Data Analytics: Foundations to Practice · Data Ethics, Privacy, and Bias
How Bias Enters Data and Analysis
Bias in data analysis rarely enters through deliberate malice; it enters through specific, identifiable mechanisms at several distinct stages of the analytics process, and recognizing those stages is the first defense against it.
Bias can enter at the very first stage, data collection, when the process used to gather data systematically over- or under-represents certain groups; a customer survey distributed only through a specific channel, such as email, systematically excludes customers who don't actively engage through that particular channel, producing a dataset that doesn't genuinely represent the full underlying customer base the analysis is actually meant to describe.
- Bias can enter at data collection when the gathering process systematically over- or under-represents certain groups, like a survey distributed through only one channel.
- Historical data reflecting past human decisions can encode existing bias, meaning models trained on it risk learning and perpetuating that same bias unintentionally.
- Even without a directly sensitive attribute, a proxy variable like zip code can correlate closely enough with race to produce similar biased effects.
- Bias can also enter purely through interpretation, when an analyst unconsciously favors a conclusion confirming an existing hypothesis or stakeholder preference.