Every research paper in the three top finance journals: its replication package, the empirical methods and data sources it uses (measured from the full text, reference list removed), and its instruments if it estimates IV. Search any phrase to chart its adoption.
Every research paper in JF, JFE and RFS, 2000–2025, whether or not it has a replication package. Design tags and the WRDS-data flag come from each paper's full text (a term mentioned at least twice); papers without full text on file carry no tags.
Links to the code-and-data package of each paper in JF, JFE and RFS. Each journal keeps packages somewhere different: JF as zip files on the Wiley article page ("Supporting Information"), JFE on Mendeley Data (linked from the article's data availability statement), RFS on Harvard Dataverse. Packages are matched to papers by the article DOI where the repository records it, otherwise by title.
Share of the journal's research papers in that year with at least one package link. Code-sharing policies started at different times (JF around 2016; JFE required it for papers first submitted from July 2021; RFS moved to Dataverse around 2021), so early years are mostly empty by design.
Phrases (up to three words) count papers mentioning them at least once; dictionary methods count at two or more mentions.
Mentions are not use: a paper that discusses DiD in its literature review still counts at one mention. Two or more mentions removes most of those cases.
These choices move levels, not trends. Across the 50 methods, yearly shares at one mention or at three correlate 0.95–0.97 with the two-mention series (levels about 40% higher and 25% lower); restricting to empirical papers correlates 0.99 (levels about 17% higher).
Empirical papers also carry labels from an LLM (Gemini Flash-Lite) that read each paper's first 12,000 characters. Agreement between those labels and the dictionary, at two or more mentions:
The LLM only saw the start of each paper (for JFE, mostly metadata), so disagreement is a prompt to look, not a verdict. Broad labels such as “var_time_series” and “thomson_reuters” cover more than the matching dictionary entry.
For every paper that mentions IV at least twice, Claude (Sonnet subagents, one paper at a time) read the passages around IV vocabulary and recorded each instrument, what it instruments, and the authors' case for relevance and the exclusion restriction, with a verbatim quote for each. Every quote is checked against the paper's text.
Journal filter below applies. Share of all research papers in the year. A paper counts only if it estimates IV itself; conditioning “instruments” in asset-pricing GMM do not count. Heckman selection models were coded inconsistently across readers (under 2% of instruments).
Share of named instruments in each family, by publication period. Journal filter below applies.