@deb_cohen @carlheneghan I'll post my analysis tonight, busy building models for actual job. But I think I might disagree based on my initial look, but need to dig a bit deeper as Chris Giles is obv very smart man.
@deb_cohen @carlheneghan So looking at the FT model, what they have done is to say – all excess deaths can be attributed to COVID. While this isn’t necessarily true, let’s go with it for the sake of argument.
The ONS excess deaths data is available weekly and shows the total deaths registered that week.
The ONS excess deaths data is available weekly and shows the total deaths registered that week.
@deb_cohen @carlheneghan Then they have assumed a 4 day lag, so they assume excess deaths for the 10th of April mostly occurred on the 6 of April. Again fine.
They then compare the excess ONS deaths figure with the NHS hospital death data for the same day (including 4 day lag) and note the difference.
They then compare the excess ONS deaths figure with the NHS hospital death data for the same day (including 4 day lag) and note the difference.
@deb_cohen @carlheneghan They then basically project this difference forward to current day to get their 40k figure. First thing to say is ONS excess death data is lumpy. Its only provided in weekly discrete totals. I've just used E+W data for speed I’ve pushed them back 4 days to show the lag 
@deb_cohen @carlheneghan First thing I would question, is to get a daily trend it looks like they have just drawn a line from peak to peak on the weekly discrete data like this. I might be wrong but it looks this way to me.
@deb_cohen @carlheneghan Compared against the NHS England data on COVID deaths on date deaths occurred you get this:
@deb_cohen @carlheneghan So this more or less how they’ve got to 40k. My trend is very crude just to make the point.
However, I would argue this is misleading, as the actual days that deaths occurred between those peaks can affect the daily trend significantly.
However, I would argue this is misleading, as the actual days that deaths occurred between those peaks can affect the daily trend significantly.
@deb_cohen @carlheneghan As an example, below shows some mock data with same peaks but different distribution of when actual deaths occurred within those peaks affecting the trend.
@deb_cohen @carlheneghan So you really need to apply some weighting to daily curve of weekly excess deaths, not just connect the peaks. This needs to be done historically but should also be used as a guide to project forward. FT say they aren't making any assumptions on future trend but implicitly are.
@deb_cohen @carlheneghan Luckily the ONS publish data on COVID deaths both in hospital and in community/care homes by the actual date the death occurred, so it would make sense to use the shape of this data to model the excess death data
@deb_cohen @carlheneghan (I’m surprised they haven’t - unless you assume the non-COVID portion of the excess death data is dramatically different from the COVID-attributed deaths from the ONS - but I’m not sure how this is reasonable). ONS all location COVID death data matches the curve of the NHS data
@deb_cohen @carlheneghan Applying this curve to the ONS excess deaths and then projecting forward assuming the ONS COVID death data maintains its relationship with the NHS data you get something like the following:
@deb_cohen @carlheneghan Total excess deaths is<30,000. But I’m not sure you can attribute all excess deaths to COVID. As the graph shows, the ONS daily COVID death data is lower than the ONS excess death data. On 6th of April ONS COVID total was 9,363 vs total excess deaths at 15,161,
40% difference.
40% difference.
@deb_cohen @carlheneghan Interesting NYT article, shows Sweden has no excess deaths once COVID deaths are removed. Suggests countries with unexplained excess deaths are due to lockdown. nytimes.com/interactive/20…
