Statistics Seminars: Estimating shared distribution shape from multiple samples
22 February 2010 14:15 in CM221
Motivated by two potentially important applications, we will examine a
related idealised methodological problem.
Suppose that we can obtain (not very large) samples from a number of
populations and we hypothesise that the populations have differing
means and standard deviations but share the same underlying unknown
shape for the probability density function. This generalises the
common modelling assumption that each population is normally
distributed, i.e. that the underlying shape is the standard normal.
We will show how to construct a suitable model and apply Markov Chain
Monte Carlo (slight variant on Gibbs sampling) to carry out Bayesian
inference for the underlying shape. Some simulated examples will
explore the limitations of learning in terms of the individual and
overall sample sizes and features of the underlying shape.
Contact firstname.lastname@example.org for more information