Spotting Bias in Data Collection

A sample can be large and still be wrong. Bias means your data collection pushes results in a consistent direction, and collecting more of the same flawed data only makes you more confident in the wrong answer.

• Selection bias: the sample leaves out part of the population. Surveying only people who visit a store’s website misses those who never go online.
• Survivorship bias: you only see what made it through a filter. Studying successful startups tells you little if you never recorded the failed ones.
• Response bias: people answer in ways that aren’t quite true, perhaps to look good or because the question was worded to nudge them.
• Non-response bias: the people who reply differ from those who stay silent.

A quick example of non-response bias

A company emails 1,000 customers. Half are satisfied, and half are unhappy, but satisfied customers reply 40% of the time while unhappy ones reply only 10% of the time.

• Satisfied replies: 500 × 0.40 = 200
• Unhappy replies: 500 × 0.10 = 50
• Total replies: 250, so the response rate is 250 / 1,000 = 25%

Among the replies, 200 out of 250 are satisfied, so the survey reports 80% satisfaction. The true figure is 50%. Nothing was miscalculated, yet the conclusion is badly off.

Where bias harms model performance

A model learns whatever pattern its training data contains, including the distortions. If a hiring dataset mostly contains past hires from one background, the model will treat that background as the signal for success.

Before touching any algorithm, check how the data was gathered; it's the cheapest way to prevent any major technical issues.