Propagating uncertainty in ecological models to understand causation
Propagating uncertainty in ecological models to understand causation
复制标题
DOI：
10.1002/fee.2610
复制
发表时间：
2023-04
影响因子：
10.3
通讯作者：
Neil A. Gilbert;H. Eyster;Elise F. Zipkin
中科院分区：
文献类型：
--
作者：
Neil A. Gilbert;H. Eyster;Elise F. Zipkin
Addicott等人（2022）证明，通过Akaike信息准则（AIC）进行模型选择可能导致研究人员选择焦点变量因果效应偏倚的模型（Tredennick et al. 2021; Arif and MacNeil 2022）。我们同意，应用于一组相关模型的模型选择方法不太可能提供因果推论，特别是当研究人员没有先验地选择变量来评估特定假设时。相反，有必要对震源系统的因果结构进行批判性思考（即“科学先于统计”；McElreath 2020）。我们感谢将这一问题引起生态社区注意的努力，但希望对分析中两阶段最小二乘的使用发表评论-其中一次回归的拟合值的点估计用于第二次回归作为数据-并强调一种替代方法，其中第一阶段的不确定性被传播到第二阶段。作者模拟（见Addicott et al.[2022]图1）鱼类种群增长作为食物（观察到的）和捕捞努力（未观察到的）的函数。捕鱼量（努力的代表）被观察到，但被增长率所混淆。包含食物和捕获物的朴素模型会产生食物的偏倚效应，但AIC比只包含食物的简单（但无偏倚）模型“更偏爱”。作者评估了第三种方法，其中混淆变量（捕获量）被建模为食物和渔网数量的函数（一个影响捕获量但不影响增长率的工具变量）。所得的渔获量（连同食物）点估计值然后用于第二次回归来估计增长率，从而对食物的影响进行无偏估计。虽然这种两阶段方法可能是有效的，并且通常被应用，特别是在经济学等领域（Angrist和Krueger 1995； McElreath 2020），但它不会传播
Addicott et al.(2022) demonstrated that model selection via Akaike information criterion (AIC) may lead researchers to choose models in which the causal effects of focal variables are biased (Tredennick et al. 2021; Arif and MacNeil 2022). We agree that model selection approaches applied to a suite of correlative models are unlikely to provide causal inferences, particularly when researchers do not choose variables a priori to evaluate specific hypotheses. Rather, critical thinking about the focal system’s causal structure (that is,“science before statistics”; McElreath 2020) is necessary. We appreciate the effort to bring this issue to the attention of the ecological community but wish to comment on the use of twostage least squares in the analysis–in which point estimates of fitted values from one regression are used in a second regression as data–and highlight an alternative approach in which the uncertainty from the first stage is propagated to the second stage.The authors simulated (see Addicott et al.’s [2022] Figure 1) fish population growth as a function of food (observed) and fishing effort (not observed). Catch (a proxy for effort) is observed but is confounded by growth rate. A naive model containing food and catch produces a biased effect of food but is “preferred” by AIC over a simpler–but unbiased–model containing only food. The authors evaluated a third approach in which the confounding variable (catch) is modeled as a function of food and the number of nets (an instrumental variable affecting catch but not growth rate). The resultant point estimates of catch (along with food) are then used in a second regression to estimate growth rate, leading to an unbiased estimate of food’s effect. While this two-stage approach can be effective and is commonly applied, particularly in fields such as economics (Angrist and Krueger 1995; McElreath 2020), it does not propagate