A Conversation with John Ioannidis


The COVID-19 pandemic has been a testing time for the already testy academic discourse. Decisions have had to be made with partial information. Information has come in drizzles, showers and downpours. The velocity with which new information has arrived has outstripped our ability to make sense of it. On top of that, the science has been politicized in a polarized country with a polarizing president at its helm.

As the country awoke to an unprecedented economic lockdown in the middle of March, John Ioannidis, professor of epidemiology at Stanford University and one of the most cited physician scientists who practically invented “metaresearch”, questioned the lockdown and wondered if we might cause more harm than good in trying to control coronavirus. What would normally pass for skepticism in the midst of uncertainty of a novel virus became tinder in the social media outrage fire.

Ioannidis was likened to the discredited anti-vax doctor, Andrew Wakefield. His colleagues in epidemiology could barely contain their disgust, which ranged from visceral disappointment – the sort one feels when their gifted child has lost their way in college, to deep anger. He was accused of misunderstanding risk, misunderstanding statistics, and cherry picking data to prove his point.

The pushback was partly a testament to the stature of Ioannidis, whose skepticism could have weakened the resoluteness with which people complied with the lockdown. Some academics defended him, or rather defended the need for a contrarian voice like his. The conservative media lauded him.

In this pandemic, where we have learnt as much about ourselves as we have about the virus, understanding the pushback to Ioannidis is critical to understanding how academic discourse shapes public’s perception of public policy.

Saurabh Jha (SJ): On March 17th, at the start of the lockdown, you wrote in STAT News cautioning us against overreacting to COVID-19. You likened our response to an elephant accidentally jumping off a cliff because it was attacked by a house cat. The lockdown had just begun. What motivated you to write that editorial?

John P.A. Ioannidis (JPA): March seems a long time ago. I should explain my thinking in the early days of the COVID-19 pandemic. Like many, I saw a train approaching. Like many, I couldn’t sense the train’s precise size and speed. Many said we should be bracing for a calamity and in many ways I agreed. But I was concerned that we might inflict undue damage, what I’d call “iatrogenic harm”, controlling the pandemic.

To answer your question specifically, I wrote the piece because I felt that the touted fatality rate of COVID-19 of 3.4 % was inflated, but we had so limited data and so much uncertainty that infection fatality rate values as different as 0.05% and 1% were clearly still possible. I was pleading for better data on COVID-19 to make our response more precise and proportionate.

SJ: We now know that the infection fatality rate (IFR) is much lower than 3.4 %.  I’m curious – why did you doubt this figure? At the time, the virus created havoc in Iran and Italy. Hospitals in the richest areas in Italy rationed ventilators. Was a fatality rate of 3.4 % so implausible?

JPA: Small changes in the fatality rate make a dramatic difference in the number of deaths. 3.4 % is an entirely different universe from 0.5 %. Imperial College epidemiologists, using an overall IFR of 0.9 %, assumed that if 60-80 % of the population were infected, as would happen without precaution or immunity, 2.2 million Americans would die.

I’m a physician and epidemiologist with a fellowship training in infectious diseases. Though I felt that COVID-19 was a serious threat, I didn’t think it was Spanish Flu redux. COVID-19 wasn’t behaving like a “3.4 % fatality rate” pandemic. I doubted that widely quoted fatality rate, which is what the Chinese public health authorities told the WHO, because by March it was apparent that COVID-19 infection comprised a clinical spectrum, ranging from mild symptoms which could be managed at home, to severe lung disease which needed ventilatory support. The crucial piece of the epidemiological puzzle was the number of people who were infected but didn’t know they were infected because they had no, or very minimal, symptoms.

The presence of asymptomatic and mildly symptomatic people who are not detected changes the shape of the pandemic and should change our response to it, too. For starters, it means that the infection fatality rate – the fatality rate amongst the infected – will, by definition, be lower than the case fatality rate (CFR) – the fatality rate amongst known symptomatic people who get tested.

The second implication is that the infection is more contagious and has spread further than what we believe, which makes testing, tracking, and isolating infected people more challenging. Testing remains important but each day we delay rolling out mass testing, testing becomes less efficacious, and even less so when there are so many asymptomatics or people with mild symptoms who won’t seek testing.

Figuring the true IFR of a virus isn’t some petty academic musing. To be clear, distinguishing between IFR and CFR for a virus like Ebola is silly because its CFR is about 50 %. But when the CFR of a virus is less than 5 %, we must ask– what’s the true IFR? Does the CFR diverge much from the IFR? How many asymptomatic carriers of the virus are there?

SJ: Few would argue against better data. But decisions must be made with the data we have rather than the data we wish we had. The consequences of a delay in acting because we wait for better data can only be guessed, rather than proven, at the time. Nevertheless, inaction has consequences. Some felt that you were advocating that we do nothing in the pandemic until we obtained data robust enough to design policy.

JPA: That wasn’t my position, though I can see why people thought I was advocating inaction when I was actually asking, begging actually, for better data to inform our actions. The two – decisions and knowledge – aren’t mutually exclusive. We can design policy on imperfect information yet keep gathering evidence so that our approach is fine tuned. A decision such as an economic lockdown should be assumed provisional, awaiting more research and better information. Of course, we can’t know everything there is to know about a novel virus in the beginning. Inaction is a false choice. What we’re choosing between is an immutable decision and a decision updated by emerging evidence, rather than between inaction and gathering evidence.

SJ: Let me ask you, frankly. Did you support the lockdown?

JPA: Let me answer, frankly. Yes. But only as a temporary measure.

SJ: So, you’re not against locking down the economy?

JPA: By February, we missed the window for nipping the novel coronavirus in the bud. Had we acted earlier, with aggressive testing, tracing, and isolating, like the South Koreans, the Taiwanese and the Singaporeans did, the virus wouldn’t have spread as wildly as it did. The biggest lesson from this pandemic is that the costs of delaying controlling the infection can be substantial. Act decisively in haste or repent at leisure.

Once we missed the boat, the lockdown was inevitable. I say “inevitable” grudgingly because I don’t think it should have reached that eventuality.

SJ: The situation would certainly have been different had the extent of the spread been identified in January, and the infection was controlled. If I understand you correctly, given our situation in March, as avoidable as it could have been, and our state of knowledge at that time, you supported the lockdown.

JPA: That’s correct.

Once the country was locked down, I felt we should be focusing in minimizing its duration. I view “lockdown” as a drug with dangerous side effects when its use is prolonged. It’s an extreme measure – a last resort, the nuclear option.

A country should be locked down not a minute longer than absolutely necessary. We have to keep assessing its risk-benefit calculus, by collecting and analyzing data, making sure we’re measuring the denominator accurately, and finding vulnerable and not vulnerable sub-groups.

SJ: I don’t mean to play “gotcha.” But isn’t what you’re saying contradictory? You didn’t believe that COVID-19 was a “3.4 % fatality rate” pandemic but you also supported the lockdown, which you, rightly, call an “extreme” measure.

JPA: If the fatality rate were truly 3.4 %, I’d have myself tied like Ulysses did, perhaps to my refrigerator to avoid ever getting out of my house. I’d want an even stricter lockdown.  One of the challenges in science communication is downgrading the threat of an infection, which you believe is inflated, without making it sound harmless. That I didn’t think that COVID-19 was that dangerous didn’t mean that I thought it was harmless.

SJ: But you compared COVID-19 to the flu. That comparison irked many doctors, particularly those in the frontline, who felt they were being gaslighted. Doctors from Lombardy, New York City (NYC), Seattle were seeing jam packed ICUs, high mortality rates in the ICU, multi-organ failure, and ventilator shortages. They were overwhelmed. They had never seen so much carnage caused by an infection, certainly not by the flu. Surely, we don’t need the denominator to figure out that COVID-19 isn’t just the flu. Surely, the numerator speaks for itself.

JPA: When conveying the severity of a novel virus, it’s useful anchoring to past infections for perspective. The seasonal flu is a natural choice for comparison. I agree “just the flu” sounds dismissive, even insulting to healthcare workers, because it sounds like the common cold. The seasonal flu isn’t “just the flu” either. It actually kills 350, 000 to 700, 000 people a year worldwide. In the USA it kills 30,000 to 70,000 people per year, and would kill even more if we didn’t vaccinate healthcare workers and half the population.

I don’t think comparing COVID-19 to the seasonal flu is unscientific, but that comparison must be nuanced. COVID-19 is a strange beast. It’s way more dangerous than the flu in the elderly and in those with comorbidities. Yet the flu is more dangerous than COVID-19 in children and young adults, even allowing for the fact that COVID-19 causes Kawasaki-disease-like syndrome in some children. Again, we face a communication challenge – how do we convey the severity of a virus which is both more dangerous and less dangerous than the flu? If I emphasize the less vulnerable group, I’ll be accused of being flippant about the virus. Yet if I focus only on its devastation in the most vulnerable group, I’m not painting the true picture.

Even though I spent lot of effort nailing the precise IFR of COVID-19, any single number IFR is misleading, because the average fatality rate hides the heterogeneity of risk. Once we figure out that the virus, on average, isn’t as bad as we thought, the next step is identifying the low-risk and the high-risk, i.e. risk stratifying.

SJ: I’m going to challenge the mortality statistics of the seasonal flu you have quoted, which are widely quoted, and was quoted by Donald Trump, too – though its source isn’t fake news but the CDC. Aren’t these figures an estimation or projection? And isn’t it true that the deaths attributable to COVID-19 is derived more from direct counting than from an estimation and, therefore, likely to be more accurate?

JPA: It’s true that mortality of seasonal flu is an estimation. But this estimation isn’t science fiction. It’s derived from sound scientific principles. The data on seasonal flu (flu-like illnesses) is robust. We know much more about the seasonal flu than COVID-19.

Now, your point that we’re literally counting, as opposed to estimating, deaths from coronavirus is correct. But I’ll push back that this may not yield mortality figures as accurately as people think. Because of the attention on coronavirus, we’re better at knowing that a deceased person had coronavirus than had the flu. This means we’re good at knowing when someone died with coronavirus – but not necessarily that they died from the infection. We assume that dying with coronavirus is dying from coronavirus.

SJ: But many have died in their homes with no documentation of being infected. We have assumed that dying without documented coronavirus is not dying from coronavirus. Surely, misattribution of deaths to coronavirus works in both directions.

JPA: I agree. Which is why we need better data to understand this virus better. One point I want to emphasize – the misattribution is paradoxically greatest in the group most vulnerable to coronavirus, i.e. those with limited life expectancy. This group is most likely to die from COVID-19. Because of their limited life expectancy, this group is also likely to die from their non-COVID morbidities.

One way to better measure the impact of COVID-19 is measuring excess deaths, which is the death rate beyond what one usually encounters annually. Excess deaths comprise several groups – e.g. people killed by COVID-19 infection and people who have died because they didn’t receive timely care because they were afraid to go to the hospital, or because healthcare resources were focused on COVID-19 patients. The magnitude of the latter group will be more evident in years to come. Another group are deaths caused by the social and economic consequences of the lockdown, such as from suicides and alcohol and drug abuse. This number, which’ll also be evident in years to come, shouldn’t be underestimated. At a global level, consequences of lockdown-induced starvation, derailment of immunizations for lethal childhood diseases, and lack of proper management of tuberculosis are tremendous threats.  

SJ: In your editorial you said that the bulk of the mortality of COVID-19 was in people with limited life expectancy, rather than young people. You said “vast majority of this hecatomb would be people with limited life expectancies. That’s in contrast to 1918, when many young people died”. Some felt you were minimizing the live of the elderly. That you didn’t think that their lives were worth the economic consequences of the lockdown was because they’d be dead soon, anyway.

JPA: That’s an unfortunate misrepresentation of my position.

There is an age gradient of fatality with COVID-19. This fact has been shown in several studies. Not only is there an age gradient but a steep inflection point with age, around 70. The hazard ratios are striking. Age predicts mortality better than even comorbidities. This scientific fact can easily be hijacked by demagogues by calling people concerned about the negative consequences of the lockdown “heartless granny killers.” That isn’t helpful.

The fact that COVID-19 disproportionately affects the elderly, i.e. older people are more vulnerable, means that they need more precise and thoughtful protection. I have been advocating for more attention to and protection for elderly people, not less. Unfortunately, that’s not what happened. For instance, Andrew Cuomo, governor of New York, told hospitals to send infected nursing home residents back to their nursing homes, which was like putting out a forest fire with kerosene.  The same happened also in other states. This act alone may have caused countless deaths amongst nursing home residents. We failed to protect our most vulnerable, in part because of our “one-size-fits-all” approach. In Lombardy, there were disproportionate deaths in nursing homes. It is estimated that 45-53% of US deaths were in nursing home residents, and similar or even higher percentages were seen in several European countries.

We needed extra precautionary effort in high-risk settings such as nursing homes, prisons, meat processing plants, and homeless shelters. The corollary of having high-risk groups is that there must be low-risk groups, and low-risk people can continue working. We can’t treat everyone as “high risk” because then the high risk won’t get the extra attention and care they deserve. In our approach to controlling coronavirus we made no distinction between teenagers partying on beaches in Florida and debilitated, frail residents living in congested nursing homes in NYC. Our uniform approach was neither scientific nor safe.

COVID-19 is a virus which unmasks our social and economic fault lines.

SJ: You’ve criticized models for using faulty data in projecting the death toll. When the lockdown started there were only 60 deaths in the US. You projected 10, 000 deaths using an IFR computed from infected passengers on Diamond princess cruise ship. Yet today there more than 132, 000 deaths – the figure would likely have been even higher were it not for the social distancing/ lockdown we employed on March 16th. Though the mortality numbers are still much lower than the doomsday predictions of Imperial college, they do make your projections overly optimistic.

JPA: I never said that I knew that the death toll was going to be “10,000 deaths in the US”. How could I, in a piece where the message was “we don’t know”! The 10,000 deaths in the US projection was meant to be in the most optimistic range of the spectrum and in the same piece I also described the most pessimistic end of the spectrum, 40 million deaths. The point I wanted to emphasize was the huge uncertainty.

Now, it’s perfectly reasonable following the precautionary principle advocated by Nassim Taleb and basing our response on the worst-case forecast. But as scientists it’s not reasonable staring such huge residual uncertainty in its face and doing nothing about it. It’s our job to reduce uncertainty by collecting more robust data.

SJ: You calculated the IFR of COVID-19 using the published fatality rates in various settings. We’ll get to the methods later. For now, I want to focus on the result. By your calculation, the IFR ranged from 0.02 to 0.86 % with a median estimate of 0.26. Let’s take just one data point: NYC. There were 18, 000 deaths. Even if we assume the entire city, population of 8.3 million, was infected – a big assumption – that yields an IFR of at least 0.21 %. The lowest bound of the fatality rate of NYC is far higher than the lowest bounds of your estimate. Does this fact not challenge your calculation and the assumptions made in the calculations?

JPA: IFR is not a fixed physical property like the Avogadro’s constant.  It’s highly variable which depends as much on the virus as it does on us. Perhaps it depends even more on us than the virus. It depends how we interact with each other, how close we are to each other, who gets infected, who gets ill. As we know the virus more, we get better at dealing with it. The IFR is a shape-shifting moving target.

Which brings me to NYC. It certainly faced the infection courageously head on. Yet neither its experience nor its IFR can be generalized. At least three factors contributed to the high death toll in NYC: the disproportionate number of deaths in nursing homes because of a catastrophic policy blunder, the very compact nature of the city, particularly where the vulnerable populations live, and nosocomial spread of infection. Also, doctors were still learning how best to manage patients in the ICU and their approach to ventilatory support was probably too aggressive, in hindsight. I’m not blaming doctors. NYC was dealt a bad hand.

SJ: If the IFR is “shape-shifting moving target” – why did you labor so hard to measure it?

JPA: It’s still important knowing the mean and range, particularly if one wants to calculate the risk-benefit of different policies in different settings. We just can’t assume that the IFR of COVID-19 in NYC in April is the same as its IFR in Houston in July or the IFR of Singapore either in April or in July. Some hotbeds in NYC must have had IFR of 1% or more. Singapore has already detected 43,000 cases and had only 26 deaths, so the upper bound of its IFR is 0.06% and may be even smaller. The IFR of Houston in July is something that we can hopefully shape and decrease with precise actions. When we learn from history, when we understand the special circumstances of the past and ensure we don’t repeat the mistakes, hopefully the IFR doesn’t repeat itself.

Also, I showed in my methods of computing the IFR the huge diversity of IFR because of wide variation in seroprevalence estimates. It’s not just the final result that’s important, it’s the individual components which make the final number which are important as well.

SJ: Could you expand more on nosocomial spread of COVID-19.

JPA: Many patients were likely infected in hospitals by infected healthcare workers. This is understandably a controversial issue which people are reluctant to broach.

We don’t know the exact scale of the nosocomial spread but in several hard-hit locations it was probably not trivial. This happened because many infected healthcare workers, particularly those < 60 had no idea they were infected. Once again, it underscores how important it was understanding the extent of the asymptomatics and people with only mild symptoms to which they naturally pay no attention. They unwittingly and unknowingly infected patients.

Like nursing homes, hospitals house the most vulnerable. Only a handful of unaware infected healthcare workers would have been sufficient to allow the virus to spread and feast on patients in hospital. This happened even more prominently early in the pandemic, when precautionary measures, such as wearing personal protection equipment, weren’t universally adopted, and we had no idea how far coronavirus had spread.

Deaths are a lagging indicator of the extent of infection. By the time the first death from COVID-19 in the US was recorded, the virus had comfortably set foot in American society. By believing the virus was deadlier than it actually was, we underestimated how far it had spread, and thus allowed the virus to be more deadly than it needed have been.

SJ: You received considerable pushback for your piece. Has it changed your opinion of academic discourse?

JPA: Appearing on Fox may have infuriated some of my colleagues – but that speaks to the polarization in the US. I’m a data-driven technocrat. It’s unlikely I would fit well with “conservative ideology” (good grief)!

I welcome academic discourse and disagreement. I have no doubt that I know very little and that I make mistakes, but I am just trying to learn a bit more and to make fewer mistakes, if possible. I consider that people who criticize me with valid scientific arguments are my greatest benefactors. But the outrage propagated by social media is a force of its own, and destroys any intelligent discourse, civil or uncivil. Once the outrage gets going, platforms for academic discourse censor and the discourse just doesn’t happen. I was unable to publish my essay about nosocomial spread of COVID-19 in nursing homes and hospitals. I submitted to many outlets. I suspect the editors feared social media backlash against my raising an uncomfortable issue. Fear isn’t healthy for science.

Saurabh Jha is an associate editor of THCB and host of Radiology Firing Line Podcast of the Journal of American College of Radiology, sponsored by Healthcare Administrative Partner.