Observational Studies vs. Experiments
Distinguish observational studies (no treatment assigned) from experiments (treatments randomly assigned). Understand why only a properly designed experiment can establish causation, and why observational studies can only show association.
A lesson of AP Statistics, in Collecting Data, about 10 minutes.
By the end of this lesson, you will be able to distinguish observational studies from experiments, explain why experiments can establish causation while observational studies cannot, and identify confounding variables.
Two Ways to Collect Data
Observational study: researchers observe and record data without imposing any treatment. Subjects self-select into groups or are classified by existing characteristics.
Experiment: researchers deliberately impose a treatment on subjects and observe the response. The experimenter controls who gets what.
The Critical Difference: Causation
Only a randomized experiment can provide evidence of a cause-and-effect relationship.
Observational studies can show association (correlation) but cannot establish causation. The reason: confounding variables.
Confounding variable: a variable that is associated with both the explanatory variable and the response variable, making it impossible to determine which is causing the change in the response.
Example: An observational study finds that people who sleep fewer hours have higher rates of heart disease. Does sleep deprivation cause heart disease?
Not necessarily — stress might be a confounding variable: stressed people sleep less AND stressed people have more heart disease. Stress, not sleep loss, might be the true cause.
In an experiment, we could randomly assign some people to sleep fewer hours (though this raises ethical issues). Random assignment distributes confounders equally across groups, isolating the effect of the treatment.
Why Random Assignment Works
When subjects are randomly assigned to treatment groups:
- Known confounders are distributed approximately equally across groups
- Unknown and unmeasured confounders are also distributed approximately equally
- Any difference in outcomes can be attributed to the treatment, not to lurking variables
This is why randomization is such a powerful tool — it handles confounders we haven't even thought of.
Census vs. Sample
Census: data collected from every member of the population. Accurate but often impractical (expensive, time-consuming).