Excess deaths, accelerated deaths, and co-morbidities

When I first mentioned excess deaths, in April 2020, I already warned that when analyzing excess deaths, there is a delicate balance between waiting for enough signal to accumulate, and waiting too long until the signal dissipates.

***

Here is an example of premature analysis, from a long-ago post on Junk Charts.

The really-short last column is short because the chart was issued when the full-year data for 2011 have not yet materialized. As presented, readers should not compare that column to the other columns. If they don't realize data are incomplete, they mistakenly think that 2011 was particularly short-changed on the metric.

Analogously, when reading the following type of chart on excess deaths, we have to wink at the last few weeks suspiciously:

The last date on the chart is August 1, and the caption said the analysis is for March to July 2020.

Notice how the lines droop toward the right end. Some readers may seize on this "trend" to argue that the worst is over. You can't blame the readers for reading out loud what is shown on the chart. But this is a fake trend - an artifact of a reporting delay. In computing excess deaths, statisticians use deaths as confirmed by death certificates, a process which takes time - from five days to a couple of months - to complete.

A responsible data journalist should omit the last month of data, and inform readers about insufficient data. An ambitious data journalist can build a model that projects what proportion of the deaths is typically reported by a certain time, and applies an inflation factor to deaths during the last month or so - and use a dotted line for these projections.

As regards excess deaths, if the analyst does not wait long enough, the analysis is biased by observation time. The more recent the data, the less complete it is. Even if the number of deaths stayed constant, the lines still droop because the last counts have not fully baked.

***

Waiting too long presents a different kind of problem. This problem is related to co-morbidities.

Let's start with the popular claim that Covid-19 deaths are being over-counted because of "co-morbidities," i.e. pre-existing conditions that raise the mortality rate of Covid-19. For instance, if someone had heart disease before catching Covid-19 and dying, the death should be attributed to heart disease and not Covid-19.

Such an argument confuses the meaning of co-morbidity. A co-morbidity makes it more likely to die from Covid-19 if someone catches it. A co-morbidity does not have to be a cause of death. Gender (being male) and race (being black) are both co-morbidities, and we do not say a black male died not from Covid-19 but from being black and being male.

A co-morbidity is a predictive factor that has a statistically significant effect on death.

What about co-morbidities such as diabetes and hypertension? This gets complicated. Every common cause of death has comorbidities (i.e., some other factors that have statistically significant effects on death). See this research study of heart disease in which 75% of those who are listed as dead from heart disease have comorbidities, such as diabetes, kidney disease, and pulmonary disease. Having diabetes, for example, raises the chance of dying from heart disease by 60%.

If one wants to challenge the official causes of deaths, one has to revamp the entire system. What one runs into is a core statistical problem: nature does not grant one the right to isolate independent single causes.

Imagine reclassifying the heart-disease deaths with comorbid diabetes as diabetes deaths. Next, if you do a comobidity analysis on diabetes deaths, you'll find that a sizable proportion has comorbid heart disease. You're now shuffling the deck chairs. Except in text books and carefully designed experiments, there is no such thing as holding all else equal. Co-morbidities are just correlated causes.

But we need a way to tame this mess. We want to label something as the primary cause of death so we can print nice tables such as Leading Causes of Death. That's why the medical examiners establish conventions to assign a primary cause of death. They are the experts and we should respect them.

***

What about the related argument that Covid-19 does not cause any excess deaths, as what we are seeing are accelerated deaths? That is clearly true, as I pointed out in April. Particularly since this virus kills older patients more efficiently. To wit, an elder who died last week is someone who would not die in two months.

What happens if we look at cumulative excess deaths over a long time, say one year or longer? That trend line will dwindle, and may even get back to zero. This analysis just informs us that everyone eventually dies. It brings no more insight than that.

The research study on heart disease, which I linked above, implements one way to get around this problem. It looks at excess loss of life, in terms of person-time. If someone died 10 months earlier than expected, that is negative in the data of person-time - even though this death is neutral after 10 months for excess deaths.

In conducting excess deaths analysis, the analyst has to catch the sweet spot. Too early, and there isn't enough signal; too late, and the signal has faded away.

***

Before ending, I'm issuing a prediction. In 2021, deaths from diabetes, dementia and heart disease will be abnormally low. This reflects the other side of comorbidities. Some of the people who would have died from those diseases have died from Covid-19. (According to this link, the top comorbidities for Covid-19 are: respiratory disease, hypertension, diabetes, dementia, heart disease, renal failure.)

 

 

 

 

Comments

[Michael Droy]

I'm issuing a prediction. In 2021, deaths from diabetes, dementia and heart disease will be abnormally low.

Maybe. Certainly the excess death counts I have seen in data have been very low or negative for some time, which perhaps confirms that, the effect already happened in the Summer in Europe.
I doubt actually that we will see the effect last long though.
For example in Nursing homes, one study found that while death occurred a mean time of 14 months after arrival, median death was just 5 months. When people are close, they are very close. I doubt if Covid advances death by more than 6 months for more than a handful of people (including some of those undiagnosed with serious illnesses by doctors, but identified by the grim reaper).

But the real issue is the effects of Covid restrictions on lives. How has the Covid effect on the Economy (lower wealth, unemployment, depression, murder) and the Health Services reduced lives.
An undiagnosed cancer might cut 20+ years of a patient's life. Life expectancy has gone up in many countries 10 years in the last 30 for good reasons.

So it may well be that we see increased excess deaths over the medium to longer term - out to 5 years I would guess.

QALYs are the proper measure. Quality Adjusted Life Years are the measure of how long a patient can expect to live in years adjusted for life quality (100% if fully active, less if not). In Budget-limited health services (NHS or most of Europe) they are used to calculate health policies. How much does this drug or operation cost and how many additional QALYs does it give? And does that match our targets and if so for which types of patients. Hip replacements do very well boosting both years and quality. Other procedures are deemed not suitable for the aged or obese.

So QALYs are the basic unit of health economics, and civil servants and economists have been calculating them for decades. There are very clear estimates for health service availability and for economic factors such as wealth or unemployment.
A QALY analysis to judge Covid policies is a natural civil servant task.
Ask why you have never heard of a QALY analysis?

When the big post-Covid inquiries happen to investigate Government Actions, failure to publish a QALY analysis will be top of the list.

[Kaiser]

MD: QALY sounds great in theory but it belongs to one of those metrics with a zillion components that are impossible to measure, and impossible to validate. So, it's hard to execute in practice. I'd support doing it, but make sure all assumptions are posted. I'm sure private insurers also do such calculations - they are just done in private in the name of profits.

The other challenge is the differential value of human life that underlies such an analysis. Some people believe their lives are worth more than others but other people believe all lives should have equal value.

Also, economic wellbeing and public health are not zero-sum.