Historical data records past decisions, not facts

What it is

Machine-learning pipelines treat historical data as "facts that happened," yet most historical data records past decisions: arrest records record where police patrolled and whom they stopped, not who committed crimes; loan denials record the bank's lending policy at the time, not who would truly have defaulted; medical-spending records record who could afford care, not who was sicker. A model trained on decisions as labels learns not a predictor of the target construct but a replay of the historical decision process — faithfully reproducing past policies, biases, and resource allocations, now wearing the objective costume of "data-driven."

Why it happens

The contamination path is construct substitution. The construct the system wants to predict (crime risk, default probability, healthcare need) is unobservable, so an available proxy (arrests, loan outcomes, medical bills) stands in; proxy = construct + systematic distortion of the collection process, and the model cannot separate the two, so it learns both. Worse, decisions have a self-confirming structure: heavily stopped populations accumulate more records, and more records trigger more policing — the decision manufactures subsequent data that "confirms" it. Time deepens the problem: norms and policies shift, and decisions from decades past (a neighborhood once deemed high-risk and denied service) sit in the data while the world has moved on — yet the model aligns to the outdated decision distribution.

Studying it

Construct-validity analysis is the core method: systematically examine "what does this label actually measure," interrogate the theoretical chain from label to target construct, and locate the label's decisive origins (who decided, under which policy). Biased-label analysis compares the proxy against measures closer to the construct — arrest records against self-report crime surveys, medical spending against health inventories — to estimate how much of the label is decision distortion. Natural experiments exploit policy changes: when patrol policy or lending rules change while underlying population behavior does not, the drift in label distribution measures the decision process's contribution. Methodological cautions: cleaner construct measures carry their own biases (self-reports under-report), so conclusions are best framed as estimates of a distortion interval, not point values.

Where it stops holding

Not all historical data is decisions: physical measurements, completed-transaction records, and user-submitted content suffer little decision contamination — "decision-type labels" concentrate in domains involving institutional discretion and resource allocation (law enforcement, finance, healthcare, education). Nor does this render historical data unusable: where decision processes have materially improved and distortion directions are known, debiasing and reweighting can partially correct. The boundary is that correction presupposes admitting the label was decided — and most failures begin with denying exactly that.

Applying it

• Write a "source anatomy" for every label: who produced this number under which rule; is it a measurement (sensor reading) or a decision (human discretion, policy threshold); for decision-type labels, record the deciding authority and the policy then in force.
• For decision-type labels, find a construct-closer comparison measure and run a one-time distortion estimate; above a preset threshold, the label may not be used alone for population-sensitive prediction targets.
• Persist decision provenance into the data: the reasons behind historical decisions (policy versions, discretion basis) stored as metadata so "construct signal" and "decision signal" can later be separated.
• Verify: after launch, track outcome distributions for populations the model's decisions affect — if the model's outputs correlate far more with historical decisions than with realized outcomes, construct substitution remains uncorrected; redesign the label.

Related

• Same group: P4.10.1 Annotator composition and judgment standards are part of the data · P4.10.2 Collection channels decide who never enters the data
• Adjacent: P4.09.3 Feedback loops reinforcing records · Q3 Construct measurement in quantitative methods
• Search terms: construct validity · biased labels · measurement versus decision