README.md

# Measuring the Intensive Margin of Work: Task-Level Labor Input Data

This dataset provides **task-level measures of labor input within occupations**, constructed from O*NET task frequency data.

For each occupation–task pair, we estimate:

* **Task flow (μ)**: expected number of times a task is performed annually per worker within an occupation
* **Task share (π)**: proportion of total labor input allocated to the task

Under a homogeneous task-duration assumption, task shares can be interpreted as **time allocations across tasks within occupations**.

Unlike standard O*NET-derived measures, this dataset provides **statistically specified estimators of task-level labor input**, including uncertainty (variance) for incumbent-based estimates.

These measures enable:

* AI exposure measurement
* workforce decomposition
* automation targeting
* task-level economic analysis

---


## Measurement Units

* **Task flow (μ)**
  Expected annual number of task occurrences **per worker within an occupation**

* **Task share (π)**
  Fraction of total labor input allocated to a task within an occupation

  Interpretable as a time share under the assumption of homogeneous task duration.


---

## Data Files

Data is organized by:

* O*NET version
* estimate type (mean vs full)
* measure (flow vs share)

### Mean estimates (point estimates only)

* Task flow (μ):
  `task_labor_input_mean_estimates/{ONET_VERSION}/ONET_{ONET_VERSION}_weight_mode_STANDARD_task_flow_mean_estimates.csv`

* Task labor input share (π):
  `task_labor_input_mean_estimates/{ONET_VERSION}/ONET_{ONET_VERSION}_weight_mode_STANDARD_task_labor_input_mean_estimates.csv`

### Full estimates (mean + variance)

* Task flow (μ):
  `task_labor_input_full_estimates/{ONET_VERSION}/ONET_{ONET_VERSION}_weight_mode_STANDARD_task_flow_full_estimates.csv`

* Task labor input share (π):
  `task_labor_input_full_estimates/{ONET_VERSION}/ONET_{ONET_VERSION}_weight_mode_STANDARD_task_labor_input_full_estimates.csv`

---

## Mean vs Full Estimates

* **Mean estimates**
  Combine incumbent and analyst task ratings and report **point estimates only**.

* **Full estimates**
  Rely exclusively on **incumbent survey data**, which provides frequency distributions with sampling uncertainty.
  This enables construction of **fully specified estimators**, including:

  * mean (μ or π)
  * variance (Var)

Full estimates therefore support **statistical inference and uncertainty quantification**, while mean estimates provide broader coverage.


---

## Dataset Structure

Each dataset is defined at the **occupation–task level**, with one row per `(onetsoc_code, task_id)` pair.

### Columns

**Mean estimates:**

* `onetsoc_code` — O*NET occupation code
* `task_id` — O*NET task identifier
* `mean` — Estimated value (μ or π)

**Full estimates:**

* `onetsoc_code` — O*NET occupation code
* `task_id` — O*NET task identifier
* `mean` — Estimated value (μ or π)
* `variance` — Estimated variance of the estimator


---

## Data Source: O*NET

O*NET task frequency data is based on:

* **Incumbents (survey respondents)**
  Provide frequency distributions with associated standard errors

* **Analysts (occupation experts)**
  Provide point estimates without measures of dispersion

This dataset distinguishes between:

* combined estimates (mean only)
* incumbent-based estimates (mean + variance)

---

## Coverage

The dataset includes multiple O*NET releases (from version 20.1 onward).

Each version may differ due to:

* updates to task definitions
* changes in occupation coverage
* new survey responses

⚠️ **Version comparability note**
O*NET releases are not strictly comparable over time. Differences across versions may reflect survey and taxonomy updates rather than true economic changes.

---

## Quick Start

```python
import pandas as pd


job_task_input_mean_estimates = pd.read_csv(
    "https://huggingface.co/datasets/MIT-WAL/job_task_input_share/resolve/main/task_labor_input_mean_estimates/30_2/ONET_30_2_weight_mode_STANDARD_task_labor_input_mean_estimates.csv"
)
```

---

## Example Applications

* AI exposure measurement at the task level
* Workforce decomposition into task bundles
* Construction of task-based production functions

---

## Limitations

* Task durations are assumed homogeneous when constructing task shares
* Task flows rely on discretized frequency bins (midpoint approximation)
* Full estimates are limited to incumbent-based tasks
* Measurement error arises from survey sampling and bin approximation
* Cross-version comparisons should be interpreted with caution

---

## Citation

If you use this dataset, please cite:

Bouquet, Pierre and Sheffi, Yossi (2026).
*Measuring the Intensive Margin of Work: Task Shares and Concentration.*
MIT Center for Transportation & Logistics Research Paper No. 2026/004.

SSRN: https://ssrn.com/abstract=6174538
DOI: http://dx.doi.org/10.2139/ssrn.6174538

---

## License

Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International
(CC BY-NC-ND 4.0)
https://creativecommons.org/licenses/by-nc-nd/4.0/