Confounding, explained.
Confounding occurs when common causes of exposure and outcome distort the association used to estimate a causal effect.
Why it happens
Background-specific control means are 20 and 60; treatment adds 10 in either group. The selected treated group contains 50–95% advantaged cases while control has the complementary share. Randomisation switches both expected shares to 50%, removing the composition difference.
Confounding occurs when common causes of exposure and outcome distort the association used to estimate a causal effect.
Read the result
Read each background row before the combined row. A changing combined contrast with unchanged within-background differences exposes selection, not a changing treatment response.
A worked example
A course with experienced applicants
At 70% advantaged treated cases, control has 30% advantaged cases; the raw gap is 26 outcome units.
Randomise assignment so each group has an expected 50% advantaged share and compare the combined means again.
The gap falls to ten, matching both within-background gaps. Sixteen units previously came from composition, not treatment.
OPTIONAL DEEPER DETAILGo deeper: inside the model
Inside this model
Background-specific control means are 20 and 60; treatment adds 10 in either group. The selected treated group contains 50–95% advantaged cases while control has the complementary share. Randomisation switches both expected shares to 50%, removing the composition difference.
Where this idea is useful
A practical use
A training programme may attract people with more prior experience; comparing graduates with everyone else can mix selection with learning.
A common misconception
“A large outcome gap proves a large treatment benefit.”
The gap can include common-cause differences between groups. Defensible causal interpretation requires an assignment argument, not just a large contrast.
What this explanation leaves out
- All outcomes are invented expected units. Exact balancing removes sampling noise. Randomisation in real studies does not guarantee perfectly balanced realised samples or eliminate every source of bias.
Can I adjust for every variable I can collect?
Adjustment should follow a causal model. Controlling colliders or post-treatment variables can introduce bias; more controls are not automatically better.
Which background condition could influence both participation and the outcome in your next comparison?
Associated thinkers
Associations marked provisional are awaiting source review.
Further reading
Explore the original research or the teaching reference behind this experiment.