Read on Twitter
Noticed yesterday that @IHME_UW’s #COVID19 projections for mean total deaths in the USA had dropped from 80k-95k in earlier models to just over 60k in their latest update. 
A look at state-level predictions reveals a correlated decline across many of the hardest-hit states. 
A look at state-level predictions reveals a correlated decline across many of the hardest-hit states.
Compared to earlier updates where state numbers would fluctuate, the change from 4/5 to4/7 saw consistent, across-the-board declines across many hard-hit states (NY, NJ, CT, MA, GA, FL, IL). This suggests a universal model parameter (e.g. mortality rate) may have been changed.
@IHME_UW do not address the ~30% decline in mean total deaths in their updates blog but do mention two key changes to the 4/7 model . So it's unclear whether additional data or changes to model may be responsible. http://www.healthdata.org/covid/updates
This is what the estimated mean total deaths looks like when you plot projections across time for different models. The 4/7 update (
) clearly stands out.
) clearly stands out.
As @CT_Bergstrom has pointed out, the @IHME_UW projections likely represent the range of outcomes given a best-case scenario - where strong physical distancing measures are enforced and the first wave is successfully suppressed. https://twitter.com/CT_Bergstrom/status/1244815009303023616
Patchwork implementation/enforcement of physical distancing means that successful suppression may be unlikely. The lowered estimate is even more concerning given the extent to which the IHME model influences policy across the USA. https://www.washingtonpost.com/health/2020/04/06/americas-most-influential-coronavirus-model-just-revised-its-estimates-downward-not-every-model-agrees/
I do applaud @IHME_UW’s efforts to make their latest raw projection data publicly available. That said, I wasn’t able to find a historical archive of their COVID19 predictions on their website. http://www.healthdata.org/covid/data-downloads
Historical @IHME_UW projections were pulled from the
𝘤𝘰𝘮𝘱𝘪𝘭𝘦𝘥_𝘦𝘴𝘵𝘪𝘮𝘢𝘵𝘦𝘴.𝘤𝘴𝘷 filein this github repo https://github.com/mccoates/ihme_covid_ests. I’ve checked the 2020_04_07.04.all numbers from github with the file on IHME’s server and they appear to be consistent.
𝘤𝘰𝘮𝘱𝘪𝘭𝘦𝘥_𝘦𝘴𝘵𝘪𝘮𝘢𝘵𝘦𝘴.𝘤𝘴𝘷 filein this github repo https://github.com/mccoates/ihme_covid_ests. I’ve checked the 2020_04_07.04.all numbers from github with the file on IHME’s server and they appear to be consistent.
I did this all last night so haven’t had a chance to clean up my spaghetti code. Happy to do so and share my notebook publicly if there’s interest. However, should be reasonably easy to replicate these results using the data contained in the github repo linked above.
For the record, I'm not an epidemiologist and have actively resisted attempts at armchair epidemiology during this crisis. All I've shown is a trend analysis of @IHME_UW projections over time. Planning to let the experts take over from here
