When priors and costs exactly cancel

Question

A screen reads X. Under "clean" (H0), X \sim 	ext{Normal}(0, 1) with prior \pi0 = 0.8. Under "flagged" (H1), X \sim 	ext{Normal}(2, 1) with prior \pi1 = 0.2. A false alarm costs 1 unit; a miss costs 4 units. You declare "flagged" when X c. Find the cost-minimizing cutoff c.

Accepted Answer

The optimal threshold. For two equal-variance normals, c = \frac{\mu0 + \mu1}{2} + \frac{\sigma^2}{\mu1 - \mu0}\,\ln\!\left(\frac{\pi0\, cI}{\pi1\, c{II}}ight). Plug in. (\mu0+\mu1)/2 = 1, \sigma^2/(\mu1-\mu0) = 1/2, and \frac{\pi0 cI}{\pi1 c{II}} = \frac{0.8 	imes 1}{0.2 	imes 4} = \frac{0.8}{0.8} = 1. Since \ln(1) = 0, the shift vanishes: c = 1 + 	frac{1}{2}	imes 0 = 1.0. Read it. The prior favors "clean" 4-to-1, which alone would push the cutoff up, but a miss costs 4 times a false alarm, which alone would pull it down. Here the two tilts are exactly equal and opposite, so the optima…

When priors and costs exactly cancel

Your answer

Discussion

Learn the concepts

Related questions