61
reads

Shinto Eguchi 
2010-04-24 
13:30 - 14:30 
308 , Mathematics Research Center Building (ori. New Math. Bldg.)

It is pointed that statistical inference for observational studies is often suffered with selection bias, which is impossible to adjust from any evidence from only the real data set.  Missingness, censoring, hidden confounder easily destroys the statistical performance for the standard methods because they impose for a true distribution to keep away from the statistical model assumed. We elucidate the distributional behavior perturbed by non-randomness from a viewpoint of information geometry.  A tubular neighborhood of the statistical model is modeled to cause undetectably small selection bias.  We propose to make a confidence interval resistant from any possible selection bias.