Thread by @dhmontgomery

Three of today’s five newly reported #COVID19 deaths date from February, raising some skeptics’ red flags.

Unfortunately we only have a month or so of data to examine this, but what I see doesn’t raise MY red flags. 1/

2/ First: we know in general that many death certificates are filed promptly, but others can take many weeks or months to file. The data has what’s called a “long tail” — it’s not super weird for there to be OCCASIONAL months-old death certificates filed.

3/ In general, when working with death data, it’s often best to exclude the most recent 3-4 weeks of data *at a minumum*.

For example, here’s CDC death data. You can see the 2021 line plunge at the end, but that’s not a real decline. It’s just data slowly trickling in.

4/ This is, of course, not very useful for those of us trying to figure out what’s happening NOW. So we resort to various expediencies, like using data by *report date* instead of death date.

This has many problems; we just sort of hope the problems smooth out over time?

5/ Like, the newly reported COVID deaths are going to include some old deaths. But if the share of old deaths holds roughly constant over time, this isn’t a *huge* problem. If 20% of this week’s deaths are 3+ months old, but back in April 20% were ALSO 3+ months old, no biggie?

6/ I check this regularly by comparing the curves of #COVID19 deaths by death date vs. by report date. As you can see, while there are some differences due to lags, the shapes are very similar. That’s a good sign.

7/ But to get beyond the gut-check level, we need more granular data.

The good news is we have that.

The bad news is we only started getting it a month ago. That’s not really enough time to speak conclusively about patterns.

8/ That granular data is (near-)daily updates in the number of COVID deaths by death date. If one were to record this data every day (as I do) you can look at the changes — how old were each day’s deaths? — and plot that over time.

9/ Over the month of data we have, you can see that (this past incomplete week excluded) well over 50% of reported deaths occurred within the past 7 days.

An average of maybe 20% (up and down) have been really old. And then a smattering in between.

10/ Again: this is like a month of data. It’s not enough to speak conclusively about any trends. But in context with other things we know, it doesn’t suggest there’s a major problem with really old death data — sometimes you get data dumps, but they’re not the norm.

11/11 Looking forward to checking back in on this later this fall, when we have more robust data. (And this is why @mnhealth should have started reporting this data last year, when I and other reporters asked…)