Investigation on Wellbore Multileakage Point Detection Technology Based on Excitation Pressure Waves.

Mendelian randomization (MR) uses genetic variants to explore the causal relationship between an exposure and an outcome of interest [, ]. It is usually implemented as a form of instrumental variable analysis. MR has become an established approach in epidemiology since first formalized by Davey Smith and Ebrahim []. This has been facilitated by the availability of large‐scale genetic databases, such as the UK Biobank, as well as the development of novel methods and accompanying software to aid researchers in conducting MR analyses [, ]. As a result, the number of publications using MR has seen a rapid increase in recent years []. If all the instrumental variable (IV) assumptions are satisfied, MR analyses can overcome biases due to unobserved confounding and reverse causation that are prevalent in standard regression analyses. However, MR studies are still susceptible to other forms of bias. In particular, selection bias has been shown to affect instrumental variable and MR analyses, including in some situations when a multivariable regression analysis would not be affected [, , ]. Selection bias exists when the value of a parameter of interest in the sample used for analysis differs from its value in the target population. When the parameter of interest is the exposure‐outcome causal effect, there are two mechanisms by which selection bias may occur. The first is as a result of collider bias: two independent variables that are both causes of a third variable (the “collider”) will become artificially correlated in an analysis conditioned on a particular value of . This is illustrated in Figure . In practice, the collider may represent participation in a study or inclusion in the analysis sample, which motivates the use of collider bias as a structural framework for analyzing selection bias []. A second type of selection bias occurs when the selection variable is not a collider, but the effect of interest is heterogeneous across levels of . This type of bias was first described by Greenland [] and has received increased attention in recent years [, , ]. Several mechanisms may give rise to this form of selection bias, including interaction, effect modification and non‐collapsibility; here, we use the term “heterogeneity” to incorporate all of them. The impact of this form of bias on MR studies has received limited attention so far. It is worth noting that selection bias due to heterogeneity cannot occur under the sharp causal null hypothesis that the exposure does not affect the outcome for any individual in the target population []; trivially, in this case, there would be no heterogeneity. These two types of selection bias are sometimes referred to as Type 1 and Type 2 selection bias respectively []. In this paper, we will focus primarily on collider bias; as we discuss later, negative control outcomes are not suited to studying selection bias due to heterogeneity.