# Lecture 9 data

## Data for alternative explanations

The lecture bundles and preloads `labor_market.csv` for the tech-expansion
exercise and optional hiring/turnover exercise. `remote_work.csv` remains
saved for later use. Their source downloads are frozen as of September 29,
2026. Recreate them with:

```sh
Rscript course_design/prepare_lecture9_context.R
```

### `labor_market.csv`

140 monthly observations, January 2015 through August 2026, with no missing
values. The source is `source/fred_labor_2026-09-29.csv`, downloaded using
[FRED's CSV export](https://fred.stlouisfed.org/graph/fredgraph.csv?id=CES6054150001,FEDFUNDS,JTSHIR,JTSQUR,JTSLDR&cosd=2015-01-01&coed=2026-08-01).
That export returned the longer available history; the preparation script
selects the stated teaching period and renames columns, preserving values.

- `tech_jobs`: [BLS CES6054150001](https://fred.stlouisfed.org/series/CES6054150001),
  employees in computer systems design and related services, thousands,
  seasonally adjusted. This industry includes workers of all ages and occupations.
- `interest_rate`: [Federal Reserve FEDFUNDS](https://fred.stlouisfed.org/series/FEDFUNDS),
  effective federal funds rate, percent, monthly average, not seasonally adjusted.
- `hires`: [BLS JTSHIR](https://fred.stlouisfed.org/series/JTSHIR), hiring rate.
- `quits`: [BLS JTSQUR](https://fred.stlouisfed.org/series/JTSQUR), voluntary quit rate.
- `layoffs`: [BLS JTSLDR](https://fred.stlouisfed.org/series/JTSLDR), layoffs and discharges rate.

The three JOLTS series are monthly national nonfarm rates, seasonally adjusted,
expressed as percentages of employment. They do not isolate new graduates or
replacement hires. The lecture's annual summaries use complete years; the
2026 summary printed by the preparation script covers January–August only.

The tech exercise filters to 2019 onward and converts thousands of employees
to millions before plotting. The hiring exercise filters the same period,
selects month and the three rates, gathers the rates into two columns, and
draws three lines. Both exercises use the supplied values directly.
The industry series peaks in March 2023; young software-developer employment
in ADP peaks in September 2022. These measure different populations.
The former interest-rate join is preserved in the archived lecture source.

### `remote_work.csv`

76 monthly observations, May 2020 through August 2026. Source:
[WFH Research JSON](https://wfhresearch.com/datafiles/WFHtimeseries_monthly.json),
saved as `source/wfh_monthly_2026-09-29.json`, release `WFH_release_2026_08`.
We extract `wfhcovid_matquestion` and preserve its observations.

`wfh_percent` is the percentage of full paid workdays worked from home.
The SWAA sample for this series requires at least $10,000 in prior-year
earnings and uses survey weights. It does not directly measure mentoring,
junior hiring, or a particular employer's working arrangements.

The source's December 2019 row is an annual ATUS estimate and is omitted
from the monthly teaching file. The authors interpolate June 2020 from
May/July and September 2023 from August/October. We retain those values;
the lecture plot begins in January 2021.

Credit: Barrero, Bloom, and Davis (2021), *Why Working from Home Will Stick*.
Data are distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

### Interpretation

These series let students examine the timing and breadth of the proposed
mechanisms. They do not estimate how much any mechanism caused young-worker
employment to decline. Meta's company statement and the formal proximity
study remain in the archived lecture; the underlying firm data from that
study have not been obtained.

## EIG employment rates

`eig_employment.csv` comes from Ozimek and Goldschlag's March 5, 2026 article,
[AI and Young-adult Jobs: The Real Mystery](https://agglomerations.eig.org/p/ai-and-young-adult-jobs-the-real).
The original [replication workbook](https://github.com/EIG-Research/EIG-ai-unemp-she-wrote/blob/82634cd1232db13fcc9ffbea13de1f2ec0150eab/output/figures.xlsx)
is preserved as `source/eig_figures.xlsx`, downloaded September 29, 2026.
The pinned repository commit is `82634cd1232db13fcc9ffbea13de1f2ec0150eab`.

The authors use weighted Current Population Survey observations, ages 15–65,
excluding the armed forces. Their older employment group is ages 26–65.
The education split is bachelor's or higher versus below bachelor's,
which includes associate degrees and some college.

Our file extracts the six `ma_er_*` columns from the workbook's `all` sheet.
Each row is a month, group, and population definition. `employment_rate` is
the authors' trailing 12-month average of the percentage employed.
`Everyone` includes students; `Out of school` excludes those attending school.
Values are rates in percent, without indexing or additional smoothing.
The article's displayed figures additionally index these rates to June 2023;
the lecture uses the published rates so students can interpret percentages directly.

There are 576 rows: six series over 96 supplied months, January 2018 through
January 2026. October 2025 is absent from the source. We preserve the authors'
supplied averages and do not insert an observation for that month.
The lecture focuses on June 2023 onward. Its EIG and ADP plots have different
end dates and cover different populations.

Recreate the CSV from the frozen workbook with:

```sh
Rscript course_design/prepare_eig.R
```

From June 2023 to January 2026, the smoothed employment-rate changes are
-1.41 percentage points for young adults with bachelor's degrees or higher,
-1.56 for those below bachelor's, and +0.56 for ages 26–65.
Excluding people in school gives -1.52, -2.06, and +0.61 respectively.
These descriptive results do not identify the cause of the changes.

## Newspaper example

`newspaper_circulation.csv` transcribes the ten labeled points in
`../images/news.png`, reused from F25 Lecture 8. Each row is a newspaper
and labeled year. The final Post point is labeled March 31, 1981 in the
graphic; the file represents dates by year. The 1979 Post value is 624,000,
correcting a 642,000 transcription in the old lecture code.

## Employment data

Source: Stanford Digital Economy Lab and ADP Research,
[Canaries Dashboard](https://digitaleconomy.stanford.edu/project/indicators/canaries-dashboard/).

Downloaded September 28, 2026. Source vintage: September 17, 2026.
Observations run from September 2021 through August 2026.

- [Software developers download](https://storage.googleapis.com/aviary-del-public/release_memos/latest/downloads/canaries_software_developers_results.zip)
- [Age and AI exposure download](https://storage.googleapis.com/aviary-del-public/release_memos/latest/downloads/canaries_age_by_exposure_results.zip)

`software_developers.csv` has 60 rows, one per month, and six age columns.
`age_by_exposure.csv` has 300 rows, one per month and AI-exposure group,
with the same six age columns. The exposure groups each contain about a fifth
of the ranked occupations; they contain different numbers of workers.
Exposure is based on estimates of which occupational tasks AI could assist
with. It does not measure actual AI adoption at each employer.

Each value is an employment index: 100 times employment divided by employment
in November 2022, separately for each series. Thus 80 means employment was
20% below its own November 2022 level. The index does not give the number
of workers, hires, job openings, or unemployed people. A group's age is
measured at each observation; this is not a cohort of the same people.

The sample includes firms with five years of ADP history and workers with
matched occupation codes. It is not representative of the whole U.S. labor
market. The dashboard describes associations between AI exposure and
employment; it does not identify a causal effect of AI.

The original CSVs and dictionaries are preserved in `source/`. Preparation
only shortens column names and removes the constant `vintage` column; no
rows, age groups, or values are removed. Run from `f26`:

```sh
Rscript course_design/prepare_canaries.R
```

The script reuses the frozen source files. It rejects a different vintage
so a later refresh cannot silently change the lecture answers.

## Why software developers?

The lecture's motivating question concerns data science careers. As checked
on September 28, 2026, the dashboard's public occupation-specific downloads
cover software developers, customer service representatives, stock clerks,
and home health aides. No separate data-scientist employment series is listed.
The age-by-exposure file pools occupations and cannot isolate data scientists.
The software-developer analysis is a case study; its results cannot be assumed
to hold for data scientists.

[ADP's May 19 report on data scientists](https://www.adpresearch.com/main-street-macro/data-scientists-high-demand-rebounding-wage-growth)
provides occupation-specific pay findings, including March 2026 median annual
pay of $130,000 across career stages and a pickup in wage growth in early 2026.
The lecture cites these findings separately. The NER Pulse workbook linked
from that article contains national employment changes, not a data-scientist
employment breakdown by age. No pay data have been substituted for employment.