The gold standard for drawing causal inference of a treatment T on an outcome Y. Randomizing individuals into treatment conditions typically creates balance on all measured and unmeasured confounders, so that the effect of T on Y can be interpreted as causal.
A variable that affects both the treatment/exposure T and the outcome Y, potentially biasing an estimate of the causal effect of T on Y. Confounders may be observed (i.e. measured in a study) or unobserved; observed confounders can be accounted for statistically when drawing causal inference.
A study where the participants are not randomly assigned to the treatment. This usually occurs in studies where it would be impossible or unethical to randomize participants (e.g., it would be unethical to assign a nonsmoker to start smoking).
Bias in causal inference that results from the fact that individuals often self-select their treatment/exposure status. For example, college attendees may have lower alcohol use in adulthood than their non-college counterparts, but we cannot infer causality because individuals self-select into college.
An individual’s probability of treatment/exposure modeled as a function of many possible confounding variables. These scores are a useful tool for comparing groups in different treatment/exposure conditions when the sample is not randomized.
A framework for measuring the impact of a treatment/exposure by considering what would have happened if the treatment had been different (i.e., a counterfactual outcome). For example, to consider the effect of marriage on substance use, you would need to observe outcomes for individuals given their actual marital status and also given the other status. Since the outcome can only be observed under one status, POTENTIAL outcomes are considered.
A potential outcome that did not occur. For example, if we study the impact of attending preschool on later academic achievement, we would need to determine the achievement of a subject who went to preschool and what his achievement would have been if he had not attended preschool. Since it is impossible for him to retroactively NOT attend preschool, this outcome is the counterfactual.
A study that was designed to be a randomized controlled trial, but after treatment assignment was completed, some individuals received a different treatment than they were assigned to. One example is randomizing children to receive a particular preschool curriculum, but parents switch their children to a different center.
Sometimes a treatment impacts the outcome in an indirect manner, via a third variable, called a mediator. Mediation models posit that the treatment/exposure T variable causes change in the mediator M, which, in turn, causes change in the outcome Y.
Building on the mediation model (see above), causal mediation is used to adjust for the fact that levels of the mediator M are not randomized. This approach allows for estimation of the causal effect of treatment/exposure T on the outcome that transmits through mediator M.
A variable that interacts with the effect of the treatment/exposure T on the outcome Y, or with the effect of the mediator M on the outcome Y. If different groups within the study sample experience different levels of the moderator, the outcomes for the groups will be different.
The aim of causal inference research is to identify the impact of exposure to a particular treatment or program. Much of the Methodology Center's work on causal inference focuses on using propensity scores to determine causality in observational studies. This work allows scientists to evaluate health interventions more accurately and will lead to more effective and efficient treatment and prevention of health and social problems.
Introductory Example: How an HIV-Prevention Intervention Works
The Reducing Risky Relationships HIV (RRR-HIV) intervention was designed to decrease incorrect and dangerous thoughts about relationships in order to reduce risky sexual behavior among women being released from prison. Results showed that women who participated in the program engaged in less unprotected sex.
For future interventions, it is important to determine why this happened. Was it because, as hypothesized, the intervention changed their beliefs about relationships? Was it because the intervention reduced their substance use? Or was it because they bonded with other participants in the intervention? With multiple factors at work, determining the mechanism through which the intervention reduced unprotected sex can be difficult. Using a causal mediation model, scientists can answer these types of questions.