user2016

ASA Grant Award Recipient

The use of realistic statistical models for complex data is often hindered by the high cost of conducting inference about the model parameters. Because of this, it is sometimes necessary to use approximate inference methods, even though the impact of these approximations on the fitted model might not be well understood. I will discuss some practical examples of this, demonstrating how to fit various Generalized Linear Mixed Models with the R package glmmsr, using a variety of approximation methods, with a focus on what difference the choice of approximation makes to the resulting inference. I will talk about some more general issues along the way, such as how we might detect situations in which a given approximation might give unreliable inference, and the extent to which the choice of approximation method can and should be automated. I will finish by briefly reviewing some ideas about how best to share and discuss challenging models and datasets which could motivate the development of new approximation methods.