Obtaining Actionable Inferences from Epidemiologic Actions.
Obtaining Actionable Inferences from Epidemiologic Actions.
复制标题
从流行病学行动中获得可行的推论。
DOI：
10.1097/ede.0000000000000960
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Naimi,AshleyI
中科院分区：
文献类型：
--
作者：
Naimi,AshleyI
实验操作的一个明显优势是可以使用随机化。在形式上，随机化提供了一种确定明确定义的因果效应的能力，而不需要求助于关于暴露(或其潜在混杂因素)是否或如何与结果相关的假设。在这种情况下，“识别”指的是在数学上将在给定研究中收集的数据等同于兴趣的主要影响的能力。16事实上，(理想的)随机试验之所以被视为金标准，恰恰是因为因果关系识别(非参数)是通过暴露随机化来实现的。10在Hutcheon等人的研究背景下，理想的(尽管不可能的)随机试验将包括将怀孕时登记的孕妇分配到增重组(例如，怀孕期间增加X1公斤与增加X2公斤)。在这样的研究中，女性将被跟踪，直到发生感兴趣的事件(活产)或发生竞争风险。然后，通过简单地对比每组婴儿的出生体重，就可以(非参数地)估计怀孕体重增加的影响。实际上，这样的研究是不可能的。然而，通过将这一分析定义为(假设的)试验，可以澄清在定义和估计妊娠体重增加对胎儿大小的影响方面的几个挑战。这将使我们能够采取措施减轻这些挑战，或者清楚地阐明对我们结果有效性的威胁。如果随机化成为可能，估计妊娠体重增加对胎儿大小的影响的关键挑战可以相对容易地解决。这些因素包括(1)预计不会出现混淆，这将使效果估计成为可能，而不需要对(可能的)高维混杂空间进行参数调整；(2)正确考虑竞争风险的能力，包括胎儿丢失和死产；以及(3)准确定义一个人如何“增加X1公斤与X2公斤”以及(相关地)怀孕体重增加的“影响”的能力。这种方法解释了怀孕之间保持不变的混杂因素，但不考虑怀孕之间变化的不受控制的混杂因素。在Hutcheon等人的研究中，两个可以说是怀孕体重增加对胎儿大小影响的最强混杂因素无法根据饮食和体力活动进行调整。由于登记数据的可获得性，关于怀孕体重增加和体重对出生结果的影响的研究经常受到这种未经测量的混淆的影响。
One clear advantage of experimental manipulation is that randomization can be used. Formally, randomization provides an ability to identify clearly defined causal effects without resorting to assumptions about whether or how the exposure (or its potential confounders) relate to the outcome. To “identify,” in this context, refers to the ability to mathematically equate the data collected in a given study with the primary effect of interest. 16 Indeed,(ideal) randomized trials are held as the gold standard precisely because causal effect identification is (nonparametrically) conferred by exposure randomization. 10 In the context of the study by Hutcheon et al, the ideal (albeit impossible) randomized trial would involve allocating pregnant women enrolled at conception into weight gain groups (eg, gain X1 versus X2 kg during pregnancy). In such a study, women would be followed until the event of interest (live birth) or a competing risk. The effect of pregnancy weight gain could then be estimated (nonparametrically) by simply contrasting the birthweight of infants in each group. Practically, such a study is not possible. However, by framing the analysis as a (hypothetical) trial, several challenges in defining and estimating the effect of pregnancy weight gain on fetal size can be clarified. This would allow us to either take steps to mitigate these challenges or clearly articulate the threats to the validity of our results. Were randomization possible, key challenges in estimating the effect of pregnancy weight gain on fetal size could be handled with relative ease. These include (1) the expected absence of confounding, which would enable effect estimation without resorting to parametric adjustment for a (possibly) high-dimensional confounder space;(2) the ability to properly account for competing risks, including fetal loss and stillbirth; and (3) the ability to define precisely how one might “gain X 1 versus X2 kg” and (relatedly) what is meant by the “effect” of pregnancy weight gain.Confounding is one of the primary issues that Hutcheon et al seek to address with a sibling-paired design. The approach accounts for confounders that remain constant between pregnancies, but uncontrolled confounders that change between pregnancies are not accounted for. In the study by Hutcheon et al, two of the arguably strongest confounders of the effect of pregnancy weight gain on fetal size could not be adjusted for diet and physical activity. Because of the availability of registry data, studies of the effect of pregnancy weight gain and body mass on birth outcomes are often affected by such unmeasured confounding.