Why Games Make a Good Statistics Classroom
Competitive online games generate enormous amounts of structured data. Every match records who played, which character or class they chose, what skill bracket they belonged to and whether they won. Community sites publish win rates for every character, patch notes cite balance statistics, and players routinely use those numbers to decide what to play next.
That makes the domain ideal for teaching. The variables are intuitive, the outcomes are binary and the stakes feel real to anyone who has spent an evening losing matches. Most importantly, the data contains a built-in grouping variable, skill level, that behaves exactly like the hidden factors that produce Simpson's paradox in medicine, economics and social science.
The Setup: Two Heroes and a Win Rate Argument
Imagine a team-based arena game with dozens of playable heroes. Teachers can let students pick the genre themselves, since a quick look through any Multiplayer Games catalog shows how many titles now track per-character statistics in some form, from shooters and strategy games to collectible battlers and online RPGs. The details of the game matter less than the structure of the data. Two heroes, which we will call Hero A and Hero B, are the subject of a heated forum debate. A popular statistics site reports that Hero B wins far more often than Hero A across all ranked matches, and many players conclude that Hero A is simply weaker and should be buffed.
Here is a simplified dataset of 1,000 matches. In the lower skill bracket, Hero A played 400 matches and won 180, a 45 percent win rate. Hero B played only 100 matches in that bracket and won 40, a 40 percent win rate. In the upper skill bracket, Hero A played 100 matches and won 65, a 65 percent win rate, while Hero B played 400 matches and won 248, a 62 percent win rate.
Where the Paradox Appears
Look at each bracket separately and Hero A wins more often in both. Yet combine the brackets and the picture flips. Hero A won 245 of 500 matches, or 49 percent. Hero B won 288 of 500, or 57.6 percent. The aggregate table says Hero B is clearly stronger. The bracket tables say Hero A is better everywhere.
| Bracket | Hero A matches | Hero A wins | Hero A win rate | Hero B matches | Hero B wins | Hero B win rate |
|---|---|---|---|---|---|---|
| Lower skill | 400 | 180 | 45% | 100 | 40 | 40% |
| Upper skill | 100 | 65 | 65% | 400 | 248 | 62% |
| Combined | 500 | 245 | 49% | 500 | 288 | 57.6% |
Both statements are arithmetically correct, and that is precisely what makes the example useful. The Stanford Encyclopedia of Philosophy notes that the paradox is unproblematic mathematically yet still strikes many people as surprising, and gaming data recreates that surprise perfectly. Students can verify every number by hand, and nothing is hidden except the distribution of matches across brackets.
The Hidden Variable Is Skill
The reversal happens because the two heroes are played by very different populations. Hero A is popular with newer players, so most of its matches take place in the lower bracket, where everyone loses more often. Hero B is favoured by experienced players, so most of its matches happen in the upper bracket, where win rates are naturally higher. The overall win rate is really measuring who plays each hero, not how strong each hero is.
Statisticians call skill a confounding or lurking variable here. It influences both the choice of hero and the chance of winning. Once it is accounted for, the comparison becomes fair, and the original conclusion collapses. Students who want to practise this step can learn to split tables properly with three-way cross tabulations, which are the natural tool for catching this kind of distortion.
Connecting the Game to the Classic Case
Once the gaming version is clear, the famous examples become much easier to understand. The best known is the 1973 graduate admissions data from the University of California, Berkeley, where overall figures appeared to show bias against women. When the data was examined department by department, the apparent bias largely disappeared, because women had applied more often to departments with low admission rates for everyone.
Course notes from the University of Pennsylvania summarise the lesson neatly, observing that the choice of major had confounded the effect attributed to sex. Replace "department" with "skill bracket" and "applicants" with "players," and the structure is identical to the hero win rate debate.
Turning the Example into a Classroom Exercise
The gaming framing works best as an active exercise rather than a lecture. A simple sequence works well:
- Show students only the overall win rates and ask which hero should be buffed.
- Reveal the bracket-level table and ask them to recalculate the win rates themselves.
- Ask them to explain the reversal in their own words before introducing the term confounding variable.
- Finish by asking what additional variables, such as map choice or team composition, might create further reversals.
Students who play online games often bring their own examples to the discussion, such as characters that look weak in public statistics but dominate in high-level play. That personal connection tends to make the concept stick far longer than abstract medical tables.
The Leaderboard Lesson Players Already Know
Experienced players often say that win rates without context are meaningless, and Simpson's paradox is the formal version of that instinct. A single number that pools very different groups can tell a story that no individual group supports.
The broader lesson reaches well beyond games. Whenever a dataset combines groups that differ in size and composition, from hospital outcomes to marketing channels, the first question should be whether a hidden variable is shaping the totals. Multiplayer match data simply makes that question vivid, relatable and easy to test with numbers students can check themselves.