impactlab-latam / README.md
Hugdrigo's picture
Upload 22 files
b9cdea1 verified
|
Raw History Blame Contribute Delete
5.28 kB
---
title: ImpactLab LATAM
emoji: 🧪
colorFrom: blue
colorTo: green
sdk: static
app_file: index.html
license: mit
short_description: Policy evaluation with DiD and synthetic controls.
pinned: false
fullWidth: true
tags:
- policy-analysis
- causal-inference
- latin-america
- data-visualization
- world-bank
- research
---
# ImpactLab LATAM · Policy Evaluation Studio
**An open-source, browser-only research workbench for transparent, exploratory policy comparisons across Latin America.**
**Author:** Rodrigo Pantoja (`@Hugdrigo`)
**Release:** 1.0.0
**License:** MIT (original code). Public indicators remain governed by their source terms.
**Hosting:** Hugging Face Static Space; no paid API key or running Python server is required.
## Why this project matters
A country comparison does not become causal merely by plotting outcomes before and after a date. ImpactLab is designed to show each analytical choice and its limitations. A user selects a treated economy, donor economies, an outcome indicator, a study window, and an assumed policy-break year. The tool then produces **exploratory**, inspectable contrasts with provenance and reproducibility outputs.
## Eight research modules
| Module | Function |
| --- | --- |
| 01 Overview | Summary, trend chart, main constraints and method availability |
| 02 Time series | Observed treated and donor-mean trajectories, per-year contrasts |
| 03 Difference-in-Differences | Explicit two-group pre/post comparison, pre-trend slope gap and year-resampling interval |
| 04 Synthetic control | Nonnegative donor weights that sum to one; fitted on pre-break years only |
| 05 Placebos | Alternative treated-country assignments and earlier break-year comparisons |
| 06 Data audit | Completeness, missing years, latest reported year and method eligibility |
| 07 Protocol | Intervention documentation, external evidence link, confounder checklist, notes |
| 08 Export | Selected-panel CSV, research report Markdown, full restorable session JSON |
## Data provenance
* **DEMO** (initial load): locally generated **fictional** observations, explicitly marked synthetic. These are not official country statistics and should never be used as evidence.
* **WORLD BANK LIVE** (after button press): directly fetched from the [World Bank Indicators API v2](https://datahelpdesk.worldbank.org/knowledgebase/articles/889392) in the user's browser. No API key required. Indicators may have different reporting lags and definitions.
* **IMPORTED**: research sessions restored from JSON. The application labels them **unverified imported data** regardless of their original metadata.
See [DATA_DICTIONARY.md](DATA_DICTIONARY.md) and [PROVENANCE.md](PROVENANCE.md).
## What is and is not estimated
**Included:** transparent DiD contrast, convex donor-weighted synthetic comparison, pre-break slope diagnostic, descriptive resampling range, year-placebo and unit-placebo exercises, completeness audit.
**Not included:** identification of a real treatment event, randomized assignment, causal proof, instrumental-variable identification, generalized synthetic-control estimation, cluster-robust standard errors, significance testing, forecasting or policy recommendations generated by a large language model. The statistics do **not** establish that an actual policy caused a change. See [METHODOLOGY.md](METHODOLOGY.md) and [RESEARCH_ETHICS.md](RESEARCH_ETHICS.md).
## How to use
1. Open the **App** tab. The project starts in **SYNTHETIC DEMO** mode.
2. Choose the indicator, treated country, comparison countries and time window.
3. Enter a break year and click **Run evaluation**.
4. Explore the eight tabs. An explicit warning appears if data are insufficient.
5. To load real indicators, click **Load World Bank data** and wait for the **WORLD BANK LIVE** label.
6. Record the actual intervention and its external documentary source in **Protocol**, if known.
7. Download a report and an analysis session from **Export**.
The default synthetic case includes an intentionally simulated shift for Chile in life expectancy after 2016, to make diagnostics testable. **It does not describe a real Chilean reform.**
## Reproduction and testing
Code is vanilla JavaScript, HTML and CSS. There are no runtime package installations or secrets.
```bash
node --test test_engine.cjs
python -m http.server 8000
```
The optional browser checks need Python Playwright and a local Chromium executable. They can also be run by providing an HTTP URL or by directly injecting local scripts into the test browser, as implemented in `test_browser.py`.
See [QA_REPORT.md](QA_REPORT.md), [ARCHITECTURE.md](ARCHITECTURE.md) and [DEPLOYMENT_ES.md](DEPLOYMENT_ES.md).
## Responsible use
* Do not upload private data, passwords or secrets to the public repository.
* World Bank series may be revised and may not share identical reporting years.
* Country selection, intervention timing and donor selection are researcher decisions and must be justified.
* A high synthetic-control fit can coexist with confounding or invalid causal identification.
* The research report labels synthetic, live and imported sources separately.
**Research prototype. Not an official economic report or a validated causal-impact estimate.**