Causal Inference with DoWhy: Working From Home & Productivity

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

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

Overview & sources

Companion data for a beginner-friendly causal-inference tutorial that uses DoWhy's four-step framework (Model, Identify, Estimate, Refute) to recover a known average treatment effect from observational data. The dataset is fully synthetic: 5,000 employees with a binary work-from-home treatment, a continuous productivity outcome, two confounders (introversion, number of children) that drive self-selection into treatment, and a subway-disruption instrument that satisfies the exclusion restriction by construction. Because the true causal effect is fixed at 1.0 productivity points, every estimator can be scored against the truth: the naive difference in means overshoots to 1.39, while the backdoor methods (regression, IPW, doubly robust) recover ~1.0 and IV (2SLS) returns 0.89 with a much wider interval. The entire data-generating process is open and reproducible.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). A Beginner's Guide to Causal Inference with DoWhy in Python [Data set]. https://carlos-mendez.org/post/python_dowhy_intro/

Sharma, A., & Kiciman, E. (2020). DoWhy: An end-to-end library for causal inference. arXiv preprint arXiv:2011.04216. Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press.

BibTeX

@misc{mendez2026pythondowhyintro,
  author       = {Mendez, Carlos},
  title        = {A Beginner's Guide to Causal Inference with DoWhy in Python},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/python_dowhy_intro/}},
  note         = {Data set}
}

@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}
}
@book{pearl2009causality,
  author    = {Pearl, Judea},
  title     = {Causality: Models, Reasoning, and Inference},
  edition   = {2nd},
  publisher = {Cambridge University Press},
  year      = {2009}
}

Variable explorer search & filter all 5 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 dataset is generated by a known structural model so that every estimator can be checked against the truth. For employee i the causal graph is subway_disruption → work_from_home → productivity, with introversion and num_children pointing into both treatment and outcome (confounders).

• Confounders. introversion ~ Normal(5, 1.5); num_children ~ Poisson(1.5).
• Instrument. subway_disruption ~ Bernoulli(0.4) — affects treatment only.
• Treatment assignment (observational, non-random): logit P(WFH) = −1.5 + 0.3·introversion + 0.2·num_children + 1.0·subway_disruption; work_from_home ~ Bernoulli(P(WFH)).
• Outcome. productivity = 50 + 1.0·work_from_home + 0.8·introversion − 0.5·num_children + ε, with ε ~ Normal(0, 2). The instrument does not enter this equation — the exclusion restriction holds by construction.

The coefficient on work_from_home in the outcome equation is the true average treatment effect (ATE = 1.0) — the target every estimator in the post is judged against. Confounder bias arises because introverts both self-select into WFH (treatment coefficient 0.3) and are independently more productive (outcome coefficient 0.8), so a naive WFH−office comparison overstates the effect (1.39 vs the true 1.0).

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 survey behind this dataset; values are drawn from the data-generating process above. Results are internally consistent with the simulation but are not empirical evidence about real-world work-from-home effects.
• Known truth by design. The true ATE is fixed at 1.0 so methods can be scored. On real data the truth is unknown — that is exactly why identification (DoWhy's Step 2) and refutation (Step 4) matter.
• Cross-section, not a panel. Each row is one simulated employee at a single point; there is no time dimension, no repeated measurement, and no row identifier.
• Exclusion restriction is assumed true here. subway_disruption is built to affect productivity only through WFH choice. In practice the exclusion restriction cannot be tested from data and must be justified by domain knowledge.
• Selection on observables is satisfiable here. The two confounders are fully observed, so backdoor methods are unbiased by construction; with a missing confounder they would be biased while IV would remain valid.