A bound, not an absence.

Atlas

The index on the map, by department and by municipality

Three views of the same data. Flat, with the metrics around it; raised, to lift one department out and look at it; and municipalities, to go down to all 1,123 units. Pinning a department lets you descend to its municipalities from the map itself, and choosing a region closes the frame in on that selection while the neighbours bleed off the edge.

• A territory without color is not a zero: it is a territory no supervised institution reported that year.
• Block height is constant and encodes nothing. A height that depended on the value would let the units in front hide the ones behind, and a unit's reading must not depend on where it falls on the map.
• The San Andrés archipelago is a chip drawn off scale: it lies 700 km offshore and covers 26 km², so inside the frame it shrank the map by a fifth to paint a dot.
• Two of the panel's 1,123 municipalities have no polygon in the 2024 national map: they appear in the ranking and the figures, but not on the map.

Abstract

• Data you can trust. Every download is fingerprinted with sha256; zeros that are really missing values are treated as missing; the splice between the supervisor's two tables is measured (median gap 0.02%).
• A method chosen by evidence, not habit. PCA was ruled out because sampling adequacy came in at 0.317 and 0.407; a denominator that manufactured correlation (−0.23) was found and fixed (−0.13).
• A result that says how much it knows. β = +0.0038 (p = 0.54) with entity and time effects: it rules out effects above 0.55 pp per identifying standard deviation of the index (about 2.0 pp per raw standard deviation) and says plainly it cannot see smaller ones, not even ±0.50 pp.

What this shows

Four capabilities, each with its proof on this page

Data

Nineteen public sources, each download fingerprinted

The index

The assumption gets measured before the method is chosen

Results

With entity and time effects, the effect is bounded

minimum detectable effect at 80%

0.60 pp

per identifying standard deviation of the index; with the bootstrap, equivalence rules out ±1.0 pp and not ±0.50 pp (p = 0.07)

wild cluster bootstrap p

0.55

999 Rademacher draws with the null imposed over 33 departments, studentised with the cluster-robust error

permutation placebo p

0.51

499 reassignments of each department's whole trajectory; shuffling within the year would narrow the cloud artificially

Pesaran CD on the residuals

2.40

weak but present cross-sectional dependence (p = 0.016); hence Driscoll-Kraay alongside the cluster

• 1The three dimensions on their own, the specification in changes, CCE, SLX and the two alternative indices agree: none is distinguishable from zero. Leaving out one department at a time, the coefficient stays between +0.000 and +0.008 and is never significant.
• 2The specification curve estimates 160 variants: without time effects, 50 of 80 come out significant; with time effects, 5 of 80. What separates the two halves is the national trend, not inclusion.
• 3The only design with a signal is the initial-exposure design (+0.017, p = 0.005), which survives its own bootstrap and its own placebo. It does not survive the contrast that matters: initial urbanisation produces an equally significant slope on its own (+0.050, p = 0.008), and with both exposures in the same equation the index falls to p = 0.10. Nor does it survive a Holm correction over the twelve contrasts published. It is a differential slope for the more urban departments, not an effect of the index, and it is published as such.
• 4The event study around 2020 cannot test pre-trends — the dependent variable starts in 2019 and the shock is 2020 — and it is published with that limitation written down.
• 5The 2026–2028 forecast layer is published as a scenario, not as a superior forecast: in the backtest it cuts the naive forecast's mean error by 38.6%, but it wins in 3 of 8 years, without 2021 the gain is +8.3%, and grouped by origin year the difference is not significant (p = 0.29).

Decisions

Every decision of substance is written before the code

The decisions that change a result — frequency, data model, index method, econometric design — go into a decision record with their alternatives, their cost and how to reverse them. Constants live in one place and no published figure exists without the test that holds it up.

Reproduce

Five commands reproduce the project

1. uv sync

   Exact dependencies from the lock file. Nothing is installed outside it.

2. uv run iif acquire all

   Downloads the nineteen sources, verifies the count against each API and writes the manifest with sha256.

3. make check

   Ruff, 242 Python tests, the dbt models and tests on DuckDB, and the site render. It exits zero or there is no commit.

4. uv run iif econ

   Runs the full econometric battery in under half a minute and writes the file every figure on this page comes from.

5. uv run iif atlas

   Generates the geometry and series this map consumes, inside the 3 MB budget the contract sets.

The local and CI engine is DuckDB. BigQuery is kept as an alternative target with a per-query cost cap, and Snowflake as a later demonstration; nothing in the project depends on either staying alive.