CLIN-06 · Proxy variables

A model that never sees race, income, or insurance status can still act on all three. In a public dataset, identify one variable that plausibly proxies for race, socioeconomic status, or access to care — then show, numerically, how it reaches the predictions. What is the model actually learning?

A public dataset with documented demographic fields and at least one candidate proxy — utilization counts, prior cost, distance to facility, insurance type, ZIP-derived measures, referral source. Read the canonical case for method before starting: Obermeyer et al., "Dissecting racial bias in an algorithm used to manage the health of populations," Science, 2019, where cost stood in for need (https://www.science.org/doi/10.1126/science.aax2342). Read it for how they built the argument, not for a number to reuse.

Correlation announced as leakage with no path shown from variable to prediction. Dropping the proxy, watching a metric barely move, and declaring the problem solved — the signal usually reappears through the remaining features. Auditing a subgroup the dataset never adequately sampled. Calling a proxy discriminatory when the underlying difference is a real clinical difference the model should use — that distinction is the hard part and must be argued, not asserted.

A proxy trace: the candidate variable and why it is suspected, with a stated mechanism; its association with the demographic attribute, reported with sample sizes; its influence on predictions shown by a simple, honest method (stratified performance, with-and-without comparison, or single-feature contribution) with that method's limits named; a re-check for whether the signal persists after removal; and a paragraph separating what the evidence supports from what it does not.

CT: distinguishing proxy encoding from legitimate clinical signal. CR: choosing a candidate worth tracing. CO: a peer argues the benign interpretation and the student answers it in writing. CM: explaining leakage to someone who does not model. CZ: who is systematically scored wrong, and in which direction.

A clinician, health-equity researcher, or informaticist reviews the proxy hypothesis and the strength of the causal language. Standing instruction: reject causal wording the design cannot support. School-supervised.

R1: one dataset, one candidate proxy, one named mechanism. R2: every association reproducible with sample sizes shown. R3: comparator is the model without the proxy. R4: thin cells and residual leakage both reported. R5: leakage explained in plain language. R6: refuses causal claims and refuses to audit subgroups the data cannot support.

(a) A correlation heatmap presented as proof of discrimination. (b) The proxy is dropped, the metric holds, and the student reports the model as fixed while the signal routes around the deletion untouched.

Lane A immediately (health-science capstone or AP Research). Strong pairing with CLIN-04 for a student who wants the audit and the mechanism. No verified credit claim.