ADR 0009: Small-study-effect test conventions

Context

The four diagnostic methods introduced in 0.8–0.9 test different aspects of small-study effects. An API freeze needs an explicit account of their inputs, inference conventions and limits, rather than treating them as interchangeable tests for publication bias.

Recorded decisions

Method Inputs and inference Independent software reference
Begg–Mazumdar Common-effect-centered standardized model-scale effects and variances; Kendall tau-b; exact two-sided inference without ties below 50 studies, otherwise tie-adjusted normal inference; optional continuity correction metafor::ranktest()
Egger Model-scale effects regressed on standard error with inverse-variance weights and multiplicative dispersion; t test with k-2 df metafor::regtest(model="lm", predictor="sei")
Harbord Retained raw binary OR event margins; efficient-score regression, multiplicative dispersion and t inference; no continuity correction meta::metabias(method.bias="Harbord")
Peters Conventional study log OR reconstructed with the source study-level correction; raw inverse total sample size and S*F/N weights; multiplicative dispersion and t inference meta::metabias(method.bias="Peters")

Each method requires at least three eligible studies and an identifiable test. Eligibility, not merely the source model's study count, determines whether a test can run. None changes the fitted pooling model or establishes or excludes publication bias. The four methods remain explicit user choices; there is no automatic diagnostic selection or inference-default change.

Egger and Begg consume the source study effects, including Fisher's z for correlations. Peto study effects are one-step approximations and carry an additional caveat. Harbord and Peters require retained binary OR counts and are unavailable for generic or correlation inputs. Peters uses conventional log OR even when the source pooled analysis uses Peto.

Consequences and evidence

Preserve the current named result fields and formulas while reviewing their 1.0 contracts. Additional interpretation warnings are intentional; consumers should not compare exact warning tuples as a statistical validity criterion. Detailed equations and applicability are in the methods reference and small-study-effect guide.

Pinned R generators and fixtures, direct regression/rank calculations, scale and permutation tests are listed in validation. The first corrective batch adds correlation composition and Peto-caveat tests and fixes Peters' included-study correction-count metadata. These are implementation evidence, not independent statistical sign-off. Gate M5/M6 reviewers must review eligibility, ties, correction reuse, numerical boundaries and tolerances on the eventual candidate commit before ratification.