Causal Inference with DoWhy: The Lalonde Job-Training Data

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

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

Overview & sources

Companion data for a hands-on Python tutorial that estimates the causal effect of a job-training program on earnings using DoWhy's four-step framework (Model → Identify → Estimate → Refute). The file is the Lalonde sample from the U.S. National Supported Work (NSW) Demonstration, a 1970s randomized employment program for disadvantaged workers (LaLonde 1986; the experimental subsample popularized by Dehejia & Wahba 1999). It records 445 participants — 185 randomly assigned to job training (treatment) and 260 to control — with eight pre-treatment covariates and real earnings in 1978 (re78) as the outcome. The post encodes the eight covariates as common causes in a causal graph, identifies the backdoor estimand, and estimates the average treatment effect with five methods (regression adjustment, IPW, doubly robust AIPW, propensity-score stratification, and propensity-score matching), all clustering near a roughly 34–38% earnings gain over the control mean. These are real data from a field experiment, not a simulation.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). Introduction to Causal Inference: The DoWhy Approach with the Lalonde Dataset [Data set]. https://carlos-mendez.org/post/python_dowhy/

LaLonde, R. (1986). Evaluating the econometric evaluations of training programs with experimental data. American Economic Review, 76(4), 604-620. Dehejia, R., & Wahba, S. (1999). Causal effects in nonexperimental studies: Reevaluating the evaluation of training programs. Journal of the American Statistical Association, 94(448), 1053-1062. Sharma, A., & Kiciman, E. (2020). DoWhy: An end-to-end library for causal inference. arXiv:2011.04216.

BibTeX

@misc{mendez2026pythondowhy,
  author       = {Mendez, Carlos},
  title        = {Introduction to Causal Inference: The DoWhy Approach with the Lalonde Dataset},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/python_dowhy/}},
  note         = {Data set}
}

@article{lalonde1986evaluating,
  author  = {LaLonde, Robert J.},
  title   = {Evaluating the Econometric Evaluations of Training Programs with Experimental Data},
  journal = {American Economic Review},
  volume  = {76}, number = {4}, pages = {604--620}, year = {1986}
}
@article{dehejia1999causal,
  author  = {Dehejia, Rajeev H. and Wahba, Sadek},
  title   = {Causal Effects in Nonexperimental Studies: Reevaluating the Evaluation of Training Programs},
  journal = {Journal of the American Statistical Association},
  volume  = {94}, number = {448}, pages = {1053--1062}, year = {1999}
}
@article{sharma2020dowhy,
  author  = {Sharma, Amit and Kiciman, Emre},
  title   = {{DoWhy}: An End-to-End Library for Causal Inference},
  journal = {arXiv preprint arXiv:2011.04216}, year = {2020}
}

Variable explorer search & filter all 10 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 post applies DoWhy's four-step framework to this data, targeting the average treatment effect (ATE), ATE = E[Y(1) − Y(0)], of treat on re78:

• 1. Model — encode a DAG in which the eight covariates (age, educ, black, hisp, married, nodegr, re74, re75) are common causes of both treat and re78; treat → re78 is the effect of interest.
• 2. Identify — the backdoor criterion returns the adjustment estimand d/d[treat] E[re78 | age, educ, black, hisp, married, nodegr, re74, re75] under the unconfoundedness assumption.
• 3. Estimate — five estimators of the same ATE: regression adjustment (models E[Y|X,T]); IPW τ = (1/n) Σ [ T·Y / e(X) − (1−T)·Y / (1−e(X)) ] with propensity score e(X) = P(T=1|X); doubly robust AIPW = regression estimate plus an IPW-weighted residual correction (consistent if either the outcome or the propensity model is correct); PS stratification (5 strata); and PS matching (nearest-neighbour, shifting the estimand toward the ATT).
• 4. Refute — placebo-treatment, random-common-cause, and data-subset tests stress the estimate (the placebo collapses the effect from ~$1,676 to ~$62).

The dataset itself contains no constructed variables: every column is an observed field from the NSW evaluation (assignment, demographics, prior/post earnings). The quantities above are computed by the analysis code, not stored in the file.

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

• Small sample. Only 445 observations (185 treated / 260 control); propensity-score methods can suffer from poor overlap and estimates carry high variance.
• Earnings are right-skewed with mass at zero. re74, re75, and re78 have a large spike at $0 (no earnings) and a long right tail, so means exceed medians substantially.
• Unconfoundedness. The backdoor estimand assumes the eight covariates capture all confounding; this is credible here because assignment was randomized, but the same data is often used to study what happens when it is not.
• ATE vs ATT. Four estimators target the ATE; propensity-score matching discards unmatched controls and shifts the estimand toward the ATT, so its estimate answers a slightly different question.
• Experimental subsample. This is the randomized NSW experimental sample (Dehejia-Wahba), not the larger observational PSID/CPS comparison versions; results are not directly comparable to those non-experimental variants.