Location data carries an aura of certainty. A coordinate with eight decimal places is precise, the device really was there. But it is also partial: each observed device represents an unknown and variable number of people who are not in the data at all. The penetration rate, the share of the real population captured in the device panel, differs across regions, demographic groups, and even from day to day. Point maps and heat maps built directly from the raw trace represent something real and useful, but they quietly present a sample as if it were the whole.

This talk presents an approach that treats the data as what it is, a sample to be modeled, and reports a calibrated probabilistic model rather than the underlying locations. The output is an estimate over an area and a time window, with stated uncertainty. In recent work estimating visitation to public parks, for instance, the model produces probable visit counts for a defined area of interest on a given day, and simultaneously attributes a probabilistic demographic profile of those visitors by linking observed devices to the demographics of their inferred home locations. The result is reported as a distribution, never as an individual's trace.

This does two things at once. For jurisdictions where data precision is now legally constrained, it offers a privacy-preserving way to recover analytical value. And even where full-resolution data remains available, it produces a more truthful picture of behavior. Consider a park where, on a typical day, one in ten visitors carries a tracked device. If on a single day that share happens to be five in ten with no real change in attendance, the raw data shows a dramatic visitation spike that never occurred. Modeling the sample, with awareness that coverage fluctuates and that genuine special events also exist, smooths this artifact and separates signal from sampling noise.

We describe the pipeline at a high level. Observed devices are weighted by penetration rate, correcting for how many people each device represents, with the correction grounded in the demographics of inferred home locations and varying by region. A dasymetric mapping step then redistributes activity within each zone according to land cover and feature type, so estimated visits land where people actually go rather than being spread uniformly across a block group. A central component is a national sampling effort spanning roughly 3,000 census block groups, a dataset on the order of hundreds of billions of records, which lets these corrections be empirically grounded rather than assumed.

As a demonstration, we apply the model to a question that is difficult to study without compromising privacy: how do communities respond to evacuation orders? The probabilistic approach lets us compare response across like events, multiple wildfires against one another and multiple hurricanes against one another, where factors such as onset speed, warning lead time, and a hurricane's shifting projection cone vary. We also examine whether response differs across demographic groups, framed as a question about the reach of public warning systems rather than about the communities themselves, and we test whether any apparent signal is real or an artifact of variable sampling, treating a null result as equally informative.

We close on policy. A growing set of state laws, including Maryland's MODPA, Oregon's amended Consumer Privacy Act, and Virginia's recently enacted ban, now restrict precise geolocation data using a common 1,750-foot threshold. In practice, compliant coarsening often discards analytical value without clearly improving privacy. Reporting calibrated models rather than raw locations suggests how to honor the intent of these laws while preserving the insight public agencies depend on.

The workflow is built on free and open source foundations, including H3 spatial indexing, distributed processing with PySpark on Spark, and Bayesian estimation in PyMC. Attendees will leave with a transferable way to think about mobility data: report the calibrated model, not the individual, for both better science and stronger privacy.