Interactive demo · Statistical Methods for Modern Biology · Chapter 1

Is the spike real? The maximum of Poisson counts

An ELISA array scores 100 positions of a protein in 50 patients. Without any epitope, each position collects false positives at a rate of 1% per patient, so its count is (close to) Poisson with λ = 0.5. One position shows 7 positive calls. The probability that one given position reaches 7 is about 10−6, but the 7 was found by looking at the largest of 100 counts, and the probability that the maximum reaches 7 is about 100 times larger. The top panel shows the data; the bottom panel shows the exact distribution of the maximum (red, log scale) and the frequencies from a Monte Carlo simulation (teal). The table gives the exact answer, its approximations and the Monte Carlo estimate with its uncertainty.

Data

Background rate λ per position

Threshold m (is the maximum ≥ m?)

Positions scanned n

Monte Carlo runs B

Try these

Each Monte Carlo run is one complete experiment under the null: n independent Poisson(λ) counts and their maximum. Because the Poisson quantile function is increasing, the maximum of the n counts is the quantile of the maximum of n uniforms, which has the distribution of U1/n; the page draws that directly, so a million runs take a moment and give exactly the same distribution as simulating every position. The generator is seeded, so everyone sees the same run.

In the lecture notes