By now you have likely seen a wide array of Hurricane Maria death estimates in Puerto Rico. The official toll stands at 64. With other private estimates ranging near 1000, 2000, and now Harvard researchers show >4000 in the New England Journal of Medicine (NEJM). Isn't statistics a rigorous science, and how can these estimates be so far apart on a simple and binary question of death? We explore the basic answers to this below.
The key here is that the death surveys vary from hurricane to hurricane. The conditions in Puerto Rico are so dissimilar from those on the mainland, that additional factors made this storm difficult to contrast on an apples-to-apples basis. Add political forces from all sides. And you have a recipe for confusion. And further exasperation.
We have very different death results, all packaged as hard statistics. All of them compiled somewhat differently, and the extreme estimates are most certainly wrong. Even if we assume a super-wide 50% confidence interval! This won't sit well with anyone.
64 is too low, and there were no mechanisms for full tabulation in the government-endorsed version. The rules that were employed here, which of course can not be replicated later, by definition have to force a severe under-reporting bias. On the other hand, the NEJM goes too far in the other direction with their >4600. Instead of following similar direct measures to answer the question, they rely more on ex-post surveys, assumptions about accuracy of historical comparisons, and gross attribution (e.g., errors in boricua leaves, seasonality, etc.) The mortality rates were then customized on the fly, and rounded-higher to be overly generous.
We wind up with a very disingenuous results that can not even be considered in the margin-of-error of the correct result. So ask yourself what you would do in this case that follows.
Say hurricane deaths are usually within one month of landfall. You have a Puerto Rican with a terminal disease in September 2017, also with a 1-month expectancy. But by November the individual is alive, but flies to Miami to receive much better medical treatment. Then in December, without additional information, how would you categorize this person? Assumed to be dead from Hurricane Maria?
On social media polls, most correctly say no.
We have very different death results, all packaged as hard statistics. All of them compiled somewhat differently, and the extreme estimates are most certainly wrong. Even if we assume a super-wide 50% confidence interval! This won't sit well with anyone.
64 is too low, and there were no mechanisms for full tabulation in the government-endorsed version. The rules that were employed here, which of course can not be replicated later, by definition have to force a severe under-reporting bias. On the other hand, the NEJM goes too far in the other direction with their >4600. Instead of following similar direct measures to answer the question, they rely more on ex-post surveys, assumptions about accuracy of historical comparisons, and gross attribution (e.g., errors in boricua leaves, seasonality, etc.) The mortality rates were then customized on the fly, and rounded-higher to be overly generous.
We wind up with a very disingenuous results that can not even be considered in the margin-of-error of the correct result. So ask yourself what you would do in this case that follows.
Say hurricane deaths are usually within one month of landfall. You have a Puerto Rican with a terminal disease in September 2017, also with a 1-month expectancy. But by November the individual is alive, but flies to Miami to receive much better medical treatment. Then in December, without additional information, how would you categorize this person? Assumed to be dead from Hurricane Maria?
On social media polls, most correctly say no.
Say hurricane deaths are usually wi a month of landfall. You've a Puerto Rican w a terminal disease in Sept-2017, also w 1-month expectancy. By Nov the individual is alive & flies to Miami for better care. In Dec wo further info, should we consider it death from Hurricane Maria?— Statistical Ideas (@salilstatistics) June 5, 2018
But it is by going to these immoderations that we can alter the definitions of what it means to have just a storm death. The truest estimate remains closer to a thousand or so, and that should be jump-off point for domestic policy discussion.
No comments:
Post a Comment