Variables, Instrumental

Confounding is a pervasive problem for drawing inferences about political causes and effects. In brief, some individuals, countries, or other units are exposed to a “treatment” or “intervention,” while others are not. Differences in outcomes may reflect the effect of treatment, or they may be due to confounders—that is, variables associated with exposure to treatment and with the outcome. For example, does civil war inhibit economic growth? Do political institutions shape development? How does service in the military affect wages after war? Such questions are difficult to settle, because a range of unobserved variables are associated with the presence of civil war, types of political institutions, or service in the military, and these may shape growth, development, or individual wages. Reverse causality can also be a problem.

Instrumental variables can be used to address the problem of confounding, in both experiments and observational studies. In randomized controlled experiments, a coin flip determines which subjects are assigned to treatment, so subjects assigned to receive the treatment are, on average, just like subjects assigned to control. However, even in experiments there can be confounding, if subjects who accept the treatment are compared with those who refuse it. Analysts should, therefore, compare subjects randomly assigned to treatment with those randomly assigned to control. Instrumental-variables analysis may be used to estimate the effect of treatment on Compliers (subjects who follow the treatment regime to which they are assigned). In experiments, treatment assignment usually satisfies two key requirements for an instrumental variable: It is statistically independent of unobserved causes of the dependent variable, and it plausibly affects the outcome only through its effect on treatment receipt.

In observational studies (those in which assignment to treatment is not under the control of the researcher), the problem of confounding is typically more severe because units self-select into the treatment and control groups. Instrumental-variables analysis can be used to recover the effect of an endogenous treatment—that is, a treatment variable that is correlated with confounders. However, strong assumptions are often required, and these can be only partially validated from data. The use of instrumental variables in observational studies is discussed below, after a benchmark application to experimental data is first described.