Simpson's Paradox is a phenomenon in which a trend or association seen in aggregated (combined) data reverses or disappears when the data are broken into subgroups, usually because a lurking variable is unevenly distributed across the groups.
Simpson's Paradox is what happens when combining data hides the truth. A relationship that holds in every subgroup can flip when you mash all the subgroups together. The classic setup looks like this: Treatment A beats Treatment B for patients with mild symptoms, and Treatment A also beats Treatment B for patients with severe symptoms, yet when you pool everyone, Treatment B looks better overall. That feels impossible until you notice the third variable doing the damage. If Treatment A got mostly severe cases (which have low success rates no matter what) and Treatment B got mostly mild cases, the overall rates get dragged in opposite directions.
In AP Stats terms, the paradox lives in two-way (contingency) tables. The overall percentages are marginal relative frequencies, and the subgroup percentages are conditional relative frequencies. Simpson's Paradox shows up when the conditional story and the marginal story disagree. The fix is always the same idea. Look at the disaggregated data, because the subgroups reveal the real pattern and identify the lurking variable (here, symptom severity) that the combined table buried.
Simpson's Paradox sits in Topic 2.2 (Representing Two Categorical Variables) in Unit 2: Exploring Two-Variable Data, supporting learning objective 2.2.A, comparing numerical and graphical representations for two categorical variables. The CED's essential knowledge builds your toolkit for spotting it. Two-way tables summarize two categorical variables (UNC-1.P.3), and segmented bar graphs or mosaic plots let you compare distributions across groups (UNC-1.P.1, UNC-1.P.2). Simpson's Paradox is the dramatic payoff of those skills. It proves that joint, marginal, and conditional relative frequencies can each tell a different story, so you have to know which one answers the question being asked. It's also your first real encounter with lurking variables, an idea that powers Unit 3's entire argument for why experiments need random assignment.
Keep studying AP® Statistics Unit 2
Correlation Does Not Imply Causation (Unit 2)
Simpson's Paradox is the most extreme proof of this principle. A lurking variable doesn't just weaken an observed association, it can flip the direction entirely. If aggregation can make the worse treatment look better, you can't trust an observed association to tell you what causes what.
Two-Way Tables and Conditional Relative Frequencies (Unit 2)
The paradox is really a statement about conditional versus marginal relative frequencies disagreeing. The overall success rate is a marginal calculation, while the per-subgroup rates are conditional. Simpson's Paradox is what you get when those two calculations point in opposite directions.
Confounding and Random Assignment (Unit 3)
Simpson's Paradox usually appears in observational data where groups weren't comparable to begin with, like severe cases piling into one treatment. Random assignment in an experiment balances lurking variables across groups, which is exactly why experiments can support causal claims and observational studies can't.
Simpson's Paradox shows up almost exclusively in multiple-choice questions built around a two-way table or a two-treatment comparison. The classic stem describes a study where Treatment A has the higher success rate for mild cases and for severe cases, but Treatment B wins overall, then asks you to explain why or to identify which treatment is actually preferable. The right move is to name the lurking variable (like symptom severity), explain that it was unevenly distributed between groups, and trust the disaggregated subgroup data over the combined rates. You may also need to compute conditional relative frequencies from a table to verify the reversal yourself. No released FRQ has used the term verbatim, but the underlying skill of interpreting conditional versus marginal proportions and flagging lurking variables shows up regularly in FRQ responses about observational studies.
These overlap but aren't the same. 'Correlation does not imply causation' is the general warning that an observed association might be explained by a lurking variable instead of a causal link. Simpson's Paradox is a specific, dramatic case where the lurking variable doesn't just create a misleading association, it makes the aggregated trend point the opposite direction from every subgroup. All Simpson's Paradox situations involve a lurking variable, but most lurking-variable problems don't produce a full reversal.
Simpson's Paradox occurs when a trend in combined data reverses or disappears once the data are split into subgroups.
It happens because a lurking variable, like symptom severity, is unevenly distributed between the groups being compared.
When the paradox appears, the disaggregated (subgroup) data give the more trustworthy comparison, not the overall rates.
On the AP exam it appears in two-way table problems, so you need to compute and compare conditional relative frequencies, not just marginal totals.
Simpson's Paradox is strong evidence for why random assignment matters in experiments, since balanced groups can't produce this kind of reversal.
If Treatment A wins in every subgroup but loses overall, the correct conclusion is that Treatment A is better and the overall rates are misleading.
It's when an association in aggregated data reverses or vanishes after you break the data into subgroups. For example, Treatment A can beat Treatment B in both the mild and severe symptom groups, yet B can have the higher overall success rate. It's covered in Topic 2.2 with two-way tables.
No, the numbers are all correct. The combined data are misleading because a lurking variable is distributed unevenly between groups, like one treatment getting mostly the hardest cases. The subgroup comparison reveals the real pattern.
A lurking variable is the cause; Simpson's Paradox is one possible effect. Lots of associations are distorted by lurking variables, but it's only Simpson's Paradox when the distortion is so strong that the aggregated trend points the opposite way from every subgroup.
Trust the subgroup (disaggregated) data. The subgroups control for the lurking variable, so they compare like with like. If Treatment A has the higher success rate for both mild and severe patients, Treatment A is the better treatment regardless of the overall percentages.
It's in Unit 2 (Exploring Two-Variable Data), specifically Topic 2.2 on representing two categorical variables, under learning objective 2.2.A. You'll see it again conceptually in Unit 3 when confounding and random assignment come up.
Connect this key term to the AP exam workflow: review the course, practice questions, and check related study tools.
Review units, study guides, and course resources.
Check this vocabulary in multiple-choice context.
Apply key concepts in written AP responses.
Estimate the exam score you are working toward.
Review the highest-yield facts before practice.
Put the full course together before test day.