Interpreting COVID Findings: Avoid These Dangerous Mistakes

F. Perry Wilson, MD, MSCE


May 13, 2020

Find the latest COVID-19 news and guidance in Medscape's Coronavirus Resource Center.

This transcript has been edited for clarity.

Welcome to Impact Factor, your weekly dose of commentary on a new medical study. I'm Dr F. Perry Wilson from the Yale School of Medicine.

This week, an absolutely huge study out of England gives us insight into who is at risk of dying with COVID-19.

But beware, folks: This is a preprint. No peer review has had its way with the study yet. Nevertheless, that hasn't stopped major news outlets from covering it, with some drawing potentially dangerous conclusions. Let's set the record straight.

This is a paper out of Ben Goldacre's lab. If you don't know Ben, he is an extremely well-regarded British physician and evidence-based medicine evangelist. His TED talk on bad science has over 500,000 views on YouTube and is well worth a watch. He knows what he's doing.

In fact, some people consider me the American Ben Goldacre. Those people are my parents. And that's it.

Source: F. Perry Wilson, MD, MSCE

This is a study of more than 17 million individuals in England who receive primary care from clinics using a standardized electronic health record system. The analysis was restricted to those with at least 1 year of data in the system to ensure relatively good capture of relevant comorbidities. The huge dataset was then linked to a registry of COVID-19 deaths (there were 5683 of those at the time the data were analyzed).

The real advantage over other studies here is clear: You're not looking at people who got tested or people with symptoms; you are looking at basically everyone. It allows you to ask the question: Who—in England, at least—is most at risk of dying with COVID-19?

The results confirm a lot of what we have seen from other studies.

The risk for death, for example, was markedly higher as age increased.


The COVID death rate was 26 times higher among those over age 80 compared with those between ages 50 and 60, for example. They also showed that the COVID death rate in men was substantially higher than that in women.


The authors honestly could have stopped here: There are people at higher risk; here are some things that make people higher-risk. But they went a step further than most other groups have gone. They put these risk factors in a multivariable model to provide a deeper level of understanding.

This is awesome, but be careful—this is not causality.

We all want to know what causes COVID-19 deaths, because it would imply that if we change that thing, we can reduce COVID-19 deaths—not just predict them. It's actually a really important difference.

Let me give an example.

Does high blood pressure cause COVID-19 death? In other words, would lowering blood pressure reduce COVID-19 deaths?

Well, people with high blood pressure were about 20% more likely to die of COVID-19 than people without high blood pressure.


But that doesn't mean it's the blood pressure that did it, right? Maybe people with high blood pressure tended to have other things that caused the death. Maybe they don't get as much exercise, maybe they have more heart disease. We can create little causal diagrams (also known as directed acyclic graphs) like this to sketch out those hypotheses.


After adjustment for comorbidities, notably including heart disease, the association between high blood pressure and COVID-19 death went away.


Does that mean that high blood pressure doesn't cause COVID-19 death?

Again, not exactly.

It depends what you adjust for. Hypertension does cause cardiovascular disease. Let's assume that having cardiovascular disease is a causal part of why people die of COVID-19. If that's the case, then if we reduced hypertension we could reduce heart disease and thus reduce COVID-19 death. That's causality. High blood pressure is causally related to COVID-19 death via its effect on cardiovascular disease. It's like how not wearing a seatbelt causes motor vehicle death via its effect on having you fly through a window.

When you are trying to figure out what causes what, it's really important that you not adjust for factors that may lie along the causal pathway. It's also really important that you do adjust for confounders—those factors associated with both hypertension and COVID death that don't lie along the causal pathway.

How do you tell the difference? It's really hard. A lot of it begins with simple clinical intuition.

Which leads us to this finding, which has garnered a lot of attention. Does smoking protect against COVID-19 death?!

This article from The Economist raises the possibility based on this very study. But actually, this result looks to be the result of adjusting for something on the causal pathway.

In the unadjusted analysis, current smoking was associated with a 25% increased risk for COVID-19 death. Makes sense to me. This is a lung virus; smoking hurts the lungs. I'm on board. But after adjustment, current smoking appears to have a 12% reduced risk for death.


A casual interpretation may be: All else being equal, smoking is protective. Or, in other words, start smoking.

This is obviously wrong, but let's go through why.

I'm going to hypothesize that smoking actually does cause an increase in COVID-19 deaths.


I'm further going to hypothesize that the mechanism is via respiratory disease.

The fully adjusted model adjusts for respiratory disease. So when we look at the adjusted risk of smoking, what we are saying is, basically, If all these smokers didn't have respiratory disease, they'd actually be more likely to survive COVID-19 than nonsmokers. Okay. But that's just like saying, If all these people who don't wear their seatbelt didn't fly through the window, they'd be more likely to survive car accidents. Fine, I guess, but definitely not a reason to take off your seatbelt.

The researchers actually looked at this, and found—sure enough—that all you have to do is adjust for respiratory disease to make smoking look protective.

The key take-home here is that you need to be super-thoughtful when you are trying to assess causality in observational data. "Adjust for everything" is not a good strategy because you run the risk of adjusting for things on the causal pathway, which totally screws up your results.

Now, Ben Goldacre (the British Perry Wilson) knows this and makes no claim to causality in the paper. But that doesn't stop newspapers, Twitter users, blogs, and the rest of the world from making that mistake.

Hopefully, now you won't.

Don't smoke if you got 'em.

F. Perry Wilson, MD, MSCE, is an associate professor of medicine and director of Yale's Program of Applied Translational Research. His science communication work can be found in the Huffington Post, on NPR, and here on Medscape. He tweets @methodsmanmd and hosts a repository of his communication work at

Follow Medscape on Facebook, Twitter, Instagram, and YouTube


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.