Statistical inference in Mendelian randomization: From genetic association to epidemiological causation

Transcription

1 Statistical inference in Mendelian randomization: From genetic association to epidemiological causation Department of Statistics, The Wharton School, University of Pennsylvania March 1st, UMN Based on joint work with Jingshu Wang, Dylan Small (Penn). Jack Bowden, Gibran Hemani (University of Bristol).

2 Epidemiology = Why certain people are getting ill? Central problem: Causal inference from observational data. Methodological challenge: Correlation does not imply causation. Conventional solution: Conditioning on confounding variables. Matched case-control studies. Cohort studies. Problem: Biased if there is unmeasured confounder. 1/44

3 Motivating example Question What is the causal effect of Body Mass Index (BMI) on Systolic Blood Pressure (SBP)? Exposure variable X : BMI. Outcome variable Y : SBP. Confounders C: age, gender, lifestyle (food, smoking,... )... Fundamental problem Where is the end of this list? 2/44

4 MR: unbiased estimation without enumerating C Mendelian randomization (MR) is an alternative design of observational studies. Instead, MR uses additional instrumental variables Z. Not necessary to enumerate C. Brief history of MR Katan (1986): first proposal. Davey Smith and Ebrahim (2003): gaining attention. Hernán and Robins (2006): An Epidemiologist s dream? Pickrell et al. (2016): Surging interest in human genetics. Publications per year: only 5 in 2003, almost 400 in /44

5 This talk: MR using summary data Individual-level data (restricted access) Exposure variable X : BMI; Outcome variable Y : SBP; Instrumental variables Z 1,..., Z p : Genetic variants (SNPs). Summary-level data (public) Dataset BMI-FEM BMI-MAL SBP-UKBB Source GIANT (female) GIANT (male) UKBB Sample size GWAS lm(x Z j ) lm(x Z j ) lm(y Z j ) Coefficient ˆγ Selection j ˆΓj Std. Err. σ Xj σ Yj Step 1 Use BMI-FEM to select significant (p-value ) and uncorrelated SNPs (p = 25). Step 2 Use BMI-MAL to obtain (ˆγ j, σ Xj ), j = 1,..., p. Step 3 Use SBP-UKBB to obtain (ˆΓ j, σ Yj ), j = 1,..., p. 4/44

6 Statistical problem Model: Errors-in-variables regression Observe mutually independent random variables: ˆγ j N(γ j, σ 2 Xj ) and ˆΓ j N(Γ j, σ 2 Yj ). For some real number β 0, Γ j β 0 γ j for almost all j. β 0 can be regarded as the causal effect of X on Y. 5/44

7 Example SNP effect on SBP SNP effect on BMI Method Simple Overdispersion + Robust loss Figure: Scatter plot of ˆΓ j versus ˆγ j in the BMI-SBP example (p = 25). The goal is to fit a straight line across the origin. Observation: effects are tiny but many are significant. 6/44

8 Reasonable model? Key assumptions 1 Measurement error is normal. 2 (ˆγ 1,..., ˆγ p ) (ˆΓ 1,..., ˆΓ p ). 3 ˆγ j ˆγ k and ˆΓ j ˆΓ k if j k. 4 (Implicitly) γ 0. Otherwise no information for β 0. 5 ˆγ j is unbiased: E[ˆγ j ] = γ j. Pre-processing guarantees these assumptions 1 Large sample size CLT. 2 Two-sample MR: ˆγ and ˆΓ are computed from independent samples (GIANT males and UKBB). 3 Selecting uncorrelated and significant SNPs using another independent dataset (GIANT females). 7/44

9 MR estimates causal effect Two remaining questions 1 Why should the model Γ j β 0 γ j be linear? 2 Why is β 0 causal? Answer: MR is using SNPs as instrumental variables (IV). Heuristic Consider the following linear model: p X = γ j Z j + η X C + E X, j=1 Y = β 0 X + η Y C + E Y = p j=1 (β 0 γ j }{{} Γ j )Z j +... Key assumption: Z (C, E X, E Y ), i.e. Z are valid instruments. 8/44

10 Background: What is an IV? 2 C 1 Z X Y 3 Figure: Causal effect of X and Y is confounded by C. Z is an instrumental variable. Core IV assumptions 1 Z is associated with the exposure (X ). 2 Z is independent of the unmeasured confounder (C). 3 Z cannot have any direct effect on the outcome (Y ). 9/44

11 MR = using genetic variants as IVs 2 C 1 Z X Y 3 Examine the core IV assumptions for MR 1 Criterion 1 is not a problem. 2 Criterion 2 almost comes free due to Mendel s Second Law of random assortment. A minor concern is population stratification 3 Criterion 3 is not clear. There is a wide-spread phenomenon in genetics called pleiotropy, a.k.a. multi-function of genes. 10/44

12 Linear model (for summary data) is very robust Surprisingly, Γ j β 0 γ j is true if the individual-level model for (X, Y, Z, C) are non-linear. Heuristic Each SNP corresponds to a very small randomized trial. Maybe the relationship between Γ and γ is nonlinear, but it is locally linear near γ = 0. Can formalize this argument (see the paper) and show: β 0 lim x 0 E[Y (X + x) Y (X )]. x Y (x) is the counterfactual or potential outcome. The approximate linear model is still true even if the ˆΓ j are computed from glm(y Z j ). 11/44

13 Recap Two-sample summary-data MR (ˆγ j, ˆΓ j, σxj 2, σ2 Yj )p j=1 Genetic associations Inference = β 0 Epidemiological causation Statistical model 1 Observe mutually independent random variables: ˆγ j N(γ j, σ 2 Xj ) and ˆΓ j N(Γ j, σ 2 Yj ). 2 For some real number β 0, Γ j β 0 γ j for almost all j. Three concrete models for α j = Γ j β 0 γ j 1 No direct effect: α j 0. 2 Random effects model: α j N(0, τ0 2 ). 3 Contamination model: α j (1 ɛ)n(0, τ0 2 ) + ɛg. 12/44

14 Outline 1 2 : No direct effect 3 : Random direct effects 4 : Random direct effects with outliers 5 More examples 6 13/44

15 Model (No direct effect) The linear model Γ j = β 0 γ j is true for every j = 1,..., p. Log-likelihood of the data (up to additive constant): l(β, γ 1,..., γ p ) = 1 [ p (ˆγ j γ j ) Profile likelihood j=1 l(β) = max γ l(β, γ) = 1 2 σ 2 Xj p j=1 Profile score (PS): ψ(β) = (d/dβ)l(β). p (ˆΓ j γ j β) 2 ]. j=1 (ˆΓ j βˆγ j ) 2 σxj 2 β2 + σyj 2. Estimator: ˆβPS = arg max l(β) solves ψ(β) = 0. β σ 2 Yj 15/44

16 Some theory Assumption σxj 2 = O(1/n) = σ2 Yj. γ 2 is bounded (even if p ). Heuristic p Linear structural model: X = γ j Z j + η X C + E X. j=1 γ 2 is like the variance of X explained by Z. Theorem (Consistency) Under, ˆβ PS p β0 if p/n /44

17 Some theory (continued) Theorem Suppose n, and further p is fixed or p but γ 3 / γ 2 0, then V 2 V1 ( ˆβ PS β 0 ) d N(0, 1), where V 1 = p j=1 γ 2 j σ2 Yj + Γ2 j σ2 Xj + σ2 Xj σ2 Yj (σ 2 Xj β2 0 + σ2 Yj )2, V 2 = p j=1 γ 2 j σ2 Yj + Γ2 j σ2 Xj (σ 2 Xj β2 0 + σ2 Yj )2. V 1 and V 2 can be consistently estimated from data. Necessity IV selection: adding Z p+1 hurts if γ p+1 = 0. But adding even very weak SNPs usually reduces variance. 17/44

18 Diagnostic plots Residual Q-Q plot Under model 1, standardized residuals follow N(0, 1) if ˆβ = β 0 : ˆt j = Influence of a single SNP Asymptotic expansion: ˆΓ j ˆβˆγ j. ˆβ 2 σxj 2 + σ2 Yj ˆβ PS = 1 + o p(1) V 2 p j=1 (ˆΓ j β 0ˆγ j )(ˆΓ j σ 2 Xj β 0 + ˆγ j σ 2 Yj ) (σ 2 Xj β2 0 + σ2 Yj )2. In practice, we can also use leave-one-out. 18/44

19 Example (continued) Same three datasets, selected 160 uncorrelated SNPs that have p-values 10 4 in BMI-FEM. Diagnostic plots: 10 rs Standardized residual 5 0 Leave one out estimate rs Quantile of standard normal Instrument strength (F statistic) Clear overdispersion! 19/44

20 The most likely explanation of the lack of fit is pleiotropy (direct effects). Heuristic (Direct effect of Z j is α j ) p X = γ j Z j + η X C + E X, j=1 Y = β 0 X + p α j Z j + η Y C + E Y = j=1 p j=1 (β 0 γ j + α j }{{} Γ j )Z j +... Model (Random direct effects) Assume α j = Γ j β 0 γ j i.i.d. N(0, τ 2 0 ) for all j. τ 2 0 should be small (scales like 1/p). 21/44

21 Connection to population genetics We assumed the normal random effects model because of The observed overdiserpsion. Statistical convenience. This is consistent (or not inconsistent) with genetics: Fisher s (1918) infitestimal model. New perspective on pleiotropy: 1 1 Boyle, E. et al. (2017). An expanded view of complex traits: from polygenic to omnigenic. Cell 169, p /44

22 Back to statistics: Failure of profile likelihood The profile likelihood of is given by l(β, τ 2 ) = 1 2 Easy to verify p j=1 (ˆΓ j βˆγ j ) 2 σ 2 Xj β2 + σ 2 Yj + τ 2 + log(σ2 Yj + τ 2 ), [ ] E β l(β 0, τ0 2 ) = 0. But the other score function is biased: τ 2 l(β, τ 2 ) = 1 2 p j=1 (ˆΓ j βˆγ j ) 2 (σ 2 Xj β2 + σ 2 Yj + τ 2 ) 2 1 σ 2 Yj + τ 2. This is not too surprising as we are profiling out p nuisance parameters γ 1,, γ p (the Neyman-Scott problem). 23/44

23 Adjusted profile score There exist many ways to modify the profile likelihood. We will take the approach of McCullagh and Tibshirani (1990). Adjusted profile score (APS) ψ 1 (β, τ 2 ) = β l(β, τ 2 ), p [ ψ 2 (β, τ 2 ) = j=1 σ 2 Xj (ˆΓ j βˆγ j ) 2 (σ 2 Xj β2 + σ 2 Yj + τ 2 ) 2 1 σ 2 Xj β2 + σ 2 Yj +τ 2 ]. Trivial root: β ± or τ 2. Let ˆβ APS be the non-trivial finite solution. 24/44

24 Some theory Theorem Assume that is true, (β 0, pτ 2 0 ) is in a bounded set B, p and p/n 2 0. Then 1 With probability going to 1 there exists a solution in B. p 2 All solutions in B are consistent: ˆβ APE β0 and pˆτ APS 2 pτ0 2 p 0. Can obtain similar asymptotic normality assuming p/n λ (0, ) (see the paper). 25/44

25 Example (continued) Same 160 SNPs. Standardized residual rs Leave one out estimate rs Quantile of standard normal Instrument strength (F statistic) A clear outlier: rs Slightly underdispersed. 26/44

26 Model (Random direct effects with outliers) α j (1 ɛ)n(0, τ 2 0 ) + ɛg for some unknown G. Recall the expansion: ˆβ PS = 1 + o p(1) V 2 p j=1 (ˆΓ j β 0ˆγ j )(ˆΓ j σ 2 Xj β 0 + ˆγ j σ 2 Yj ) (σ 2 Xj β2 0 + σ2 Yj )2. Problem: a single SNP can have unbounded influence (same for APS). Our solution: robustify the adjusted profile score. 28/44

27 A method robust to outliers Standardized residual: t j (β, τ 2 ) = ˆΓ j βˆγ j σ 2 Xj β2 + σ 2 Yj + τ 2. Robust adjusted profile score (RAPS) ψ (ρ) 1 (β, τ 2 ) = ψ (ρ) 2 (β, τ 2 ) = p ρ (t j ) β t j, j=1 p j=1 σ 2 Xj t j ρ (t j ) δ σxj 2 β2 + σyj 2 + τ 2. ρ is robust loss, δ = E[T ρ (T )] for T N(0, 1). Reduces to solution 2 (APS) when ρ(t) = t 2 /2. 29/44

28 Some theory General theory is very difficult because we are solving nonlinear equations. Theorem (Local identifiability) In, E[ψ (ρ) (β 0, τ 2 0 )] = 0 and E[ ψ (ρ) ] has full rank. Asymptotic normality can be established assuming consistency and additional technical conditions (see the paper). 30/44

29 Example (continued) Same 160 SNPs. Huber s loss function. Standardized residual rs Leave one out estimate rs Quantile of standard normal Instrument strength (F statistic) Influence of the outlier is limited. 31/44

30 Example (continued) Same 160 SNPs. Tukey s biweight function. Standardized residual rs Leave one out estimate rs Quantile of standard normal Instrument strength (F statistic) Influence of the outlier is essentially 0. 32/44

31 Comparison of all the methods Method Estimate Standard error Profile score (PS) Adjusted PS (APS) Robust APS (RAPS, Huber) Robust APS (RAPS, Tukey) Inverse variance weighted (IVW) Weighted median MR Egger Existing meta-analysis methods for MR 1 IVW: weighted average of ˆΓ j /ˆγ j. 2 Weighted median of ˆΓ j /ˆγ j. 3 MR Egger: weighted least squares of ˆΓ j against ˆγ j. They all ignore measurement error in ˆγ j and weak IV bias. 33/44

32 Simulation setting Create simulated datasets to mimic the BMI-SBP example with known β 0. ind. ˆγ j N(γ j, σxj 2 ) where (γ j, σxj 2 ) come from the real data. Γ j are generated in 6 different ways (β 0 = 0.5): : Γ j = γ j β 0 (i.e. α j 0); : α j i.i.d. N(0, τ 2 0 ), where τ 0 = 2 (1/p) p σ Yj ; j=1 : Same as 2 but add 5 τ 0 to α 1. i.i.d. Heavy-tailed α: α j τ 0 Lap(1). Heteroskedastic α: Var(α j ) γj 2 and normal. Many outliers: Add 5 τ 0 to 10% randomly selected α j. 35/44

33 Simulation results 1 (p = 25) IVW Egger W. Median method PS APS RAPS β^ Black: boxplot of point estimate over 1000 realizations. Red: quartile using median standard error. 36/44

34 Simulation results 1 (p = 160) IVW Egger W. Median method PS APS RAPS β^ Black: boxplot of point estimate over 1000 realizations. Red: quartile using median standard error. 37/44

35 Another real data example: BMI-BMI BMI-GIANT: full dataset from the GIANT consortium (i.e. combining BMI-FEM and BMI-MAL), used to select SNPs. BMI-UKBB-1: half of the UKBB data, used as the exposure. BMI-UKBB-2: another half of UKBB data, used as the outcome. Because exposure and outcome, γ j Γ j. So true β 0 = 1 and there is no direct effect. 38/44

36 BMI-BMI results (GIANT, UKBB-1, UKBB-2) p sel # SNPs Mean F IVW W. Median W. Mode 1e (0.026) (0.039) (0.042) 1e (0.024) (0.039) (0.044) 1e (0.024) (0.036) (0.041) 1e (0.022) (0.034) (0.038) 1e (0.019) (0.033) (0.039) 1e (0.017) (0.031) (0.035) 1e (0.015) (0.027) (0.231) 1e (0.014) (0.023) (7.130) p sel # SNPs Median F Egger PS RAPS 1e (0.055) (0.023) (0.026) 1e (0.050) (0.023) (0.025) 1e (0.048) (0.021) (0.025) 1e (0.043) (0.019) (0.023) 1e (0.036) (0.018) (0.020) 1e (0.031) (0.017) (0.018) 1e (0.027) (0.016) (0.016) 1e (0.022) (0.015) (0.015) 39/44

37 Selection bias (UKBB-1, UKBB-1, UKBB-2) p sel # SNPs Mean F IVW W. Median W. Mode 1e (0.02) 0.83 (0.025) (0.046) 1e (0.017) 0.8 (0.022) (0.053) 1e (0.016) (0.019) (0.058) 1e (0.015) (0.019) (0.079) 1e (0.013) (0.016) (0.12) 1e (0.012) (0.015) (0.122) 1e (0.011) 0.57 (0.014) (0.096) 1e (0.01) (0.013) (0.093) p sel # SNPs Median F Egger PS RAPS 1e (0.051) (0.015) (0.021) 1e (0.046) (0.014) (0.018) 1e (0.043) (0.012) (0.016) 1e (0.041) (0.011) (0.016) 1e (0.037) (0.01) (0.015) 1e (0.033) (0.009) 0.66 (0.014) 1e (0.03) (0.008) (0.013) 1e (0.025) (0.008) (0.012) 40/44

38 This talk Paper: Statistical inference in two-sample Mendelian randomization using robust adjusted profile score. arxiv: Software: R package mr.raps is currently on CRAN. Can be directly called from the TwoSampleMR platform ( 42/44

39 Additional references I Overview of MR (Epidemiology): Davey Smith, G. and Ebrahim, S. (2003). Mendelian randomization : can genetic epidemiology contribute to understanding environmental determinants of disease? International Journal of Epidemiology. Davey Smith, G. and Hemani, G. (2014). Mendelian randomization: genetic anchors for causal inference in epidemiological studies. Human Molecular Genetics. Overview of MR (more Stats): Hernán, M. and Robins, J. (2006). Instruments for causal inference: an epidemiologist s dream?. Epidemiology. Didelez, V. and Sheehan N. (2007). Mendelian randomization as an instrumental variable approach to causal inference. Statistical Methods in Medical Research. Human genetics: Gamazon, E. et al. (2015). A gene-based association method for mapping traits using reference transcriptome data. Nature Genetics. Pickrell, J. et al. (2016) Detection and interpretation of shared genetic influences on 42 human traits. Nature Genetics. 43/44

40 Additional references II Robust MR methods (individual-level data): Kang, H. et al. (2016). Instrumental Variables Estimation with Some Invalid Instruments and its Application to Mendelian Randomization. JASA. Tchetgen Tchetgen, E. et al. (2017). The GENIUS approach to robust Mendelian randomization inference. arxiv: Robust MR methods (summary-level data): Bowden, J. et al. (2015). Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. International Journal of Epidemiology. Bowden, J. et al. (2016). Consistent estimation in Mendelian randomization with some invalid instruments using a weighted median estimator. Genetic Epidemiology. Pleiotropy: Solovieff, N. et al. (2013). Pleiotropy in complex traits: challenges and strategies. Nature Reviews Genetics. Boyle, E. et al. (2017). An expanded view of complex traits: from polygenic to omnigenic. Cell. 44/44