A0599
Title: Independent Metropolis sampler without rejection
Authors: Florian Maire - Universite de Montreal (Canada) [presenting]
Abstract: In Bayesian statistics, approximating the posterior distribution of the model parameters is key to multiple methods of estimating posterior expectations. We consider the situation where (i) simulating i.i.d. samples from the approximate posterior is doable but computationally expensive and (ii) evaluating the unnormalized posterior density is also computationally expensive. We show that there exists a modification of the Independent Metropolis sampler which is particularly useful in such contexts: central to this construction is the idea that proposed candidates cannot be rejected but rather always accepted at the ``right'' time in the dynamic of the Markov chain. We will illustrate the benefit of such an algorithm for the problem of identifying therapeutic strategies that bolster antitumor immunity. In this model, the likelihood evaluation involves solving a system of ordinary differential equations that has to be numerically integrated.
