Skip to content
 

Considerate Swedes only die during the week.

Bar chart of coronavirus deaths in Sweden, by date.

Reported Coronavirus deaths in Sweden, by date.

This post is by Phil Price, not Andrew.

A lot of people are paying attention to Sweden, to see how their non-restrictive coronavirus policies play out. Unlike most other countries in Europe, they have instituted few mandatory measures to try to slow the spread of the virus. Instead, they’ve taken a ‘softer’ approach, telling people the risks and asking people to make good choices. And people are certainly changing behavior: according to Google’s mobility reports , as of April 5 the use of transit stations was down 37%, workplaces were down 10%, and retail establishments were seeing 25% less traffic. So there’s definitely some ‘social distancing’ going on, although not nearly as much as, say, Norway (retail down 60%, workplaces down 32%).

So…what’s the result? How do deaths in Sweden compare to other countries? Well, on paper they’re doing OK, with deaths doubling every 5 days. That’s not as good as their neighbors (Denmark, Finland, and Norway are all around 6-7 days, and that’s a difference that adds up, or rather multiplies up, over the course of a month or two) but it’s about the same as Belgium and, well, hard decisions have to be made and conceivably the Swedes could feel that this is the right balance of economy versus deaths.

But: I don’t trust the numbers coming out of Sweden. See the attached plot of coronavirus deaths by date. I think we can all agree that people are dying on weekends, they’re just not being reported. If this were just some clerical thing, like deaths not being counted until the clerks show up at the office on Monday, then we would expect that either the numbers would be corrected over time, or that there would be a big spike on Mondays when the weekend deaths are counted, but we don’t see either. (The plot is from Worldometers and the numbers match those that are reported daily in the New York Times).

As the world tries to figure out how to manage an end to the shutdown that’s in place in many countries, data from countries like Sweden that are doing things differently should be very valuable. But only if we can trust the data. Anyone have any idea what is going on with the death count in Sweden? I don’t.

This post is by Phil.

34 Comments

  1. Steve says:

    Haven’t we seen the same pattern in the data here in NY? Whoever is collecting data or reporting it, doesn’t keep up with it over the weekend? Isn’t that one factor?

    • Dale Lehman says:

      Well that sure looks like an authoritative site. And, when I look at the death data (and intensive care data), neither appears to have weekend dips. For that matter, I’ve looked at the US data and don’t see weekend dips either. It would seem logical to expect weekend reporting to be lower (for a variety of reasons) normally, but these are not normal times. For just one example, children are usually home on weekends, so hospital staffing is reduced – neither of these appear to be true right now. And, the data I’ve seen do not generally show drops on weekends.

      • Johannes B says:

        Well, here in Sweden children are at home on weekends. Or more to the point, only high schools and universities have moved to “emergency remote teaching”. Our kids will go to school again tomorrow after 4½ days of easter weekend.
        Easter is traditionally a major travel weekend, but trips from Stockholm (the worst stricken area) are down by around 90%.

      • jim says:

        I plotted the first two columns on the first tab (Antal per dag region: Statistikdatum & Totalt_antal_fall) which I suspect is total fatalities?

        I plotted the data and colored the Sundays (4/12, 4/5, 3/29, 3/22, 3/15). There’s definitely an aversion to death on Sunday. To me the weekend / weekday signature looks pretty strong.

        • Swede says:

          Note that Totalt_antal_fall is total number of cases, ie, not deceased. To see the number of deceased per day, use the data explorer and choose the data set Tidsserie: Avlidna per dag (“time series: deceased per day”). The graph does not have any weekend pattern.

          Go to https://experience.arcgis.com/experience/09f821667ce64bf7be6f9f87457ed9aa . Find the graph labelled “Tidsserie: Sjukdomsfall per dag”. Under the graph there is the text “Sjukdomsfall/dag” (“cases of illness / day”), click the right arrow next to it to cycle through other data: “Sjukdomsfall / dag kumulativt” (“cases / day cumulative”), “Intensivvårdade / dag” (“in intensive case / day”), “Avlidna / dag” (“deceased / day”).

          Translations of other words on the page:
          Sjukdomsfall = number of cases of illness
          intensivvårdade = have been in intensive care
          avlidna = deceased
          antal = number
          laboratoriebekräftade = laboratory-confirmed
          åldersgrupp = age group
          uppgift saknas = n/a

          At the top, it says that data is updated daily at 11:30 and published at 14:00 (CET).

  2. Andrew says:

    Phil:

    That graph has an absolutely horrible x-axis. Also, the comb-like nature of the plot is unfortunate. The spaces between the bars have no meaning at all! But, on the plus side, the y-axis doesn’t go below zero.

    • Alex C. says:

      Andrew, can you elaborate on why you think the x-axis is horrible (and how it can be improved)? Also, would you recommend keeping the bars and simply eliminating the spaces between them, or would you recommend using a line plot, or…? I’m asking because I have a number of graphs that are very similar to the one in the post, and I’m always looking for ways to improve the way I report data.

      • jim says:

        Well I’m sure phil just did a quick plot, but:

        2/3 of x-axis has no data, which means the data that is there has to be shrunk to make space for the blanks. If you just plot from 3/2 onward, that’s pretty good. That also allows every bar to have a date label, rather than having the somewhat bizarre 4-day interval between labels.

        I like the bar chart viz, it makes the daily variation stand out a lot more than a scatter chart.

        When I plotted in excel I colored the weekends. Yep, that’s the kinda expertise I bring to the table, heh heh. (tongue in cheek)

      • Michael Nelson says:

        My 10 cents: There’s no data prior to March 10, but the x-axis starts on Feb 15; there’s no logic to having a 3-day gap between each label; there are no tick marks; the angled labels are difficult to read and make it difficult to tell which bar has which label. The y-axis labels (and horizontal lines) jump in such wide increments and across such a large range that a bar can be twice as tall as the one before it but look nearly identical–all that blank space on the left hurts readability without adding anything. Take out the entire area to the left of Mar 10 and also the one above ~120. (If it’s important to note that tracking started earlier, then put that in a note.) In addition to taking out blank dates and adding tick marks, the graph might be improved by putting the dates on the x-axis, which would allow for a 1:1 alignment between labels and ticks without rendering less legible; only indicating the month on the first day of the month and dropping the leading zeroes; using colors to add information, like indicating weekends and/or holidays.

    • Phil says:

      It’s not my plot, it’s a screenshot from Worldometers.

  3. Henning says:

    I don’t know what those numbers are. The first bar graph in the link below shows the daily number of deaths in Sweden over the last ten days, as reported by the Swedish public health agency FHM. (I just checked it against FHM’s excel sheet linked to by Johannes above, and they match.) No bar goes above 100 deaths (as they do in the OP’s graph). Maybe the numbers are distorted somewhere between FHM and the international reporting. One thing we are continually told here is that reporting lags behind by a number of days. That is, the bar for one day potentially keeps growing for a week or more.
    Here’s the link

  4. Tom says:

    Cross country comparison (and even interstate comparisons) are hampered by inconsistent definitions so all these numbers should be taken with a grain of salt.

    In Italy, they don’t consider co-morbidity. “The way in which we code deaths in our country is very generous in the sense that all the people who die in hospitals with the coronavirus are deemed to be dying of the coronavirus.”
    https://www.telegraph.co.uk/global-health/science-and-disease/have-many-coronavirus-patients-died-italy/

    “Belgium counts deaths occurring both in hospitals and elsewhere, like in care homes. Deaths with clinical symptoms of the coronavirus disease are considered as “suspicious” and are included in the figures of COVID-19 deaths, even if they were not tested.” (The source link is in French, but suppose for the sake of argument that Wikipedia summarized it accurately…)
    https://en.wikipedia.org/wiki/2020_coronavirus_pandemic_in_Belgium#Deaths

    FWIW, I’m not sure why they don’t express all these statistics as excess deaths. “Curious” looked at NYC this way in a comment on a different post:
    https://statmodeling.stat.columbia.edu/2020/04/13/america-is-used-to-blaming-individuals-for-systemic-problems-lets-try-to-avoid-that-this-time/#comment-1296821

    I imagine you would have large deviations from average daily/weekly/monthly deaths early on as susceptible populations died and the error would gradually decrease to the marginal mortality from Covid being introduced to the population. Given all the uncertainty in mortality coding in addition to the transmission, prevalence, and mortality of Covid, an excess deaths time series would seem to make the fewest assumptions.

    • Johannes B says:

      The Swedish time series counts only confirmed (positive PCR) cases, though testing is focused on patients in health care and elderly care. Cause of death does not matter, on the other hand. The idea is to have a precise, well-defined measurement, rather than a comprehensive but more subjective one. For the latter you can look at the EU’s excess death tracker http://euromomo.eu.

  5. jim says:

    ‘nuther good point from Taleb:

    “It’s not the reaction that wrecked the economy. It’s the LACK of INITIAL reaction that wrecked the economy.”

    • Brent Hutto says:

      So a month of shutdown wrecked the economy but three or four months would NOT have wrecked it?

      How do you reconcile that with the oft-stated “expert” opinion that if we relax the shutdown even a little bit this summer there will be an even worse outbreak as a result?

      I’m just having a hard time with the attitude that, absent some miraculous vaccine appearing out of nowhere, the only way to save to world is to shut everything down and keep it shut down for months, years, indefinitely. At some point the reaction does more damage than the virus. I suspect we have already crossed that point, if not it’s a matter of (a few) weeks away at most.

      • Andrew says:

        Brent:

        You write, “I’m just having a hard time with the attitude that, absent some miraculous vaccine appearing out of nowhere, the only way to save to world is to shut everything down and keep it shut down for months, years, indefinitely.” I don’t think anyone’s saying this. I think people are recommending some combination of (a) enough shutdown to keep the rate of new cases low enough to not overwhelm the health-care system, and (b) containment of cases to reduce spread among people who go back to work. Shut-down-everything was just a starting point to make it easier to get to (a) and (b).

        • Brent Hutto says:

          The difficulty will come if the infrastructure, workers and political will to do (b) containment of cases is not forthcoming in any sort of weeks-to-months timeframe. The elite/expert position seems to be that your item (b) is a non-negociable prerequisite of any relaxing of the shutdown.

          If so, I fear we’re still looking at many months. In fact, given our experience these last few months I’m not even sure a truly effective screening/testing/tracing/quarrantine regimen could be implemented faster than a vaccine.

          But I devoutly hope to be proved wrong or we’re looking at a choice between two horrible alternatives. Completely crater the economy by keeping the shutdown tight for months to years versus letting the virus spread through and opened-up society because we can’t/won’t do the hard work of containment.

          If faced with that choice come August, I would rather we take our chances with COVID-19.

          • jim says:

            “If faced with that choice come August, I would rather we take our chances with COVID-19.”

            Even if the current administration just started working on this now, it should aiming for June 1 at the latest. It really shouldn’t be that complicated. You don’t need medical professionals to track people. By that time we will have had three months to gear up testing. Seems like that should have been doable.

            I share your concerns about the economy. I really feel like much of the talk at this point should be about what’s being done to get to open.

      • jim says:

        Andrew’s points, plus:

        1) “Initial reaction” includes more than shut-down. It also includes quarantine, track and trace, and testing

        2) I think in your focus on the cost of shutting down you’re under-estimating the cost of staying open.

        The death toll would have been much higher without a shutdown given that no other action was taken. Tech companies were advising people to work from home even before the initial breakout in the nursing home in eastern King County. Probably a wise move considering that this disease disproportionately impacts senior people with more knowledge of their businesses.

        This is speculation, but from what we’ve seen in nursing homes I suspect that the disease has a concentration multiplier effect: the more people in an enclosed area that have it, the higher the concentration in the environment and the more deadly it becomes, so working normally would increase both the infection rate and severity. I speculate that regular exposure to higher concentrations would also drive more fatalities into the lower the age brackets. Total speculation, but hardly out of the question.

  6. Dzhaughn says:

    Washingtonians are comparably considerate by only testing positive for the virus. (See graph on “Status of Testing” at link below.)

    https://www.seattletimes.com/seattle-news/health/coronavirus-daily-news-updates-april-13-what-to-know-today-about-covid-19-in-the-seattle-area-washington-state-and-the-nation/

  7. Dan F. says:

    Countries like Spain and Italy and France and Sweden robust bureaucracies, and counting deaths is the sort of thing they do well generally. There are often legal requirements related to certification of deaths. In overloaded systems these may mean that more deaths are certified on weekdays than on weekends. Is it the date of death that is reported or the date of the reporting of the death?

    Newspaper reports indicate that excess mortality (mortality above normal) in places like Madrid and Castilla could be 2 or 3 times greater than the number of deaths attributed to coronavirus in these communities. The Spanish government has made clear that it reports as a coronavirus death only the death of someone who has been confirmed by a test to have had coronavirus. Many (apparently possibly several thousand) who died in residence homes were never tested, for whatever reasons (there are several), so their deaths are not officially attributed to coronavirus.

    For back of the envelope thinking, reported death counts should be taken as quite reliable lower bounds, and reasonable indications of numbers of deaths up to maybe a factor of 2 in countries with robust institutions, and up to an order of magnitude (or two) in countries with less robust institutions. The numbers are probably better in countries known for transparency. Those needing to justify possibly stupid decisions (e.g. Sweden, UK) might have incentives to underreport or obfuscate.

    A careful analysis of excess mortality should give a better idea of deaths due to coronavirus than official figures.

    • CraigM says:

      It seems like only focusing on deaths directly caused by the virus is missing a lot of the impact of the pandemic and response.

      In any sort of policy analysis involving the costs and benefits of this episode, we need to include the lives saved (and lost) as a result of societal responses, but which are not directly linked to covid. I have unfortunately had to be at a major metropolitan Children’s Hospital with an inpatient sick child over a couple of intervals in the last three weeks. It is so empty that it is almost creepy – it turns out that moderate social distancing is really effective at preventing kids from getting influenza and other potentially serious illnesses, and so some kids aren’t dying. There are, however, a huge number of non-urgent but life-enhancing procedures which are being deferred, and an insane amount of stress on all of the staff and patients, which will have impacts on life quality and quantity which will only be realized in the future. There are many ways lives are saved from the people’s recent behavioral changes – vehicle fatalities are down, and decreased pollution from reduced unnecessary travel has an impact on population death rates and quality of life. It’s complicated, and even without official efforts at obfuscation, pinning down the deaths due to covid is hard.

      Probably the only meaningful measure of deaths will come from looking at the change in total mortality (disaggregated demographically to the extent possible) over a year or longer.

      However, if we fail to use this natural experiment as an opportunity to collect and exploit data about the real costs of decisions made gradually and without much collective thought, ranging from air quality standards to social practices such as hand-shaking, we will be making a mistake every bit as consequential as the multilateral official denialism that has impaired our response to the virus.

  8. MichaelS says:

    The data from Worldometers is based on the date a death was reported, not on the date it actually happened. Looking at the Swedish Public Health Agency (SPHA) official data on deaths per date does not give this picture at all. The reason, as this has been discussed here in Sweden, that has been given by the SPHA is that files are typically retroactively changed over several days (based on reading the journals). That’s why there’s no obvious spike on Mondays, rather, this graph overstates the deaths on most/all weekdays compared to weekends.

  9. Robert says:

    My sense is that the issue is how the Worldometers website is compiling the data. If you go to the actual site Worldometers pulls Sweden’s data from (https://experience.arcgis.com/experience/09f821667ce64bf7be6f9f87457ed9aa), you see different daily trends. They look more like a traditional bell curve. This gets to MichaelS’s comment above. Note that the overall numbers from Worldometers and arcgis match. Having been tracking data in WA, though, it appears that there is a lag to daily reporting, meaning data is backfilled as it becomes available. From the trends on the arcgis site, it looks like Sweden is turning the corner but again we want to wait until the data is backfilled (if that is an issue).

    • Aleksandra says:

      Worldometer reports the deaths per day of reporting (i.e. lower number on weekends when reporting is not as efficient). The Swedish official website reports deaths per day of occurrence (the numbers for each day are updated when the information comes in, sometimes days after the actual death date). I think that this is the main reason for differences in how the the “deaths per day” figures look like at Worldometer and on webpage from Swedish officials.

Leave a Reply