Causal Machine Learning and the Resource Curse

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_EconML/data/"
use "${BASE}sim_resource_curse.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_EconML/data/"
df = pd.read_stata(BASE + "sim_resource_curse.dta")

# load every dataset at once
files = ["sim_resource_curse"]
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 + "sim_resource_curse.dta", "sim_resource_curse.dta")
df, meta = pyreadstat.read_dta("sim_resource_curse.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_EconML/data/"
df <- read_dta(paste0(BASE, "sim_resource_curse.dta"))

Overview & sources

Companion data for a hands-on Python tutorial that estimates heterogeneous causal effects of mining and mineral prices on economic development using EconML's CausalForestDML — a Double Machine Learning causal forest. The dataset is fully synthetic with a known data-generating process, so estimates can be checked against ground truth. Its structure mirrors Hodler, Lechner & Raschky (2023): 300 districts across 8 countries observed annually over 2003–2012 (3,000 district-year rows). Treatment has four levels — no mining (0) and mining at low (1), medium (2), and high (3) mineral prices — and is heavily imbalanced at 85%/5%/5%/5%. The outcome is log nighttime lights. The forest recovers an ATE of 0.240 for the basic mining effect (true 0.250), a non-linear price gradient, and the finding that institutions moderate the mining margin but not the price margin. The entire DGP is open and reproducible.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). Causal Machine Learning and the Resource Curse with Python EconML [Data set]. https://carlos-mendez.org/post/python_EconML/

Hodler, R., Lechner, M., & Raschky, P. A. (2023). Institutions and the resource curse: New insights from causal machine learning. PLoS ONE, 18(6), e0284968. Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., & Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1), C1–C68. Wager, S., & Athey, S. (2018). Estimation and inference of heterogeneous treatment effects using random forests. Journal of the American Statistical Association, 113(523), 1228–1242.

BibTeX

@misc{mendez2026pythoneconml,
  author       = {Mendez, Carlos},
  title        = {Causal Machine Learning and the Resource Curse with Python EconML},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/python_EconML/}},
  note         = {Data set}
}

@article{hodler2023institutions,
  author  = {Hodler, Roland and Lechner, Michael and Raschky, Paul A.},
  title   = {Institutions and the resource curse: New insights from causal machine learning},
  journal = {PLoS ONE},
  volume  = {18}, number = {6}, pages = {e0284968}, year = {2023},
  doi     = {10.1371/journal.pone.0284968}
}
@article{chernozhukov2018dml,
  author  = {Chernozhukov, Victor and Chetverikov, Denis and Demirer, Mert and Duflo, Esther
             and Hansen, Christian and Newey, Whitney and Robins, James},
  title   = {Double/debiased machine learning for treatment and structural parameters},
  journal = {The Econometrics Journal},
  volume  = {21}, number = {1}, pages = {C1--C68}, year = {2018},
  doi     = {10.1111/ectj.12097}
}
@article{wager2018estimation,
  author  = {Wager, Stefan and Athey, Susan},
  title   = {Estimation and inference of heterogeneous treatment effects using random forests},
  journal = {Journal of the American Statistical Association},
  volume  = {113}, number = {523}, pages = {1228--1242}, year = {2018},
  doi     = {10.1080/01621459.2017.1319839}
}

Variable explorer search & filter all 18 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

The estimator is EconML's CausalForestDML, a Double Machine Learning (DML) causal forest applied to the four-level treatment with log nighttime lights as the outcome.

• Partially linear model (Robinson 1988; Chernozhukov et al. 2018): Y = τ(X)·T + g₀(X,W) + ε and T = m₀(X,W) + v, where g₀ and m₀ are nuisance conditional means estimated by Gradient Boosting.
• Residualization: Ỹ = Y − E[Y|X,W], T̃ = T − m₀(X,W), then Ỹ = τ(X)·T̃ + ε — the Frisch–Waugh–Lovell logic. The causal forest is the second-stage learner that estimates the covariate-dependent slope τ(X).
• CATE: τ(x) = E[Y(1) − Y(0) | X = x] — the per-profile treatment effect (here estimated for each pairwise treatment contrast).
• GATE: GATE_g = (1/n_g) · Σ_{i∈g} τ̂(Xᵢ) — the CATE averaged over a pre-specified subgroup; SE = sqrt(mean(seᵢ²)/n_g) from the per-observation Bootstrap-of-Little-Bags standard errors.
• ATE: E[τ(X)] — the overall average, reported per pairwise contrast with 90% BLB confidence intervals.
• Honest causal forest (Wager & Athey 2018): each tree uses one subsample to choose splits and a disjoint subsample to estimate leaf means; 5-fold cross-fitting via GroupKFold on district_id blocks within-district leakage.

Synthetic data-generating process (ground truth, from script.py): the log-NTL effect of mining is 0.25 at mean institutions with institutional moderation 0.15; the medium-price premium is 0.05 (small) and the high-price premium is 0.30 (large), yielding ground-truth ATEs of 1-0 = 0.25, 2-0 = 0.30, 3-0 = 0.55, 2-1 = 0.05, 3-1 = 0.30, 3-2 = 0.25. Outcome noise SD is 0.25. A parallel conflict process (mining base 0.70, institutional dampening −0.50, price premia 0.15/0.50, base rate 0.12) generates the conflict dummy. Because every confounder is built in, the Conditional Independence Assumption holds by construction.

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. There is no real data behind this tutorial; results are internally consistent with the calibration but are not empirical evidence about real-world mining, prices, or development.
• Heavily imbalanced treatment. 85% of rows are untreated (no mining) and only 5% fall in each mining level (150 rows each); within-mining price contrasts (e.g. 3-1) draw on just 300 observations, so price-effect standard errors are large.
• No clustered standard errors. EconML's inference=True reports forest-level Bootstrap-of-Little-Bags SEs that treat observations as independent; with this panel (same district across years) true clustered SEs would typically be larger. GroupKFold by district prevents first-stage leakage but does not cluster second-stage variance.
• Contemporaneous outcomes. Treatment and outcome are measured in the same year; the structural template (Hodler, Lechner & Raschky 2023) lags the outcome to t+1 to rule out reverse causality.
• Identification rests on the CIA. The Conditional Independence Assumption holds here by construction; in real data it is untestable and a causal forest does not relax it.