Estimating the Early Death Toll of COVID-19 in the US

May 22, 2020

Information

Daniel Weinberger, PhD Associate Professor of Epidemiology (Microbial Diseases)

ID5230

To CiteDCA Citation Guide

00:00I would not let you juice our next Speaker.
00:04Doctor Daniel Weinberger.
00:05Doctor Weinberger joined the faculty at
00:08the Yale School of public health in 2013.
00:11He earned his PhD in Biological Sciences from
00:14the Harvard School of public health in 2009.
00:17His research is at the intersection
00:19of Microbiology and Epidemiology
00:21and focuses on understanding the
00:23biological and epidemiological
00:24drivers of respiratory infections,
00:26including pneumococcus RSVP
00:27influenza and Legionella.
00:28Doctor Weinberger thank you for being here.
00:38Thank you very much. I can share my
00:40screen if it's OK. If you're able to.
01:05OK, thank you very much for the
01:08invitation to share this work with you.
01:10So this is a collaborative project I've
01:13been working on since late March with a
01:16large group that includes plugs from the
01:19NIH in New York City Department of Health,
01:22UMass Amherst, alidade health, and Russ.
01:24Killed the University.
01:26So when we started this project.
01:29We were interested in trying to extend some
01:32work that we've been doing as a group.
01:35I'm trying to estimate the burden of
01:37influenza during previous pandemics
01:39and thought that it would be a
01:41fairly straightforward project.
01:42Didn't anticipate that it would get
01:44a lot of attention in particular,
01:47but it turns out that this has become a.
01:51Sort of radioactively hot topic.
01:53Since we started doing this project.
01:56There's been.
01:58A lot of controversy.
02:02Sort of about the official kovit death tool.
02:06And a lot of questions about the
02:08reliability of sort of the official.
02:11That's that are being reported by ACDC,
02:13which I think makes this work especially
02:16important because we try to sidestep
02:18some of the issues around testing
02:20which had sort of played some of the.
02:23The official numbers said basically
02:25the controversy is been around
02:27sort of weather deaths are being
02:29accurately recorded as due to Cove it,
02:31or whether sort of of.
02:34People are sort of inappropriately
02:36coding deaths as Judah covered
02:38when people are dying.
02:40Of other causes,
02:41and maybe you're infected with so good,
02:43but not necessarily dying due to Cove in.
02:47So we have been working on this
02:49group that I mentioned at the
02:51beginning and have been partnering
02:52with the Washington Post on trying
02:54to try to estimate excess death.
02:56So this the idea here is instead of
02:59looking at the number of deaths that are
03:01actually recorded as being due to Kobe,
03:04just looking at sort of the
03:05changes in total deaths and deaths
03:07due to pneumonia or influenza,
03:09which are less likely to be biased
03:11by some of these coding issues.
03:16So in the United States you know the
03:18cause of death is decided by individuals
03:20by individuals all around the country
03:23who have their own sort of criteria,
03:25and it's typically done by by physician
03:28when the death is due to natural causes
03:31or by metal medical examiner or coroner,
03:33when it's when it's an unattended death.
03:38And the CDC and state helps parents are
03:40recording typically both an underlying cause
03:43of death and a contributing cause of death.
03:45So, for instance, you could have current
03:48virus listed as the underlying cause,
03:50an ammonia listed as a contributing
03:52cause or or heart attack.
03:54Let's just contributing cause.
03:56Or vice versa.
03:57Or you might have one of these things listed,
04:00so you might have pneumonia listed,
04:02but current viruses left off,
04:04or any combination of those possibilities.
04:07So the official death registration in
04:08the US is done at the state level.
04:11Minutes reported up to the CDC
04:13National Center for health statistics.
04:15And there is some like that data.
04:19So we have good reason to think that the
04:22number of reported deaths is an undercount.
04:25This is sort of a typical feature of sort
04:27of pathogen specific deaths in general,
04:30so this is, you know,
04:32been staying in previous influenza pandemics,
04:34where it's it's typically assume
04:35that just some fraction of the
04:37deaths that are due to influenza
04:39during an influenza pandemic or even
04:42during sort of typical seasonal flu,
04:44or actually recorded as such,
04:45so it's much more common to look at
04:48pneumonia and influenza together.
04:50And looking at sort of increases
04:52above a typical seasonal baseline to
04:55try to estimate the full burden of.
04:57Of death and of course this is
04:59particularly important with current virus,
05:01especially early in the epidemic where
05:04the testing was really slow to ramp up,
05:07and in many states testing was
05:09really inadequate at a time when
05:11we think that the virus might have
05:14been circulating at a high level.
05:19Complicating things further.
05:20The way that CDC records the data
05:23and it gets reported from the states
05:26there is optimal lag in the data,
05:29so the data that we're seeing from one
05:31or two or three or even four or five
05:34weeks ago is going to be incomplete
05:36in this varies quite a bit by state,
05:39so this plot is just showing our
05:41estimates for the proportion of
05:42deaths that are reported based on
05:44how far we are from the death,
05:46so we can see sort of study on
05:49the left hand side.
05:50Is starting at 2 weeks after the death.
05:53There's a huge amount of variability.
05:55Some states are only capturing maybe
05:5730% of the deaths or reporting 30%
05:59of the deaths that will eventually
06:01be reported two weeks out.
06:03Other states,
06:04like New York or quite good and we're
06:06getting something like 95% of the
06:08deaths even after just two weeks.
06:11It tends to ramp up fairly rapidly
06:13in most states,
06:14so you know after three or four weeks,
06:16we're getting more than 90% of the deaths.
06:18It will eventually be reported
06:20getting reported on,
06:21but there are some states like
06:22this is Kentucky here,
06:24which seems to have a particularly
06:25slow reporting where even out to
06:27sort of 10 weeks after the death,
06:29we're sort of in the 80% range
06:31of death that are reported.
06:34So our analysis is goals were quite simple.
06:37It was to quantify the excess
06:39burden of death.
06:41Student ammonia,
06:41or influenza,
06:42or do the code 19 to quantify the
06:45excess burden of deaths due to any
06:47cause and then to try to compare
06:50the excess Destin reported deaths.
06:52And we're also hoping to adjust for
06:55reporting delays and variations in
06:58deaths that are related to influenza.
07:00We're using data that are publicly
07:03available that are reported by the
07:05National Center for health statistics.
07:07Were there sort of updating these data
07:10on a daily basis in reporting the
07:12number of deaths in each week that
07:15were due to any cause or due to pneumonia,
07:18influenza,
07:18or current virus as a as a grouping?
07:22We also have some information on.
07:26Testing another source of data on deaths,
07:28which tends to be a little bit
07:30more up to date than the NCHS data
07:32from kobetracking.com,
07:33which is the data source that Ginny
07:35pits are mentioned in the previous stuff.
07:39And essentially, we're doing a fairly
07:42simple regression model where we're
07:45trying to model the number of deaths
07:48that occur in a given state in each week.
07:51We're adjusting for seasonality.
07:53We're adjusting forward influenza
07:55activity during the previous week to take
07:58into account through the lag between.
08:00Who activity in deaths?
08:02We're allowing the baseline to
08:04vary year to year to account
08:07for changes in population size,
08:09and we're adjusting for this reporting delay,
08:12which is estimated separately using a
08:15Bayesian Nowcasting algorithm called knobs,
08:17which was recently described
08:19by Nick Menzies Group.
08:24So we're basically fitting this
08:25regression through data today
08:27through the beginning of February.
08:28So this is a period when we don't think
08:31there was much coronavirus there,
08:33so we're sort of fitting to the data
08:35to try to get sort of a sense for
08:38with the typical seasonal pattern
08:40looks like and then extrapolating
08:42that baseline forward for the
08:44period from February to April.
08:46Then we're generating uncertainty
08:47intervals by resampling scheme that Nick
08:49breaks group is previously developed.
08:51I try to get an estimate for sort
08:53of the uncertainty in those.
08:55Baseline estimates
08:59So this is just sort of a simple top
09:02line picture of the excess or the total
09:05deaths occurring in each state overtime.
09:08So on the upper left corner this is New York,
09:12including New York City, Which.
09:15His, as everyone knows, experienced
09:17the most severe epidemic in the US.
09:19So the typical seasonal baseline
09:21is shown with this red line.
09:23Here the variation in deaths in
09:25previous years is shown in Gray,
09:27which you can't even really see.
09:29Do the baseline for New York there and then.
09:32The observed number of deaths.
09:35Of or 2020 is shown with this black line.
09:37So really what we're looking at
09:39is the difference between the
09:40black line in the red light.
09:42Here for each state,
09:43and so you can see pretty
09:44sizable increases in deaths.
09:46Above the seasonal baseline for New York,
09:48New Jersey, Massachusetts,
09:49District of Columbia,
09:50the data or a bit more noise of it.
09:53You can see this clear increase about it.
09:56Above historical patterns,
09:57Maryland sort of across the board
10:00we see pretty strong increases.
10:03There are a number of states
10:05where we do not see increases and
10:08these tend to be sort of smaller,
10:11more rural states.
10:12Vermont a number of states in the Midwest.
10:16New Hampshire we got Minnesota.
10:17Here are the Minnesota is starting
10:19to increase in more recent weeks,
10:20so I think if we were to extend this
10:22at another couple of weeks we probably
10:24would see something for Minnesota here.
10:30We are also interested in looking at
10:33the estimates for excess deaths in
10:35relation to the reported cova death.
10:37So here we're just looking at the
10:39excess pneumonia and influenza deaths,
10:41which are shown in red.
10:42So the trajectory for so this is just
10:45basically subtracting off the baseline
10:47from the previous plot and the reported
10:49number of covad deaths for the same week.
10:52So you can see in New Jersey those
10:54two curves basically line up very
10:56well in the great dash line.
10:59Here is showing us the increase in testing.
11:01So this is basically.
11:03Saying he was increasing testing
11:04at about the same time we were,
11:07they were increasing.
11:09In cases and very strong agreement
11:12between those two curves.
11:14You can contrast that with Florida,
11:16where we see this earlier increased
11:18sort of in early March of pneumonia
11:21and influenza deaths.
11:23And then the reported code.
11:25The deaths don't actually increased until.
11:28Several weeks later and one possible
11:30explanation for that is if you
11:32look at the Great Dash Line here,
11:34they're testing levels were quite
11:35low and they really didn't even
11:37start until a couple weeks after
11:39the epidemic had taken off.
11:41So there are quite a few deaths we think.
11:44Sort of during early to merge that were
11:46missed in Florida in the official tallies.
11:49Louisiana pretty good concordance between
11:51the observed and reported deaths,
11:53and likewise for Washington.
11:55There just seems to be a relationship
11:58between sort of wind testing started
12:01relative to the epidemic and the amount of.
12:05Sort of unexplained cases that we see.
12:09We're also interested in looking
12:10at the increase in deaths in
12:13relation to influence like illness,
12:15so we look if we lineups for the.
12:18Unexplained increase in pneumonia,
12:20ones adepts again and explain
12:22increases and influence like illness.
12:23Basically what we see is that the
12:26influence like illness increases.
12:28Its the blue line here and then about
12:30a week later we see an increase in
12:33the pneumonia and influenza deaths,
12:35suggesting that what we're seeing in
12:38terms of excess jets is related to
12:40the virus and not necessarily due
12:42to lock down measures which would
12:45have sort of more diffuse effect.
12:47Enough necessarily have such
12:49a sort of temporarily.
12:50Related Increase similar to Iowa.
12:56So just looking through the
12:58top line estimates here,
13:00this is data through April 25th
13:03where we have an estimate for the
13:06entire US of about 51,000 kuva dots.
13:09During this period,
13:11if we look at the excess
13:13pneumonia influenza deaths,
13:15we have about 57,000 deaths,
13:17so just a little bit more pneumonia,
13:20influenza deaths nationally.
13:22Then we have reported cova deaths
13:25and then about 8083 thousand.
13:27All cause deaths,
13:29so overall about sort of 40 to 50% higher.
13:34Death toll rather than we get from
13:38what's reported in the data in
13:40the sort of official coded data.
13:42This does vary quite a bit
13:44by state and overtime,
13:45so if you were to look at your
13:47New York City in particular,
13:49I'm early in the epidemic.
13:51The reported number of Kobe
13:53deaths is about 3 times higher
13:55than sorry that the access code.
13:57It's about 3 times higher than
13:59reported number of coded deaths as
14:01they have sort of increased testing
14:03and change the reporting guidelines.
14:05That gap is narrowed to about so now the.
14:09Exodus that's about 50% higher
14:10than the Coca dots,
14:11and in other states there
14:13is not much of a gap at all.
14:17So we were also interested in just
14:20trying to do something very simple
14:23without any sort of model behind it.
14:27Just because you know,
14:28every every analysis approach has
14:30assumptions that we wanted to see if we
14:32sort of took an independent approach.
14:34If we get a similar answer and what we see.
14:37So basically what we did was we took the
14:39provisional data that are reported that
14:41were reported this year and week 19th.
14:44This is reported last week.
14:46And we know that those data
14:48are highly incomplete.
14:49Over the last month or so,
14:51and there's this lag in the data,
14:53and we can see that here you
14:55can see sort of the yellow line
14:58sort of would be trailing down.
15:00And So what we did was we looked
15:02at the data from this year.
15:04That were reported in Week 19 and also
15:07looked at the at the data reporting
15:09Week 19 from the previous year.
15:12And what we see?
15:14Is that?
15:15The data in this year were sort
15:18of well aligned up until.
15:20So in mid to late March and then
15:22you see this sort of very sharp
15:25divergences between those lines.
15:26So the provisional data for 2020 are much
15:29higher than the original data for 2019,
15:31and this is sort of a very crude way
15:33of adjusting for the reporting delays.
15:36If we just look at the difference
15:38between these curves,
15:39we have about 79,000 excess deaths.
15:41You can compare that to the 83,000 that
15:43we estimated with the regression model.
15:45We have some other approaches we've been
15:48using as well where we don't adjust for.
15:50Blue where we get a slightly smaller effect,
15:53but you know we're sort of in the range to.
15:57You know 70 to 80,000 excess deaths,
15:59regardless of the method that we're using.
16:02So just to conclude,
16:04the estimated death tool related
16:07the pandemic is about 50% higher
16:09than the reported number of deaths.
16:12This that sort of estimate for
16:15how much higher the excess deaths
16:18is has been changing overtime and
16:21has been narrowing so so the gap
16:24has been narrowing as.
16:26Sort of,
16:27the standards for reporting deaths for Cove.
16:29It have changed and the recommendations
16:31have changed so you know with the current
16:34data were probably closer to about 40%,
16:37but in any case I think this
16:39puts to bed the idea that were.
16:42Start over accounting the cova
16:44deaths when we're.
16:46Certainly can't be official data.
16:47There's no evidence that we see that.
16:50The official data are in any way inflated,
16:53and if anything we think that they
16:56are substantially under reported.
16:58And if you just seem to be a lot of it,
17:01a pretty sizable gap between states
17:03in how much of a difference there is
17:05between the reported number of Copa
17:07deaths in the total number of death
17:09suspect this has something to do with
17:11testing practices as well as sort of
17:13culture around how deaths are coded.
17:15And we still don't know what's
17:17driving that unexplained increase,
17:19so we know that.
17:21You know, pneumonia,
17:22influenza are sort of accounting for some.
17:26Get code is pneumonia,
17:27influenza or accounting.
17:29For some of that increase that we're seeing,
17:31but there's still a large
17:33unattributed increase and future
17:35work will try to understand.
17:36Sort of what's driving them with
17:39somebody 'cause specific factors out.
17:41So thank you very much.
17:42We have a previous version of
17:44this work is unmet archive.
17:46If you'd like to learn more or I'd be
17:48happy to answer any questions by email.
17:54Thank you very much doctor Weinberger.