Hypothesis testing with the Poisson
A Poisson test asks whether an observed count is too extreme for a claimed mean rate. The null model must use the claimed rate, scaled to the full observation interval.
Put the claimed rate in H-zero
- Write and choose , or from the claim.
- Scale the null rate to the total time before calculating.
- The alternative hypothesis chooses the tail.
Under H₀, how surprising is 4 or more?
red tail = P(X ≥ 4) = 0.0383
H₁ chooses the upper tail; observed 4 chooses where it begins.
Key idea
Calculate the exact tail under H-zero
Worked example
Current-paper late-arrival test
Before a timetable change, girls and boys arrive late at mean rates 0.10 and 0.15 per day. In 5 days after the change, 4 students are late. Test at 5% whether the total mean rate has increased.
Since , reject .
Conclusion: There is sufficient evidence, at the 5% level, that the mean number of late students has increased. (9709/62/F/M/24 Q5(d))
Common mistake
For a large null mean, use a corrected normal tail
If the null mean exceeds 15, use with a continuity correction. The variance still comes from , not from the observed count.
Machine faults occur at a mean rate of 1.5 per week. After an overhaul, 28 faults occur in 26 weeks. Use a suitable approximation to test at 5% whether the mean rate has fallen.
Show worked answer
Since , reject .
Conclusion: There is sufficient evidence that the mean weekly fault rate has fallen.
Examiner note
