A proxy metric is an observed variable that indirectly encodes information about an attribute or process of interest. It can support modeling when the original variable is latent, unavailable, or intentionally excluded, but its meaning depends on the causal structure around it. section 11.6.4 section 1.3.4
Proxy metrics play two different roles:
- Compress latent confounders into a lower-dimensional latent vector Z, allowing neural decoders to estimate engagement and purchases while relying on front-door information flow. section 11.6.4
- Carry unwanted information about protected attributes into predictions, even after explicit gender or ethnicity fields are removed. section 1.3.4 section 8.1.3
Loan decisions and counterfactual fairness
A loan model may exclude gender yet use a person’s name, neighborhood, social network, or shopping habits as proxies. Counterfactual analysis asks whether the decision would change if the person’s ethnicity changed, while examining corresponding proxy changes. section 1.3.4 section 8.1.3
Latent confounder encoding
A VAE-inspired causal model represents several latent confounders with a one-dimensional Z. Confounders2Engagement maps Z and guild membership to engagement probability; PurchasesNetwork maps Z, won items, and guild membership to purchase parameters. section 11.6.4
