By Gordon Hull
Early on in the Covid-19 pandemic, I dedicated a post (and a short follow-up) to the idea that our knowledge of Covid-19 is mediated by the indicators we have to represent it, and that those indicators are themselves epistemically tricky. In particular, there’s a difficulty in understanding “Covid incidence,” because of difficulties in translating from “positive Covid tests” to how many Covid cases are actually present in a given population. The standard shorthand way of addressing this has been to look at percentage of positive tests, with the guideline that unless this percentage is low enough, the number of positive tests likely significantly under-represents the number of cases. The situation reminded me of the ambiguity of “malaria cases” in parts of Africa, where the dashboard tally of number of cases does not transparently communicate the number of people who actually have malaria.
The last couple of weeks have indicated the extent to which there’s a further STS point lurking. The standard test for Covid-19 is a PCR, which has the advantage of being highly sensitive. It has the disadvantage of requiring complicated reagents and a lab, and so one reason (and the only even faintly forgivable reason) for the lack of testing in the U.S. is bottlenecks at the lab and supply level. A number of folks have been arguing that there isn’t enough capacity in the system, even if it were done well, to test as many people as need testing. Certainly in the status quo, where lots of people have to wait a week to get their test results, the test is useless for a lot of purposes, since they’ll probably no longer be contagious by the time they get the test result. The test, in other words, doesn’t produce any actionable information.
One solution would be to embrace cheap saliva testing. This testing costs a tiny fraction of PCR testing, can be done in any clinical setting (and there’s a push to develop at-home-ready ones), and generates results very quickly. The only catch is that it’s not nearly as sensitive – catching maybe 60% of cases. There’s thus a debate over whether the lower sensitivity will be made-up for by the increased capacity. My sense is that it is – if you can test everybody in a classroom or a dorm a couple of times a week, you’re going to catch a lot more cases of Covid, a lot earlier, than in the status quo, where you aren’t catching much of it at all, and much of that too late.
A piece in the NYT the other day underscores the further point that the sensitivity question is baked into the PCR test itself. In addition to being too slow, that is, perhaps the PCR is too sensitive:
“Some of the nation’s leading public health experts are raising a new concern in the endless debate over coronavirus testing in the United States: The standard tests are diagnosing huge numbers of people who may be carrying relatively insignificant amounts of the virus. Most of these people are not likely to be contagious, and identifying them may contribute to bottlenecks that prevent those who are contagious from being found in time.”
As the article explains:
“The PCR test amplifies genetic matter from the virus in cycles; the fewer cycles required, the greater the amount of virus, or viral load, in the sample. The greater the viral load, the more likely the patient is to be contagious. This number of amplification cycles needed to find the virus, called the cycle threshold, is never included in the results sent to doctors and coronavirus patients, although it could tell them how infectious the patients are. In three sets of testing data that include cycle thresholds, compiled by officials in Massachusetts, New York and Nevada, up to 90 percent of people testing positive carried barely any virus, a review by The Times found. On Thursday, the United States recorded 45,604 new coronavirus cases, according to a database maintained by The Times. If the rates of contagiousness in Massachusetts and New York were to apply nationwide, then perhaps only 4,500 of those people may actually need to isolate and submit to contact tracing.”
This is jarring – and one possible solution depends on the test calibration. Most tests return a positive if it can find the virus within 40 cycles. If you reduce that to, say, 30, you’ll find a lot fewer cases and, potentially, generate more useful information because you'll not be swamped with so many false positives. The problem is endemic to reducing highly granular information about viral load (and nb we are not really sure what viral load triggers either infection or contagiousness) to a binary dashboard indicating infected/not-infected. A lot of the discussions of saliva tests compare them to pregnancy tests, where a test strip would give you a "+" or a "-." That's appealing and probably describes how the test would look, but of course there's a significant disanalogy: you're either pregnant or not in a way that's considerably less complicated than "you have Covid" or not. More generally, though, you can have a more specific test (fewer false positives) if you decrease the sensitivity (home pregnancy tests have to be confirmed, after all), and that calibration is generally unavoidable in diagnostic testing. Indeed, debates about the relative merits of sensitivity vs. specificity in testing are well-known from mammography. It seems to me that there’s a couple of points to underscore here in the context of Covid PCR tests.
First, to harp on the indicators point, the number of “Covid-19 cases” we have - the data behind all those graphs of the pandemic - depends on how you calibrate the machine for sensitivity vs specificity, and that information isn’t even generally transmitted with test results and may not be consistent between labs. It’s hard to think of a better example of the STS point about how even biological concepts are inextricable from our sociotechnical systems.
Second, what counts as a “Covid-19 case” for you depends in part on what you want to do with the information. If you want to know who’s contagious in order to quarantine or contact trace, then you’re going to set the PCR apparatus differently, look for how many cycles it took to find the virus (i.e., use a different indicator), or use more widespread but less sensitive testing. Excess sensitivity is the enemy here, because you'll both swamp labor-intensive contact tracing systems and overburden people with unnecessary quarantine. The less sensitive saliva tests, perversely, may more accurately indicate when somebody is contagious, because there needs to be more virus there for them to trigger a positive. And, even if they miss some cases, that may be recompensed by the shorter list of people to contact trace. If you don’t have access to that better information in some manner, you have to give up on contact tracing and default to the old maximin strategy of quarantining large numbers of people (perhaps up to a general lockdown), many of whom probably don’t need it. Such strategies as general lockdowns and blunt quarantines are born of ignorance. That in turn matters because the blunter the measure, the less socially sustainable it is. Anybody who’s got kids at home doing parent-directed homeschooling exciting asynchronous independent learningTM sees that point immediately.
On the other hand, there are also times when “Covid-19 case” needs the sensitive version. One of the Times article’s main sources, Michael Mina, in fact, wrote a Twitter thread clarifying his position, underscoring that “I do NOT want to change the PCR threshold to call someone with low viral load negative. This was unclear in the article. I simply want to use all the information in the most informative way to guide downstream actions.” That is, we should use the data about viral load to inform decision-making; I just want to underscore that it matters what you're trying to decide. Certainly there are clinical situations where you’d want to know for sure if someone was suffering from Covid-19 (vs., say, the flu) for example. But other times, you need the best available indicator of current contagiousness. And the PCR test as currently configured may not be the best way to learn that, even if we could scale up its use sufficiently.
Recent Comments