Student Seminars

Dao Nguyen
Department of Statistics, University of Michigan

Title: A hybrid second order iterated smoothing approach

¡°Plug and play¡± inferences, also known as ¡°derivative free¡± or ¡°likelihood free¡± inferences are receiving great attention recently due to the fact that in many practical problems, the likelihood is intractable to compute directly. The attractive properties of these methods are that unobserved process enters the algorithm only through the requirement that realizations can be generated at arbitrary parameter values. In the same line with plug and play approach, this paper introduces a hybrid second-order iterated smoother. To reduce the high variance of these estimators, we use fixed lag smoother. To improve the speed of convergence, an approximation of the observed information matrix is also proposed. While enjoying greater convergence rate, most observed information matrix approximation are computational expensive, especially in plug and play approaches. Therefore, to relax the intensive computation, we use Newey-West covariance estimator for a few initial iterations before adapting sequential Monte Carlo approximations of the variance. Due to the special structure of iterated smoothing, we also bypass the sequential Monte Carlo approximations of the variance by using the last estimated smoother value at the beginning of the next filter iteration. In a toy example and in a challenging inference problem of fitting a malaria transmission model to time series data, we find substantial gains for our methods over current alternatives.

 

Student Seminar Archive

For questions regarding the Statistics Student Seminar or if you are interested in presenting, please contact Joonha Park (joonhap@umich.edu) or Jun Guo (guojun@umich.edu).