A random sample of a state's adults finds that those who eat breakfast daily have lower average BMI. What can be concluded?
From association to causation
The three explanations for an observed association, observational studies versus randomized experiments, and what random sampling and random assignment each let you claim.
Key ideas
- Three explanations
An observed association can be (1) random chance (handled by a test or interval), (2) a causal relationship, or (3) the result of a confounder. Telling (2) from (3) depends on how the data were collected.
- Observational study
Researchers record variables as they occur. Shows associations; always vulnerable to confounding, even after adjusting for some variables.
- Randomized experiment
Researchers randomly assign the explanatory variable. Random assignment balances confounders between groups, so a difference can be attributed to the treatment.
- Random sampling
Every population member has a chance to be selected; results generalize to that population.
- Controlling one confounder
Accounting for one variable controls that variable, not all confounders. Only random assignment handles them all.
What a design allows (adapted from the Lec 14 diagram)
| Random assignment | No random assignment | |
|---|---|---|
| Random sample | Causal conclusions that generalize to the population | Associations that generalize to the population; no causal claim |
| Not a random sample | Causal conclusions for the units in the study only | Associations among the units studied only |
Four designs for one question
Does a mindfulness app lower stress scores? Classify what each design supports. (A) A random sample of a university's students; each chooses whether to use the app. (B) 200 volunteers randomly assigned to the app or a control. (C) A random sample of the university's students, randomly assigned. (D) App users who answered an online poll.
A: random sampling, no assignment: an association that generalizes to the university's students, not a causal effect. Students who choose the app may differ in other ways.
B: random assignment of volunteers: a causal conclusion for these 200 volunteers (and people like them), but not a generalization to all students.
C: both: a causal conclusion that generalizes to the university's students.
D: neither: an association among self-selected respondents only.
Only C supports "the app lowers stress for this university's students." B supports causation for the volunteers; A supports a generalizable association.
Check yourself
Sixty volunteers are randomly assigned to two keyboard layouts and their typing errors are recorded. The layouts produce strong evidence of a difference in mean errors. What is justified?
Select every statement that is true about random assignment.
An observational study finds an association and then shows it persists after accounting for income. A reporter writes: "Since income was controlled, the effect is causal." What is wrong?
Which explanation for an observed association does a hypothesis test address directly?
Which design feature is what lets a study's results be extended to a larger population?