著者
鈴木 越治 小松 裕和 頼藤 貴志 山本 英二 土居 弘幸 津田 敏秀
出版者
一般社団法人日本衛生学会
雑誌
日本衛生学雑誌 (ISSN:00215082)
巻号頁・発行日
vol.64, no.4, pp.786-795, 2009 (Released:2009-10-02)
参考文献数
53
被引用文献数
2 1

A central problem in natural science is identifying general laws of cause and effect. Medical science is devoted to revealing causal relationships in humans. The framework for causal inference applied in epidemiology can contribute substantially to clearly specifying and testing causal hypotheses in many other areas of biomedical research. In this article, we review the importance of defining explicit research hypotheses to make valid causal inferences in medical studies. In the counterfactual model, a causal effect is defined as the contrast between an observed outcome and an outcome that would have been observed in a situation that did not actually happen. The fundamental problem of causal inference should be clear; individual causal effects are not directly observable, and we need to find general causal relationships, using population data. Under an “ideal” randomized trial, the assumption of exchangeability between the exposed and the unexposed groups is met; consequently, population-level causal effects can be estimated. In observational studies, however, there is a greater risk that the assumption of conditional exchangeability may be violated. In summary, in this article, we highlight the following points: (1) individual causal effects cannot be inferred because counterfactual outcomes cannot, by definition, be observed; (2) the distinction between concepts of association and concepts of causation and the basis for the definition of confounding; (3) the importance of elaborating specific research hypotheses in order to evaluate the assumption of conditional exchangeability between the exposed and unexposed groups; (4) the advantages of defining research hypotheses at the population level, including specification of a hypothetical intervention, consistent with the counterfactual model. In addition, we show how understanding the counterfactual model can lay the foundation for correct interpretation of epidemiologic evidence.
著者
鈴木 越治 小松 裕和 頼藤 貴志 山本 英二 土居 弘幸 津田 敏秀
出版者
日本衛生学会
雑誌
日本衛生学雑誌 (ISSN:00215082)
巻号頁・発行日
vol.64, no.4, pp.796-805, 2009 (Released:2009-10-02)
参考文献数
56
被引用文献数
1 1

Confounding is frequently a primary concern in epidemiological studies. With the increasing complexity of hypothesized relationships among exposures, outcomes, and covariates, it becomes very difficult to present these hypotheses lucidly and comprehensively. Graphical models are of great benefit in this regard. In this article, we focuse on directed acyclic graphs (DAGs), and review their value for confounder selection, categorization of potential biases, and hypothesis specification. We also discuss the importance of considering causal structures before selecting the covariates to be included in a statistical model and the potential biases introduced by inappropriately adjusting statistical models for covariates. DAGs are nonparametric and qualitative tools for visualizing research hypotheses regarding an exposure, an outcome, and covariates. Causal structures represented in DAGs will rarely be perfectly “correct” owing to the uncertainty about the underlying causal relationships. Nevertheless, to the extent that using DAGs forces greater clarity about causal assumptions, we are able to consider key sources of bias and uncertainty when interpreting study results. In summary, in this article, we review the following three points. (1) Although researchers have not adopted a consistent definition of confounders, using DAGs and the rules of d-separation we are able to identify clearly which variables we must condition on or adjust for in order to test a causal hypothesis under a set of causal assumptions. (2) We also show that DAGs should accurately correspond to research hypotheses of interest. To obtain a valid causal interpretation, research hypotheses should be defined explicitly from the perspective of a counterfactual model before drawing DAGs. A proper interpretation of the coefficients of a statistical model for addressing a specific research hypothesis relies on an accurate specification of a causal DAG reflecting the underlying causal structure. Unless DAGs correspond to research hypotheses, we cannot reliably reach proper conclusions testing the research hypotheses. Finally, (3) we have briefly reviewed other approaches to causal inference, and illustrate how these models are connected.
著者
津田 敏秀 頼藤 貴志 土居 弘幸 鹿嶋 小緒里
巻号頁・発行日
2014-04-01 (Released:2014-04-04)

大気汚染物質の健康影響が全国的に注目を集めているが、国内での大気汚染の健康影響を評価した疫学研究は少ない。本年度は、短期曝露の死亡への影響と大気汚染曝露の周産期・小児期への健康影響についての研究を行った。短期曝露の死亡への影響については、東京都の乳児と高齢者を対象に、大気汚染曝露と乳児死亡・高齢者死亡との関連について評価した。乳児死亡に関しては、最近話題の微小粒子状物質だけでなく、浮遊粒子状物質の中の微小粒子状物質以外の粒子(海外ではCoarse particlesと呼ばれている)も死亡のリスクを上昇させていた。また、高齢者においても同様に、微小粒子状物質とCoarse particlesが全死因死亡・死因別死亡(心血管系死亡や呼吸器系死亡)のリスクを上昇させていた。大気汚染曝露の周産期・小児期への健康影響に関しては、厚生労働省が実施している21世紀出生児銃弾調査のデータを対象とし、妊娠中の大気汚染曝露と出生後の行動発達との関連を評価した。更に、平成28年度には東海地方の産科で出生した母児ペアを対象にして、個人レベルの大気汚染曝露濃度を予測するモデルを作成する予定である。具体的には、Land Use Regression modelと言われるモデルで、対象者の居住環境近くにある道路情報や人口情報、地形情報などを利用し、個人の妊娠期間中の大気汚染濃度を予測するモデルを作成する。自治体の平均濃度を割り振るのではなく、この方法を用いることにより、対象者の曝露情報の誤分類を減らすことが可能となる。平成27年度は、これらの研究準備のため、個人の住所情報の整備を行った。