Definition
The target construct is the abstract, often unobservable concept of interest (illness, intelligence, crime). The measurement procedure produces the numbers we store (costs, IQ score, arrests).
Formula
Machine learning does not learn the target construct, it learns the measurement.
- Examples: health care costs as a proxy for illness (racial bias, Obermeyer et al. 2019); 5-year survival as a proxy for life-years gained (UK liver transplants disadvantage the young); arrests as a proxy for crime (sampling bias).
- The mismatch is most harmful for the output variable. More data removes unbiased noise, not bias.
- Good practice: question the measurement, check who is missing, write datasheets (Gebru et al. 2018).
Appears in
- Lecture 7, construct vs. measurement
- Lecture 7, health insurance and liver transplants
- Lecture 7, harms in data
- Lecture 7, good data practices
- Lecture 9.1, sources of unfairness: sampling bias, features, target variable