← Back to research
Robust Inference / Anti-trafficking application

Robust Bayesian Methods for Quasi-Sparse Counts

Treating hard-to-measure counts — like human trafficking — as a quasi-sparse modeling problem, and testing robust Bayesian priors such as the Rescaled Beta and Gauss Hypergeometric frameworks against standard count models.

Field
Bayesian methods & count modeling
Status
Active research
Origin
SMU · DHS collaboration
Core tools
Robust priors · Count models

The problem

Counts of hard-to-measure phenomena — like human trafficking — are quasi-sparse: dominated by zeros and small numbers, punctuated by occasional larger values, and recorded with heavy uncertainty and under-reporting. Standard count models such as Poisson or negative binomial can be surprisingly fragile here. A few unusual observations, or a poorly chosen prior, can pull estimates around and overstate confidence.

The question this project pursues: how do you estimate rates and effects from such data in a way that stays robust — stable under outliers, sparsity, and prior misspecification?

Quasi-sparse, defined

Lots of zeros and small counts, a few larger ones, and real measurement uncertainty — exactly the regime where default priors tend to mislead.

The approach

The core idea is to replace fragile default priors with robust Bayesian alternatives whose tail behavior is designed to resist being dominated by a handful of extreme observations.

Rescaled Beta and Gauss Hypergeometric priors

These families offer heavier, more flexible tails than conjugate defaults, letting the data — not the prior — drive conclusions when the two disagree. The project benchmarks them head-to-head against standard count models to see how each behaves as data grow sparser and noisier.

Grounded in real anti-trafficking work

The motivation comes from collaborative work originating at Southern Methodist University, alongside the Department of Homeland Security and North Texas law enforcement, on agricultural labor trafficking — including geospatial tools that refine search areas and a review of the push and pull factors behind international trafficking. Reliable estimates from sparse, sensitive counts directly affect where limited investigative resources are sent.

Why it matters

Robust priors yield estimates that degrade gracefully when data are thin and uncertain: fewer wild swings, better-calibrated uncertainty, and conclusions that hold up under scrutiny. Although motivated by trafficking, the same quasi-sparse problem recurs across public health, ecology, and rare-event modeling.