Debiased Inverse-Variance Weighted Estimator in Two-Sample Summary-Data Mendelian Randomization

1.1. Motivation: Many Weak Instruments in MR.

• Weak or null SNP effects, small effects of common variants, and small genetic variances of rare variants can all produce many weak instruments in MR.The paper describes two-sample summary-data MR using genetic variants from GWASs as instruments.Source: §1.1 · p. 1
• When an SNP’s estimated exposure association is near zero, its ratio estimate can be biased and unstable; IVW aggregates these estimates by meta-analysis.The paper identifies IVW as the most popular such estimator.Source: §1.1 · p. 2
• Under many weak instruments, the consistency and asymptotic normality of IVW and its screening counterpart were unknown, despite their widespread use.Prior work on weak IVs in MR was mostly numerical, and econometric results focused on one-sample individual-data settings.Source: §1.1 · p. 2, §1.2 · p. 2

1.2. Prior Work and Our Contributions.

• The paper analyzes the asymptotic behavior of IVW with and without screening, then proposes dIVW, a bias-corrected IVW modification that is generally consistent and asymptotically normal without screening.The introduction contrasts dIVW with more complex optimization-based estimators and notes that one such estimator may lack unique estimates in some data-generating scenarios.Source: §1.2 · p. 2
• The authors assess how estimated rather than known variances affect IVW and dIVW, and propose two screening methods to improve dIVW efficiency when a third GWAS is available.One method can eliminate instruments with no exposure association; the other iteratively selects a threshold to target efficiency within a class.Source: §1.2 · pp. 2–3
• The paper also extends dIVW to balanced horizontal pleiotropy and presents simulation and real-data results.Technical proofs and additional results appear in the Supplementary Material.Source: §1.2 · p. 3

• The MR models make the outcome depend on exposure, while both exposure and outcome may also depend on an unmeasured confounder.The instruments are assumed independent of that confounder, and the model errors are mutually independent and independent of the instruments and confounder.Source: §2 · p. 3
• MR requires relevant instruments, independence from unmeasured confounding, and no direct instrument effect on the outcome; the last assumption may be violated in some studies.An instrument is relevant when its exposure effect γ_j is nonzero.Source: §2 · p. 3
• In two-sample summary-data MR, the exposure and outcome datasets are independent, and only their summary statistics—not individual-level data—are analyzed.The exposure estimates γ_j, while the outcome estimates β_0γ_j.Source: §2 · pp. 3–4
• Because many SNPs can be weak or null, IVW screening uses a third independent dataset and thresholds SNP z-scores to select instruments.The commonly used λ ≈ 5.45 corresponds to screening at the genome-wide significance level 5×10^-8.Source: §2 · p. 4
• The asymptotic analysis lets sample sizes and the number of SNPs grow, and assumes independent summary statistics with normal distributions and bounded variance ratios.The text presents these distributional and boundedness conditions as plausible in the stated GWAS setting.Source: §2 · p. 4
• Average instrument strength κ summarizes normalized SNP exposure effects; its limiting value, and the screened counterpart κ_λ, help characterize IVW estimator properties.The paper also describes estimating these strength parameters and the number of selected instruments for checking theoretical conditions.Source: §2 · p. 5

3. Properties of the IVW Estimators.

• The analysis assumes known outcome and selection-dataset standard deviations; the more general case with estimated standard deviations is deferred to Section 4.Source: §3 · p. 5
• κ/p→∞ is required for unscreened IVW consistency and asymptotic normality when β0 ≠ 0, a rate the authors describe as unlikely in typical MR studies.The authors note that approximately κ/p ≤ nX/p²; with 1,000 SNPs and nX = 1 million, nX/p² = 1.Source: §3 · pp. 6–7
• With screening, asymptotic normality requires κλ/pλ→∞ and κλ√pλ/λ²→∞; screening reduces bias but increases the IVW estimator's standard deviation.The authors state that κλ/pλ→∞ is weaker than κ/p→∞ for unscreened IVW.Source: §3 · p. 7
• When average IV strength κλ diverges, IVW can be consistent without asymptotic normality; when κλ stays bounded, it is inconsistent and biased toward 0.The bounded-strength case is described as common with many weak and null IVs.Source: §3 · p. 7
• When β0 = 0, the weak-IV bias is absent, and unscreened IVW is consistent and asymptotically normal under reasonable conditions.The asymptotic bias equals zero in this special case.Source: §3 · p. 7
• The 5×10−8 screening threshold corresponds to λ ≈ 5.45, but may not satisfy the conditions for IVW consistency or asymptotic normality.If all IVs are very weak, κλ remains bounded regardless of λ; if every IV strength equals λ, κλ/pλ may be small.Source: §3 · p. 7

4.1. Debiased IVW Estimator.

• The dIVW estimator multiplies IVW by an explicit bias-correction factor, yielding negligible asymptotic bias without requiring κ/p → ∞.The correction addresses IVW’s bias toward zero under weak instruments.Source: §4.1 · p. 8
• Under κ√p → ∞ and max_j(γ²_Xj)/(κp + p) → 0, dIVW is consistent and asymptotically normal under a weaker strength condition than IVW requires.The stated comparison is with IVW’s κ/p → ∞ requirement.Source: §4.1 · p. 8
• dIVW has larger asymptotic variance than IVW, reflecting a bias–variance trade-off when weak-instrument bias can dominate the estimator’s standard deviation.Both asymptotic variances have order (κp)^−1 when the estimators are asymptotically normal.Source: §4.2 · p. 8

4.3. Choice of λ in Screening.

• Screening dIVW is intended to improve efficiency, and λ = √(2 log p) can remove null IVs with probability close to 1 while avoiding overly aggressive screening.The paper notes that dIVW without screening may be asymptotically more efficient when many instruments are weak.Source: §4.2 · p. 8, §4.3 · pp. 9–10

4.2. Improving Efficiency With Screening.

• With screening, dIVW is consistent and asymptotically normal under Theorem 4.1 without IVW’s stringent κλ/pλ → ∞ condition.The theorem requires κλ√pλ/max(1, λ²) → ∞, along with its other stated conditions.Source: §4.2 · p. 9
• When p/nX → 0, estimated standard errors yield the same asymptotic result as true standard deviations; simulations also find good approximation at p/nX = 20%.The p/nX → 0 condition is described as sufficient rather than necessary.Source: §4.2 · p. 9

4.3. Choice of λ in Screening.

• MR-EO selects λ by minimizing estimated asymptotic variance, but its chosen estimator is not directly guaranteed to be asymptotically normal by Theorem 4.1.Simulations report good performance and nominal confidence-interval coverage; estimated variance may be too optimistic from winner’s curse, an issue expected to be largely moot when p and nX are large and p/nX is bounded, ideally small.Source: §4.3 · pp. 10–11

4.4. Extension to Balanced Horizontal Pleiotropy.

• Under balanced horizontal pleiotropy, dIVW remains consistent and asymptotically normal but has larger variance; if pleiotropy is not balanced, dIVW is biased.The extension models pleiotropic effects as independent random effects and replaces Assumption 2 with Assumption 2′.Source: §4.4 · p. 11

5.1. A Simulation with the BMI-CAD Dataset as Population.

• Across Cases 1–3, dIVW with or without screening shows negligible bias and nominal coverage; MR-EO is most efficient among dIVW variants without losing coverage or large bias gains.Screening reduces SD in Cases 1–2 with many null IVs, but does not improve dIVW in Case 3, where all IVs are non-null but many are weak. Unscreened IVW is biased toward zero and has poor coverage; screening improves its performance.Source: §5.1 · p. 14
• MR-raps performs comparably to dIVW without screening, while MR-raps-shrink is comparable to MR-EO-screened dIVW but is more computationally complicated and may lack a unique solution.MR-Egger, MR-median, and MR-mode can be biased at low average IV strength; even screened, they have larger biases than dIVW.Source: §5.1 · pp. 14–15
• In Case 1, genome-wide screening selects far fewer than the 20 true non-null IVs, while √2log p and MR-EO select close to 20; MR-EO also selects more null IVs.In Cases 2–3, threshold-based procedures do not adequately select many weak non-null IVs, and removing too many weak IVs may reduce dIVW efficiency.Source: §5.1 · p. 15
• When balanced horizontal pleiotropy is added to Case 3, results are nearly identical; when that assumption is violated, dIVW is biased, often mildly, as large αj/γj terms tend to be downweighted.The reported bias is a weighted average of αj/γj values, with weights that tend to downweight larger values.Source: §5.1 · p. 15

5.2. Empirical Guidelines for Asymptotics.

• Coverage ranges from 93.5% to 96.0% across screening thresholds, and estimates appear less dispersed as the estimated effective sample size grows; asymptotics appear to begin above 20.The authors recommend checking that the effective-sample-size diagnostic is greater than 20.Source: §5.2 · p. 16, Figure 1 · p. 16

5.3. Empirical Evaluation of the Effect of Using SEs not SDs.

• With estimated SEs, dIVW maintains nominal coverage and outperforms IVW; MR-EO screening performs similarly to unscreened dIVW when s/p is not small.The SEs are generally close to simulated SDs, except in Case 7 at λ = 3.90, where the effective sample size is 1.11 and two outliers drive the discrepancy. Results are almost identical when known SDs are used as SEs.Source: §5.3 · p. 17
• At s = 1,000 or 2,000, the genome-wide threshold λ ≈ 5.45 often selects no SNPs, while estimator performance tends to improve as sample size increases in Cases 5–6.These observations reinforce that the usual threshold may be too large in MR studies with many weak IVs.Source: §5.3 · p. 17

• MR-mode was very unstable with or without screening.Source: §6 · p. 18
• dIVW with MR-EO had the smallest SE except for unscreened IVW, but matched unscreened dIVW’s SE, so screening was probably unnecessary for this dataset with many weak IVs.Source: Table 4 · p. 19, §6 · p. 19
• 0.315 was the IVW estimate without screening versus 0.365 for dIVW; the difference was not statistically significant.Source: §6 · p. 18, Table 4 · p. 19
• 0.278–0.287 were the point estimates across methods after genome-wide-significance screening (λ = 5.45), with larger SEs, most likely because too many IVs were screened out.Source: Table 4 · p. 19, §6 · p. 19
• Estimators accounting for balanced horizontal pleiotropy were similar to those without it, apart from an expected SE increase from random-effect terms; dIVWα denotes dIVW under that setting.Source: §6 · pp. 17, 19
• Residuals close to the 45-degree line suggested Assumption 2 likely holds for this example.Source: §6 · p. 19

7. Summary and Discussion.

• The dIVW estimator is consistent and asymptotically normal under weaker conditions than IVW, with or without screening.The paper describes dIVW as a simple IVW modification robust to many weak instruments.Source: §7 · p. 19
• The generalized IVW class uses a general weighting function that can encompass soft thresholding, but does not include dIVW because its weights do not sum to 1.The paper identifies extending theory for this broader class under many weak IVs as future work.Source: §7 · pp. 19–20
• The paper recommends unscreened dIVW as the baseline; with an independent selection dataset, MR-EO screening may improve efficiency, while the genome-wide significance threshold tends to remove too many IVs.The recommendations also call for diagnostics using standardized-residual Q-Q plots and checking κ̂λ.Source: §7 · p. 20

Software and Reproducibility.

• The proposed methods’ R code and examples for reproducing the numerical results are available in the mr.divw package.The passage gives the package’s GitHub location.Source: Software and Reproducibility · p. 20

Supplementary Material: Debiased Inverse-Variance

• The title fragment identifies the supplementary material as concerning a weighted estimator in two-sample summary-data.The supplied text is only the ending of the title.Source: Supplementary Material: Debiased… · p. 23

1.1. Comparison of Estimators under Balanced Horizontal Pleiotropy.

• Under balanced horizontal pleiotropy, dIVW and MR-raps have good statistical properties, unlike the reported shortcomings of MR-Egger, MR-median, and MR-mode.Table S1 reports Case 3 simulations based on 10,000 repetitions with β0 = 0.4.Source: §1.1 · p. 23, Table S1 · p. 23

1.2. Empirical Evaluation of the Effect of Using SEs not SDs.

• The feasible and infeasible dIVW point estimates and standard deviations are very similar, supporting their asymptotic equivalence when p/nX → 0.Table S3 reports simulations using true SDs for Cases 4–7, based on 10,000 repetitions.Source: §1.2 · p. 24, Table S3 · p. 25

1.3. Q-Q Plot of BMI-CAD under Balanced Horizontal Pleiotropy.

• For BMI-CAD, standardized residuals close to the 45-degree line make Assumption 2′ appear plausible.Figure S2 plots the standardized residuals in (S1) against a standard normal.Source: §1.3 · p. 24, Figure S2 · p. 24

2. The dIVW Estimator Beyond Balanced Horizontal Pleiotropy.

• When balanced horizontal pleiotropy fails, dIVW is biased, but its bias may remain small when large αj/γj values receive low weights, as can occur when weak IVs have more pleiotropy.The authors note that the simulation study does not cover all possible violations, and bias may be larger in other settings.Source: §2 · pp. 26–27
• 75% of IVs violating balanced horizontal pleiotropy still produced empirical dIVW bias that was not too severe compared with most estimators in the literature.The simulations varied the proportion of nonzero pleiotropic effects across 25%, 50%, and 75%.Source: §2 · pp. 26–27

3.1.1. When λ = 0.

• For λ=0, the proof establishes IVW consistency as κ→∞ and shows ˆβIVW=oP(1) when κ→0 and p→∞.It also derives asymptotic normality results for IVW under the theorem’s conditions.Source: §3.1.1 · pp. 30–32

3.2. Proof of κ/p approximately upper bounded by nX/p2.

• The proof addresses the stated approximate upper bound κ/p by nX/p^2, relating it to genetic-variance contributions that are small relative to Var(X).Source: §3.2 · p. 35

3.3. Proof of κλ is approximately increasing in λ.

• The proof establishes that κλ is approximately increasing in λ.The argument uses a monotonicity result and a Cauchy–Schwarz inequality.Source: §3.3 · pp. 35–36

3.4. Proof of Theorem 4.1.

• Theorem 4.1’s proof establishes dIVW asymptotic normality using a central limit theorem and consistency when κ√p→∞, extending across threshold and variance-estimation settings.The proof treats λ=0 and general λ≥0, both with known and unknown variances.Source: §3.4 · pp. 36–37
• Under κλ√pλ/max(1,λ2) → ∞, Var(−ψ′/V4) = o(1), as established using Markov’s inequality.Here V4 = Θ(κλpλ).Source: §3.4 · p. 38

3.4.3. Without assuming knowing SDs and λ = 0.

• The proof setup assumes independent exposure and outcome datasets, with their individual-level data unavailable while the corresponding summary statistics are used.Source: §3.4.3 · pp. 38–39
• It defines hat matrices for centered-IV regressions and uses marginal-regression residual sums of squares, including a decomposition of RSSYj into total centered variation minus the IV projection term.Source: §3.4.3 · p. 39

2. From Taylor expansion,

• Under Assumptions 1–2, p/nX→0, and κ√p→∞, dIVW is consistent and asymptotically normal as p,nX→∞.The proof combines lemmas for the estimator’s numerator and denominator.Source: §2 · p. 45
• The proof also treats general λ≥0 without assuming known σXj and σ*j, under additional growth conditions.The section states these results for unknown standard deviations; the displayed assumptions specify the growth conditions.Source: §2 · p. 45

3.5. Proof of the dIVW estimator under balanced horizontal pleiotropy.

• Under balanced horizontal pleiotropy, the proof uses a normal distribution for the outcome association and establishes dIVW consistency by analogy with Theorem 4.1.The normal distribution has mean β0γj; the proof also modifies a lemma to account for pleiotropy.Source: §3.5 · pp. 48, 50