Every dataset is a picture of whoever got counted. Today: whose rows are missing, why, and what that does to every answer you compute — no matter how correct your code is.
No reading — this page is the reading. Runestone doesn’t cover sampling bias; it comes back in Unit 8 with the AI chapters.
Big Data, reaching toward Impact. 5.9 asked what data can’t prove. Today asks what data never captured. Both are the bridge to Unit 8.
A sample is the part of a population that actually ends up in your data. A sample is representative when it matches the whole population on the things that matter for your question. When the way data is collected systematically leaves some people or situations out, the data has bias — and every result computed from it inherits that bias.
| Source of bias | Example |
|---|---|
| Who has access | An app-based survey can’t hear from people without the app or a phone. |
| Who volunteers | Online polls hear from people who feel strongly. |
| Where and when you ask | A survey at 2 PM on a weekday misses people at work. |
| What was never recorded | No column for it → no answer about it. Ever. |
| Who survived | Studying only the companies still in business hides the ones that failed. |
Four neighborhoods, same size, and in this simulation the same number of real incidents — 20 each. The only difference: how many residents have the app. Flip between what the app sees and what actually happened.
R = representative by design · B = biased sample.
Five questions. For every dataset: who’s missing?
In April you’ll study AI systems that make decisions — who gets a loan, which résumé gets read, where patrols go. Those systems learn from datasets like the ones in this unit. When the data leaves people out, the system does too. You’re not meeting that idea fresh in April — you’re meeting it today.