Case study
Causal Inference of Wildfire Impact
Quantifying the "economic blast radius" of wildfire events by tying NIFC perimeters to SafeGraph visit streams, correcting for local shocks, seasonality, and staggered treatment timing.
Objective
Translate raw satellite perimeters and anonymized mobility data into actionable metrics that show how far, how quickly, and for how long business activity is depressed after a wildfire alarm.
- Generate concentric spatial buffers for every 2018–2024 wildfire in CONUS.
- Attach SafeGraph weekly visits for POIs falling inside each buffer band.
- Construct an estimator that respects staggered adoption + spatial spillovers.
Data engineering
PostGIS and PySpark pipelines orchestrate raster-to-vector conversions, dissolve perimeters, and index concentric buffers (0–2/2–5/5–10/10–25 km). I matched 1.3M POI-week observations to the closest fire perimeter and fused NOAA drought, AQI, and census controls to guard against correlated shocks.
- Spatial joins accelerated with quad-tree indexing and tiled ingestion.
- Seasonality removed via POI-class × DMA fixed effects and holiday dummies.
- Quality checks ensured <2% orphaned POIs per weekly pull.
Modeling approach
Estimated a Poisson Pseudo-Maximum Likelihood difference-in-differences with an event-study structure to absorb anticipation and persistence. Errors clustered by fire ID × DMA, and placebo alarms confirmed the design was well-behaved.
- ATT shows 35–42% visit declines inside the 0–2 km band immediately post alarm.
- 2–5 km and 5–10 km rings briefly spike (+8%) as shoppers substitute outward.
- Most zones normalize within 24 weeks; outer 10–25 km rings stay neutral.
Applications
Results feed city planning dashboards that flag vulnerable commerce corridors and help insurance partners price business-interruption products with more granularity.
- Supports rapid allocation of SBA emergency loans based on observed losses.
- Offers retailers a playbook for pop-up services in substitution bands.
- Roadmap: extend to labor outcomes (Homebase shift data) and property damage.