Causal Effects via DAGs

Breaking down the Back and Front Door Criteria

This is the 4th article in a series on causal effects. In the last article of this series, we explored the question of identifiability. In other words, can the causal effect be evaluated from the given data? There we saw a systematic 3-step process to express any causal effect given a causal model where all variables are observed. The problem, however, becomes much more interesting when we have unmeasured confounders. In this article, I discuss two quick-and-easy graphical criteria for evaluating causal effects.

Identifiability

Identifiability is a central question in causal analysis i.e. can the causal effect be evaluated from the given data? In the previous blog, we saw a systematic 3-step process for answering this question for so-called Markovian causal models. These causal models satisfy two conditions: 1) no cycles and 2) no unmeasured noise terms that simultaneously cause two or more variables. This type of model can be represented by a directed cyclic graph i.e. a DAG. Examples of Markovian and non-Markovian DAGs are shown below.