Paper Application: County Unemployment in April 2020

This vignette is the MSC.jl version of the empirical application in Shen, Song, and Abadie: the effect of COVID-19 stay-at-home orders on county unemployment in April 2020.

Target From The Paper

The paper's application is a high-dimensional synthetic-control problem:

• outcome: county monthly unemployment rate
• controls: counties in Arkansas, Iowa, Nebraska, North Dakota, South Dakota, Utah, and Wyoming
• treated units: counties in states with stay-at-home orders
• post-treatment period: April 2020
• target estimand: ATT across treated counties

The draft reports:

The text says the data start in January 2010, but also reports T0 = 147. With April 2020 as the post period, T0 = 147 corresponds to January 2008 through March 2020. The example script therefore defaults to January 2008. It also excludes Alaska and the District of Columbia, which aligns the BLS county file with the paper's reported treated-unit count.

Data

The public data source is the BLS Local Area Unemployment Statistics flat-file archive:

https://download.bls.gov/pub/time.series/la/

BLS sometimes blocks automated bulk downloads. For that reason, the package does not download hundreds of megabytes during documentation builds. Instead, download the county LAUS data file manually from the BLS flat-file archive and save it as data/la.data.64.County from the directory where you run the example. You can also set BLS_LAUS_COUNTY_FILE to an explicit file path.

Run The Example

From the package root:

julia --project=. examples/03_covid_sah_orders.jl

The script is examples/03_covid_sah_orders.jl. It does the full data construction:

1. reads the BLS county unemployment-rate series;
2. keeps complete county histories through April 2020;
3. excludes Alaska and the District of Columbia to match the reported sample;
4. uses the seven no-order states as controls;
5. builds the common-adoption panel;
6. estimates MSC with the application value lambda = 0.03;
7. reports ATT, selected lambda, and pre-treatment RMSE.

The core estimation call is:

est = msc_estimate(panel; lambdas = [0.03], standardize = false)

Set MSC_CV=true to run a cross-validated penalty path instead of the fixed application value. The full application has 438 control counties and 2,674 treated counties, so the example defaults to MSC_MAX_ITER=100 and MSC_TOL=1e-4 to keep the script runnable on a laptop.

What We Replicated

Using the BLS flat file downloaded under data/, the example constructs the same-sized county panel as the paper:

Controls: 438 counties in Arkansas, Iowa, Nebraska, North Dakota, South Dakota, Utah, Wyoming
Treated:  2674 counties
T0:       147 pre-treatment months

The script now prints the paper comparison directly:

Paper application replication check
--------------------------------------------------------------------
quantity                             MSC.jl          paper
--------------------------------------------------------------------
control counties                        438            438
treated counties                       2674          2,674
pre-treatment months                    147            147
April 2020 ATT, pp                   4.9552           5.06
--------------------------------------------------------------------
Sample counts match the paper. The ATT is an application replication
using the current BLS flat-file vintage and the script's solver settings.

The important distinction is that MSC.jl currently replicates the application workflow and the reported sample counts. The numerical ATT is close under the default bounded run, but it is not a frozen exact reproduction of the published number because the paper does not provide a replication archive with the exact BLS vintage, county inclusion file, and optimizer/tuning settings.

Interpreting Differences

If exact equality with 5.06 is the goal, the next step is not more prose in the vignette. It is to pin the original paper's replication archive or tune the optimizer against an explicitly stated target. Without that archive, this example should be read as a transparent application replication: same design, same sample size, public BLS data, and an openly reported estimate.