowing for rapid calculation of confidence intervals. Hence, given cumulative hazard estimates based on, for example, Aalen’s additive hazard model, we can obtain many other parameters without much more effort. We give several examples and the associated estimators. Coverage and convergence speed are explored via simulations, and the results suggest that reliable estimates can be obtained in real-life scenarios.

2018

Three essays on access to knowledge

Patrick Gaulé

EPFL2009