Data in Our Lives: Collection, Use and Consequences

Data is information that has been recorded, usually as numbers, words, images or locations. Big data is the name for data sets so large, so fast-growing or so mixed in type that ordinary tools like a single spreadsheet cannot handle them. People often describe big data with three words that start with V: volume (a huge amount), velocity (it arrives fast, often every second) and variety (numbers, text, photos, GPS points and more, all mixed together). A quick calculation shows how fast it adds up. A smartwatch that records your heart rate once every second makes $60\times 60\times 24=86400$ readings in one day, for one person. Multiply that by millions of users and you have big data.

Personal information is any information about an identifiable person: your name and address, but also your location history, your purchases or a photo of your face. Consent means you have agreed, knowing what you are agreeing to. In Canada, private businesses must follow the federal Personal Information Protection and Electronic Documents Act (PIPEDA, passed by Parliament in 2000), which requires organizations to get meaningful consent and to collect only the information they need for the purpose they state. A real Canadian case shows why this matters. In June 2022, the Office of the Privacy Commissioner of Canada, together with the privacy commissioners of Quebec, British Columbia and Alberta, reported that the Tim Hortons ordering app had recorded users' locations every few minutes, even when the app was not open, and found that this broke Canadian privacy law. The investigators also found that the company had collected this location data for targeted advertising but never actually used it for that purpose: the data was collected and stored anyway.

A population is the whole group you want to know about. A sample is the part of it you actually collect data from. A sample is biased when it leans in one direction because of how it was chosen or measured, so its results do not match the population. The most famous example is a 1936 poll by the American magazine Literary Digest. It mailed about 10 million ballot cards, got about 2.4 million back, and predicted that Alf Landon would easily beat Franklin Roosevelt for president. Roosevelt won with about 61% of the popular vote. The names came largely from telephone directories, car registrations and the magazine's own subscriber lists, which in 1936 over-represented wealthier people, and the people who chose to mail a card back were not typical either; historians still debate which problem mattered more. The lesson: a huge sample does not fix a biased one.

Bias also enters when computers learn from old data. In 2018 the news agency Reuters reported that Amazon had abandoned an experimental tool for rating job applicants after finding that it downgraded résumés containing the word "women's" (as in "women's chess club captain"). The tool had learned from about ten years of past applications, most of which came from men, so it copied the pattern in the old data. A computer that learns from biased data can repeat that bias faster and on a larger scale than a person could.

• Sampling bias: the way the sample is chosen leaves some groups out (surveying only people at a gym about how often people exercise).
• Voluntary response bias: people choose whether to answer, so those with strong opinions are over-represented (an online "rate this movie" poll).
• Non-response bias: many people who were asked do not reply, and they differ from those who do.
• Measurement or wording bias: the question or tool pushes answers one way ("Don't you agree that homework is too long?").
• Historical bias: past data reflects past unfairness, and a model trained on it repeats that unfairness.

Sampling Methods and Bias with Surveys: Crash Course Statistics #10

A graph can show true numbers and still give a false impression. The most common trick is a truncated axis: the vertical axis starts above zero, which makes small differences look huge. Suppose (invented data) 52% of people in a survey prefer Brand A and 48% prefer Brand B. If the bar graph's axis starts at 46%, Brand A's bar is $52-46=6$ units tall and Brand B's is $48-46=2$ units tall. The picture says "A is 3 times as popular," but the real ratio is $52÷48\approx 1.08$, only about 8% more. Other tricks include pictures that grow in two dimensions when only one number grew, time windows chosen to hide the overall trend, and missing labels or units.

How to spot a misleading graph - Lea Gaslowitz