Job Training Partnership Act

This dataset was constructed for the empirical application in Han (2021), "Optimal Dynamic Treatment Regimes and Partial Welfare Ordering." It combines data from the Job Training Partnership Act (JTPA) study with auxiliary data on local high school density used as an instrumental variable for educational attainment. The JTPA was a large-scale US randomized evaluation of subsidized job training for disadvantaged adults conducted in the early 1990s. The dataset covers disadvantaged individuals and records their schooling decisions, job training participation (and assignment), pre-program and post-program earnings, and demographic information. It is used to study the optimal sequential policy of first assigning a high school diploma and then assigning job training, with the goal of maximizing employment outcomes. It can be used to validate instrumental variable (IV) methodology.

Task: The collection can be used to study causal inference algorithms.

Summary: 

Size of dataset: 9,224 x 11

Task: Causal Inference

Data Type: Mixed Data

Dataset Scope: Standalone Dataset

Ground Truth: Known Graph

Temporal Structure: Static Data

License:  CC BY-NC 4.0

Missing Values: No Missing Values

Features:

sex: Binary covariate: 1 if male, 0 if female.

n_hs2: Continuous variable: number of high schools per square mile in the individual's local area. Used to construct the instrument for schooling.

Z1: Binary instrument: indicator that local high school density is at or above 35 per square mile. Serves as an IV for D1 (high school diploma).

edu: Continuous variable: years of completed education.

D1: Binary treatment indicator: whether the individual has a high school diploma, i.e., 12 or more years of education (derived from `edu`).

prevearn: Continuous intermediate outcome: pre-program earnings.

Y1: Binary intermediate outcome: indicator that pre-program earnings are at or above the 80th percentile (derived from `prevearn`).

Z2: Binary instrument: whether the individual was randomly assigned to job training under the JTPA program (1 = assigned, 0 = not assigned). Serves as an IV for D2.

D2: Binary treatment indicator: whether the individual actually received job training (1 = received, 0 = did not).

earnings: Continuous outcome: total earnings (in USD) in the 30 months following the job training program.

Y2: Binary terminal outcome: indicator that 30-month post-program earnings are at or above the sample median (derived from `earnings`).

Files: 

jtpa.txt: dataset

Graph:

The dataset has a natural and well-motivated causal structure that can be described as a standard DAG (Abadie et al., (2002)). The two-stage sequential IV treatment is:

Z1 -> D1 -> Y1Z2 -> D2 -> Y2

D1 <-> Y1D2 <-> D2

D1 -> Y2

sex -> Y1sex -> Y2

It is valid to substitute the coarsened variables `Z1`, `D1`, `Y1`, `Y2` in the graph by `n_hs2`, `edu`, `prevearn` and `earnings`, respectively.

本数据集专为Han（2021）发表的"Optimal Dynamic Treatment Regimes and Partial Welfare Ordering"一文的实证应用所构建。该数据集整合了职业培训伙伴法案（Job Training Partnership Act, JTPA）研究的相关数据，以及用于作为受教育水平工具变量的本地高中密度辅助数据。JTPA是20世纪90年代初在美国开展的一项针对弱势成年人的补贴式职业培训大型随机评估项目。本数据集覆盖弱势成年人群体，记录了他们的就学决策、职业培训参与情况（及分配情况）、项目前后的收入数据与人口统计学信息。该数据集被用于研究“先授予高中文凭、再开展职业培训”的最优序贯政策，目标是最大化就业相关成果，同时可用于验证工具变量（Instrumental Variable, IV）方法的有效性。

研究任务：
本数据集可用于因果推断（causal inference）算法研究。

数据集摘要：
数据集规模：9,224 × 11
研究任务：因果推断
数据类型：混合数据
数据集范围：独立数据集
真实因果结构：已知因果图
时间结构：静态数据
许可证：CC BY-NC 4.0
缺失值情况：无缺失值

特征说明：
sex：二元协变量，男性取值为1，女性取值为0。
n_hs2：连续型变量，代表受访者所在区域每平方英里内的高中数量，用于构建就学相关的工具变量。
Z1：二元工具变量，当本地高中密度大于等于35所/平方英里时取值为1，该变量作为D1（高中文凭）的工具变量。
edu：连续型变量，代表受访者完成的教育年限。
D1：二元处理变量指示符，代表受访者是否拥有高中文凭，即受教育年限达到12年及以上（由`edu`字段推导得出）。
prevearn：连续型中间结果变量，代表受访者参与项目前的收入水平。
Y1：二元中间结果变量，当受访者项目前收入达到或高于样本80分位数时取值为1（由`prevearn`字段推导得出）。
Z2：二元工具变量，代表受访者是否被随机分配至JTPA项目的职业培训组（1为被分配，0为未被分配），该变量作为D2的工具变量。
D2：二元处理变量指示符，代表受访者实际是否接受了职业培训（1为接受，0为未接受）。
earnings：连续型结果变量，代表受访者在职业培训项目结束后30个月内的总收入（单位：美元）。
Y2：二元最终结果变量，当受访者培训后30个月的收入达到或高于样本中位数时取值为1（由`earnings`字段推导得出）。

数据文件：
jtpa.txt：本数据集文件。

因果结构图说明：
本数据集具备自然且符合研究动机的因果结构，可通过标准有向无环图（Directed Acyclic Graph, DAG）进行描述（Abadie等，2002）。该两阶段序贯工具变量处理流程的因果关系为：
Z1 → D1 → Y1
Z2 → D2 → Y2
D1 ↔ Y1
D2 ↔ D2
sex → Y1
sex → Y2
可分别将图中的粗化变量`Z1`、`D1`、`Y1`、`Y2`替换为`n_hs2`、`edu`、`prevearn`与`earnings`。