Are US Covid-19 Fatalities Declining? Probably

I read some triumphalist claims that Covid-19 fatalities are declining. I want to remind readers about the hazards of interpreting (1) administrative data, and (2) data revisions.

First, there are official tabulations of fatalities due to Covid-19. We should worry about suppression of data in, for instance, Florida, but let’s take the CDC data (which compiles the data provided by authorities) at face value. It’s not clear that all fatalities attributable to Covid-19 are caught by the tracking system. Then we might use “excess fatalities” as a check on the administrative tabulation. Figure 1 below shows how the CDC data on Covid-19 deaths matches the deviation from the fatalities we expect (“excess fatalities”).

Figure 1: [Updated 9/13] Weekly fatalities due to Covid-19 as reported to CDC for weeks ending on indicated dates (black), excess fatalities calculated as actual minus expected (teal), fatalities as tabulated by The Covid Tracking Project/Atlantic (dark red). Light green shading denotes CDC data that are likely to be revised. Source: CDC  9/5/2020 vintage, OurWorldinData accessed 9/13/2020 [updated] and author’s calculations.

Note that excess fatalities far exceed the officially designated Covid-19 fatalities for most of the sample.

Further note that both of the CDC series – Covid-19 Fatalities and Excess Fatalities – drop off dramatically in recent weeks. If you didn’t read the notes attached to the CDC spreadsheet, you’d conclude that we’ve won! But inspection of the spreadsheet reveals notes that indicate that the most recent data is incomplete. In fact, as I show in this post, about the four most recent weeks worth of data are going to be substantially revised. I shade this period in green in the above graph. A hint that this is a substantial problem is provided by comparing the trajectory of the unofficial tally compiled by the Our World in Data project of the Atlantic group, which indicates a much smaller decline.

That means it is possible that “excess fatalities” — our proxy measure for Covid-19 fatalities  — is still increasing (although officially designated Covid-19 fatalities are probably declining, as the Our World in Data series not subject to really large revisions).

 

43 thoughts on “Are US Covid-19 Fatalities Declining? Probably

  1. baffling

    menzie, i would caution you on plotting the excess fatalities curve in the green, where it turns negative. some of our more astute observers on this blog (think bruce hall or corev) will begin to use this as evidence that covid is now saving lives by reducing deaths! i can hear it now, with headlines screaming on faux news, covid saves lives!

    Reply
    1. pgl

      Trump said we have turned the “final corner”. When asked about this – Dr. Fauci said this claim was not true at all. Of course who listens to Fauci when we have the real experts – CoRev and Bruce Hall – to rely on!

      Reply
    2. Ivan

      Yes, the problem is that “Excess fatalities” is calculated as a fully completed (1 year ago) number minus the incomplete current year number. Simply noting that these numbers are likely to be substantially revised is the best you can do. There is no medicine against willful misinterpretation.

      Reply
      1. Menzie Chinn Post author

        Ivan: I think that is not a correct characterization of the data I have plotted in the Figure. Please consult them.

        Counts of deaths in the most recent weeks were compared with historical trends (from 2013 to present) to determine whether the number of deaths in recent weeks was significantly higher than expected, using Farrington surveillance algorithms (1). The ‘surveillance’ package in R (2) was used to implement the Farrington algorithms, which use overdispersed Poisson generalized linear models with spline terms to model trends in counts, accounting for seasonality. For each jurisdiction, a model is used to generate a set of expected counts, and an upper bound threshold based on a one-sided 95% prediction interval of these expected counts is used to determine whether a significant increase in deaths has occurred. Estimates of excess deaths are provided based on the observed number of deaths relative to two different thresholds. The lower end of the excess death estimate range is generated by comparing the observed counts to the upper bound threshold, and a higher end of the excess death estimate range is generated by comparing the observed count to the average expected number of deaths. Reported counts were weighted to account for potential underreporting in the most recent weeks.

        In other words, your characterization of the data plotted is incorrect (although it may apply to some other group’s estimate of excess fatalities).

        If you want to see what the current deaths as reported looks like relative to the preceding one year’s, please see Figure 2 in this post.

        Reply
        1. Ivan

          Thanks; I misunderstood that – they do seem to at least try modeling in the expected future corrections. But somethings “fishy” about that drop from +8K to -5K.

          Reply
          1. Menzie Chinn Post author

            Ivan: Note, I use my own “excess fatalities” series, as actual-expected. CDC cites several, and I think the bound excess fatalities from below at zero (you can’t have negative excess fatalities).

  2. pgl

    Will these deaths continue to decline as we enter the fall season? I listened to an interview with Dr. Fauci who said we may experience another rise in the daily death count. But what does he know?

    Reply
    1. Ivan

      Its worth noting that the number of daily infections in Spain and France have gone higher than at the worst days back at their peak in April-May. Yet current death numbers in those countries are only about 10% of the peak death back then. So we could have a several fold increase in cases this fall, yet only a bump in deaths. Current numbers certainly suggest that getting diagnosed back in the spring was much more likely to end in death than being diagnosed in late summer. However, right now big increases in cases are mostly in rural counties, and they don’t have very robust health systems.

      Reply
      1. Ulenspiegel

        “Its worth noting that the number of daily infections in Spain and France have gone higher than at the worst days back at their peak in April-May. Yet current death numbers in those countries are only about 10% of the peak death back then. So we could have a several fold increase in cases this fall, yet only a bump in deaths.”

        Deaths tails infections by 4 weeks. ATM most infected people are young, the protection of the older seems to work better than in April, but I may be wrong.

        Reply
    2. Ooe

      He is only doing this for decades. Also, you should not trust the CDC under Trump. According to news reports, the mal-administration has been editing the CDC MMR reports since April! (to make Trump look good).

      In fact, Dr. (death) Redfield, claimed that deaths would go down . that was 2 weeks ago. Guess what? they are about 1000 per day for the last two weeks.

      Reply
  3. CoRev

    Covid data is atrocious. The Atlantic Project says this about it s data.
    “Help us get better data

    The COVID Tracking Project can’t offer official guidance to state health departments or compel states to report missing data. We’ve built a state grading system so state governments and their constituents can see who’s doing well and who is underperforming. We’ve built a volunteer outreach team who are in contact with authorities in every state and territory. But in the end, all we can do is ask for better and more complete data—and then ask for it again. If you’d like to see better data, you can sign up for our email list—we’ll send out guidelines on contacting your state officials to push for more transparency and better data reporting.”

    So it collects US data from the states using volunteers. It’s still lower than the official CDC results. I guess its useful for comparative purposes. Maybe!

    Reply
    1. 2slugbaits

      CoRev So it collects US data from the states using volunteers.

      No. That’s not what the Atlantic Project is saying. They are asking volunteers to work with state agencies to identify better sources of data. Not at all the same thing. Let me give you an example. Like a lot of COVID data gathering sites, the Atlantic Project takes today’s total count from a state’s public health department and subtracts yesterday’s total count. That gives them the net increase. If you look at the data it’s even referred to as a “net” increase. Sometimes when the state corrects an error you will even see negative values posted as a net increase. If all you’re interested in are the cumulative totals, then what the Atlantic Project and others are doing is fine. Where it gets to be a problem is if you want to track new cases or new deaths by day as a kind of time series analysis. For example, the rt.live site uses daily data to estimate the effective rate of transmission (Rt). The problem is that when calculating Rt what you need is the actual date of the infection, or at a minimum the date when the test was taken. What gets fed to most websites isn’t the date of infection or the date of the test, but the date that the test result is reported to the state’s health department, which might be several weeks later than when the test was taken. With some states you can actually recover the date of the actual test and match it up with positive and negative results for that day; however, it’s not easy and you generally cannot get this data below the state level. And when you do that you will see daily numbers bounce around for a few weeks. Trying to get it at the county level is very difficult. Getting the net increase at the county level is easy; but getting the actual date of the positive test is another matter. But it’s really important because a lot of states and counties are basing their action plans on local positivity rates, and to get a meaningful positivity rate you need to know when the test was taken, not when the result was report to the state’s health department.

      Hope that helps.

      Reply
      1. CoRev

        2slugs, I interpreted this sentence: “But in the end, all we can do is ask for better and more complete data—and then ask for it again. to meant the were back filling data holes through volunteers. Your interpretation may be correct. Another Ho Hum attempt at what?

        We both agree the states and lower levels data are bad, my term is atrocious.

        Reply
        1. Menzie Chinn Post author

          CoRev: But with time series, and different vintages, we can determine the attributes of each of the series. And CDC data gets revised toward the OWID data, as newer and newer vintages come out. That suggests something to me.

          Reply
        2. 2slugbaits

          CoRev to meant the were back filling data holes through volunteers.

          Then you misunderstood. I already explained to you why that is NOT what they are doing. Let me try again, once more with an example. In my area local news organizations report daily death and case counts using a mix of data sources. In some cases the county health department provides the latest results of case counts and deaths to local news outlets at the same time they report those updates to the state. But it takes the state a couple of days to update its database with the new county numbers. Other counties don’t provide the most current data to news organizations, so data from those counties isn’t available until the state processes it. So what local news organizations report is a hybrid, with county level data reflecting the most recent data and some county data lagging a few days behind. As a result the cumulative numbers from news organizations are always a little higher than what the state reports. The data isn’t “atrocious”, it’s simply a case of some data sources being a few days ahead of others due to reporting procedures.

          That said, there are some problems with the data. For example, how should positive COVID cases be reported for incoming college students? If they contracted the virus before they arrived on campus but didn’t show symptoms and didn’t get tested until on campus, should that case be associated with the college town’s numbers or with the student’s home town? It’s not obvious how that should be handled. You could argue it both ways. That’s the kind of thing that the Atlantic Project is asking about.

          Reply
          1. Barkley Rosser

            And as I noted on another thread, there may be even more serious misreporting at the county level due to students. It is not just where did they get it, it is where are they now? In some cases students are being reported as being on their campuses whereas in some cases they are being reported as being at home, with that up in the air as in some cases students are being sent home (or some of them) as happened recently on my campus.

            As noted previously, this confusion is at least partly getting washed out at the state level as one county’s error is offset by another’s although that is not taken care of when the students are from out of state, a bigger issue for private schools than state ones.

          2. CoRev

            2slugs, why the long winded overly complex example of local versus state versus news…? The Atlantic was talking specifically of its state level data collection. I’ll concede your interpretation is probably more correct than mine. I’ll also concede there is another word to describe untimely, incomplete, inaccurate data. You pick the word

          3. 2slugbaits

            CoRev why the long winded overly complex example

            Evidently not long winded enough because it’s obvious that you still don’t get it. The Atlantic Project is looking for volunteers to identify better data sources.

            The Atlantic was talking specifically of its state level data collection.

            If you understood my “long winded overly complex example” then it should have been obvious that using a hybrid of the most recent county level data (if available) alongside data provided by the state yields a more current statewide total. If you still don’t understand why that’s the case, then let the beer wear off and think about it when you’re sober. Of course, the timeliness of the data isn’t the only concern. There are also ambiguous problems with the data, such as the example of college students. The Atlantic Project is looking for ways to get around that problem.

            If you’re genuinely concerned about COVID-19 data quality issues, then you should be angry as hell at McConnell and Trump for sitting on funding that would improve that data.

          4. noneconomist

            “Evidently not long winded enough because it’s obvious you still don’t get it.”
            Game. Set. Match.

  4. Bruce Hall

    CDC provisional death counts:
    Note: Provisional death counts are based on death certificate data received and coded by the National Center for Health Statistics as of September 11, 2020. Death counts are delayed and may differ from other published sources (see Technical Notes). Counts will be updated periodically. Additional information will be added to this site as available.

    So, yes, most recent CDC data will be revised. That’s why I track the weekly reports to see the pattern emerging. Other sources attempt to aggregate data, but not necessarily reviewing the data in as much depth. There are “confirmed” and there are “probable” COVID-19 cases/deaths.

    Counts of deaths from all causes of death, including COVID-19, are presented. As some deaths due to COVID-19 may be assigned to other causes of deaths (for example, if COVID-19 was not diagnosed or not mentioned on the death certificate), tracking all-cause mortality can provide information about whether an excess number of deaths is observed, even when COVID-19 mortality may be undercounted. Additionally, deaths from all causes excluding COVID-19 were also estimated. Comparing these two sets of estimates — excess deaths with and without COVID-19 — can provide insight about how many excess deaths are identified as due to COVID-19, and how many excess deaths are reported as due to other causes of death. These deaths could represent misclassified COVID-19 deaths, or potentially could be indirectly related to the COVID-19 pandemic (e.g., deaths from other causes occurring in the context of health care shortages or overburdened health care systems). https://www.cdc.gov/nchs/nvss/vsrr/covid19/excess_deaths.htm

    I suppose one set of educated estimates may be as reasonable as another. After all, the CDC is still revising death totals as far back as the week of March 28.
    https://www.dropbox.com/s/8oe6ggutc461a65/Covid-19%20Deaths%2C%20Cases%2C%20and%20Hospitalizations%20-%209-8-20.pdf?dl=0

    Reply
  5. CoRev

    Bruce, I just download the entire data set daily to capture ALL the latest updates. For some reason Menzie thinks that missing data can be found at other sites. Most of which use the CDC data.

    Reply
      1. CoRev

        Menzie, spent a career doing analysis. Can you tell me why you will not recognize that the OWID data for the US is sourced from the US CDC? It is on their list of sources on the OWID site: “United_States_of_America https://www.cdc.gov/coronavirus/2019-ncov/cases-in-us.html

        It amazes the hubris when things like: ” And CDC data gets revised toward the OWID data, as newer and newer vintages come out. That suggests something to me.” To me also. With the CDC data the parent, then the OWID will converge with the CDC data. Do you actually believe that the US provisional data will go down to the OWID data for 0911/20? CDC’s data – 09/11/20 daily- 1035 7 day rolling avg. 747. OWID shows 09/11/20 daily- 974 7day rolling avg – 713.1.

        Of these two, which is most likely to change to match the other. I TRULY HOPE YOU ARE CORRECT AND THE US NUMBERS GO DOWN TO OWID’S.
        CDC – https://covid.cdc.gov/covid-data-tracker/#trends_dailytrends
        OWID – https://ourworldindata.org/coronavirus-data-explorer?zoomToSelection=true&year=latest&time=2020-09-11&country=~USA&region=World&deathsMetric=true&interval=daily&smoothing=0&pickerMetric=total_deaths&pickerSort=desc

        PGL has shown he can not use these data sites. I hope you can do better than he. If not just ask instead of making another sarcastic comment about doing analysis.

        Reply
          1. CoRev

            Menzie, asks: “CoRev: Wasn’t that the entire point of our multi-comment exchange? That OWID was not relying on solely CDC as you repeatedly (and erroneously) asserted?” The entire point? NO! But I concede your point re: excess deaths/mortality.

            Since you included CDC 7 day AVG. fatalities I compared that set of OWID data to CDC to see how well they track.
            “Of these two, which is most likely to change to match the other. I TRULY HOPE YOU ARE CORRECT AND THE US NUMBERS GO DOWN TO OWID’S.
            CDC – https://covid.cdc.gov/covid-data-tracker/#trends_dailytrends
            OWID – https://ourworldindata.org/coronavirus-data-explorer?zoomToSelection=true&year=latest&time=2020-09-11&country=~USA&region=World&deathsMetric=true&interval=daily&smoothing=0&pickerMetric=total_deaths&pickerSort=desc

          1. CoRev

            Menzie, sorry, I missed the 2nd source in the list. Now, why is OWID lower than the CDC data? Do you still believe CDC will converge with OWID?

          2. Menzie Chinn Post author

            CoRev: Wasn’t that the entire point of our multi-comment exchange? That OWID was not relying on solely CDC as you repeatedly (and erroneously) asserted?

            To answer your question,fFor the most recent month’s data, yes CDC data will converge towards OWID, because the CDC series is slowly incorporating data provided by local health authorities. (I don’t understand your assertion that OWID is lower than CDC figures — for the green shaded area, that is not true in a substantial way.)

          1. pgl

            Thanks for reminding us of CoRev’s April 2019 insanity. To suggest FRED reports different labor market data than the BLS is pretty funny since FRED’s footnotes clearly reflect that their source is BLS and that they update their series the same day BLS releases new data.

            There is a pattern here. Someone reports reliable data that undermines Trumpian spin and ANGRY CoRev goes on the attack with dozens if not hundreds of sheer rants that only shown one thing – CoRev never passed 1st grade reading.

      2. CoRev

        Menzie, spent a career doing analysis. Can you tell me why you will not recognize that the OWID data for the US is sourced from the US CDC? It is on their list of sources on the OWID site: “United_States_of_America https://www.cdc.gov/coronavirus/2019-ncov/cases-in-us.html

        It amazes the hubris when things like: ” And CDC data gets revised toward the OWID data, as newer and newer vintages come out. That suggests something to me.” To me also. With the CDC data the parent, then the OWID will converge with the CDC data. Do you actually believe that the US provisional data will go down to the OWID data for 0911/20? CDC’s data – 09/11/20 daily- 1035 7 day rolling avg. 747. OWID shows 09/11/20 daily- 974 7day rolling avg – 713.1.

        Of these two, which is most likely to change to match the other. I TRULY HOPE YOU ARE CORRECT AND THE US NUMBERS GO DOWN TO OWID’S.
        CDC – https://covid.cdc.gov/covid-data-tracker/#trends_dailytrends
        OWID – https://ourworldindata.org/coronavirus-data-explorer?zoomToSelection=true&year=latest&time=2020-09-11&country=~USA&region=World&deathsMetric=true&interval=daily&smoothing=0&pickerMetric=total_deaths&pickerSort=desc

        PGL has shown he can not use these data sites. I hope you can do better than he. If not just ask instead of making another sarcastic comment about doing analysis.

        Reply
        1. 2slugbaits

          CoRev Menzie, spent a career doing analysis.

          Now that’s scary. You never took calculus. You don’t know the basics of time series analysis. You don’t know basic statistics. I’d be shocked as hell if you understood ANOVA/MANOVA. You don’t understand min/max, or stocks vs flows, or local versus global optimizations. You refuse to read any economics textbook. As best I can tell your idea of “analysis” is dumping data to a spreadsheet and creating a graph.

          Reply
          1. CoRev

            2slugs, was originally trained as an intel. analyst. Different set of tools, especially then.Then went into SA another set of tools. I took calc and stat classes, but they were seldom needed so didn’t pursue. If/when I needed calc, I had a bunch of young college kids to do the work. Never really needed stats.

            To Barkleys distress, I actually did work in the space program doing tracking. One of my tasks was to create a program to track the Sun with C and or X Band radar antennas to measure its background noise. That effort was needed to engineer the satellite xmitters to power through the Sun’s noise when satellites were placed at Lagrange points. Never needed calc.

            It takes logic and understanding of the big picture for a manager to use the tool(s) or tool handlers. So what was the point you were trying to make?

          2. noneconomist

            Yet here you are, CR, with an admitted “different set of tools” trained in alternate(?) analysis arguing economic analysis with the blog’s author, a renowned economist as well as other prominent economists who post here.
            Not only arguing but often insisting they have no clue what they’re talking about while you–bereft of any admitted understanding of econometrics or statistical analysis–blabber on. And on. And on. And on.

          3. pgl

            “Now that’s scary. You never took calculus. You don’t know the basics of time series analysis. You don’t know basic statistics.”

            Well yea but Lawrence Kudlow never took a class in economics and he made it to be the chief economic advisor for President Donald J. Trump.

          4. baffling

            ” I took calc and stat classes, but they were seldom needed so didn’t pursue. If/when I needed calc, I had a bunch of young college kids to do the work. Never really needed stats.”
            then you never really did analysis. many years ago, when i first started working, the work environment was different. it was bloated by a bunch of baby boomers who rode the economic tailwinds of demographics. there were some good managers, who got stuff done. but a lot of them were very poor, having risen to their ranks via the peter principle. they were put out to pasture, and only managed noncritical and unimportant tasks. we had a phrase for those folks. those who can’t, “manage”. i would imagine this was an accurate description of corev, especially in the final couple decades of his work life.

          5. CoRev

            Baffled, envious?

            You might be interested in these analyst job openings in DOD https://www.usajobs.gov/GetJob/ViewDetails/578066000
            Go ahead and check the education requirements for degrees and calculus. Also note the reliance on experience over education.

            I know when you enter a room you must be the smartest person with the highest salary due to your study in calculus. I also know the smartest, most intelligent person I ever knew, never finished HS. He won the AA Fuel Spring Nationals, took Nuclear Physics in college (still no diploma or GED), and was an accomplished photographer all in his 20s. When I knew him he was a computer analyst/programmer in the Apollo Program. He was the closest thing to a renaissance man I’ve known.

  6. Ivan

    I note that he difference between excess fatalities and reported Covid deaths is getting bigger. That suggest an increased problem with political pressure on those who collect the cause of death data (Florida being the most blatant example). But at least the overall trends are still looking the same (even if some states have done creative writing on the actual numbers).

    Reply
  7. sammy

    The old saying is “The devil is in the details.” In this case the detail is the definition of “New Case.” According to the CDC, the “presence of antibodies” is counted as a “New Case.” This is not a New Case, but an Old Case. In fact the “presence of antibodies” means the person has fought off the disease and is no longer susceptible or contagious. It is a positive indicator, reported as a negative indicator.

    Reply
    1. pgl

      As usual Sammy is babbling incoherently. WTF are you talking about and how is it relevant to this post? Oh you do not know. It is like those monkey banging on their typewriters.

      Reply
    2. baffling

      if you are measuring cumulative totals, then it is a new case and an important contribution to understanding how much of the population has been impacted. it simply gives historical context, not real time context.

      Reply

Leave a Reply

Your email address will not be published.