Source
Prashant Garg and Thiemo Fetzer, “Causal Claims in Economics” (OSF 2024; arXiv v2, February 2026), LIT-079, read in full in NOTE-083, with v1 compared for the reframing.
What was actually shown
The measure. Each paper becomes a directed graph over JEL concepts. An edge is a stated source → sink relation, and it is “causal” when the paper supports it with DiD (including TWFE and event studies), IV, an RCT, an RDD or synthetic control. Structural estimation and causal language without such a design count as non-causal (§2.2). GPT-4o-mini reads the first 30 pages three times, extracts edges three times from each summary, and an edge is kept if it appears in at least 4 of the 9 lists. That leaves 42,404 papers, not 44,852 (Table A10).
The trend (§3, Fig. 2a). 23.8% of baseline edges are causal. The mean per-paper causal share is 7.7% (1990), 8.6% (2000), 21.5% (2010), 31.7% (2020) and 32.6% (2023). By field, pre- to post-2000: Health 9.7 → 41.9%, Urban 9.0 → 37.4, Development 6.9 → 37.2, Macro 6.3 → 12.9. Paper-level method shares move the same way: DiD 10.5% (1980) → 20.3% (2023), RCT 0.8% (2000) → 8.0%. The field ordering agrees with Goldsmith-Pinkham (2024), which the paper cites.
What could have come out otherwise. The share could have been flat, or its direction could have depended on the overlap threshold. It rises at every threshold (Fig. A1), though its level rises with the threshold too: 0.152 at EO = 1, 0.225 at EO = 4, 0.270 at EO = 9 (Table A10).
What this does not say
- That the edges are accurate. No edge was checked against a human label. The external checks are of paper-level method and field labels (n = 285): DiD, RDD and IV reach F1 0.68–0.74, but RCT per-run F1 is 0.03 and field precision 0.05–0.43 (Tables A6–A7). The snippet check is self-consistency and stops at EO = 3. The v1 validation figures (F1 0.83–0.99) were rewritten in v2 to these.
- That the rise is in economics rather than in the extractor. Modern papers name their designs; older ones may not. A DiD rate of 10.5% and an RDD rate of 1.3% in 1980 look like label noise (NOTE-083’s inference). With method-mix and graph-volume controls the trend falls from +0.0187 to +0.0011 per decade. The promised robustness table is absent, and the per-decade unit is hard to square with the headline levels.
- That causal claims pay. The +1.36 pp top-five probability per doubling of causal edges comes from bivariate regressions with year fixed effects only, no field controls, on the roughly 28% of papers with a causal edge, against non-causal measures estimated on about 83%. The paper itself calls β “a conditional association, not a causal effect”.
- That the claims are true. The pipeline records what papers claim, not whether the identification works (fn. 1).
Connections
- Cartwright’s account of evidence for singular causal claims (THEORY-451, from LIT-213) already names this paper as documenting the narrowing to RCT-, DiD- and IV-backed claims she describes. Neither cites the other.
- NOTE-083 reads the paper mainly as a cautionary example of LLM-assisted extraction, the record’s own method. Validating the unit of analysis one uses, here edges rather than paper labels, is the lesson; it is machine-learning practice and is not filed here.
Narrowed and promoted (2026-10-20)
Batch 074 of the promotion audit narrowed the title to the measurement and promoted it under ADR-047 v3. The numbers it keeps are in the paper’s running text (§3): 7.7% (1990), 8.6% (2000), 21.5% (2010), 31.7% (2020), and the per-decade slopes +0.0187 unadjusted, +0.0174 with field controls and +0.0011 with method-mix and graph-volume controls. The units of the slopes are the paper’s and do not obviously match the levels (see above). The wider claim, that economics papers increasingly back their claims with design-based identification, is the paper’s own reading of the series. It is held as THEORY-824, Proposed until the edges are audited against human labels.