Fit a hierarchical logistic regression with AUs nested in watersheds. Outcome: binary impairment status per AU per biennial cycle. AU-level predictors: water type, AU size, lagged impairment status (autoregressive). Watershed-level predictors: enforcement action density (EEADP_Enforcement aggregated by municipality→watershed per 2-year window), CSO discharge volume, population density (Census ACS). State-level temporal predictor: DEP staff FTE as budget proxy, lagged one cycle.
Expected findings: the autoregressive term will dominate (94% persistence → large log-odds). Cause type and water type will be the strongest structural predictors — bacterial impairments are ~3× more likely to resolve than non-bacterial; lakes almost never improve. Enforcement/budget effects will likely be small and have wide posteriors, because the binding constraint on delisting is physical tractability of the impairment cause (invasive plants and legacy mercury are essentially irreversible on decadal timescales), not regulatory pressure. The most credible result may be a well-quantified null.
Requires PySTAN and full conda env; not suitable for CI.
The April 2025 EPA statewide pathogen TMDL formally covered ~679 bacterial Category 5 AUs. Infrastructure investment (septic upgrades, sewer extension, CSO controls) will drive eventual delisting — the question is timing. Use the 2024/2026 cycle (expected ~2027) as the first post-TMDL data point, then track the bacterial cohort across subsequent cycles. Compare against the pre-2025 delisting rate (2.8% per cycle) to estimate whether TMDL coverage accelerates infrastructure deployment or whether the lag is long enough that impairment rates barely budge for a decade. Cross-reference with CSO capital spending records and permit compliance timelines where available.
The EJ/EJSCREEN correlation analyses and CSO map scripts are slow due to shapefile
loading and per-feature spatial joins. Consider pre-simplifying geometries, caching
dissolved boundaries, or switching to vectorized geopandas.sjoin.