| Objective:Drawing causal conclusions from observational epidemiological studies can be problematic due to the limitations of reverse causation and unmeasured confounding.Relying on the principle that the genetic information is randomly assorted at conception,MR(Mendelian randomization)utilizes genetic variants as IVs(Instrumental variables)and mimic randomization of the study population,thus overcome the downsides that compromise standard association study techniques and offer the novel potential for reliable causal inference within the observational design.In response to the advent of the GWASs(Genome-Wide Association Studies)for a variety of phenotypes and the availability of summary statistics for thousands of trait-associated loci,summary-data MR schemes incorporating multiple IVs have been widely applied to medical research.The statistical power of the MR analysis is highly dependent on the proportion of the variance in the exposure that explained by the IVs,whereas the genetic variants can generally explain only a small proportion of the variance,therefore result in inadequate statistical power for the model.The inclusion of a dramatic number of genetic variants as IVs can improve the explanatory ability for the variance in the exposure and therefore lead to the increased power of MR for validating the causal hypotheses,but care has to be taken that it is more likely to contain pleiotropic IVs within the enlarged set of genetic variants in the model.The issue of pleiotropy—where a genetic variant influencing the outcome through biological pathways other than the exposure of interest—can invalid IVs due to violations of the core assumptions of MR models,and bias the estimation.Given such considerations,this study develops an improved MR model that aims to increase the statistical power and simultaneously correct for the pleiotropic bias,thus provide more accurate and credible causal estimations.Methods:The remodeled MR framework can be described as follows: firstly,set the significance threshold used to select IVs to be at a less stringent level to make more genetic variants that associated with exposure to be comprised into the IV set,hereby the proportion of the variance in the exposure that can be explained by the IVs increases and thus the power of the model improved.Subsequently,make use of the outlier detection model to divide the IV set into a subset of IVs without pleiotropy and another subset of pleiotropic IVs,and adopt the IVW(Inverse Variance Weighting)and Radial MR-Egger models to estimate the effect that corresponds to the IV subset without pleiotropy and the pleiotropic IV subset,respectively.The theoretical basis for the innovation lies in that the valid IV subset can be exploited with maximum estimation efficiency and simultaneously the bias originated from the pleiotropic IV subset can be utterly corrected.Lastly,combine the estimators that derive from the two IV subsets and obtain the causal estimator of exposure with the outcome.Compared with previous MR models that directly exclude the pleiotropic IVs,this improved model retains the proportion of the variance in the exposure that explained by the pleiotropic IVs.By this means,the exposure is instrumented by all the IVs rather than the valid IVs merely,and the statistical power is further increased.Traditional MR approaches assume that the IV-exposure associations are estimated without error(the NOME assumption),which will be violated when weak IVs are involved in the analytical procedure,the causal estimation will then bias towards null and the type I error rate for detecting the pleiotropic IVs will severely inflate owing to the violation.Through extensive theoretical and simulation analysis,the study evaluates and compares the statistical performance of a series of outlier-detection models and MR causal effect estimation models whenever they cooperate with many weak IVs.Accordingly,the proper methods can be picked to avoid the misidentification of the pleiotropic IVs and avert the underestimation of the causal conclusions.Results:Theoretical and simulative results show that the Cochran’s Q statistic test with modified weights is robust to the strength of the IVs,and it preserves the type I error rate at the nominally 5% significance level whilst maintaining the statistical power in terms of heterogeneity quantification,despite in the presence of weak IVs.While other methods are sensitive to the strength of the IVs and exhibit extremely inflated or conservative type I error rate in the respect of detecting pleiotropic IVs when there are weak IVs.Regarding the causal estimation,IVW and Radial MR-Egger with modified weights can remove the bias in MR estimates when the NOME assumption is violated as a result of weak IVs.In contrast,the IVW and Radial MR-Egger using first-and second-order weights produce estimations that bias towards the null when NOME is violated,and the stronger the violation,the stronger the attenuation.In conclusion,the MR models with modified weights outperform other MR methods in terms of identifying pleiotropic IVs and estimating causal effect when the strength of the IVs is weak,so they are best suited to the situation for our analytical framework that possibly involving in plenty of exposure associated genetic variants but appears to have weak strength.For this reason,the Cochran’s Q statistic test is selected to detect the pleiotropic IVs,the IVW model and Radial MR-Egger model with modified weights are selected to estimate the causal effect basing on valid IVs and pleiotropic IVs,separately.Through evaluating the performance of several meta-analyses,the Bayesian meta-analytic estimation using weak information prior,which is superior to the traditional frequency-family methods in small sample meta-analysis,is eventually determined for the utilization of combining the effect estimators derived from two IV subsets.The empirical analysis illustrates the application of the new proposed MR framework for assessing the causal role of some certain sleep-related traits on Late-onset Alzheimer’s disease(LOAD)but finds no significant causal effects of genetically determined sleep traits on the risk of LOAD.Further evaluating the performance of the statistical power,the proposed MR model has much sufficient statistical power to detect a certain effect size than the traditional MR models.Meanwhile,under the identical size of power,the proposed MR model should allow for the identification of small causal associations,conversely,there would need to have moderate to large effects on disease risk for identification in other MR models.Conclusion:MR is a major step forward in understanding the pathogenesis of disorders,and the findings of MR could help prioritize clinical trials,drug development and inform public health decision making.MR is susceptible to theoretical limitations of low statistical power due to the small amount of variation in a phenotypic trait that is typically instrumented by genetic variants.An increase of the expected power can be achieved by including more IVs into the model,but investigators will have to account for the estimation bias when bringing into more weak IVs and/or pleiotropic IVs.The improved MR framework in this study has been verified to have a preferable property in that it could identify the true causal effect with ideal power and effectively correct for the potential bias,thus further overcomes the limitations of the existed MR methods,thereupon offers more precise and reliable estimations. |
PDF Full Text Request