Double Machine Learning with 401(k) Pension 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_doubleml_pension/data/"
use "${BASE}pension_raw.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_doubleml_pension/data/"
df = pd.read_stata(BASE + "pension_raw.dta")

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

Overview & sources

Companion data for a hands-on Python tutorial that estimates the causal effect of 401(k) eligibility and participation on household net financial assets using Double Machine Learning (DoubleML). The dataset is the real 401(k) sample drawn from the 1991 U.S. Survey of Income and Program Participation (SIPP) — the same nationally representative cross-section of 9,915 households used by Chernozhukov et al. (2018) and distributed with the DoubleML package via fetch_401K (originally constructed from the SIPP by Poterba, Venti & Wise). About 37 % of households are eligible for a 401(k) and 26 % participate. The post fits three estimators — Partially Linear Regression (PLR) and the doubly robust Interactive Regression Model (IRM), both targeting the Average Treatment Effect of eligibility, and the Interactive IV Model (IIVM), which instruments participation with eligibility to recover the Local Average Treatment Effect on compliers — each fit with four ML learners (Lasso, Random Forest, Decision Tree, XGBoost) under cross-fitting.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). Double Machine Learning with 401(k) Data: From Eligibility Effects to Complier Analysis [Data set]. https://carlos-mendez.org/post/python_doubleml_pension/

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. https://doi.org/10.1111/ectj.12097 — Poterba, J., Venti, S., & Wise, D. (1995). Do 401(k) contributions crowd out other personal saving? Journal of Public Economics, 58(1), 1–32. https://doi.org/10.1016/0047-2727(94)01462-W

BibTeX

@misc{mendez2026pythondoublemlpension,
  author       = {Mendez, Carlos},
  title        = {Double Machine Learning with 401(k) Data: From Eligibility Effects to Complier Analysis},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/python_doubleml_pension/}},
  note         = {Data set}
}

@article{chernozhukov2018double,
  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{poterba1995do,
  author  = {Poterba, James and Venti, Steven and Wise, David},
  title   = {Do 401(k) contributions crowd out other personal saving?},
  journal = {Journal of Public Economics},
  volume  = {58}, number = {1}, pages = {1--32}, year = {1995},
  doi     = {10.1016/0047-2727(94)01462-W}
}

Variable explorer search & filter all 14 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 estimates the causal effect of 401(k) access on net financial assets with three Double Machine Learning (DML) models. In every model the nuisance functions are fit by ML learners (Lasso / Random Forest / Decision Tree / XGBoost) under cross-fitting, and the treatment parameter is recovered from a Neyman-orthogonal score.

• Naive benchmark (biased): difference in mean net_tfa between groups, Δ̂ = Ȳ(e401=1) − Ȳ(e401=0) = $19,559 for eligibility (and $27,372 for participation) — conflates the causal effect with confounding from income, education, etc.
• PLR — Partially Linear Regression (estimand: ATE of eligibility): Y = θ₀·D + g₀(X) + ε with D = m₀(X) + V. Partial out g₀ (outcome model) and m₀ (treatment model), regress the residuals; assumes a constant additive effect θ₀. Mean ATE ≈ $8,730.
• IRM — Interactive Regression Model (estimand: ATE of eligibility, doubly robust / AIPW): θ₀ = E[ g₀(1,X) − g₀(0,X) + D·(Y−g₀(1,X))/m₀(X) − (1−D)·(Y−g₀(0,X))/(1−m₀(X)) ], where m₀(X)=P(D=1|X) is the propensity score; consistent if either g₀ or m₀ is correct; trimming_threshold = 0.01. Mean ATE ≈ $8,213.
• IIVM — Interactive IV Model (estimand: LATE on compliers): treatment = p401 (participation, endogenous), instrument Z = e401 (eligibility). Wald-type ratio θ_LATE = (E[Y|Z=1] − E[Y|Z=0]) / (E[D|Z=1] − E[D|Z=0]). Mean LATE ≈ $11,746.

Covariates (X). The base specification uses 9 covariates (age, inc, educ, fsize, marr, twoearn, db, pira, hown); a flexible specification adds quadratic terms for the four continuous covariates (age, inc, educ, fsize) so the Lasso can capture nonlinear confounding. nifa and tw are alternative wealth measures present in the source file but not used as outcome or covariates in the tutorial.

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

• Real survey data, single snapshot. These are real 1991 SIPP households, not simulated; but the cross-section is one point in time, so dynamic effects of 401(k) participation over time are not identified.
• Conditional exogeneity. The ATE estimates assume eligibility is as good as randomly assigned after conditioning on the observed covariates. Unobserved factors (e.g. financial literacy) that affect both eligibility and savings would bias the estimates.
• Extreme asset values. net_tfa ranges from −$502,302 to $1,536,798 (and tw up to $2,029,910); a handful of outliers influence the mean-based ATE estimates.
• IIVM identifies the LATE, not the ATE. The ≈$11,746 IIVM estimate applies only to compliers (households who participate because eligible) and should not be generalized to the full population without the monotonicity assumption.
• nifa / tw are extras. The tutorial's outcome is net_tfa; nifa (non-401(k) financial assets) and tw (total wealth) ship in the source file but are not modeled — included here for completeness.