Observational Study
Statistics & ProbabilityAn observational study collects data by watching or measuring subjects without any manipulation or intervention.
Definition
An observational study is when researchers watch and record what happens without changing or interfering with anything. They just observe.
Example
Watching how many cars run a red light at an intersection over a week is an observational study. The researchers do not change traffic signals; they just count.
Key Insight
Observational studies can show that two things are related, but they cannot prove that one thing causes the other.
Definition
An observational study is a study in which the researcher observes and records data without assigning treatments or manipulating variables. Unlike controlled experiments, observational studies cannot establish causation because of potential confounding variables.
Example
Researchers find that people who drink coffee live longer on average. This is observational: coffee drinkers might also exercise more or have other healthy habits (confounders) that explain the association.
Key Insight
The key limitation of observational studies is confounding. Association does not imply causation without a controlled experiment or careful statistical adjustment.
Definition
An observational study is one in which treatment assignment is not under the investigator's control. Causal inference from observational data requires assumptions about no unmeasured confounders (ignorability or unconfoundedness). Methods include propensity score matching, instrumental variables, and difference-in-differences.
Example
To estimate the causal effect of education on earnings using observational data, an economist may use proximity to a college as an instrumental variable (IV), exploiting exogenous variation in education not driven by individual characteristics.
Key Insight
The potential outcomes framework (Rubin causal model) formalizes observational study challenges: the fundamental problem of causal inference is that we observe only one potential outcome per unit, never both.