How Deadly Is Florida’s 2021 Summer Wave?
Although the COVID-19 situation in Florida has dramatically improved over the past month or so – despite still being among the worst states in terms hospitalization rates (44th as of September 21) – the toll of the deadly summer wave will not be exactly known for many more weeks after all the data has been processed and reported. Although the picture is far from complete, there is enough information to give fairly accurate estimates of how many people have died. Several methods of estimation will be used and combined to give a rough, final estimate.
1. Reported Deaths
First, let’s start by using reported deaths. As you may recall, there is a renewed ‘controversy’ surrounding Florida’s death reporting. In Figure 1, we can see that deaths reported and deaths by date of death are highly correlated, generally tracking very closely over time. In the two previous waves, the curves of reported deaths have had lower peeks with wider spread than actual deaths, though they have a similar shape. Although not quite symmetric, we can use the wave’s peak as an estimate of its entire shape. Assuming the death reporting process has not changed dramatically for the most recent wave, it will likely be similar to previous ones.
In Figure 2, we see the 7-day average of reported deaths (along with a smooth line) with delineations showing each of three large waves (‘1’ being Summer 2020, ‘2’ being Winter 2020, and ‘3’ being Summer 2021). The delineations were generally determined by the derivative of the smoothed 7-day average and when the curve returned to the pre-wave level. That is, a wave begins when the rate of change in deaths reported becomes positive, ending when the rate of change in deaths reported becomes negative (some small bumps within the larger wave were ignored). The period between waves 2 and 3 was not assigned to wave 2 because deaths had already returned to the ‘baseline’ level by late April. As can be seen in Figure 1, these time periods match almost exactly with the actual death curve.
For the previous waves 1 and 2, there were 14,306 and 17,359 total deaths reported, respectively. In wave 3, there have already been 13,483 deaths reported with no definitive peak. Naively assuming a symmetric curve that is currently at its peak, there could be an additional 13,000 deaths reported for this wave, putting the cumulative total at more than 66,000 deaths. This could even be an optimistic estimate as the curve of deaths reported is likely skewed.
The percent of total deaths within the wave at peak reporting (i.e. cumulative deaths reported less than equal to the curve’s mode) was 34% and 48% for waves 1 and 2, respectively. It is conceivable that Summer 2021, which dwarves the wave in Summer 2020, could have caused even greater backlogs in reporting, making a curve that extends even further out. Let’s assume that wave 3 is currently at its peak and that deaths reported thus far are an average of the previous waves’ peak to total ratio, or 41%. This would result in a total of 32,900 (or 13,483/0.41) deaths for wave 3, of which 13,483 have been reported, making an additional 19,500 deaths and over 72,000 overall.
These estimates make plenty of assumptions and, admittedly, draw conclusions from just two ‘observations’ – the previous waves. However, if the current curve in Florida follows nearly every jurisdiction’s pattern in death reporting, this wave of reported deaths is, at best, only about halfway through – and it’s a long way down.
2. Hospital-Reported Deaths Only
As you may know, Florida hospitals report covid deaths to HHS daily (you can download the JSON file directly here). Not only is this a very strong predictor of the overall deaths on a given day, but it can also be used as a minimum value of the total deaths (although there are a handful of singular days where hospital-reported deaths exceed total deaths reported).
In this method, we simply take the difference between the number of deaths that hospitals have reported and the total deaths that have been reported. This can be visualized in Figure 3 with the 7-day average of each. Obviously, the exact number of hospital deaths is only known through the current day (as of September 23 for this writing) so some extrapolation is required to complete the curve. All that was done (which may be conservative) is to reverse the prior, increasing component beginning at the day where hospital deaths are less than the most recent known day. The extrapolation extends for six weeks (42 days) and ends at the baselines, pre-wave level.
Using this method, the minimum estimate of additional deaths is about 3,150, for a total of over 56,000 deaths. This could be considered an absolute minimum, 99.999% level of confidence – despite some guesswork over the next six weeks. Not only is the extrapolated portion likely an underestimate, it also assumes no covid deaths outside of hospitals which has never happened (when averaged over 7 days, removing individual daily anomalies mentioned previously). In fact, the ratio of total deaths to hospital deaths is rather large, at about 1.4. For the most recent 50-day period (6/20/2021 to 8/9/2021) of ‘complete’ death reporting (so, excluding the most recent 45 days when death reporting is assumed to be incomplete), daily total deaths reported were 40% greater than hospital deaths. It is possible that this 40% figure will decrease, but extremely unlikely to approach anywhere near zero.
There is also the possibility that it may exceed this 40% estimate if the threshold for hospitalization increased due to limited hospital capacity. Remember, Florida led the country in hospitalization rates (the most accurate non-hospital death predictor of total deaths) for much of the summer, with rates higher than at any point in the pandemic for any state since the New York/New Jersey initial wave in Spring 2020. It would not be surprising for more severe cases to go without inpatient-level of care that may have been admitted in previous waves.
3. Hospital Deaths Model
Recognizing that the number of hospital-reported deaths is highly correlated with the total deaths, why not build a simple statistical model? A general linear model with Poisson distribution using data from November 20, 2020 through August 9, 2021 (which excludes the past 45 days) was fit. In addition to the 7-day average of hospital deaths, the rate of change of 7-day average hospital deaths along with the 7-day average of proportion of total population fully vaccinated were also included. The fitted model is:
Total Deaths =exp[1.16 + 0.98 (log(7-Day Average Hospital Deaths))– 0.55 (Rate of Change 7-Day Average Hospital Deaths) – 0.27 (7-Day Average Proportion Fully Vaccinated)]
For example, the peak 7-day average of hospital deaths was 288 on 9/3/2021, with the proportion of fully vaccinated of 0.5406 and rate of change of hospital deaths of 1.003 (found by taking average on 9/3 divided by average on 9/2). Plugging these values into the equation gives a value of 400 expected total deaths on 9/3/2021 (for reference, only 303 deaths have been reported for 9/3 as of 9/23). Using this model, September 3 is expected to be the peak of deaths at about 400. For dates beyond September 23, the linear trend of the last seven days of predicted total deaths was extrapolated out 42 days (values were truncated at 20, simulating a return to the pre-wave level of average deaths).
As in the previous section, we can take the difference between what is expected and already reported (the difference between the red and black lines in Figure 4). Using this method yields an estimated 8,100 additional deaths to occur in this wave, making the total about 61,200 by November (not necessarily reported by then, but having occurred).
4. Method Averages
This table shows each of the methods discussed, some with variations. The range of estimates for total deaths from all methods is 56,251 to 72,588 with an average of about 62,800, or 9700 additional deaths to be reported.
|Method||Additional Deaths||Total Deaths|
|Hospital Deaths Only (Section 2)||3146||56251|
|Hospital Deaths with 1.2 Total to Hospital Deaths Ratio (Section 2)||4288||57393|
|Hospital Deaths Model (Section 3)||8125||61230|
|Reported Deaths Curve with 50% Reported (Section 1)||13483||66567|
|Reported Deaths Curve with 40% Reported (Section 1)||19483||72588|