Replication package for Administrative Flood-Map Records and Flood-Insurance Trajectories in the United States

# JFRM results replication package This package reproduces the results for Administrative Flood-Map Records and Flood-Insurance Trajectories in the United States. It is organised for applied researchers: the processed analysis data are included, one command rebuilds every reported table and figure, and a separate guide explains how the processed files can be reconstructed from public agency data. ## Reproduce the article and Supporting Information From the package directory, create the supplied environment: ```bash conda env create -f environment.yml conda activate jfrm-replication ``` Then run: ```bash python run_replication.py ``` The command rebuilds 22 publication-numbered CSV tables in `results/tables/` and 11 PNG figures in `results/figures/`. These are Tables 1--4 and Figures 1--3 in the article, together with Tables S1--S13 and Figures S1--S8 in Data S1. Tables with lettered panels, such as S3A and S3B, are separate CSV files. The repository-status row in Table S1 is adapted to describe this self-contained release; the statistical entries reproduce the reported version. The program replaces the existing `results/` directory only after a successful run. Runtime depends on the computer; on a recent multi-core machine the full estimation route generally takes about 5--15 minutes. At least 8 GB of memory is recommended. ## What is included - `data/processed/tract_year_panel.parquet`: the balanced analysis panel, with 239,120 tract-years for 34,160 tracts in California, Florida, Louisiana, New Jersey, New York, North Carolina, South Carolina and Texas, 2018--2024. - `data/processed/map_record_timing.parquet`: tract-level FEMA administrative map-record dates and footprints. - `data/processed/tract_geometry.parquet`: display-only 2020-tract geometry for Figure 1, projected to EPSG:5070 and simplified to 500 metres. It is not used for the underlying overlays, which use unsimplified TIGER/Line geometry. - `data/processed/nfip_source_rows_2022_2024.parquet`: 11,406 NFIP source tract-years that require the official Census 2010-to-2020 relationship allocation, plus six zero-valued direct keys needed to reconstruct the source support exactly. - `data/processed/census_relationships/`: the official Census relationship rows used for that bounded reallocation. - `code/`: the estimators and publication-output code called by `run_replication.py`. - `data/variable_dictionary.csv`: definitions, units and missing-value rules for the principal processed fields. - `data_download_and_processing.md`: public-source download locations, vintage limitations and the optional raw-data route. The small NFIP source file complements the panel. Its six zero-valued direct keys repeat identifiers but no outcome mass; the other rows are absent from the linked panel geography. The replication program combines these records in memory before applying the Census relationship allocation. ## Empirical scope The observed event is the first retained FEMA administrative map record in 2018--2024, pooled across LOMR, FIRM-panel and panel-revision records. The preferred sample also requires at least 1% Special Flood Hazard Area in the later current NFHL snapshot. This is a trajectory design, not a measure of the direction or magnitude of a historical flood-zone change. NFIP outcomes are annual tract aggregates. The principal HMDA application universe uses action codes 1--5; approval and denial shares use decided applications, action codes 1--3. The panel gives equal weight to tract-years and is a purposive eight-state sample rather than a national probability sample. The two noncentral county-year pricing sensitivities follow the specification printed in Data S1: their 2018 volume-control vector includes the all-retained HMDA record count. All principal HMDA count and share estimates use the action-defined universes above. ## Reading the result files - Tables 1 and S1--S5 document data construction, treatment support and sample scope. Table 2 reports the three NFIP outcomes; Table 3 collects scale, margin, RR2 and common-sample sensitivities; Table 4 reports the five action-defined HMDA outcomes. Tables S6--S12 provide the detailed timing, transformation, robustness, comparability and pricing evidence, while Table S13 maps the earlier display set into the revised presentation. - In Tables 2, 4, S6A and S6B, count-outcome estimates are log(1+x) points, HMDA share estimates are percentage points, and pricing estimates are native percentage points. Table 3 identifies the unit of every row explicitly. - Figures 1--3 show the map-record year, NFIP record-relative paths and controlled/common-sample NFIP estimates. Figures S1--S8 show support, pseudo-date, trajectory, transformation, RR2 and overlap-weighting diagnostics. ## Reconstructing the processed data The included processed route is the route for reproducing the reported numerical results. The optional public-source route applies the documented construction definitions to agency files available when it is run. Follow `data_download_and_processing.md`, then run: ```bash python code/prepare_data.py --raw-dir data/raw python run_replication.py ``` Agency downloads are periodically revised or retired. A later public-source reconstruction may therefore differ from the frozen processed files even when the same definitions are used. The downloader selects the current NOAA annual revisions and records their vintages in the source filenames. Use the included processed data to reproduce the article exactly; use the public-source route to inspect or update data construction and interpret any changed estimates as a data-vintage update. ## Licence Component-level reuse terms are stated in `LICENSE.txt`. Citation information is supplied by the repository record.