cmpatino/acs-income-2018

ACS Income 2018 (Folktables ACSIncome)

Dataset Description

This dataset contains individual-level records from the 2018 American Community Survey (ACS) Public Use Microdata Sample (PUMS), prepared for the binary income prediction task defined in the folktables benchmark. The goal is to predict whether a person's total annual income exceeds $50,000.

The dataset includes 1,611,572 individuals across the United States with 10 demographic, employment, and socioeconomic features. Each record is labeled with a binary outcome indicating whether the individual's total personal income (PINCP) exceeds $50,000 per year.

This task was introduced as a modern replacement for the UCI Adult/Census Income dataset (which uses 1994 Census data) in:

Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt. "Retiring Adult: New Datasets for Fair Machine Learning." Advances in Neural Information Processing Systems 34 (NeurIPS 2021).

Usage

Loading with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("cmpatino/acs-income-2018")
df = ds["train"].to_pandas()

# Features and target
X = df.drop(columns=["PINCP > 50k"])
y = df["PINCP > 50k"]

print(f"Samples: {len(df)}, Features: {X.shape[1]}, Positive rate: {y.mean():.3%}")
# Samples: 1611572, Features: 10, Positive rate: 37.139%

Loading directly with pandas

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/cmpatino/acs-income-2018/data/train-00000-of-00001.parquet"
)

# Quick overview
print(df.shape)            # (1611572, 11)
print(df.dtypes.value_counts())
# float64    10
# bool        1

# Target distribution
print(df["PINCP > 50k"].value_counts())
# False    1013075
# True      598497

# Income rate by sex
print(df.groupby("SEX")["PINCP > 50k"].mean())
# SEX
# 1.0    0.445   (Male)
# 2.0    0.289   (Female)

Dataset Summary

Property	Value
Individuals	1,611,572
Features	10 (demographic, employment, socioeconomic)
Target	PINCP > 50k (binary: True = income > $50K, False = income ≤ $50K)
Positive rate	37.1% (598,497 individuals with income > $50K)
Geographic scope	United States (all 50 states + DC + Puerto Rico)
Survey year	2018 (1-year estimates)
Source	U.S. Census Bureau, American Community Survey PUMS

Prediction Task

The task is binary classification: predict whether an individual's total personal income exceeds $50,000 per year.

Filtering Criteria

The following filters are applied to the raw ACS PUMS data to construct this dataset (following the folktables ACSIncome definition):

• AGEP > 16 — individuals older than 16 years
• PINCP > 100 — individuals who reported meaningful income (> $100)
• WKHP > 0 — individuals who worked at least 1 hour per week in the past 12 months
• PWGTP ≥ 1 — individuals with valid person weight

Features

The 11 columns are organized as follows:

Target Variable

Column	Full Name	Type	Description
PINCP > 50k	Total Person's Income > $50,000	bool	True if total personal income exceeds $50,000/year, False otherwise

The target is derived from the raw PINCP variable (total person's income), which aggregates wage/salary income, self-employment income, interest/dividends, Social Security, public assistance, retirement income, and other sources.

Feature Variables (10 columns)

AGEP — Age

• Type: Continuous (integer values stored as float64)
• Range: 17–96
• Description: Person's age in years. Filtered to individuals older than 16.

---

COW — Class of Worker

• Type: Categorical (integer codes stored as float64)
• Description: Type of employment arrangement.

Code	Meaning
1	Employee of a private for-profit company or business, or of an individual, for wages, salary, or commissions
2	Employee of a private not-for-profit, tax-exempt, or charitable organization
3	Local government employee (city, county, etc.)
4	State government employee
5	Federal government employee
6	Self-employed in own not incorporated business, professional practice, or farm
7	Self-employed in own incorporated business, professional practice, or farm
8	Working without pay in family business or farm

---

SCHL — Educational Attainment

• Type: Categorical (integer codes stored as float64)
• Description: Highest level of education completed or highest degree received.

Code	Meaning
1	No schooling completed
2	Nursery school / preschool
3	Kindergarten
4–11	Grades 1 through 8
12	Grade 9
13	Grade 10
14	Grade 11
15	12th grade, no diploma
16	Regular high school diploma
17	GED or alternative credential
18	Some college, less than 1 year
19	1 or more years of college credit, no degree
20	Associate's degree
21	Bachelor's degree
22	Master's degree
23	Professional degree beyond bachelor's
24	Doctorate degree

---

MAR — Marital Status

• Type: Categorical (integer codes stored as float64)
• Description: Current marital status.

Code	Meaning
1	Married
2	Widowed
3	Divorced
4	Separated
5	Never married or under 15 years old

---

OCCP — Occupation

• Type: Categorical (4-digit integer codes stored as float64)
• Description: Census occupation code based on the Standard Occupational Classification (SOC) system. Encodes the type of work performed. Over 500 distinct occupation codes are grouped into major categories including Management (0010–0430), Business/Financial (0500–0960), Computer/Mathematical (1005–1240), Engineering (1305–1560), Science (1600–1965), Community/Social Services (2000–2060), Legal (2100–2160), Education (2200–2550), Healthcare (3000–3540), Food Preparation (4000–4160), and many more.

Full code list: ACS PUMS Occupation Codes

---

POBP — Place of Birth

• Type: Categorical (3-digit integer codes stored as float64)
• Description: Respondent's place of birth. Codes 001–059 correspond to U.S. states and territories (FIPS codes); codes 100+ correspond to foreign countries and regions worldwide. Approximately 215 distinct values.

Full code list: ACS PUMS Data Dictionary

---

RELP — Relationship to Householder

• Type: Categorical (integer codes stored as float64)
• Description: Relationship of the person to the household reference person.

Code	Meaning
0	Reference person (householder)
1	Husband / wife
2	Biological son or daughter
3	Adopted son or daughter
4	Stepson or stepdaughter
5	Brother or sister
6	Father or mother
7	Grandchild
8	Parent-in-law
9	Son-in-law or daughter-in-law
10	Other relative
11	Roomer or boarder
12	Housemate or roommate
13	Unmarried partner
14	Foster child
15	Other nonrelative
16	Institutionalized group quarters population
17	Noninstitutionalized group quarters population

---

WKHP — Usual Hours Worked per Week

• Type: Continuous (integer values stored as float64)
• Range: 1–99
• Description: Usual number of hours worked per week over the past 12 months. Value 99 indicates 99 or more hours. Filtered to individuals with WKHP > 0.

---

SEX — Sex

• Type: Binary categorical (integer codes stored as float64)
• Description: Respondent's sex as recorded in the survey. Protected attribute for fairness analysis.

Code	Meaning
1	Male
2	Female

---

RAC1P — Race (Recoded)

• Type: Categorical (integer codes stored as float64)
• Description: Recoded detailed race variable. Protected attribute for fairness analysis.

Code	Meaning
1	White alone
2	Black or African American alone
3	American Indian alone
4	Alaska Native alone
5	American Indian and Alaska Native tribes specified; or American Indian or Alaska Native, not specified and no other races
6	Asian alone
7	Native Hawaiian and Other Pacific Islander alone
8	Some Other Race alone
9	Two or More Races

Protected/Sensitive Attributes

The following features are considered protected attributes under U.S. anti-discrimination law and are central to algorithmic fairness research:

Attribute	Variable	Relevant Legislation
Race	RAC1P	Civil Rights Act (Title VII), Equal Credit Opportunity Act
Sex	SEX	Civil Rights Act (Title VII), Equal Pay Act
Age	AGEP	Age Discrimination in Employment Act

The folktables paper demonstrates significant performance disparities across demographic groups for income prediction models, highlighting the importance of fairness-aware evaluation.

Feature Distributions

Variable	Unique Values	Most Common Value	Notes
AGEP	80	—	Mean: 43.4, Std: 15.3
COW	8	1 (Private for-profit, 65.8%)	—
SCHL	~23	—	Ranges from no schooling to doctorate
MAR	5	1 (Married, 54.6%)	—
OCCP	~500+	—	High cardinality
POBP	~215	—	High cardinality
RELP	18	0 (Householder) and 17 (Non-institutional GQ) most common	—
WKHP	98	40 (standard full-time)	Mean: ~40
SEX	2	1 (Male, 52.0%)	Protected attribute
RAC1P	9	1 (White alone, 77.6%)	Protected attribute

Benchmark Results

From Table 1 of Ding et al. (2021), evaluated on the ACSIncome task (2018, nationwide):

Model	Accuracy
Constant predictor (majority class)	63.1%
Logistic Regression	77.1%
Gradient Boosted Decision Tree	79.7%

Background

The American Community Survey (ACS) is an ongoing survey conducted by the U.S. Census Bureau that provides vital demographic, social, economic, and housing information about the U.S. population every year. Unlike the decennial census (which counts all residents), the ACS samples approximately 3.5 million addresses annually and provides detailed socioeconomic data used for government planning, research, and policy decisions.

The Public Use Microdata Sample (PUMS) provides individual-level records from the ACS with privacy protections applied (geographic detail limited to Public Use Microdata Areas of ~100,000+ people). The 2018 1-year PUMS dataset covers the period January 1 – December 31, 2018.

This dataset was created as part of the folktables project, which proposed the ACSIncome task as a modern, annually-updatable replacement for the widely-used UCI Adult/Census Income dataset (based on 1994 Census data). Key advantages over UCI Adult:

• Much larger: 1.6M+ samples vs. ~49K in UCI Adult
• Current: 2018 data vs. 1994 data
• Updatable: New data released annually, enabling temporal distribution shift studies
• Geographically granular: Enables state-level fairness analysis
• Well-documented: Official Census Bureau documentation and data dictionaries

Validation Protocols

The folktables paper proposes evaluating models along multiple axes:

1. Standard train/test split — Random split for overall accuracy estimation
2. Temporal shifts — Train on one year, evaluate on a different year (e.g., train on 2014, test on 2018) to study distribution shift
3. Geographic shifts — Train on one state, evaluate on a different state to study geographic generalization
4. Subgroup evaluation — Evaluate accuracy and calibration separately for each protected group defined by SEX, RAC1P, and AGEP

Ethical Considerations

This dataset is intended for research purposes in machine learning, algorithmic fairness, and socioeconomic analysis.

Important considerations:

• Fairness implications: Models trained on this data may exhibit disparate performance across demographic groups. The folktables paper documents significant accuracy gaps between racial and gender groups. Any deployment of income prediction models should include thorough fairness audits.
• Historical bias: The data reflects existing societal inequalities in income distribution by race, sex, and other attributes. Models trained to predict income will inevitably encode these historical patterns.
• Privacy: While PUMS data has privacy protections (geographic coarsening, top/bottom-coding of extreme values, data swapping), researchers should handle the data responsibly and avoid re-identification attempts.
• Not for individual decisions: This dataset and models trained on it should not be used to make decisions about individual people's creditworthiness, employment eligibility, or other consequential outcomes.

Data Sources

• Primary data: U.S. Census Bureau, American Community Survey 2018 1-Year PUMS
• Data dictionary: 2018 ACS PUMS Data Dictionary (PDF)
• Documentation: ACS 2018 Microdata Documentation
• Task definition: folktables (GitHub)

Citation

@article{ding2021retiring,
  title   = {Retiring Adult: New Datasets for Fair Machine Learning},
  author  = {Frances Ding and Moritz Hardt and John Miller and Ludwig Schmidt},
  journal = {Advances in Neural Information Processing Systems},
  volume  = {34},
  year    = {2021},
  url     = {https://arxiv.org/abs/2108.04884}
}

License

This dataset is released under the CC-BY-4.0 license. The underlying ACS PUMS data is in the public domain (produced by the U.S. federal government).