betaFeedback

WorksEconomics, Econometrics and Finance

Comparing Two-Sample IV Estimators under Weak Instruments

Abstract

We compare three estimators in a two-sample summary-data instrumental variables (IV) model: inverse-variance weighting (IVW), debiased inverse-variance weighting (dIVW), and the profile score (PS) estimator. The framework assumes independent samples and instrument-specific summary estimates, with two-sample Mendelian randomisation as a leading application. We compare existing fixed- and many-weak-instrument results and the sufficient conditions for the three estimators. Under the variance scaling used in the profile-score literature, the principal collective-strength conditions for PS and dIVW are equivalent, although their additional restrictions on instrument concentration and growth differ. We then compare their asymptotic variance formulae for fixed and diverging numbers of instruments. For fixed LL, IVW and dIVW have the same first-order asymptotic variance. Under a common variance-ratio condition, the PS variance formula exceeds the IVW variance by a positive term of order n−2n^{-2}, compared with the order-n−1n^{-1} leading variance. For diverging LL, we establish variance orders across instrument-strength regimes and the exact ordering VIVW<VPS≤VdIVW\mathsf{V}_{\mathtt{IVW}} < V_{\mathtt{PS}} \leq \mathsf{V}_{\mathtt{dIVW}} under the same variance-ratio condition. Absolute variance gaps vanish when Lκ→∞\sqrt{L}κ\to\infty, whereas relative equivalence holds if and only if κ→∞κ\to\infty. These comparisons concern asymptotic variance formulae rather than finite-sample efficiency or mean squared error.

Introduction

Two-sample instrumental variables (IV) methods are useful when the treatment or exposure and the outcome are observed in separate samples, while a common set of instruments is available in both. The first sample identifies instrument–treatment associations and the second identifies instrument–outcome, or reduced-form, associations. This design arises in several applications and is especially prominent in two-sample Mendelian randomisation (MR), where genetic variants serve as instruments and genome-wide association studies provide summary association estimates [9].

This paper studies a summary-data two-sample IV model in which the instrument-specific first-stage and reduced-form estimates are independent across instruments. This diagonal covariance structure is natural, for example, in MR after instruments have been pruned or clumped to be approximately uncorrelated. It is more restrictive than a general two-sample IV model with correlated instruments and full covariance matrices; our results should therefore be interpreted as theory for the independent-summary-association setting rather than unrestricted two-sample IV.

We compare IVW, dIVW, and PS estimators. Two-sample IV estimation has a longstanding econometric literature [3, 8], while IVW, dIVW, and PS have been developed and studied extensively in the summary-data MR literature [12] (Ye et al. 2021). We place these results in a common summary-data IV framework and compare the sufficient conditions required by the three estimators. As a preliminary step, we establish fixed-LL consistency and asymptotic normality for IVW and dIVW directly within the common summary-data framework, providing the variance expressions needed for our subsequent comparisons. Our first contribution is to relate their sufficient conditions under many weak instruments: under the variance scaling used for PS, its collective-strength condition is equivalent to the corresponding dIVW condition, while the two procedures differ in their additional restrictions on the concentration of instrument strength and the growth of the instrument set. Our second contribution is to compare the asymptotic variance formulae of IVW, dIVW, and PS under both fixed and diverging numbers of instruments. For fixed numbers of instruments, under a common variance-ratio condition, the leading PS variance component coincides exactly with the IVW variance, whereas the remaining positive component is associated with profile-score information bias. This additional component is of smaller order than the leading variance under fixed-LL asymptotics, so PS and IVW/dIVW are first-order asymptotically equivalent despite the higher-order variance difference. For diverging numbers of instruments, we establish the variance orders across different instrument-strength regimes and show that, under the same variance-ratio condition, the IVW variance formula is strictly smaller than the PS variance formula, which is no greater than the dIVW variance formula. We further show that the absolute variance gaps vanish when Lκ→∞\sqrt{L}\kappa\to\infty, whereas relative equivalence of all three variance formulae holds if and only if κ→∞\kappa\to\infty. These comparisons concern the asymptotic variance formulae and do not, by themselves, establish an ordering of finite-sample efficiency or mean squared error. The arguments depend on the two-sample summary-data model and valid-IV proportionality, rather than on the instruments being genetic variants. MR provides the principal motivating example, but the theoretical results apply to any two-sample IV setting satisfying the stated assumptions.

Two-Sample Summary-Data IV Framework

Let β\beta be the structural effect of a treatment or exposure XX on an outcome YY, and let Z1,…,ZLZ_1,\ldots,Z_L denote valid instruments. We work directly with the first-stage and reduced-form coefficients. Let γj\gamma_j denote the association of instrument jj with XX and Γj\Gamma_j its association with YY. Validity and a common linear structural effect imply the proportionality restriction.

Assumption 1 (Valid-IV proportionality). For all j=1,…,Lj=1,\ldots,L, Γj=βγj\Gamma_j=\beta\gamma_j. Additionally, the γj\gamma_j’s are not all zero.

This restriction identifies β\beta from the first-stage and reduced-form coefficients. It is the summary-data analogue of the familiar reduced-form restriction in a linear IV model.

We consider two distinct, independent samples. Let nXn_X and nYn_Y be the treatment and outcome sample sizes. From the treatment sample we obtain estimates γ^j\hat{\gamma}_j of the first-stage coefficients γj\gamma_j; from the outcome sample we obtain estimates Γ^j\hat{\Gamma}_j of the reduced-form coefficients Γj\Gamma_j. In the simplest implementation these are marginal regression coefficients,

Xi=γ0,j+γjZij+eX,ij,i=1,…,nX.X_i=\gamma_{0,j}+\gamma_j Z_{ij}+e_{X,ij},\qquad i=1,\ldots,n_X.
Yi=Γ0,j+ΓjZij+eY,ij,i=1,…,nY.Y_i=\Gamma_{0,j}+\Gamma_j Z_{ij}+e_{Y,ij},\qquad i=1,\ldots,n_Y.

We assume nX/nY→an_X/n_Y\to a, where 0<a<∞0<a<\infty, as nX+nY→∞n_X+n_Y\to\infty.

For fixed LL, suppose the summary estimators satisfy

nY(Γ^j−Γj)→dN(0,σΓj2),\sqrt{n_Y}(\widehat{\Gamma}_j-\Gamma_j) \xrightarrow{d} \mathcal{N}(0,\sigma_{\Gamma_j}^2),
nX(γ^j−γj)→dN(0,σγj2),\sqrt{n_X}(\widehat{\gamma}_j-\gamma_j) \xrightarrow{d} \mathcal{N}(0,\sigma_{\gamma_j}^2),

with variance estimators satisfying nYσ^Yj2→pσΓj2n_Y\widehat{\sigma}_{Yj}^2 \xrightarrow{p} \sigma_{\Gamma_j}^2 and nXσ^Xj2→pσγj2n_X\widehat{\sigma}_{Xj}^2 \xrightarrow{p} \sigma_{\gamma_j}^2. Define σYj2=σΓj2/nY\sigma_{Yj}^2=\sigma_{\Gamma_j}^2/n_Y and σXj2=σγj2/nX\sigma_{Xj}^2=\sigma_{\gamma_j}^2/n_X.

For the many-instrument theory we use the following stronger working model.

Assumption 2 (Independent summary-association model). (1) The pairs {(Γ^j,γ^j)}j=1L\{(\widehat{\Gamma}_j,\widehat{\gamma}_j)\}_{j=1}^{L} are mutually independent.

(2) For all jj,

(Γ^jγ^j)∼N((Γjγj),(σYj200σXj2)),\begin{pmatrix} \widehat{\Gamma}_j\\ \widehat{\gamma}_j \end{pmatrix} \sim\mathcal{N}\left( \begin{pmatrix} \Gamma_j\\ \gamma_j \end{pmatrix}, \begin{pmatrix} \sigma_{Yj}^2 & 0\\ 0 & \sigma_{Xj}^2 \end{pmatrix} \right),

where the ratios σYj2/σXj2\sigma_{Yj}^2/\sigma_{Xj}^2 are uniformly bounded above and bounded away from zero.

The within-instrument independence follows from the independence of the two samples. Independence across instruments is an additional restriction: it holds directly for independent instruments under the corresponding marginal summary regressions and is a useful approximation in two-sample MR after LD pruning or clumping. General two-sample IV with correlated instruments instead requires the full covariance matrices of the first-stage and reduced-form coefficient vectors and is outside the scope of the present results.

A canonical application is two-sample MR. There, ZjZ_j is a genetic variant, γj\gamma_j is the variant–exposure association, and Γj\Gamma_j is the variant–outcome association. Under valid instruments and no direct effect of the variants on the outcome, Γj=βγj\Gamma_j=\beta\gamma_j. Thus the MR model used in the motivating literature is a special case of the summary-data IV framework above.

We measure average instrument strength by

κ=1L∑j=1Lγj2σXj2.\kappa=\frac{1}{L}\sum_{j=1}^{L}\frac{\gamma_j^2}{\sigma_{Xj}^2}.

and estimate it by

κ^=1L∑j=1Lγ^j2σ^Xj2−1.\widehat{\kappa}=\frac{1}{L}\sum_{j=1}^{L}\frac{\widehat{\gamma}_j^2}{\widehat{\sigma}_{Xj}^2}-1.

Remark 1 (Scope of the generalisation). All results below are stated for the summary-data IV model in Assumptions 1–2. Consequently, the fixed-LL propositions and the many-instrument results do not require the instruments to be genetic variants. The many-instrument results do, however, rely on independence across instrument-specific summary estimates; they are not claims about arbitrary correlated-instrument two-sample IV.

Inverse-Variance Weighted (IVW) Estimator

For each instrument jj, the corresponding individual-instrument estimator is β^j=Γ^j/γ^j\widehat{\beta}_j = \widehat{\Gamma}_j/\widehat{\gamma}_j. This is the summary data analogue of the individual-instrument estimator introduced in classical 2SLS theory. However, an estimator based on a single instrument may be unstable when γ^j\widehat{\gamma}_j is close to zero [10]. The inverse-variance weighted (IVW) estimator combines the LL individual-instrument estimators using inverse-variance weighting [1]:

β^IVW=∑j=1Lw^jβ^j∑j=1Lw^j\widehat{\beta}_{\mathrm{IVW}} = \frac{\sum_{j=1}^{L}\widehat{w}_j\widehat{\beta}_j}{\sum_{j=1}^{L}\widehat{w}_j}

where w^j=γ^j2/σ^Yj2\widehat{w}_j = \widehat{\gamma}_j^2/\widehat{\sigma}_{Y_j}^2 is the usual inverse-variance weight, obtained by treating γ^j\widehat{\gamma}_j as fixed. The IVW estimator can thus be written as

β^IVW=∑j=1LΓ^jγ^j/σ^Yj2∑j=1Lγ^j2/σ^Yj2.\widehat{\beta}_{\mathrm{IVW}} = \frac{\sum_{j=1}^{L}\widehat{\Gamma}_j\widehat{\gamma}_j/\widehat{\sigma}_{Y_j}^2}{\sum_{j=1}^{L}\widehat{\gamma}_j^2/\widehat{\sigma}_{Y_j}^2}.

Applying the classical 2SLS framework reveals the connection between IVW and 2SLS. For simplicity, consider individual-level data with exposure vector x\mathbf{x}, outcome vector y\mathbf{y}, and instrument matrix Z\mathbf{Z}. Let γ^=(ZTZ)−1ZTx\widehat{\boldsymbol{\gamma}} = (\mathbf{Z}^{T}\mathbf{Z})^{-1}\mathbf{Z}^{T}\mathbf{x} and Γ^=(ZTZ)−1ZTy\widehat{\boldsymbol{\Gamma}} = (\mathbf{Z}^{T}\mathbf{Z})^{-1}\mathbf{Z}^{T}\mathbf{y}. The 2SLS criterion can be written as

β^2SLS=arg⁡min⁡β(y−xβ)TZ(ZTZ)−1ZT(y−xβ)=arg⁡min⁡β(Γ^−γ^β)T(ZTZ)(Γ^−γ^β).\begin{aligned} \widehat{\beta}_{\mathrm{2SLS}} &= \arg\min_{\beta}(\mathbf{y}-\mathbf{x}\beta)^{T}\mathbf{Z}(\mathbf{Z}^{T}\mathbf{Z})^{-1}\mathbf{Z}^{T}(\mathbf{y}-\mathbf{x}\beta)\\ &= \arg\min_{\beta}(\widehat{\boldsymbol{\Gamma}}-\widehat{\boldsymbol{\gamma}}\beta)^{T}(\mathbf{Z}^{T}\mathbf{Z})(\widehat{\boldsymbol{\Gamma}}-\widehat{\boldsymbol{\gamma}}\beta). \end{aligned}

Under homoscedasticity,

References

  1. [1]Burgess, S., A. Butterworth, and S. G. Thompson (2013): “Mendelian Randomization Analysis With Multiple Genetic Variants Using Summarized Data,” Genetic Epidemiology, 37, 658–665.DOI
  2. [2]Hansen, L. P., J. Heaton, and A. Yaron (1996): “Finite-Sample Properties of Some Alternative GMM Estimators,” Journal of Business & Economic Statistics, 14, 262–280.DOI
  3. [3]Inoue, A. and G. Solon (2010): “Two-Sample Instrumental Variables Estimators,” The Review of Economics and Statistics, 92, 557–561.
  4. [5]McCullagh, P. and R. Tibshirani (1990): “A simple method for the adjustment of profile likelihoods,” J. Roy. Statist. Soc. Ser. B, 52, 325–344.DOI
  5. [6]Newey, W. K. and F. Windmeijer (2009): “Generalized Method of Moments With Many Weak Moment Conditions,” Econometrica, 77, 687–719.DOI
  6. [7]Neyman, J. and E. L. Scott (1948): “Consistent Estimates Based on Partially Consistent Observations,” Econometrica, 16, 1–32.
  7. [8]Pacini, D. and F. Windmeijer (2016): “Robust inference for the Two-Sample 2SLS estimator,” Economics Letters, 146, 50–54.pubmed.ncbi.nlm.nih.gov/27667880
  8. [9]Pierce, B. L. and S. Burgess (2013): “Efficient Design for Mendelian Randomization Studies: Subsample and 2-Sample Instrumental Variable Estimators,” American Journal of Epidemiology, 178, 1177–1184.DOI
  9. [10]Sawa, T. (1969): “The Exact Sampling Distribution of Ordinary Least Squares and Two-Stage Least Squares Estimators,” Journal of the American Statistical Association, 64, 923–937.DOI
  10. [12]Zhao, Q., J. Wang, G. Hemani, J. Bowden, and D. S. Small (2020): “Statistical inference in two-sample summary-data Mendelian randomization using robust adjusted profile score,” The Annals of Statistics, 48, 1742–1769.arxiv.org/abs/1801.09652

Paper details

vlnc:W_7yehzxmj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents