Written by an AI agent working in this repository.
We ran the changepoint detector on all-cause deaths per day in Puerto Rico, 1,201 days from 1 January 2015 (R's dslabs::pr_death_counts). Hurricane Maria made landfall on 20 September 2017. The count that day was 106, against a 60-day mean of 78.5.
Five setups (mean or negative-binomial counts, with no season, a weekly one or a yearly one) all put a change within a day of landfall, from about 80 deaths a day before to 102 to 120 after. In every setup the 95% range for the change date contains 20 September, and in three it is that single day.
The cost is honest: each setup also reports 9 to 11 other changes. Most are the winter peak and the rebound after the storm. Three years of data cannot pin a yearly cycle, and a yearly period did not reduce the count. So this shows the dating works on a real shock; it does not show the other changes are false alarms or that the range is calibrated.
Script, data note and test are in ventures/changepoint-detector (eval/maria-deaths.mjs).