Talk:Mendelian Randomization

MR SOP — outstanding items (selection/collider bias work, 2026-07-18)

Context: §4 "Selection and Collider Bias" was added to the MR SOP after the smoking→AD project found its headline result was most plausibly a survival-collider artefact.

Sections still missing

All three are cheap, and all three bit us on smoking→AD.

1. Steiger directionality / reverse causation — absent entirely. One line (TwoSampleMR::directionality_test()). Important for the disease→AD portfolio, where reverse causation is plausible.
2. Sample overlap — absent. Two-sample MR assumes independent samples; UK Biobank is in nearly every GWAS we use (it was in both our exposure and outcome). Tool: MRlap (needs full genome-wide sumstats + LD scores).
3. Dataset provenance verification — absent. We caught two mislabelled OpenGWAS accessions: ieu-a-1001 labelled "parental lifespan" is actually Okbay EA; ieu-b-5067 labelled "Wightman" is Woolf 2022 with 954 cases. Rule: verify N, case/control split and cohort composition via gwasinfo() before use; never trust the label.

MR-PRESSO / MR-Egger imbalance (raised, not yet drafted)

Observation: we report MR-PRESSO outlier-corrected far more often than MR-Egger. This is structural, not accidental — Egger intercept tests are underpowered so Scenario 4 rarely fires, while Cochran's Q is almost always significant so Scenario 2 fires constantly.

Why it matters — our own Layer 0: IVW b=−0.065 p=0.296 → MR-PRESSO b=−0.112 p=0.0015. Outlier removal manufactured the significant result later shown to be collider-driven. It removed only 2/22 SNPs (9%), so the existing ">30% removed" tripwire never fired.

Proposed amendments (Scenario 2 currently has no conclusion-change tripwire):

• Add a conclusion-change tripwire to Scenario 2: if outlier removal moves the estimate outside the original IVW 95% CI or flips significance → report corrected + uncorrected, with the corrected value as sensitivity only. Key on effect on the conclusion, not SNP count.
• Stop treating Egger intercept p>0.05 as evidence of no directional pleiotropy — it is an underpowered null, the same error pattern we rejected in §4.2. Report the intercept estimate with CI.
• Elevate Weighted Median to co-primary (better power and robustness than MR-Egger; currently only "support" in Scenario 2).
• Report the full estimator panel with a concordance flag — our estimates spanned −0.065 to −0.208 with p from 0.0015 to 0.647; single-winner adjudication hides that spread.