Description
Statistics and Probability seminar.
Speaker: Nilanjan Chatterjee.
Affiliation: Johns Hopkins University.
Date and Time: Wednesday 26 February, 11:30 am - 12:30 pm.
Venue: Ramanujan Hall, Department of Mathematics.
Title: Statistical model building using data fusion.
Abstract: Many applications require development of complex statistical
models involving many variables. It may not, however, be possible to train
such model in a single dataset of adequately large sample size that has
measured all the variables. Instead, data may be available across multiple
studies, where any individual study may not measure all the variables, but
the different studies altogether cover all the variables. In this talk, I
will describe how to fit popular non-linear models, such as logistic
regression models, by combining information from such multiple disparate
data sources. In fact we will show it is possible to fit such models only
using "summary-level" information, i.e. estimates of parameters from
fitted simpler models, from individual studies and thus overcoming some
of the logistical and ethical issues related to sharing of individual
level data across studies. Methods will be illustrated through extensive
simulation studies and real data examples