Letter in response to “C-reactive protein levels and risk of dementia—Observational and genetic studies of 111,242 individuals from the general population” by Hegazy and colleagues.

Mendelian randomization is the use of genetic variants analogously to randomized treatment allocation in a trial to make causal inferences from observational data. In their recent article, Hegazy and colleagues assessed associations between genetic variants in the CRP gene region—which they treated analogously to random allocation to increased levels of C-reactive protein (CRP)—and risk of dementia.¹ However, rather than assessing associations in the population as a whole, they assessed associations in strata of the population defined according to body mass index (BMI): those with BMI ≤25 kg/m², and those with BMI >25 kg/m². They observed positive associations between variants and dementia risk in the low BMI stratum, and inverse associations (although mostly not statistically significant) in the high BMI stratum. The authors interpreted the positive associations as evidence that CRP has a deleterious effect on dementia risk in low BMI individuals. However, there are reasons to be skeptical of this interpretation.

The genetic “randomization” exploited in Mendelian randomization occurs at conception. Therefore, all subgroup analyses in a Mendelian randomization investigation are improper subgroup analyses, as stratification is always based on a post-randomization covariate—that is, a variable measured after randomization.² The only exceptions are subgroup analyses based on variables that logically cannot be caused by the genetic variants, such as age, sex, and ancestry.

When stratification in a randomized trial is based on a post-randomization covariate, exchangeability is not guaranteed for the stratified trial arms; in other words, randomization is broken within the strata. This can be understood as an example of collider bias: stratifying (or equivalently, conditioning or adjusting) on a common effect of randomization and a variable typically leads to randomization being associated with the variable within strata, even if they were independent in the overall population.³

In this case, BMI is a plausible collider, as it is potentially causally downstream of inflammation and pre-clinical dementia. Stratifying on BMI could therefore lead to CRP variants being associated with dementia within strata of BMI in the absence of a causal effect of inflammation. The pattern of genetic associations observed by Hegazy and colleagues is typical of that occurring due to collider bias: a null overall association, which is positive in one stratum and negative in the other.⁴ There is no clear biological reason that the direction of effect of CRP would be in the positive in low BMI individuals but in the inverse direction in high BMI individuals.

Generally speaking, adjustment for covariates other than age, sex, ancestry, and technical covariates (such as study center) is discouraged in Mendelian randomization analyses due to the possibility of inducing collider bias.⁵ Stratification on covariates is possible, but requires careful application of suitable methodology so as not to induce collider bias.⁶ In both randomized trials and Mendelian randomization, analysts and reviewers should be particularly skeptical of results from subgroup analyses where the overall association is null, and estimates for subgroup analyses are in opposing directions. Although there may be some cases where the direction of causal effect of a risk factor truly differs between subgroups, these are likely to be rare, and collider bias is often a more plausible explanation.

Hegazy 及其同事对“C 反应蛋白水平和痴呆症风险——对 111,242 名普通人群的观察和遗传研究”的回应。

孟德尔随机化是在试验中使用类似于随机治疗分配的遗传变异，以根据观察数据进行因果推断。在他们最近的文章中，Hegazy 及其同事评估了CRP基因区域的遗传变异与痴呆症风险之间的关联——他们将其处理为类似于随机分配以增加 C 反应蛋白 (CRP) 水平。¹然而，他们没有评估整个人群中的关联，而是评估了根据体重指数 (BMI) 定义的人群分层的关联：BMI ≤ 25 kg/m 2 的人群和 BMI > 25 kg/m^的人群米². 他们在低 BMI 阶层观察到变异与痴呆风险之间存在正相关，而在高 BMI 阶层则呈负相关（尽管大多不具有统计学意义）。作者将这种正相关解释为 CRP 对低 BMI 个体的痴呆症风险具有有害影响的证据。然而，有理由怀疑这种解释。

孟德尔随机化中利用的遗传“随机化”发生在受孕时。因此，孟德尔随机化调查中的所有亚组分析都是不正确的亚组分析，因为分层总是基于随机化后协变量——即随机化后测量的变量。²唯一的例外是基于逻辑上不可能由遗传变异引起的变量的亚组分析，例如年龄、性别和血统。

当随机试验中的分层基于随机化后协变量时，不能保证分层试验组的可交换性；换句话说，分层内的随机化被打破了。这可以理解为碰撞偏倚的一个例子：根据随机化和变量的共同影响进行分层（或等效地，调节或调整）通常会导致随机化与层内的变量相关联，即使它们在总体人群中是独立的. ^3个

在这种情况下，BMI 是一个似是而非的碰撞因素，因为它可能是炎症和临床前痴呆症的潜在因果下游。因此，在没有炎症因果效应的情况下，对 BMI 进行分层可能导致CRP变体与 BMI 层内的痴呆相关。Hegazy 及其同事观察到的遗传关联模式是由于对撞机偏差而发生的典型模式：无效的整体关联，在一个层中为正，在另一个层中为负。⁴没有明确的生物学原因表明 CRP 的作用方向对低 BMI 个体呈正向，但对高 BMI 个体呈相反方向。

一般来说，在孟德尔随机化分析中不鼓励对年龄、性别、血统和技术协变量（如研究中心）以外的协变量进行调整，因为可能会引起碰撞偏差。⁵对协变量进行分层是可能的，但需要谨慎应用合适的方法，以免引起对撞机偏差。^6个在随机试验和孟德尔随机化中，分析人员和审稿人应该特别怀疑总体关联为零的亚组分析结果，并且亚组分析的估计值是相反的。尽管在某些情况下，风险因素的因果效应方向在亚组之间确实存在差异，但这种情况可能很少见，碰撞偏差通常是更合理的解释。