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 ✓, and every published decile cutoff reproduced to the peso.

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.

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
tests/      pytest suite for pipeline/ — offline, synthetic fixture, 100% statement+branch coverage
data/       checksums.txt · percentiles.v1.json   (raw/ is gitignored)
web/        index.html · src/{main,charts,stats,format,usd,types}.ts · styles.css · test/ (vitest, 100%)
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), 91 tests.
• web/src/ — vitest + jsdom, every render path exercised against the committed artifact, 140 tests.

make test runs both; make lint runs ruff + mypy + tsc. Four GitHub Actions enforce it on every push/PR, path-filtered so a web-only change never reaches for INDEC:

Workflow	What it checks
ci.yml	web typecheck + vitest 100% gate + production build
python.yml	ruff + mypy + pytest 100% gate (offline, fast)
data.yml	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

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).

Deploy

Static bundle + one JSON → ideal for Cloudflare Pages (web/dist/). Long cache on the content artifact, short cache on index.html. Live at argentina-income-analyzer.pages.dev, published manually with make deploy (there's no auto-deploy workflow).

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.
   • 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.