POLISY Lab: technical report

Version 0.2.0 of polisy_lab, 28 September 2026. Research theme 1: Political Ideology × AI × Innovation.

This report explains how to run the lab in Google Colab, what each stage does, which datasets it uses and how they are linked, what the outputs mean, what the first run found, and how to rebuild and publish the website. The interactive lab's Methods tab shows this same text. The research argument (what is new, what is not, and the research agenda) is in the research report, site/index.html, built from polisy_lab/report/report.md.

1. What POLISY is

POLISY (Politics, Organizations, Leadership, Strategy & Innovation) is a research lab built as code. It links VRscores measures of workforce partisanship (the share of an employer's, occupation's, industry's or metro's workers registered or modelled as Republican) to open data on AI exposure, AI adoption, migration, state policy and innovation. It then keeps every pattern it finds as a graded finding: the claim, the numbers behind it, the research question it raises and the data that would settle it.

Three rules shape the design:

Nothing is hard-wired to a file name. Every input is found by what its name looks like and what its header contains, inside folders or inside zip files.
Every join is measured. Each link between two datasets records how much matched, so a result that rests on half the data says so.
The website is generated. Analyses register findings, charts and tables; the site draws whatever is registered. Adding an analysis never requires touching the site code.
Headline findings are stress-tested. A stress-test module re-estimates each headline finding against its first rival explanation and re-grades it (Section 7).

2. What you need

The lab runs in a Colab notebook (notebooks/POLISY_AI_Innovation_Lab.ipynb, or the same code as a script in notebooks/polisy_lab_colab.py). It needs:

The code: the polisy repository (https://github.com/mitdesai01/polisy), which holds POLISY_DA/ (the file finder polisy_core.py and modules 01-10) and lab/ (the polisy_lab package, this report, the notebook and tools). On GitHub, Code → Download ZIP gives it as one zip (polisy-main.zip); POLISY_lab_code.zip holds the same two folders.
Your data files, under any names, in one Drive folder (for example MyDrive/POLISY/data). The first run used these:

Dataset	Files used in the first run	Needed for
VRscores	`vrscores_report.html` (the report generated 22 Sep 2026), or the canonical Parquet files POLISY_DA writes	all partisanship measures
AIOE	fetched from GitHub by the lab	occupation, industry and county AI exposure
DAIOE v1.0.0	`daioe-v1.0.0-scores.zip`	dynamic exposure and the comparison of exposure measures
IRS SOI county migration	`IRS_SOI_County_Migration.zip` (Inflow and Outflow folders, 2012-13 to 2021-22)	migration findings
Correlates of State Policy	`cspp_data_2026-09-24_academic and policy addition.csv` (in a zip)	state political and policy environment

Optional sources the lab uses when present: county presidential returns and the Census CBSA delineation file (found by POLISY_DA), the BTOS data downloads (fetched automatically in Colab), IRS state-to-state files (fetched automatically) and the four PatentsView tables (deferred in this version). The stress tests use three small files from GitHub, which the fetch stage downloads: county context (counties.json), telework shares by occupation, industry and metro (Dingel and Neiman), and county presidential returns for 2016-2024 (used only when POLISY_DA has no MIT file).

3. Run it in Google Colab, step by step

Step 1. Put the files in Drive

Create MyDrive/POLISY/ with a data/ folder inside. Put your downloads in data/ as they came: zips do not need unpacking and names do not need changing. Upload the code zip (polisy-main.zip from GitHub, or POLISY_lab_code.zip) to MyDrive/POLISY/ or to the Colab file panel, as it is. Without a zip, the notebook clones the repository (a private repository needs a token in the URL).

Step 2. Open the notebook and run Setup

Open POLISY_AI_Innovation_Lab.ipynb in Colab and run the first cell. It installs the few packages Colab lacks (DuckDB, Polars, RapidFuzz, naics, py7zr, markdown) and mounts Drive.

Step 3. Load the code and point the lab at your files

The second cell unzips the code into /content/polisy_code, loads polisy_core and polisy_lab, and prints both versions. It sets:

POLISY_LAB_ROOT = MyDrive/POLISY/lab: everything the lab writes (kept in Drive, so it survives the Colab session);
pc.CONFIG["SEARCH_DIRS"] = your data/ folder first, then /content.

If the inventory later picks the wrong file for an input, point it at the right one, for example pc.CONFIG["LAB_CSPP_DATA"] = "/content/drive/MyDrive/POLISY/data/cspp.csv". The key is LAB_<SOURCE>_<ROLE>; source and role names appear in the inventory table.

Step 4. Run the stages

One cell per stage, so you can stop and inspect:

fetch downloads what is open and missing: the AIOE repository, DAIOE from Zenodo, BTOS and CSPP files from their pages, and IRS state and county files. Anything already found is skipped.
inventory lists, for every source and role, how many files were found and which.
profile samples each file and records its columns, fill rates and example values.
adapt turns each source into canonical tables (one Parquet file per table, one declared grain each).
link joins the canonical tables into five panels and records the match rates.
analyze runs every analysis module and registers findings, charts and tables.
site writes the website.

A full run on the first run's files takes about a minute after downloads.

Optional first step: if you have the raw VRscores files and want the richer panels (occupation and metro panels by year, employer histories), run POLISY_DA modules 01 to 04 first (pc.run("01") ... pc.run("04"), or the POLISY_Lab notebook). Run module 05 too: it downloads the Census CBSA delineation file and O*NET titles, which let the lab link metros to CBSAs and occupations by more titles. Without them the lab reads the VRscores report instead and skips metro-level AI exposure.

Step 5. Look at the results

The notebook prints the inventory, the link rates and the findings table, and serves the site inside Colab (a link opens it in a new tab). Everything is also on Drive under POLISY/lab/.

Step 6. Publish on GitHub Pages

Copy the folder POLISY/lab/site/ into a GitHub repository, for example as docs/. In the repository go to Settings, then Pages, choose "Deploy from a branch", pick the branch and the /docs folder, and save. The site appears at https://<user>.github.io/<repository>/ a minute later. index.html is the research report: static figures and tables, no scripts, fonts from Google Fonts with system fallbacks. lab.html is the interactive lab, which loads Plotly from a public CDN; polisy_lab_offline.html carries Plotly inside and opens without internet.

4. The pipeline, stage by stage

All paths below are relative to the lab root (POLISY_LAB_ROOT).

Stage	Code	Writes
fetch	`sources.fetch_all`	`raw/` (downloads only; your own files are never moved)
inventory	`sources.inventory`	the data catalog in `results/results.json`
profile	`sources.profile_all`	`results/profiles/profiles.json`
adapt	`adapters/*.py`	`canonical/*.parquet`
link	`link.py`	`panels/*.parquet`, linkage diagnostics
analyze	`analyses/*.py`	findings, views and datasets in `results/results.json`, `results/findings.csv`, `results/linkage_diagnostics.csv`
site	`site.py` with `site_assets/`	`site/index.html`, `site/polisy_lab_offline.html`, `site/data/`

Every stage survives a missing source. An analysis whose data are absent registers a "needs data" finding naming the files it wants, so the site always shows what the lab is waiting for. Running only some stages keeps the earlier stages' results: re-running analyze does not erase the catalog or the file profiles.

Discovery: how files are found

Each source in sources.SOURCES declares its roles (for example the IRS source has county and state). Each role gives a regular expression for file names (tested on a cleaned name: lower case, copy markers such as "(1)" removed), the accepted extensions, and header tokens that must appear in the first rows. A zip is opened and every table inside is tested on its own name, so IRS_SOI_County_Migration.zip yields its 20 yearly files. The inventory prints what was found for each role.

Canonical tables

Adapters write one table per concept, each with a declared grain (core.GRAINS); a duplicate key raises a warning in the log. The main tables:

Table	Grain	From
`occ_exposure`	SOC 2010 occupation	AIOE, rebuilt from abilities and the application matrix (matches the published scores at r = 0.99999)
`ind_exposure`, `county_exposure`, `state_exposure`	NAICS-4, county, state	AIOE (AIIE, AIGE)
`oes_staffing`, `qcew_area_total`, `qcew_area_industry`	industry x occupation; area; area x industry	AIOE repository (OES staffing, QCEW 2019)
`occ_daioe`	SOC 2010 x year, 2010-2024	DAIOE 2024 refresh: levels, subdomains and within-year z-scores
`occ_soc2018_panel`, `occ_measures_soc2018`	SOC 2018 x year; SOC 2018	DAIOE SOC 2018 panel, with the comparison measures
`vr_occupation_year`, `vr_industry_year`, `vr_metro_year`, `vr_state_year`, `vr_employer_summary`, `vr_sorting`	as named	VRscores (POLISY_DA panels, else the report)
`irs_county_year`	county x year	IRS: households, people and income moving in, out and staying, with rates
`irs_state_year`	state x year	the counties' interstate totals
`irs_flows_county`, `irs_flows_state`	origin x destination x year	IRS county pairs; states summed from them (or IRS state files)
`cspp_state_year`, `cspp_catalog`, `cspp_state_profile`	state x year; variable; state	CSPP
`geo_state`, `geo_county`	state; county	lookup tables, CBSA delineation when available

5. Datasets

VRscores (Kagan, Frake and Hurst, Organization Science 2026). Workforce partisanship from voter files matched to employment profiles. The lab reads POLISY_DA's canonical panels when they exist; otherwise it parses the figures inside the VRscores HTML report: occupations (2024), industries with their 2012-2024 change, 359 metros by year, states, 6,000 large employers and the sorting indices. Partisanship is party registration where states register party, and primary participation or L2's model elsewhere; several findings are about this difference.

AIOE, AIIE, AIGE (Felten, Raj and Seamans 2021). Exposure of occupations, industries and counties to ten AI applications. The lab clones the repository, rebuilds occupation exposure from O*NET abilities and the application matrix (so per-application exposures and ability shares are available), and expands the "combined industry" codes of the OES and AIIE files into their 4-digit members.

DAIOE v1.0.0 (Engberg et al. 2026; Zenodo 10.5281/zenodo.21873968). Occupation-year exposure built from measured progress on 140 AI benchmarks, 2010-2024, for nine capability subdomains and a generative-AI composite. The lab uses the 2024 refresh on SOC 2010 (it prefers refresh-2024 over frozen-2010-2023 and reads the fastest format present) and the SOC 2018 panel. The index is cumulative: levels rise about fifty-fold between 2012 and 2024 for every occupation alike, so the lab keeps within-year z-scores next to the levels. The SOC 2018 panel also carries published comparison measures: AIOE (Felten, Raj and Seamans 2018, 2021), Webb's (2020) AI, software and robot patent scores, the GPT exposure measures of Eloundou et al. (2024) and Frey and Osborne's (2017) computerisation probability.

IRS SOI migration data. For each county and pair of filing years, the files count tax returns (households), exemptions (people) and adjusted gross income moving between counties. Summary rows carry codes in the "other county" columns: 96 total, 97 domestic (000 all, 001 same state, 003 other states), 98 foreign, 57-59 foreign detail and "other flows"; the row whose origin equals its destination counts non-movers. Cells under the disclosure threshold hold -1 and are treated as missing, and county pairs with fewer than 20 returns are folded into "other flows". The lab labels each file by its second year (2021-22 is 2022). Header case and zero-padding differ between years; the adapter normalises both.

Correlates of State Policy (Grossmann, Jordan and McCrain, IPPSR). About 3,000 state-year variables, 1900-2020. The file carries no descriptions, so the lab uses a curated set whose meaning was checked against known states (for example propgoppres for Texas reads 55.5 in 2012, McCain's 2008 share, and 57.2 from 2014, Romney's), keeps all others in a searchable catalog, and builds a 2012-2016 state profile, because most political variables end between 2014 and 2017.

Curated CSPP variable	Meaning
`propgoppres`	Republican share of the last presidential vote (%)
`propgopleg`	Republican share of state legislators (%)
`ranney4_control`	Democratic control of state government (Ranney index, 0-1)
`inst6014_nom`	State government ideology (0 conservative to 100 liberal)
`policyeconlib_est`, `policysociallib_est`	Economic and social policy liberalism (Caughey and Warshaw)
`masseconlib_est`, `masssociallib_est`	Public economic and social liberalism (Caughey and Warshaw)
`grtw`	Right-to-work law
`perc_college`	Adults with a college degree (%)
`x_top_corporateincometaxrate`	Top corporate income tax rate (%)
`hincomemed`, `incomepcap`, `unemployment`	Median household income, income per capita, unemployment rate

County context (JsonOfCounties, compiled by E. Gambit from the Census ACS 2019, BEA, NOAA, County Business Patterns and the MIT Living Wage Calculator). One record per county: land area, population, bachelor's degree share, average income, housing costs, January temperature, the 2016 vote and employment by sector. The stress tests use it for density, education, income, housing costs, climate and industry mix. It is a compilation, so official files (Census Gazetteer, ACS, USDA codes) should replace it before publication.

Telework shares (Dingel and Neiman 2020, Journal of Public Economics). The share of jobs that can be done at home, by O*NET occupation, 2-digit NAICS and metro area. The county measure is the county's sector mix weighted by the national sector shares.

County presidential returns, 2016-2024 (compiled by T. McGovern on GitHub from news-organisation tallies). Used only when POLISY_DA has no MIT Election Lab file; they give the 2016 vote for the stress tests and the county panel's vote shares.

Waiting: Census BTOS (actual AI use by firms; the adapter reads the question-by-answer-by-period workbooks and the Colab fetch downloads them) and PatentsView (AI patents by CPC codes, inventor locations and moves; deferred in this version).

6. How the datasets connect

Keys, and how much matched in the first run:

Join	Key	Matched
AIOE rebuilt from abilities -> published AIOE	SOC 2010	774 of 774 occupations (r = 0.99999)
VRscores occupations -> AIOE	O*NET title (exact 63.4%, fuzzy 20.2%)	83.6% of workers
VRscores occupations -> DAIOE	SOC 2010 code from the AIOE link	83.4% of workers
VRscores occupations -> SOC 2018 measures	SOC 2018 title (exact 85.6%, fuzzy 0.2%), else the same SOC code (5.9%)	88.1% of workers
VRscores industry titles -> NAICS-6	NAICS 2017 titles	99.98% of workers
VRscores industries (NAICS-4) -> AIIE	NAICS-4, combined codes expanded	72.6% of workers
IRS county totals -> AIGE	county FIPS	3,130 counties
IRS county pairs -> interstate totals	state, year	51.4% of interstate moves
CSPP -> states	state FIPS	51 of 51

Two joins deserve care. Occupations are linked by title because the VRscores report shows titles, not codes; the unlinked 12-16% of workers sit in occupations whose titles changed between SOC vintages. And IRS county pairs show only about half of all interstate moves, because the IRS suppresses small county pairs; the lab computes state and county rates from the summary rows (which include everyone) and uses pairs only for directions and corridors.

Metro-level AI exposure (the AIIE of a metro's industry mix, from QCEW by MSA) needs the CBSA delineation file, which POLISY_DA module 05 downloads; it was not available in the first run.

7. Methods

Weighted correlations use workers, households or jobs as weights, as each finding states; 95% intervals are percentile bootstraps (1,000 draws, seed 20260925).
Regressions are weighted least squares with heteroskedasticity-robust (HC1) standard errors. Continuous regressors are standardised, so coefficients read as "per standard deviation"; 0/1 indicators stay raw. Fixed effects (occupation group, sector, state) enter as dummies.
Spatial statistics: Moran's I with 8-nearest-neighbour weights on great-circle distance, 999 permutations; local indicators give the high-high and low-low clusters on the metro map.
Structure: principal components; k-means with the number of groups chosen by silhouette; isolation forests for unusual profiles; the migration network uses PageRank and greedy modularity communities (networkx, fixed seeds).
Panels: two-way fixed effects (state and year) with standard errors clustered by state, and Benjamini-Hochberg false-discovery control across all tested variable-outcome pairs.
Stress tests (analyses/stress.py) re-estimate a headline finding with the controls that stand for its obvious rival and re-grade it (core.regrade). Migration: net domestic migration on AIGE with state-by-year fixed effects, weighted by households, errors clustered by county, adding density, education, income, housing costs, January temperature, the 2016 vote and telework in turn, by period and by year. Exposure: the Republican share of an occupation's workers on each of ten exposure measures, adding education and wage, telework, gender and race shares, and occupation-group fixed effects in turn, and pairs of measures in one model. Structure: industries' actual partisanship and political balance (1 - |2p - 1|) against the values their occupational staffing predicts.

Each finding carries a grade:

Grade	Meaning
Robust	Holds across weights, controls and samples
Suggestive	Consistent sign and a t-statistic of 2 or more in the preferred model; not yet stress-tested
Fragile	Depends on weighting or controls; a lead, not a result
Descriptive	A pattern worth describing; no test implied
Artifact	Most likely produced by how the data were built
Needs data	An analysis that runs as soon as its dataset is added

8. What the first run found

The first run used the VRscores report, the AIOE repository, DAIOE v1.0.0, IRS county migration 2012-13 to 2021-22 and the CSPP extract; the run of 28 September 2026 added the stress tests and their controls. The interactive lab always shows the latest run. The research report (site/index.html) sets every finding against the 2024-2026 literature; in brief:

Stress tests.

The "AI exodus" is mostly the partisan and telework geography of migration. One SD more county AI exposure meant -0.86 points of households a year in 2020-22 within the same state (t = -13.6). With density, education, income, housing costs, January temperature, the 2016 vote and telework held equal, it is -0.11 (t = -2.2): 88% of the gradient goes. The controlled coefficient is small and flat from 2013 to 2022. What grew is the coefficient on the 2016 Republican vote share, from +0.33 (2013-16) to +0.74 (2020-22). The finding mig-ai-exodus is re-graded fragile.
AI exposure's Democratic lean is composition; automation risk's Republican lean is not. Generative-AI exposure (DAIOE) leans Democratic alone (-2.0 points of Republican share per SD, t = -3.3), but the lean disappears with gender and race shares (-0.1) and occupation groups (+0.1). Frey and Osborne's computerisation risk leans Republican with every control (+2.4, t = 4.0). Felten, Raj and Seamans's AIOE turns from a slight Democratic lean to a clear Republican one. The finding occ-waves is re-graded fragile.
Most of an industry's partisanship, and half of its political balance, is the occupations it employs. Across 178 industries, occupational staffing explains 72% of the variance in the Republican share and 54% of the variance in balance.

Against the literature. The migration findings (movers to Republican places, high earners leaving large counties) are documented; the Democratic lean of AI exposure is documented and runs through education; education pulling occupations Democratic and pay Republican is the "Brahmin left, merchant right" pattern; rising partisan sorting across employers is in Frake, Hurst and Kagan (2026) and the VRscores report; every VRscores drift is cohort replacement (party is fixed at 2024) and, where party is inferred, partly a measurement effect. The report argues that the frontier is at the level of the firm and sets out four research seams with the data and designs each needs.

Waiting for data: AI adoption (BTOS) and AI patenting (PatentsView).

9. Outputs and file formats

results/results.json: everything the site shows, in six parts. findings is a list of records (id, title, claim, theme, level, datasets, strength, stats, question, next_data, caveats, views, rank). views are chart specifications keyed by id, each naming a dataset. datasets are tables (columns and rows). diagnostics are the joins. catalog holds the sources and files found. profiles hold column summaries per file. run holds versions, times and the adapters' status.
results/findings.csv and results/linkage_diagnostics.csv: the same findings and joins as flat tables.
canonical/*.parquet and panels/*.parquet: every intermediate table, for your own analyses (pandas.read_parquet, DuckDB or Polars).
site/data/<table>.csv: every table behind the charts.
site/index.html: the research report; site/lab.html and site/polisy_lab_offline.html: the interactive lab.

10. Regenerating the site

The site is rebuilt from results.json in a second:

from polisy_lab.site import build_site
build_site()                                     # latest results -> LAB/site
build_site("path/to/results.json", out="docs")   # any results file, any folder

or run_all(stages=("site",), fetch=False). index.html (the report) is self-contained apart from the Google fonts; its figures are static SVG drawn with matplotlib. lab.html is self-contained apart from Plotly (pinned to version 4.1.1 on jsDelivr) and the fonts; polisy_lab_offline.html inlines Plotly from the installed Python package.

Editing the report. The argument lives in polisy_lab/report/report.md and is yours to edit. Inside it,  and  place a numbered figure or table (report/figures.py, report/tables.py), a line Table: Caption numbers the Markdown table after it, {fig:NAME} and {tab:NAME} become linked references, [@key] and @key become citations from report/references.py, and {{finding-id.stat:format}} inserts a number from a finding (for example {{stress-mig.attenuation:.0%}}). The novelty verdict of each finding is in report/audit.py. Figures, tables and those numbers are rebuilt at every build; the prose is not.

11. Extending the lab

Add a dataset. Add an entry to SOURCES in sources.py (title, publisher, url, grain, keys, and for each role the name pattern, extensions and header tokens). Write an adapter in adapters/ that reads the discovered files with pc.read_table(path, member), uses find_col for columns (so a renamed column is logged, not fatal) and writes canonical tables with write(df, name). Register it in adapters/__init__.py and its grain in core.GRAINS. Then join it in link.py with a diagnostic(...) line.

Add an analysis. Create a module in analyses/ with a run() function. Read panels with read(name, "PANELS"); register tables with dataset(id, df), charts with view(id, kind, title, dataset, **spec) and results with finding(...); return needs(...) when inputs are missing. Add the module to ANALYSES and its findings to PRIORITY in analyses/__init__.py.

Chart kinds the site draws, and their fields:

Kind	Fields
`scatter`	`x` and `y` (lists offered as axis choices), `size`, `color`, `text`, `filter`, `labels`
`bar`	`x`, `y`, `orientation`, `color`, `group`
`coef`	`x` (estimate), `y` (term), `group` (model), `lo`, `hi`, `xlabel`
`line`, `area`	`x`, `y`, `group`, `facet`, `text`, `palette` ("sequential" for ordered groups)
`network`	node table with `x`, `y`, `size`, `color`, `text`; `edges` names the edge table (a, b, weight)
`choropleth`	`location` (state), `year`, `value` (list), `text`
`points`	`lat` and `lon`, or `county_fips`; `size`, `year`, `value` (list), `text`
`table`	`columns`

A new chart kind is one renderer function in site_assets/app.js (RENDER.<kind>).

12. Limitations

Nothing here is causal. The designs (within-state comparisons, controls, fixed effects) rule out some explanations, not all.
VRscores from the report are a 2024 cross-section for occupations. Metro and state series come from the report's figures, and coverage varies by state.
Partisanship where party is inferred drifts differently from registered party; findings 11 and 18 measure this.
AIGE describes a county's 2019 jobs, not the people who move; the stress tests show it stands mostly for density, politics and telework.
The county controls of the stress tests come from a compiled source (JsonOfCounties), and telework is measured through each county's sector mix.
IRS counts households that file tax returns and misses non-filers; small county pairs are suppressed.
Most CSPP political variables end between 2014 and 2017.
Titles link occupations across classifications; about 12-16% of workers stay unlinked.

13. Reproducibility

Code versions: polisy_lab 0.2.0 (28 Sep 2026), polisy_core "2026-09-24 file finder". Tested with Python 3.11 (pandas 3.0) and 3.12 (pandas 2.2).
No synthetic data are used anywhere; every test runs on real files and skips when they are missing.
Random seeds are fixed (bootstrap 20260925; clustering, layouts and isolation forests seeded).
Every run records its stages, versions and times in results.json (run), and every join's match rate.
The site pins Plotly 4.1.1.

14. Sources

VRscores: Kagan, Frake and Hurst (2026), Organization Science 37(2): 444-465, https://ideas.repec.org/a/inm/ororsc/v37y2026i2p444-465.html; data at https://politicsatwork.org/download-data
AIOE: Felten, Raj and Seamans (2021), Strategic Management Journal; https://github.com/AIOE-Data/AIOE
DAIOE v1.0.0: Engberg, Görg, Hellsten, Javed, Lodefalk, Längkvist, Monteiro, Kyvik Nordås, Pulito, Schroeder and Tang (2026), "AI Unboxed: Capability Arrival and the Clerical Decline"; data https://zenodo.org/records/21873968; code https://github.com/Magnus-L/daioe-pipeline
Comparison measures in the DAIOE SOC 2018 panel: Eloundou, Manning, Mishkin and Rock (2024), "GPTs are GPTs", Science; Webb (2020), "The Impact of Artificial Intelligence on the Labor Market"; Frey and Osborne (2017), "The Future of Employment", Technological Forecasting and Social Change
IRS SOI migration data: https://www.irs.gov/statistics/soi-tax-stats-migration-data
Correlates of State Policy: https://ippsr.msu.edu/public-policy/correlates-state-policy; policy liberalism from Caughey and Warshaw (2016), American Journal of Political Science
Census Business Trends and Outlook Survey: https://www.census.gov/hfp/btos/data_downloads
PatentsView: https://patentsview.org/download/data-download-tables
Map shapes: US Census Bureau cartographic boundaries via us-atlas, https://github.com/topojson/us-atlas
Telework: Dingel, J. I., and Neiman, B. (2020), How many jobs can be done at home?, Journal of Public Economics 189, 104235; https://github.com/jdingel/DingelNeiman-workathome
County context: JsonOfCounties, https://github.com/evangambit/JsonOfCounties
County presidential returns 2016-2024 (compiled): https://github.com/tonmcg/US_County_Level_Election_Results_08-24