We don’t need scare tactics to develop a proper response to COVID-19.
NOTE: See a follow-up to this post on the IAPHS blog.
This will surprise some who know me, but about 20 years ago I was pursuing a future in full-time ministry.¹ At the time that meant a number of mentoring relationships, many of which consisted of jointly reading books on church leadership, then discussing how to make use of the ideas. But I kept noticing something in many of the books we were reading. The first few chapters would include some terrible “data analysis” that the authors would then leverage to justify whatever recommendations the book offered in its remaining chapters.
In most cases, while I’d find the aims of the books interesting, I’d invariably get hung up on the faults in the opening chapters used to justify those aims. I thought that there had to be a better way to gather and analyze data than what I was reading. And that lead me to grad school, hoping to learn how to better analyze organizational characteristics and change.
Why am I recounting those times in a post apparently about the novel corona virus (SARS-CoV-2)? Because the more information that’s out there about COVID-19, the more often it seems people are motivating interest in preparedness based on faulty information. As we were discussing this week in my class on diffusion, intervention efforts can be counter-productive when the basis of their arguments rely on fear-mongering tactics and inaccuracies.²
Pundits — and Twitter feeds — are too often leaning on misinterpreted data on epidemic prevalence and combining those with biased estimates of case fatality rates, to justify why COVID-19 preparedness should be prioritized. I really agree with that conclusion, but we don’t need faulty logic to get us there.
Uncertainty in Projections
There are lots of groups making projections about how widespread COVID-19 is going to become. The one source I’ve been paying the most attention to on this is Marc Lipsitch, and his frequently updated Twitter feed. In work you can find sourced all over the web by now, he estimates 40–70% of the population could ultimately become infected. However, a key component of those estimates is often left out by others quoting them. That figure assumes no control efforts. That’s where modeling should start, so I’m not criticizing using that assumption, I’m just pointing out that people have left it out of their use of his model, and that it likely isn’t the scenario we’re going to face.³
While Lipsitch’s models are important and well constructed, the data necessary to make his estimates are changing rapidly and dramatically at this stage of the outbreak. In fact, with relatively modest changes to modeling assumptions, that 40–70% range can be pushed down to 20%⁴ or lower, and even up to 80%. Some of those same modest alterations can be leveraged into substantially improved containment. As one of my friends put it when I shared those ranges with him, “So you’re telling me it’s going to end up somewhere between rare and ubiquitous?” Lest you think I’m just adding uncertainty onto Lipsitch’s work, here’s his own caution about such projections:
“Right now the quality of the data is so uncertain that we don’t know how good the models are going to be in projecting this kind of outbreak.”
Unfortunately, people prefer to cling to point estimates rather than the confidence intervals around those estimates. But modelers know that this uncertainty matters.
A Note about R₀
Discussions also seem to be over-relying on “R-nought” (R₀) as a means to summarize how contagious this novel corona virus is. R₀ indicates the number of additional cases each current infection can be expected to generate, on average. But R₀’s a tricky thing. It’s useful because it’s easy to interpret, conveys some elements of how contagious a disease is, and is simple to translate into control (drive it below 1, and an outbreak declines). However, it’s also widely known to have little association to the ultimate extent of an outbreak. And it’s complicated because it includes things beyond just the virus’s infectiousness.
This is because R₀ contains 3 elements — (1) the rate of contact between infectious and susceptible individuals, (2) the chance of transmission in each contact, and (3) the duration of infectiousness for each case. Because each of these components vary across — and even within — populations,⁵ R₀ is basically never constant for a particular virus. So, while WHO and others have been trying to estimate R₀ across outbreak locales (for good reason), it alone is just not sufficient for modeling projections.
Biases in Case Fatality Rate Estimates
Even if there’s substantial uncertainty in how many people are likely to get COVID-19, can we make sense of the severity of those infections that will occur? The answer to that one is complicated as well. The good news is that the vast majority of cases (as many as 80%) are going to be asymptomatic or have mild severity.⁶ Another substantial proportion demonstrate modest severity, though quite a few of those do require treatment, perhaps even hospitalization. However, the number I see getting the most attention is the estimate of how many of those infected will ultimately die from COVID-19 or its complications.
The case fatality rate (CFR) estimates for COVID-19 have been all over the place. Earlier this week, the WHO reported an estimate of 3.4%. Donald Trump caught a lot of flack for saying that number was an over-estimate.⁷ But, in reality, he’s probably right about that. In fact, here’s the WHO technical lead making that very point convincingly:
Why is that number likely too high, or why is the CFR so tough to estimate? CFR relies on confirmed cases. We can likely reasonably capture CFR’s numerator — the number of deaths. The devil is in the denominator. The number of people actually infected is going to be underestimated, thus inflating CFR. In fact, most infections are going to go undetected altogether (i.e., aren’t even going to be identified as cases). Even among those that are identified, we know that confirmed cases are likely to be biased towards those that are experiencing more severe symptoms. That’s why epidemiologists differentiate between the infection fatality rate — how many of those infected die — and the CFR. What’s more, the 3.4% estimate is actually the sCFR — CFR only among those who are symptomatic.⁸
Misunderstandings like this have lead many to assume this sCFR applies to those 40–70% prevalence estimates from above. It doesn’t. The 40–70% is an estimate of total infections. CFR applies to the subset of identified cases, and the estimated 3.4% sCFR apples to an even smaller subset of that (those cases who are symptomatic). If we incorrectly assumed the sCFR applied to the full infection scenarios we’d vastly overestimate the number of expected deaths. Even if we applied this number only to symptomatic cases, we’d still be overestimating likely mortality.
Why if we applied it to the right group is it still likely an overestimate? Well, we aren’t testing very broadly yet, and the more broadly testing is conducted, the lower those CFR estimates are. For example, in South Korea they’ve tested well over 100,000 people, and report a CFR of 0.6–0.7%.⁸
What’s more, there’s substantial age variation on the severity of symptoms and likelihood of mortality among COVID-19 cases. For example, while 20–29 year old residents account for nearly 30% of the cases in South Korea, none of the recorded deaths so far have been among that age group (source), and mortality estimates elsewhere are similarly skewed towards older populations (especially those above 70).⁹ While early deaths in the US were more common and among people slightly younger (50+), that’s because the earliest documented outbreak was among residents of a nursing home, who were already in poorer health than the general population.
All that to say, as things unfold, and the transmission becomes more broadly community-based, the mortality rate is likely to continue to come down. Most experts seem to think it will settle in around, or just below, 1%.¹⁰
How does this Matter for Preparedness?
The question then is how this shapes response(s) and preparations for reducing the spread of this corona virus and the impacts of COVID-19. In other words, simply because modeling and projection efforts likely contain biases and interpretations of those efforts have often been inaccurate, I am not suggesting that we can’t implement sensible efforts to slow new infections and mitigate effects for those who get COVID-19. In fact, some of that uncertainty banks on the reality that we will intervene and that those interventions will matter.
First, and foremost, it doesn’t really change what individuals people should be doing yet. The CDC and other outlets have provided guidelines that include the recommendations you’ve likely heard by now. Wash your hands regularly, and thoroughly. Stop touching your face. Consider social distancing practices, but it’s important to consider which ones will be effective. Stay home if you’re sick. Etc.
One thing we’ve talked about a number of times in my diffusion course this semester is what infectious disease interventions aim to do, especially in a wide-spread outbreak. While people often assume that prevention and containment efforts are primarily targeted at limiting total prevalence, that’s actually only part of what they seek to accomplish. Once community-based spread (i.e., transmission among the general population, not just from “external” sources) begins, interventions are often more effectively targeting “flattening the curve.” This simply means reducing the efficiency of spread to slow the timing of new infections, not necessarily assuming we can prevent them altogether.
That is, even if we can’t stop spread altogether, can we slow the timing of new cases? Why would that be helpful? A primary reason is that it can bring the overall caseload at any one time down to levels that may be more manageable (e.g., by the healthcare system) than if all those cases occurred at the same time. A vaccine isn’t going to be available any time soon, so this is the best option we have at the moment.
Importantly, there is a lot that governments and public health can do to flatten the curve. The good news is that we’re already seeing the effectiveness of a range of options. For example, a recent pre-print¹¹ suggests that aggressive non-pharmaceutical interventions (social distancing and self isolation, in particular) played an important role in keeping the situation in Wuhan from becoming much worse. What might others entail?
Testing is key. While the US botched the early roll-out, things are improving. Broad testing can provide modelers with the better information they need, and can help to calibrate responses (e.g., whether we should attempt link-tracing based containment, or focus on general population mitigation). Here in Colorado, they can now run 100+ tests per day, while the University of Washington was up to speed much quicker than elsewhere in the US and has 10-times that capacity. In South Korea, they’ve gone so far as to provide drive-thru testing. Surveillance in Singapore has been particularly aggressive. These efforts need to continue being scaled up.
Given that testing is increasingly showing growing caseload, and potentially already community-spread in some locations, social distancing is becoming even more important. One approach that’s picking up speed in my world is that some universities are ceasing in-person classes, recommending strong travel restrictions, and canceling conferences.¹² While facilitating the transition to online courses has obvious potential benefits, universities should absolutely consider who this might disadvantage. Again, I’m not arguing against disruptions — they’re going to be necessary — I just recommend that we enact them in ways that carefully consider the full scope of their potential impacts.
Just this week in my diffusion class, we were talking about the role of “health scares” in generating out-sized responses to outbreaks of infectious diseases.¹³ One of the key things I hope my students took away from that is how accurate information can even lead people to react in ways other than what’s optimal. This is especially true when the facts are changing as quickly as they are in an outbreak of something like COVID-19. But in a case with the potential for such broad-sweeping societal impacts as COVID-19, it’s even more important that we avoid facilitating undesirable responses by propagating false information. Our responses are better without that.
In a rapidly evolving situation, it can be especially important to to find sources to keep yourself informed. Here are some I find myself going back to regularly.
 Honestly, that fact perplexes even me at this point.
 There’s tons of literature out there on “health scares,” but I found this chapter was really useful for teaching many of the key concepts.
 That’s actually a re-assessment from Lipsitch himself.
 For contacts, see https://www.youtube.com/watch?v=vwVDJVbw10k. For infectiousness, see https://wwwnc.cdc.gov/eid/article/25/1/17-1901_article. For duration of incubation and infectiousness, see https://www.nejm.org/doi/full/10.1056/NEJMoa2001316.
 Though it may be the case that asymptomatic carriers can still transmit the virus.
 Trump tried to justify the fact that it’s a likely overestimate to tamp down projection scenarios and justify the administration’s slow response. As noted later in this piece. I’m not in any way agreeing with that implication. Or said another way, Trump can simultaneously be right about that being an overestimate and wrong about what it should mean for intervention efforts. Also, a “hunch” shouldn’t be used as a source in any of this.
 Unfortunately, the CDC stopped reporting the number of tests conducted on their information page.
 Because of age-based differences in mortality from particular conditions, things like CFR are often age-standardized, to account for different age structures across populations. Or, it can help explain substantially different CFR estimates.
 That’s still approximately 10-times the CFR for seasonal flu.
 This means it hasn’t been through peer review yet, and should be interpreted with that in mind.
 This approach works a lot better if announced before most of the attendees have already traveled to the conference location.
 The timing was completely coincidental, with the scheduling of the syllabus having been completed in early December.