Do Industrial Parks Work? Evaluating Place-Based Policy in Ethiopia with Difference-in-Differences

Downloads

Each dataset is available as a labeled Stata .dta and its source file.

⇩ Download all data (ZIP)stata_codebook.do

Run stata_codebook.do in Stata once to attach long-form per-variable notes to the .dta files.

Load directly in code

Every file loads straight from GitHub (raw URLs). Swap the file name to load any dataset.

Stata

* Stata 14+ : `use` reads an https URL directly
global BASE "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/python_did_industrial_park/data/"
use "${BASE}industrial_park_district_panel.dta", clear
describe
notes

Python

!pip install -q pyreadstat
import pandas as pd
BASE = "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/python_did_industrial_park/data/"
df = pd.read_stata(BASE + "industrial_park_district_panel.dta")

# load every dataset at once
files = ["industrial_park_district_panel", "industrial_park_household_rcs", "industrial_park_individual_rcs"]
data = {f: pd.read_stata(BASE + f + ".dta") for f in files}

# pyreadstat (richest metadata) reads LOCAL files -> download first
import pyreadstat, urllib.request
urllib.request.urlretrieve(BASE + "industrial_park_district_panel.dta", "industrial_park_district_panel.dta")
df, meta = pyreadstat.read_dta("industrial_park_district_panel.dta")

Copy and paste this snippet in Google Colab app. https://colab.research.google.com/notebooks/empty.ipynb

R

# R : haven::read_dta auto-downloads an https URL
library(haven)
BASE <- "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/python_did_industrial_park/data/"
df <- read_dta(paste0(BASE, "industrial_park_district_panel.dta"))

Overview & sources

Companion data for a beginner-friendly, end-to-end staggered difference-in-differences (DiD) tutorial that asks whether Ethiopia's industrial parks raised local economic activity, household living standards, and women's economic agency — and for whom. The tutorial replicates Huang, Wang & Xu (2026) on fully synthetic data spanning three grains: a satellite district-year panel of 139 woredas (2005–2020), and two Ethiopia DHS repeated cross-sections of households and individuals across five survey rounds. The estimand throughout is the average treatment effect on the treated (ATT) — the effect on the 17 park-hosting woredas relative to 122 propensity-score-matched never-treated controls — identified under parallel trends in an explicitly observational setting. The analysis runs a static two-way fixed-effects DiD and an event study with pyfixest, cross-checks them against the Sun-Abraham, Borusyak/Gardner and Callaway-Sant'Anna estimators plus a Goodman-Bacon decomposition, and runs survey-weighted repeated-cross-section DiD with Conley spatial standard errors.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). Do Industrial Parks Work? Evaluating Place-Based Policy in Ethiopia with Difference-in-Differences [Data set]. https://carlos-mendez.org/post/python_did_industrial_park/

Huang, G., Wang, M., & Xu, H. (2026). The socioeconomic impacts of industrial parks in Ethiopia. Journal of Urban Economics. https://doi.org/10.1016/j.jue.2026.103867

BibTeX

@misc{mendez2026pythondidindustrialpark,
  author       = {Mendez, Carlos},
  title        = {Do Industrial Parks Work? Evaluating Place-Based Policy in Ethiopia with Difference-in-Differences},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/python_did_industrial_park/}},
  note         = {Data set}
}

@article{huang2026industrial,
  author  = {Huang, Gordon and Wang, Mengjie and Xu, Hangtian},
  title   = {The socioeconomic impacts of industrial parks in Ethiopia},
  journal = {Journal of Urban Economics},
  year    = {2026},
  doi     = {10.1016/j.jue.2026.103867}
}

Variable explorer search & filter all 56 variables

Type to filter by name or label, or use the chips to filter by type. Each row shows a mini distribution. Click a header to sort.

Cross-file variable index

Which file each variable appears in (● = present).

Construction & formulas

All estimators target the ATT — the average park effect on the treated woredas, identified under parallel trends.

• Static TWFE DiD (satellite panel): Y_dt = β·D_dt + α_d + γ_{r(d),t} + ε_dt — α_d is a woreda fixed effect, γ_{r(d),t} a region×year fixed effect, and β the ATT. The "with-trends" spec adds t_* interactions of centred time (year − 2012) with 2007 baseline characteristics, letting each woreda follow its own linear trend.
• Event study: Y_dt = Σ_{k≠−1} δ_k·1[t−g = k] + α_d + γ_{r(d),t} + ε_dt — one coefficient per event time k (years since opening g), normalized to k = −1.
• Repeated-cross-section DiD (DHS household/individual): no respondent panel, so identified off district×round means — Y ~ treatment | district_id + region_id^survey_round, DHS survey-weighted, district-clustered; only coarse event phases.
• Goodman-Bacon: decomposes TWFE into weighted 2×2 comparisons (treated-vs-never, earlier-vs-later, forbidden later-vs-earlier).
• Conley spatial-HAC SE: hardens the panel SE for spatial (nearby woredas in a year) and serial dependence; the point estimate is unchanged.
• IHS transform: ihs_light = asinh(light_intensity) — a log-like transform that admits zeros; coefficients read approximately as proportional changes.

Synthetic data-generating process. The three CSVs are 100% synthetic, calibrated so re-running the paper's regressions reproduces its findings (signs, significance stars, approximate magnitudes). A deliberate bright-base device keeps treated park-cities intrinsically much brighter than rural controls (a level the district fixed effect absorbs); spatial and serial shocks are injected so standard errors behave realistically without moving the point estimates.

The datasets

Switch datasets with the tabs. Each shows the full variable dictionary plus a sortable statistics table with mini distributions and data coverage.

Expand all datasets expand to search (Ctrl/⌘+F) or print across all datasets

Known limitations & caveats

• Synthetic data. The three CSVs are 100% synthetic, built for teaching. They are calibrated to reproduce the paper's findings, not the paper's real (confidential) inputs — harmonized nighttime lights, the GISD30 impervious-surface product, Ethiopia DHS micro-data, and the official park list. Use them to learn the methods, not to draw conclusions about Ethiopia.
• Tiny treated group. Only 17 treated woredas against 122 matched controls, so several effects are borderline and the small treated sample cannot make all moderators precise at once (e.g. the primary-road interaction is correctly signed but insignificant).
• Observational, not randomized. Parks were not randomly placed (they went to denser, more urban, better-connected woredas), so identification rests on parallel trends, not randomization; the fixed effects and trend terms are confounding controls, not precision-only adjustments.
• Documented synthetic gaps. The raw-light coefficient runs high (~1.6 vs the paper's 1.276, a bright-base artifact); light levels are intentionally not PSM-matched (the EDA figure is baseline-normalized); the decision-power mean sits a touch below the paper's because of a linear-probability clipping ceiling. See the post's Section 13 reproduction audit.
• Repeated cross-sections. The household and individual files have no within-respondent panel key, so they admit only coarse event phases and survey-weighted regressions, never unit fixed effects.