We show that TSLS is inadmissible under squared error loss whenever there are at least three instruments, for any covariance and weight matrix, and whether or not treatment effects are constant.
constructive proof for the stated known-covariance Gaussian limit experiment
Simple estimators that stabilize TSLS when the observed first stage is close to zero improve on it at every parameter value.
uniform risk inequalities under the stated Gaussian experiment and tuning conditions
In designs calibrated to published specifications, however, the resulting improvements in mean squared error are small except in the two specifications with the smallest first-stage Wald statistics.
consistent calibrated results from 70 specifications drawn from only two source papers
In simulations calibrated to published specifications, the variance reductions when the reduced form and first stage are zero are large with three instruments but negligible with 21 or more
two-paper calibration supports the reported contrast but provides a narrow empirical illustration
Derived from the full evaluation — not a separate score.
Strengths
Corollary 1 removes the bounded-loss and bounded-identification restrictions of the earlier result for TSLS. Theorem 1 and Section 4 provide explicit estimators rather than an existence-only inadmissibility argument.
↳ Section 1; Theorem 1; Corollary 1; Section 4
The proof architecture exposes the conditional-risk and convexity arguments underlying dominance. Exact-arithmetic certificates and independent implementations provide additional checks on the numerical tuning and risk calculations.
↳ Lemma 1; Theorem 1; Appendices A, B, and D
The paper explains why unbounded loss prevents automatic transfer to feasible estimators. It also reports the small calibrated gains and labels the two-paper simulation base illustrative.
↳ Section 1, limitation paragraph; Section 5, Table 1
Limitations
The guarantee assumes known covariance and fixed weights in the Gaussian limit experiment. It does not establish risk dominance when these quantities are estimated in ordinary applications.
↳ Section 1, limitation paragraph
The 70 specifications come from only two source papers, limiting what the exercise establishes about application settings. At the estimated coefficients, average MSE improvements are below 0.1%.
↳ Section 5, Table 1
The power-series interchange and certificate construction are terse, and the numerical evaluation of the reported ψ_k and c_k quantities is not fully described. These presentation gaps make reproducing parts of the supporting analysis more demanding.
↳ Section 4, Proposition 2 discussion; Appendix A, Proposition 1 proof; Appendix B
The central advance is the constructive extension from bounded-loss, bounded-identification inadmissibility to squared-error loss on an unrestricted identification space. Lemma 1, Theorem 1, and the exact tuning certificates make the main argument traceable, although parts of the appendices remain compressed. The known-covariance Gaussian experiment is a substantive boundary on applicability, not an error in the theorem. Section 5's mostly negligible MSE gains and two-paper calibration weigh more heavily on practical impact than on the validity of the theoretical result.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Strong4.3
Confidence mediumThe score reflects a constructive resolution of the documented inadmissibility gap under squared-error loss without bounded identification strength, including heterogeneous effects. Its scope remains an exact Gaussian experiment, and the illustrated practical gains are mostly small.
Explicit assumptions, a traceable proof sequence, exact tuning certificates, and independent numerical checks support the score. Compressed appendix arguments and incompletely described evaluation of the reported bound constants keep it below the highest execution band.
The progression from the TSLS decomposition to the local stabilization argument and calibrated comparisons makes the central reasoning traceable. Dense notation and terse appendix exposition impose additional demands on readers without obscuring the stated scope.
The paper distinguishes its result from related weak-IV and decision-theoretic work and explains the consequences of its principal scope limitation. Engagement with practical alternatives such as LIML and Fuller is comparatively thin.
No concerns4 of 4 checks
The reported numbers and statistics are consistent with the stated results.
AI assistance and author responsibility are disclosed. The unavailable formalization and computational materials limit independent inspection but do not establish a reporting inconsistency.
37 references in manuscript 30 of 30 checkable references found in an index 7 have no canonical index record — counted, but not index-checkable 2 references confirmed by manual review
References not verified automatically are resolved manually within 24 hours.
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
Explicit estimators and tuning rules provide a starting point for implementation research, but dominance has not been established with estimated covariance and weights. The application pathway therefore requires further methodological work.
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Astrid Van Kampen, Matthew James Phillips, Scott Devenport · 2024 · Psychology
Due to the pervasiveness of this unsafety, young women had learned to understand the safety of the ‘self’ as their responsibility and modified and limited their self-representation accordingly.
SupportedKei Kawai, Jun Nakabayashi, Juan M. Ortner · 2025 · Economics and business
a cartel can attain higher payoffs by having the pre-selected winner randomize its bid and keep it secret from other members.
Well-supportedCasini, Deng, Z · 2024 · Economics and business
We show that time-variation in the mean induces low frequency contamination, meaning that the sample autocovariance and the periodogram share features that are similar to those of a long memory series
Supported