About

Policy Learning argues about retirement policy, labour markets, and the applied econometrics underneath both — from the primary documents rather than from the coverage of them.

Who writes this

I am an economist applying causal inference and machine learning to policy and commercial problems — difference-in-differences, synthetic control, regression discontinuity, doubly robust estimation, survival analysis, and gradient boosting. I hold a PhD in Political Economy from the Stanford Graduate School of Business and an MS in Statistics from Stanford University; my research spanned best-arm identification in multi-armed experiments and survey experiments on political behaviour. I am not a financial adviser, an accountant, or a lawyer, and nothing published here is advice — see disclosures.

Elsewhere:tomhsyu.com·linkedin.com/in/tom-hs-yu

What this site publishes

• Retirement

  Retirement policy read from the primary documents — Federal Register rules, IRS notices, PBGC rate tables, and the data underneath them.

• Economy

  Applied macro and public finance, argued from the releases rather than the coverage of the releases.

• Work

  Labour markets, occupations, and what the task-level data actually shows about automation.

• Research

  Working papers and methods, read closely enough to disagree with.

The standard every post is held to

Anyone can assert a standard on an About page. These four are written down so that you can hold a specific post against them and tell me when one of them fails.

• It starts from a primary document. A Federal Register rule, an agency notice, a working paper, a statistical release — not a news story about one. The document is linked, so you can disagree with my reading of it rather than take it on trust.
• It argues a claim that could turn out to be wrong. A thesis that cannot be falsified is not worth your time or mine. If a post has nothing at stake, it does not get written.
• It contains at least one thing made here rather than quoted — a chart built from raw data with the underlying CSV linked, a figure computed rather than repeated from someone else's summary, or a specific disagreement with a named author.
• Volume is capped rather than maximised. There is a ceiling on how much appears in a week, and it exists so that nothing gets published merely because the schedule asked for something.

Corrections

If something here is wrong I want to know, and the smaller the error the more I want to know about it — a misread base year does more damage than a bad argument, because a bad argument announces itself and a wrong number does not. Corrections are made in the post itself, and the post then carries a visible updateddate saying so. Nothing is quietly edited after publication. Reach me through thecontact page.