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/stata_cate2/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/stata_cate2/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/stata_cate2/data/"
df <- read_dta(paste0(BASE, "sim_resource_curse.dta"))

Overview & sources

Companion data for a hands-on Stata 19 tutorial that replicates the three core findings of Hodler, Lechner & Raschky (2023) on a fully synthetic panel with known ground-truth causal effects. The panel has 3,000 district-year observations (300 districts × 10 years, 2003–2012) across 8 fictional countries, with log nighttime lights (ntl_log) and a conflict indicator (conflict) as outcomes, a four-level mining/price treatment, and executive constraints and quality of government as institutional moderators. The post estimates average (ATE), group (GATE), and individualized (IATE) treatment effects for six pairwise binary contrasts via generalized random forests, comparing the Partialing-Out (PO) and doubly robust Augmented IPW (AIPW) estimators with 5-fold cross-fitting, supported by formal heterogeneity tests. Because the data-generating process is known, every estimate can be checked against the truth.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). Causal Machine Learning and the Resource Curse with Stata 19 [Data set]. https://carlos-mendez.org/post/stata_cate2/

Hodler, R., Lechner, M., & Raschky, P. A. (2023). Institutions and the resource curse: New insights from causal machine learning. PLoS ONE, 18(5), e0284968. Athey, S., Tibshirani, J., & Wager, S. (2019). Generalized random forests. Annals of Statistics, 47(2), 1148–1178. Nie, X., & Wager, S. (2021). Quasi-oracle estimation of heterogeneous treatment effects. Biometrika, 108(2), 299–319.

BibTeX

@misc{mendez2026statacate2,
  author       = {Mendez, Carlos},
  title        = {Causal Machine Learning and the Resource Curse with Stata 19},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/stata_cate2/}},
  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 = {5}, pages = {e0284968}, year = {2023}
}
@article{athey2019grf,
  author  = {Athey, Susan and Tibshirani, Julie and Wager, Stefan},
  title   = {Generalized random forests},
  journal = {Annals of Statistics},
  volume  = {47}, number = {2}, pages = {1148--1178}, year = {2019}
}
@article{nie2021quasi,
  author  = {Nie, Xinkun and Wager, Stefan},
  title   = {Quasi-oracle estimation of heterogeneous treatment effects},
  journal = {Biometrika},
  volume  = {108}, number = {2}, pages = {299--319}, year = {2021}
}

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 Conditional Average Treatment Effect is the treatment effect for units with covariate profile x:

• CATE: τ(x) = E{ y_i(1) − y_i(0) | x_i = x } — a function of x, not a single number.
• ATE: E{ τ(X) } — the CATE averaged over the whole sample (the headline number).
• GATE: τ(g) = E{ τ(x) | G = g } — the CATE averaged within a pre-specified group (e.g. each executive-constraints level).
• IATE: τ(x_i) — one estimated effect per observation.

Stata 19's cate uses a partial linear model with cross-fitting (here xfolds(5)): y = d·τ(x) + g(x,w) + ε and d = f(x,w) + u, where g/f are machine-learning nuisance functions, x are CATE variables (potential moderators) and w are controls (here i.country_id i.year).

Two estimators. PO (Partialing-Out) residualizes both y and d on (x,w) and regresses the residuals via a generalized random forest. AIPW (Augmented IPW) builds doubly robust scores from an outcome model and a propensity model; it stays consistent if either nuisance model is right.

Multi-valued treatment via pairwise binaries. The 4-level treatment (0=none, 1=low, 2=med, 3=high price) is split into six binary contrasts (1-0, 2-0, 3-0, 2-1, 3-1, 3-2), subsetting to the two relevant groups each time, e.g. treat_1v0 = (treatment == 1).

Ground truth (NTL contrasts). The DGP fixes the true ATEs: 1-0 = 0.25, 2-0 = 0.30, 3-0 = 0.55, 2-1 = 0.05, 3-1 = 0.30, 3-2 = 0.25 — the small 2-1 step vs the large 3-1 step encodes the price non-linearity.

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; values are simulated with known ground-truth effects, so results validate the method but are not empirical evidence about real-world mining or conflict.
• Direction differs from the paper. Hodler et al. (2023) found stronger institutions amplify mining benefits (upward GATE slope); this DGP produces the opposite sign (weaker institutions, larger mining effect). The reproduced structural finding is that institutions systematically moderate mining, not prices — not the sign.
• Severe treatment imbalance. ~85% of rows are controls; each treated price level is only ~5% (~150 obs). Within-mining price contrasts (2-1, 3-1, 3-2) use only ~300 observations, so confidence intervals are wide.
• Overlap failure (3-2). The high-vs-medium price contrast has propensity scores near 0/1 on its tiny subsample; AIPW fails and even PO with pstolerance(1e-8) returns an unreliable ATE. It is excluded from the summary.
• Finite-sample variability. Several estimates deviate from the ground truth (e.g. 1-0 = 0.149 vs 0.25) due to small treated samples, 5-fold cross-fitting, and the random seed; directional patterns are robust, point estimates are seed-dependent.
• Conflict ground truths. The DGP does not specify ground-truth ATEs for the conflict outcome; conflict results are interpreted directionally only.
• Requires Stata 19+. The cate command does not exist in Stata 18; the do-file aborts on older versions.