The Study On Instrumental Variable Approach To Controlling Both Measurement Bias And Confounding Bias

Background:A series of data-driven observational studies gradually become the focus of epidemiological research.Comparing with experimental studies,observational studies have several inherent defects in demonstrating causality.Firstly,causal inference in observational studies is usually susceptible to confounding,especially unmeasured confounding,which leads to confounding bias due to the lack of randomization.Besides,the difference between the measured value and the true value may result from the imprecision of measuring techniques,the changeable measurement condition,unavoidable recall biases,miscoding by the collectors of the data,or incorrect transformation of the initial variable,etc,leading to measurement bias in causal effect estimation.Instrumental variable(?)methods such as Mendelian randomization(MR)method using genetic determinant(s)as instrumental variable(s)have gained increasing recognition and popularity in controlling unobserved confounding.Nevertheless,traditional ? methods rely on all the exposure and outcome being measured precisely,which is almost difficult to be satisfied due to the ubiquitous measurement errors.The ignorance of varying levels of measurement errors always results in bias in ? estimate or reduces its power.Although some studies have developed methods correcting for measurement error in ? model,some studies'attention are limited to measurement error in exposure or outcome,or only obtained the direction or sharp bounds of the causal effect,these studies did not focus on the point identification of exposure effect.other studies required information about the distribution of the unmeasured true variable as well as the measurement error,which are rarely available in practice.When both exposure and outcome are subjected to measurement errors and the information about measurement error is unavailable,method dealing with both measurement bias and confounding bias in the ? model is scarce.Therefore,establishing an instrumental variable method controlling both measurement bias and confounding bias is one of the key issues to be solved in observational studies.Methods:Accounting for error-prone exposure and outcome as well as widely existed covariates that confound exposure-outcome association,we first examined the asymptotic bias in traditional ? estimator caused by ignoring the non-differential measurement errors in both exposure and outcome through theoretical derivation.Furthermore,we proposed instrumental variable approaches to controlling unobserved confounding and measurement errors in both exposure and outcome,in virtue of two repeat measurements of both exposure and outcome through theoretical derivation,simulation,and application.To be specific,we first constructed three causal diagrams according to the possible association between measurements of exposure and outcome(the observed measurements may be independent,partially independent or dependent when conditional on the unobserved true exposure and outcome);Three scenarios were considered in this study:(1)Continuous exposure and outcome subject to independent or partially independent non-differential measurement error in continuous ? model;(2)Continuous exposure and outcome subject to dependent non-differential measurement error in continuous ? model;(3)Binary exposure and outcome subject to misclassification in binary ? model.Based on the corresponding causal diagram and basic criteria for causal inference(including do-calcules and backdoor criteria,etc),we proposed instrumental variable approaches for controlling unobserved confounding and measurement errors in both exposure and outcome in each scenario through theoretical derivation.Furthermore,we assessed the finite sample performance of four strategies(including crude association,naive ? method,generalized method of moments three-stage least squares method,and our proposed approach)in extensive simulation studies through bias,precision,mean square error,etc.Finally,we illustrated these approaches with an application to the causality between relative gene expressions of three target genes in immune thrombocytopenia patients.The proposed methods were implemented in R functions and an R package.Results:1.Theoretical derivation results:(1)The classical non-differential measurement errors in exposure or outcome did not bias the causal effect estimate.When exposure and outcome subjected to general non-differential measurement errors and the measured values were positively correlated with the true values,the magnitude of bias depended on the difference in the slopes of measurement error models of exposure and outcome:when the slopes were equal,the causal effect estimate was unbiased,the bias became larger as the difference of the slopes increased.(2)When exposure and outcome subjected to independent or partially independent non-differential measurement errors in continuous ? model,we developed an instrumental variable approach to controlling measurement bias and confounding bias in virtue of two repeat measurements of both exposure and outcome.This approach relied on the basic assumptions of instrumental variable as well as the independent or partially independent non-differential measurement error assumption.(3)When exposure and outcome subjected to dependent non-differential measurement error in continuous ?model,we developed an instrumental variable approach to controlling measurement bias and confounding bias in virtue of two repeat measurements of both exposure and outcome.This approach relied on the basic assumptions of instrumental variable as well as the dependent non-differential measurement error assumption.(4)When exposure and outcome subjected to independent non-differential measurement error in binary IV model,we developed an instrumental variable approach to controlling measurement bias and confounding bias in virtue of two repeat measurements of both exposure and outcome.This approach relied on the basic assumptions of instrumental variable,the independent non-differential measurement error assumption as well as the equally and informative misclassification assumption.2.Simulation results:The simulation results varying across one parameter(including sample size,causal effect between variables,and parameters in exposure and outcome measurement error model)at a time and keeping all others at the levels of its initial value showed that our approaches in both three scenarios could obtain unbiased estimates of the exposure effect,with high statistical powers.Besides,our approaches were robust to the magnitude of measurement errors in exposure and outcome.3.Application results:Application results showed that in patients with immune thrombocytopenia,the relative expression of NF-?B has negative causal regulation on the relative expression of IL-18.4.We also provided R functions in this study.Moreover,the R package RCMIV was freely available on GitHub(https://github.com/lxinhui/RCMIV)to implement the proposed approaches.Conclusions:1.The classical non-differential measurement errors in exposure or outcome do not bias the causal effect estimate.When exposure and outcome subject to general non-differential measurement errors and the measured values were positively correlated with the true values,the magnitude of bias depended on the difference in the slopes of measurement error models of exposure and outcome.2.When continuous exposure and outcome subject to independent or partially independent non-differential measurement error in continuous IV model,if the basic identification assumptions of instrumental variable and independent or partially independent non-differential measurement error assumption hold,our correction approach can obtain unbiased estimates of the total causal effect.3.When continuous exposure and outcome subject to dependent non-differential measurement error in continuous ? model,if the basic assumptions of instrumental variable and symmetric dependent non-differential measurement error assumption hold,our correction approach can obtain unbiased estimates of the total causal effect.4.When binary exposure and outcome subject to independent non-differential measurement error in binary IV model,if the basic identification assumptions of instrumental variable,independent non-differential measurement error assumption,and equally and informative misclassification assumption hold,our correction approach can obtain unbiased estimates of the complier average causal effect.