Exam MAS-I Practice Question 160
Study guide: Exam MAS-I
An insurer wants to classify claims as fraudulent or legitimate from p = 30 quantitative claim characteristics. The training data hold 60 confirmed fraud claims and 2,000 legitimate claims. An analyst proposes quadratic discriminant analysis (QDA). The insurer's criterion is the lowest expected error rate on new claims.
Which critique of the proposal is most valid?
- A) The fraud class alone needs 465 covariance terms estimated from 60 claims, so QDA's added variance likely outweighs its lower bias
- B) QDA assumes the 30 characteristics are independent within each class, which claim characteristics seldom are in practice
- C) QDA draws a linear boundary between the two classes, so it cannot use the difference in spread between fraud and legitimate claims
- D) QDA ignores the roughly 3% share of fraud in the training data, so it will classify far too many new claims as fraudulent
- E) QDA requires both classes to share one covariance matrix, which is implausible when fraud claims behave so differently
Answer this question and see the worked solution