Cleaning Data and Bias in Collection

A hundred thousand survey answers sound convincing, but if every answer came from the same narrow slice of people, the dataset describes that slice, not the population you meant to study. Size tells you nothing about who was left out. That is the core trap in this topic: a dataset can be huge, clean, and perfectly organized, and still be wrong about the world because of how it was collected.

Cleaning data is a separate job from fixing bias, and it happens first. Real-world data is messy: the same city typed as NYC, New York, new york, and N.Y. are four strings but one fact. Cleaning picks a single standard form and rewrites every variant to match it, so a computer counting by city doesn't split one place into four fake ones. Done right, cleaning never changes what an entry means, only how it's written, so the underlying information is preserved while becoming usable for sorting, grouping, and counting.

Bias is a different problem, and it lives upstream of cleaning. It comes from the sampling: which people, places, or conditions got recorded at all. A fitness app that only hears from owners of one expensive smartwatch is missing everyone else, and doubling its data just means more of that same narrow group, not a broader one. The same logic applies to sensors: if a thermometer fails on hot days, every hot day is underrepresented in the data, so conclusions drawn from it will quietly understate heat no matter how carefully the recorded numbers are cleaned.