Answering causal questions using observational data

Most applied science is concerned with uncovering causal relationships. In many fields, randomized controlled trials (RCTs) are considered the gold standard for achieving this. The systematic use of RCTs to study causal relationships — assessing the efficacy of a medical treatment for example — has resulted in tremendous welfare gains in society. However, due to financial, ethical, or practical constraints, many important questions — particularly in the social sciences — cannot be studied using a controlled randomized experiment. For example, what is the impact of school closures on student learning and the spread of the COVID-19 virus? What is the impact of low-skilled immigration on employment and wages? How do institutions affect economic development? How does the imposition of a minimum wage affect employment? In answering these types of questions, researchers must rely on observational data, i.e., data generated without controlled experimental variation. But with observational data, a fundamental identification problem arises: the underlying cause of any correlation remains unclear. If we observe that minimum wages and unemployment correlate, is this because a minimum wage causes unemployment? Or because unemployment and lower wage growth at the bottom of the wage distribution leads to the introduction of a minimum wage? Or because of a myriad of other factors that affect both unemployment and the decision to introduce a minimum wage? Moreover, in many settings, randomized variation by itself is not sufficient for identification of an average treatment effect.