The NBER Audit
Where the z-statistics fall
Just below against just above each threshold
Where the excess is largest
Share of results within 0.2 of the 5% threshold that sit above it. An even split is 50%.
The result a paper leads with
Stars the printed standard error cannot support
Where the unsupported stars sit
Authors with three or more papers checked
Who writes the series
Most papers in the top 10% of their year
Most papers
Widest circle of co-authors
Institutions
Programs
Literatures still written, less read
Later papers rank lower than early ones
Later papers rank higher
One literature, read paper by paper
How the data decide that someone is retired
Where the outcome variable comes from
Who cites whom inside the series
Instrument strength
Difference-in-differences after 2018
Referee cards
Method
A table enters the arithmetic only when its own notes say what the parentheses hold. A star claim is checked only when the same table also defines its stars, uses two-sided tests and reports conventional standard errors. Printed numbers are rounded, so each z-statistic is a range. A star is listed here only when the top of that range, taken at normal critical values, falls short of the level the star claims. Small-sample critical values are larger and would widen the shortfall.
Threshold counts use one de-rounded draw per result with a fixed seed, following Brodeur, Le, Sangnier and Zylberberg (2016) and Brodeur, Cook and Heyes (2020). The instrument threshold of 104.7 is from Lee, McCrary, Moreira and Porter (2022). Referee cards are written by a language model reading the full text; a point is kept only when its quotation appears in the paper.