Argentina Income Percentile

Where do you stand in Argentina's income distribution? This tool answers that from the real INDEC EPH microdata — it downloads the household survey, applies the official sample weights, and computes the weighted income distribution itself. No eyeballed deciles, no decorative curves.

▶ Try it live: argentina-income-analyzer.pages.dev

Validated against INDEC (Evolución de la distribución del ingreso, 4º trim. 2025): Gini IPCF 0.427 ✓, mediana $450.000 ✓, media $635.996 ✓, población 30.032.540 ✓, every published decile cutoff reproduced to the peso, and every decile's income share to the published 0.1 point.

It has two parts:

1. pipeline/ — an offline, reproducible Python pipeline (uv + pandas + numpy) that turns the raw EPH base into one small, provenance-stamped JSON artifact.
2. web/ — a static Vite + TypeScript site that reads that JSON and renders honest graphics (empirical CDF, weighted histogram, Lorenz curve + data-derived Gini) plus a poverty-line check.

What makes it "real"

• Right weights. Individual income (P47T) is weighted by PONDII; household per-capita income (IPCF) by PONDIH — INDEC's non-response-corrected factors, never the plain PONDERA.
• Computed, not assumed. Percentiles come from the weighted empirical CDF; the Gini from the data's own Lorenz curve.
• Honest charts. The distribution you see is the data — a step CDF, real weighted bins, a real Lorenz curve. Nothing is smoothed into a pretty shape.
• Two distinct measures. "Tu ingreso personal" (P47T) and "ingreso por persona del hogar" (IPCF) answer different questions and are labeled as such. The poverty overlay uses IPCF, as INDEC does.
• Provenance. The source zip is pinned by SHA-256; the artifact embeds the period, checksum, and sample size. A validation gate fails the build if it stops matching INDEC's published figures — Gini, mean, median, population, every decile's share and mean, and every decile cutoff to the peso.
• Your income, in survey pesos. The survey holds nominal pesos of its reference month (October 2025 for the 4º trimestre). The app asks which month your income is from and brings it back to that month with INDEC's monthly IPC (data/cpi_monthly.json, generated monthly from the spliced series of peso) before the lookup, and says so next to the result.

Quickstart

make setup        # uv sync (python deps) + npm install (web deps)
make data         # fetch → verify(sha256) → build(JSON) → validate(vs INDEC)
make test         # python + web unit tests, each gated at 100% coverage
make lint         # ruff + mypy + tsc
make up           # vite dev server  →  http://localhost:5179
make build        # static bundle in web/dist/
make deploy       # publish web/dist/ to Cloudflare Pages (needs wrangler auth)

make (or make help) lists every target. make data downloads EPH_usu_4_Trim_2025_txt.zip (~2.9 MB, public INDEC data), computes the statistics, writes data/percentiles.v1.json and web/public/percentiles.v1.json, and asserts the reproduction of INDEC's official numbers. The pipeline steps can also be run one at a time with uv run python -m pipeline.{fetch,verify,build,validate}.

How it works

EPH microdata (T425, ;-delimited, comma decimals)
   └─ pipeline/  fetch → verify → load → weighted stats → validate → emit
        └─ web/public/percentiles.v1.json   (versioned, provenance-stamped)
             └─ web/  Vite + TS + Observable Plot  (lookups + honest charts only)

The frontend does no statistics — only lookups and trivial interpolation against the precomputed artifact. Everything quantitative is computed once, offline, and validated.

Methodology (short)

• Universe. IPCF: the whole population (zero-income persons sit at the start of decile 1, per INDEC). Individual income: perceptores only. Income non-response (-9, decile codes 12/13) is excluded.
• Weighted quantiles. Inverted weighted empirical CDF (type-1), cross-checked against numpy.quantile(method='inverted_cdf', weights=...).
• Gini. Trapezoidal area between the 45° line and the weighted Lorenz curve.
• Poverty. Per-person household income vs the Canasta Básica (CBA/CBT per adulto equivalente, October 2025 — matched to the income vintage so inflation doesn't skew the comparison). The UI counts each household member as one adult; INDEC's full age/sex equivalence scale would set a somewhat lower threshold for households with children.
• Full notes: docs/metodologia.md, and the "Metodología" section in the app.

Project structure

pipeline/   config (pinned facts) · fetch · verify · load · weighted · build · validate · artifact_check · watch · cpi_sync
tests/      pytest suite for pipeline/ — offline, synthetic fixture, 100% statement+branch coverage
data/       checksums.txt · percentiles.v1.json   (raw/ is gitignored)
web/        index.html · styles.css · test/ (vitest, 100%) · src/:
              main.ts (init, controls, renderAll) · app.ts (state, artifact, deflated income) · ui.ts
              headline · ladder · splits · geo · time · poverty · cost · visuals · about · classes
              charts · stats · format · usd · types
docs/       metodologia.md

Tests & CI

Both layers are gated at 100% coverage — statements and branches:

• pipeline/ — pytest against a tiny synthetic EPH fixture (no network, no real microdata).
• web/src/ — vitest + jsdom, every render path exercised against the committed artifact.

make test runs both; make lint runs ruff + mypy + tsc. GitHub Actions enforce it, three on every push/PR (path-filtered so a web-only change never reaches for INDEC) and one on a monthly schedule; every action is pinned to a commit SHA and Dependabot (.github/dependabot.yml) keeps the pins, npm and uv current:

Workflow	What it checks
deploy.yml	web typecheck + vitest 100% gate + production build, then (only if that passed) the Cloudflare Pages deploy: production on master, preview per PR
python.yml	ruff + mypy + pytest 100% gate (offline, fast)
data.yml	the two committed copies of the artifact are identical, the pipeline still reproduces INDEC, and the committed artifact matches the rebuild
data-update.yml	monthly watch — opens a draft PR when a newer EPH quarter is published
cpi-sync.yml	monthly sync — regenerates data/cpi_monthly.json from peso's published artifact and, when the months changed, rebuilds, tests and pushes to master (which deploys)

Reproducibility

The exact source file is pinned by SHA-256 in data/checksums.txt (trust-on-first-use). The verify step (part of make data, or uv run python -m pipeline.verify) fails if the input ever changes, so the build is reproducible against a known input. To update to a newer quarter, edit pipeline/config.py (ZIP_URL, *_FILE, QUARTER) and delete data/checksums.txt. The monthly poverty lines live in their own POVERTY_LINES block (they update more often than the EPH). The site uses only the latest 2º or 4º trimestre: INDEC's 1º and 3º trimestre include the aguinaldo, which inflates a typical month's income. The monthly watch (data-update.yml) proposes only those quarters, from a 4º trimestre pin the next 2º and from a 2º the next 4º.

Deploy

Static bundle + one JSON → ideal for Cloudflare Pages (web/dist/). web/public/_headers sets the security headers (CSP, X-Frame-Options, nosniff, referrer and permissions policies), caches the content-hashed JS/CSS/data under /assets/ for a year (immutable) and revalidates index.html and the fixed-name percentiles.v1.json (kept for direct links; the page loads the hashed copy) on every request. Live at argentina-income-analyzer.pages.dev. deploy.yml publishes it from CI once the web tests pass: every push to master that touches web/ goes to production and PRs get a per-branch preview (needs the CLOUDFLARE_API_TOKEN / CLOUDFLARE_ACCOUNT_ID secrets; without them the job still builds and just skips the deploy). make deploy does the same by hand.

Two tiers of data (important)

The app deliberately separates two kinds of figures:

1. Rigorous core — computed from INDEC EPH microdata. The IPCF distribution (percentiles/deciles, Gini, Lorenz) is validated to INDEC's published Q4-2025 figures, and the individual-income deciles are cross-checked against INDEC's own DECINDR labels. The regional, aglomerado and structural-split breakdowns (build_regions, build_aglomerados, build_splits) are computed from the same microdata and weights, but INDEC doesn't publish matching per-cell figures to anchor them against — treat them as derived, not separately validated. This is the trustworthy spine.
2. Reference layer — external sourced estimates (clearly flagged as such in the UI), stored as dated, cited blocks in pipeline/config.py:
   • HISTORY — Gini (quarterly) + poverty/indigence (semestral) + nominal median, from INDEC press reports.
   • CPI_MONTHLY — INDEC's monthly IPC from the survey's reference month on, loaded from data/cpi_monthly.json, which pipeline/cpi_sync.py regenerates from peso's published artifact (monthly, cpi-sync.yml); deflates the typed income to survey pesos. Never edit it by hand.
   • POVERTY_LINES — CBA/CBT per adulto equivalente (period-matched to the income vintage).
   • COST_OF_LIVING — rent, utilities, food, transport, internet, health + SMVM/jubilación, gathered mid-2026 from Zonaprop, IIEP-UBA/CONICET, AySA, telco comparators and press. These vary widely by case and are labeled in the app as estimates, not microdata.

To refresh any reference block, edit its dict in config.py and run make data (which rebuilds the JSON artifact). Note make build builds the web bundle, not the data.

License

Code: MIT. · Data: see NOTICE.

Data: Source: Encuesta Permanente de Hogares (EPH), INDEC — public base usuaria microdata. INDEC permits republishing aggregate/derived statistics with attribution; individual records are never shipped (only aggregates), satisfying the secreto estadístico (Ley 17.622). The cost-of-living layer cites its own external sources inline.

Elaboración propia en base a microdatos de la Encuesta Permanente de Hogares (EPH), INDEC. Fuente: INDEC, www.indec.gob.ar.