Instrumental Variable Analyses and Selection Bias

Canan, Chelsea E.; Lesko, Catherine R.; Lau, Bryan

doi:10.1097/ede.0000000000000639

Cited by 50 publications

(43 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, the full practical implications for MR estimates of a time lag between randomization at conception and study recruitment often in late adulthood [14] have not been fully considered. Although several scenarios relevant to selection bias have been addressed, such as selective survival on exposure [15,16], on exposure and outcome [17], on exposure and other causes of the outcome [18][19][20] or on instrument and other causes of the outcome [12,19], i.e. competing risk.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Use of multivariable Mendelian randomization to address biases due to competing risk before recruitment

Schooling

Lopez

Zhao

et al. 2019

Preprint

View full text Add to dashboard Cite

BackgroundMendelian randomization (MR) provides unconfounded estimates. MR is open to selection bias particularly when the underlying sample is selected on surviving the genetically instrumented exposure and other conditions that share etiology with the outcome (competing risk before recruitment). Few methods to address this bias exist.MethodsWe use directed acyclic graphs to show this selection bias can be addressed by adjusting for common causes of survival and outcome. We use multivariable MR to obtain a corrected MR estimate, specifically, the effect of statin use on ischemic stroke, because statins affect survival and stroke typically occurs later in life than ischemic heart disease so is open to competing risk.ResultsIn univariable MR the genetically instrumented effect of statin use on ischemic stroke was in a harmful direction in MEGASTROKE and the UK Biobank (odds ratio (OR) 1.33, 95% confidence interval (CI) 0.80 to 2.20). In multivariable MR adjusted for major causes of survival and ischemic stroke, (blood pressure, body mass index and smoking initiation) the effect of statin use on stroke in the UK Biobank was as expected (OR 0.81, 95% CI 0.68 to 0.98) with a Q-statistic indicating absence of genetic pleiotropy or selection bias, but not in MEGASTROKE.ConclusionMR studies concerning late onset chronic conditions with shared etiology based on samples recruited in later life need to be conceptualized within a mechanistic understanding, so as to any identify potential bias due to competing risk before recruitment, and to inform the analysis and interpretation.

show abstract

Section: Introductionmentioning

confidence: 99%

“…Figures 1d and 1e show survival on both instrument and common causes of the outcome (U2) [12,19]. The time lag between randomization (at conception) and typical recruitment into genetic studies of major diseases in middle-to old-age means some MR studies may inevitably recruit on surviving both the genetic [11,13,16,[18][19][20].…”

Section: Introductionmentioning

confidence: 99%

Use of multivariable Mendelian randomization to address biases due to competing risk before recruitment

Schooling

Lopez

Zhao

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

“…In keeping with the results of our simulation study, non-trivial levels of selection bias were demonstrated via simulations [24,23,22,19,21,18]. [18] investigated two selection mechanisms in the context of Mendelian randomisation, and the remaining papers only considered a specific selection scenario.…”

Section: Discussionmentioning

confidence: 72%

“…Although selection bias is understood in the methodological literature (e.g., [6,12,5]), it is seldom acknowledged in IV analyses or discussed in guidelines for IV analysis (e.g., [13,14,15,16,17]). However, recent exceptions include examples where selection depends on the: exposure plus confounder, or outcome [18], exposure [19,18], instrument plus measured and unmeasured confounders (of the outcome-exposure association) [20], exposure and measured variable (which causes the outcome) [21], missing values of measured covariates [22,23], and unmeasured confounder plus measured covariates [24].…”

Section: Introductionmentioning

confidence: 99%

“…In the IV literature a small number of papers have used directed acyclic graphs (DAGs) [25,26,27] to illustrate when selection violates the assump-tions of an IV analysis [19,20,21,18]. However, these papers cover a limited range of selection scenarios, with Gkatzionis and Burgess [18] confining their discussion to Mendelian randomisation, and Ertefaie et al, and Canan et al [20,21] provide an incomplete explanation of the consequence of selection. Only one paper [18] considered if the effects of selection differed according to a null and non-null causal exposure effect, and none of these papers investigated if the consequences of selection differed according to a linear and nonlinear exposure-instrument association.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Selection bias in instrumental variable analyses

Davies

Smith

et al. 2017

Preprint

View full text Add to dashboard Cite

AbstractParticipants in epidemiological and genetic studies are rarely true random samples of the populations they are intended to represent, and both known and unknown factors can influence participation in a study (known as selection into a study). The circumstances in which selection causes bias in an instrumental variable (IV) analysis are not widely understood by practitioners of IV analyses. We use directed acyclic graphs (DAGs) to depict assumptions about the selection mechanism (factors affecting selection) and show how DAGs can be used to determine when a two-stage least squares (2SLS) IV analysis is biased by different selection mechanisms. Via simulations, we show that selection can result in a biased IV estimate with substantial confidence interval undercoverage, and the level of bias can differ between instrument strengths, a linear and nonlinear exposure-instrument association, and a causal and noncausal exposure effect. We present an application from the UK Biobank study, which is known to be a selected sample of the general population. Of interest was the causal effect of education on the decision to smoke. The 2SLS exposure estimates were very different between the IV analysis ignoring selection and the IV analysis which adjusted for selection (e.g., 1.8 [95% confidence interval −1.5, 5.0] and −4.5 [−6.6, −2.4], respectively). We conclude that selection bias can have a major effect on an IV analysis and that statistical methods for estimating causal effects using data from nonrandom samples are needed.

show abstract

Association Between Metabolically Different Adiposity Subtypes and Osteoarthritis: A Mendelian Randomization Study

Mulugeta

Eshetie

Kassie

et al. 2022

Arthritis Care & Research

View full text Add to dashboard Cite

Objective. In this Mendelian randomization (MR) study, the objective was to investigate the causal effect of metabolically different adiposity subtypes on osteoarthritis.Methods. We performed 2-sample MR using summary-level data for osteoarthritis (10,083 cases and 40,425 controls) from a genome-wide association using the UK Biobank, and for site-specific osteoarthritis from the Arthritis Research UK Osteoarthritis Genetics consortium. We used 3 classes of genetic instruments, which all increase body mass index but are associated with different metabolic profiles (unfavorable, neutral, and favorable). Primary analysis was performed using inverse variance weight (IVW), with additional sensitivity analysis from different MR methods. We further applied a nonlinear MR using UK Biobank data to understand the nature of the adiposity-osteoarthritis relationship.Results. Greater metabolically unfavorable and metabolically neutral adiposity were associated with higher odds of osteoarthritis (IVW odds ratio [OR] 1.56 [95% confidence interval (95% CI) 1.31, 1.85] and OR 1.60 [95% CI 1.15, 2.23], respectively). The estimate for the association between metabolically favorable adiposity and osteoarthritis was similar, although with notable imprecision (OR 1.55 [95% CI 0.70, 3.41]). Using site-specific osteoarthritis, metabolically unfavorable, neutral, and favorable adiposity were all associated with higher odds of knee osteoarthritis (OR 1.44 [95% CI 1.04, 1.98], OR 2.28 [95% CI 1.04, 4.99], and OR 6.80 [95% CI 2.08, 22.19], respectively). We found generally consistent estimates with a wider confidence interval crossing the null from other MR methods. The nonlinear MR analyses suggested a nonlinear relationship between metabolically unfavorable adiposity and osteoarthritis (P nonlinear = 0.003).Conclusion. Metabolic abnormalities did not explain the association between greater adiposity and the risk of osteoarthritis, which might suggest that the association is largely due to a mechanical effect on the joints.

show abstract

Instrumental Variable Analyses and Selection Bias

Cited by 50 publications

References 21 publications

Use of multivariable Mendelian randomization to address biases due to competing risk before recruitment

Use of multivariable Mendelian randomization to address biases due to competing risk before recruitment

Selection bias in instrumental variable analyses

Association Between Metabolically Different Adiposity Subtypes and Osteoarthritis: A Mendelian Randomization Study

Contact Info

Product

Resources

About