Curious how accurate multimodal LLMs are at digitizing diversely formatted historical tables?

➡️ pretty accurate!
➡️ helps a lot with harmonization
➡️ far cheaper than manual digitization
➡️ but iterative, disciplined prompt training is important!

Check out our new WP - arxiv.org/abs/2505.11599

5⃣ Finally, how reasonable is assumption ii? It's ok, but the CPS does appear to undersample younger recent immigrants and female recent immigrants. However, these differences don't seem extreme.

4⃣ This does explain 20%-30% of the difference between CES and CPS measures of employment growth, though, bringing them more in line with each other.

3⃣ Then we calculate adjustments to headline employment statistics. They don't change too much, tho, bc immigrant employment doesn't look very different from native employment (at least for these headline stats):

- UR ⬆️ 0.05pp (5bp)
- LFPR ⬆️ 0.03pp (3bp)
- EPOP flat

Doing so suggests that the CPS normally (pre-COVID) undercounts about 1 million immigrants, but this has increased to more than 2 million by July 2024 (the latest data I could... publishing delays... 🤷)

This pattern is consistent -- it's shows up averaging CPS data on imputed arrivals over 1994-2019:

New 🚨 Research Brief on adjusting for undercounting of recent immigrants in the CPS (Bsky post!)

1⃣ I show that the CPS undercounts recent immigrants. For ex., the sample population of the 1994 cohort continues to rise until 1998 or even 2002. This shouldn't be...

tinyurl.com/ms4pvk67

We turn to a structural model to explore. The model's pretty cool, but...

tl;dr no, work and school are no substitutes for teens. GDLs decrease access to work and decrease access to distracting activities that take away from schooling. (5/5)

This teens also decrease their work quite a bit... So are these substitution effects?

We made a picture! (4/5)

These translate into greater high-school completion later in life! Impacted teens are more likely attend school as 17yo's, and more likely to have a regular HS degree or GED in their 20s and 30s. (3/5)

paper link: https://vkbostwick.weebly.com/gdlanddropouts.html

Using variation in the adoption of GDLs over time, we find that GDLs increase 16yo schoolgoing in states that allow 16yo's to drop out of school. (Results are pretty robust across designs + no pre-trends). (2/5)

Research blurb :: In our paper on teen driving restrictions, we show **long-run** effects of graduated driver licensing on education --> increasing minimum driving age increases high school completion when teens can otherwise drop out.

abstract ⬇️

thread and links follow... (1/5)