Auditing a Dataset for Sampling, Measurement and Label Bias Before Relying on It

You are a product manager at Genentech. How would you find out whether the data you are working with is biased? ```hint Biased relative to what? Bias only has meaning against a target. Pin down which population and which question the data is supposed to represent, then follow each step where records were collected, filtered or labeled. ``` ### Clarifying Questions - Will the data support a descriptive analysis or a business decision, or will it train or evaluate a predictive model? - Where does the data come from: a clinical trial, real-world sources such as electronic health records, insurance claims or registries, or usage data from a digital product for patients or physicians? - Which population should the conclusions apply to, for example everyone diagnosed with a condition in the US, or only patients already receiving a particular therapy? - Does the dataset include demographic fields such as age, sex, race and ethnicity, and insurance type, or would representativeness have to be checked indirectly? ### What a Strong Answer Covers - Fixing the target population and the decision the data must support, so bias has a reference point - The main ways bias enters (who is included, which fields go missing, how values are measured, how outcomes are labeled, when data was collected) and where each arises - Concrete checks for each, including error analysis by subgroup when a model is involved - Judging whether a detected skew is large enough to change the decision, and how to respond - Edge cases such as genuine population differences, small subgroups and missing demographic fields ### Follow-up Questions - The patients in your data turn out to be younger and more often commercially insured than the population you care about. What do you do before the readout is due? - A model trained on this data performs well overall but misses noticeably more cases in one subgroup. How do you decide whether it can be used? - How would you catch bias that appears after launch, as the people using the product change?