318 Hanes Hall, CB #3260 Chapel Hill, NC 27599-3260
919-962-1329
Loading Events

BIOS/STOR Student Seminar: Sam Rosin, BIOS and Alex Murph, STOR

October 29 @ 3:30 pm - 4:30 pm

Sam Rosin
PhD Student in BIOS
Advisor: Michael Hudgens

Estimating Seroprevalence of SARS-CoV-2
During the COVID-19 pandemic, governments have used
seroprevalence studies to estimate the proportion of persons
within a given population who have antibodies to SARS-CoV-2.
However, serologic assays are prone to false positives and
negatives, and non-probability sampling methods may induce
selection bias. Our work considers seroprevalence estimators
that address both challenges by leveraging validation and
covariate data. We study the estimators in simulations and apply
them to SARS-CoV-2 seroprevalence studies in North Carolina
and Belgium from 2020.

Alexander Murph

PhD Student in STOR
Advisors: Jan Hannig and
Jonathan P Williams

Bayesian Change Point Detection for Mixed Data
with Missing Values
Many models in-production assume that incoming data follows the same
distribution of the data on which the model was trained. In practice, however,
this distribution will often not remain static. We develop a change-point
detection system on data over time to locate major changes in the underlying
data distribution. This system has a strong practical motivation, since it can be
used to warn data scientists that an in-production model’s performance may
drop: a direct consequence of the training data no longer representing the
current data. This can drive decisions such as when to retrain a model or when
to investigate discrepancies in the data.
We will use a real-world example (the Mayo Clinic’s Palliative Care model) to
motivate the need for this system. Our solution takes a hierarchical Bayesian
approach that learns a vector of regimes as an unknown parameter. These
regime vectors are assumed to come from a Hidden Markov Model (HMM);
given a regime vector, data within a regime are modeled using random graphs
with a varying edge set. Missing values and mixed variable types are handled
using a Bayesian latent variable approach; parameters are learned using a
variation of the Split-Merge algorithm and a Double Reversible Jump
Metropolis-Hastings algorithm.

Share This Event

  • This event has passed.

Details

Date:
October 29
Time:
3:30 pm - 4:30 pm
Event Category:

Venue

133 Rosenau Hall
Rosenau Hall
Chapel Hill, NC 27516 United States
+ Google Map