Bias Detection CRAGE Practice Questions (Page 1)

Questions 1–5

1. 1expert · hard

   A model predicts whether a patient will miss an appointment. The team computes demographic parity and equalized odds. Demographic parity is satisfied, but equalized odds is not. Which conclusion is most defensible?

   • AThe model is unbiased because it predicts the same proportion of positives for each group
   • BThe model is biased against the group with the lower false-positive rate
   • CThe model may still be unfair because error rates differ across groups even though selection rates are similar
   • DThe model is fair because demographic parity is the gold standard
   Reveal Solution Discussion Select an answer first
2. 2application · medium

   An AI team is evaluating a hiring model for bias across gender groups. They compute the demographic parity metric and find that the selection rate for female candidates is 0.4, while for male candidates it is 0.6. Which interpretation is most accurate?

   • AThe model satisfies equalized odds because the true positive rates are equal.
   • BThe model is unbiased because the selection rates are within a 20% margin.
   • CThe model satisfies demographic parity because the difference is less than 0.5.
   • DThe model does not satisfy demographic parity because selection rates differ across groups.
   Reveal Solution Discussion Select an answer first
4. 3expert · hard

   A government agency is developing a model to allocate social services. The dataset includes demographic information, but the agency is concerned about bias. The data was collected from multiple sources: some sources used self-reported income, others used tax records. The model shows a disparity in service allocation across ethnic groups. The team must determine whether the disparity is due to measurement bias or selection bias. Which analysis would best distinguish between the two?

   • AIncrease the model's complexity to better capture income patterns.
   • BRetrain the model without the income feature and observe the disparity.
   • CCompare the distribution of income values between self-reported and tax-record sources for the same individuals.
   • DUse a different fairness metric to see if the disparity disappears.
   Reveal Solution Discussion Select an answer first
5. 4foundation · easy

   Which of the following is an example of a data-related source of bias in the AI lifecycle?

   • AA developer who unconsciously favors one algorithm over another
   • BA training dataset that underrepresents certain demographic groups
   • CA model that uses a proxy variable that correlates with a protected attribute
   • DAn algorithm that optimizes for accuracy over fairness
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   A hospital is building a model to predict patient readmission risk. The dataset was collected from a single urban hospital that primarily serves a young, tech-savvy population. The model performs well in that hospital but poorly when tested on data from rural hospitals with older patients. Which source of bias is most directly indicated?

   • AMeasurement bias due to inconsistent recording of readmission outcomes.
   • BConfirmation bias in the model's feature selection process.
   • CSelection bias due to non-representative sampling of the patient population.
   • DAlgorithmic bias caused by the model's optimization function.
   Reveal Solution Discussion Select an answer first