The problem
Counts of things nobody can measure directly are quasi-sparse. Mostly zeros and small numbers, a few larger values, and heavy under-reporting throughout. Human trafficking counts look exactly like this.
Standard spatial models borrow strength from neighboring places, which helps where the data are thin. The cost is oversmoothing. A real local spike gets pulled back toward its surroundings, and for trafficking the spikes are the point.
Lots of zeros and small counts, a few large ones, and real measurement error.
The data
I lead this project, which combines FBI National Incident-Based Reporting System incident data with American Community Survey covariates and Office of Refugee Resettlement records on unaccompanied children.
The approach
Global-local shrinkage priors pair a global parameter, which pulls every estimate toward a shared structure, with local parameters that let individual estimates escape when the data insist. I build that idea into Bayesian spatiotemporal models in R, so estimates borrow strength across places and years where the data are quiet and genuine hotspots stand out where they are not.
Where the problem came from
I started on trafficking at Southern Methodist University in 2024, where I built a Python Dash application that narrows the search area for agricultural labor trafficking sites. It turns a survivor’s escape location, travel time, and reported crop into a map of candidate farms using CroplandCROS imagery and OpenStreetMap road data.
I refined it with feedback from contacts at the Texas Attorney General’s Office and the Homeland Security North Texas Trafficking Task Force. It placed second among forty presenters at the program’s summer symposium and remains in use by SMU’s Human Trafficking Data Research team.
Why it matters
Estimates from sparse counts help decide where limited investigative and support resources go. A model that smooths away a real hotspot points those resources in the wrong direction. Trafficking is the application that drives this work, but the same shape of data shows up across public health, ecology, and rare-event monitoring.