Technology sounds like an attractive solution to contact tracing, but apps are at best a minor supplement to a large effort. In this opinion piece, Lyle Ungar writes that “we should be taking best practices from call centers, where human callers are supported by chatbots and information systems, supplemented with privacy-respecting apps on people’s phones that allow them to share information more easily and accurately. In the end, contact tracing is not an app, but a combined effort between technology, human tracers, and the general population.” Ungar is a machine learning researcher and professor of computer and information science at the University of Pennsylvania.

Contact tracing is key to reopening society. Best estimates put widespread vaccination in the U.S. more than two years in the future, and we can’t safely resume public life until we can identify who has been exposed to COVID-19, test them for the disease, and isolate them if they are sick. The U.S. has far too few human contact tracers, with states planning to hire only a tiny fraction of the estimated 180,000 contract tracers needed. Contract tracing apps have been proposed as one way to mitigate this problem. People are worried about their privacy; they should be even more worried about whether the apps will help. Even expert articles in the Journal of the American Medical Association (JAMA), underestimate the challenges.

Most of the discussion of contact tracing focuses on exposure notification apps (i.e., TraceTogether, PwC), which use Bluetooth signals to identify individuals who may have been exposed to the novel coronavirus. The most widely supported protocols (by Google and Apple) respect privacy; they broadcast and receive random numbers from your phone, but don’t reveal your name or phone numbers. Google and Apple do not allow the apps to share geolocation or other private information. Such apps are only effective in relatively tight communities, such as universities, where high adoption rates can be achieved. In a general community, where adoption is voluntary, adoption rates are vastly lower (the highest adoption rates are 32% in Australia and 38% in Iceland) and so the apps are virtually useless.

If 30% of the people have an app, for any encounter, the chance of both the exposer and the exposed person having the app is 30% * 30% = 9%. This means that only one in 10 transmissions would be detected. Experts estimate that for app-based tracing to be useful, at least 60% of the population needs to use the apps, corresponding to 60% * 60% = 36% of the instances of transmission being detected. A little more than 30% of Americans say they would install such an app.

Location Tracking Apps

Less discussed are location tracking apps, (e.g. Resolve to Save Lives’ Citizens, Twenty’s Healthy Together, Care19, and the University of Washington’s Covidsafe). These apps require you to turn on location sharing. Then, if you are contacted by a tracer, the apps help you remember where you were over the preceding days, often (in a way which should respect privacy) looking up the locations you have been to and identifying them (“the Giant supermarket” or “your friend Alice”). The apps may also provide an interface to allow you to share contacts (name and phone number) of people whom you might have exposed with the contact tracers. (A different person will contact the exposed person than the one who talked to the infected person, to prevent information leakage.)

This sounds great, but it only helps if the person contacted trusts the contact tracer. Distrust comes in many forms, from African Americans not trusting the medical establishment to Trump supporters thinking the coronavirus is fake news. Location tracking apps are thus only as effective as the fundamental trust between the citizens allows.

Neither exposure- nor location-tracking apps do what countries from Israel to China do: take location data directly from cell phone providers. In spite of cell phone location data being widely sold by U.S. cell phone providers, it seems unlikely to be acceptable to the broader American public and is proving problematic in Israel. If privacy concerns were not dominant, this sounds like an effective approach of tracking location, as your cell phone provider knows exactly where you have been, and who else (with a phone) has been there. But in a big city or even a big office building, hundreds or thousands of other people are within the spatial accuracy of the phone GPS, and so would be flagged as possible contacts, even though they were never in the same room. Contact tracers then need to drill down to the contacts that might be real.

But things are worse than that. Contact tracing is useless without fast testing. The goal is to keep R — the number of people each contagious person infects — below 1, so that the population infection rate goes down exponentially, rather than up exponentially. The faster you find infected people, the less they will infect others. Speed matters more than accuracy. This dictates several choices that go against the intuition of most medical professionals. Each extra day needed to get test results for a person increases the number of people they will infect by about 30% (1 day/3.5 infections/day). (This is actually way more complicated since there are different rates for symptomatic and asymptomatic people, and you can go get “friends of friends.”)

Fast Reporting

People are infectious for roughly three to four days. If it takes that long to get test results, they are useless for contact tracing. If a test gives 10% false negatives, contact tracers will miss 10% of the infected people — except the ones who develop symptoms — which is far better than being a day late. (The tests generally don’t give false positives.)  Similarly, it is better to have fast reporting with some fraud than to have perfect but slow reporting. If 10% of the people who test positive lie and claim that they didn’t, it is the same as the test missing 10% of the cases; not nearly as bad as waiting a day to get the test results and start tracing. Many medical professionals are trained to do everything in their power never to be wrong. These intuitions do not align with the design decisions needed in a pandemic emergency. A sloppy list of people who may have been exposed to COVID today is useful; a precise list of people who were exposed last week is not.

“In the end, contact tracing is not an app, but a combined effort between technology, human tracers, and the general population. Without good faith and trust, contact tracing will fail.” –Lyle Ungar

To summarize, Bluetooth exposure notification apps are useless outside of communities with higher adoption rates — tight communities with similar social norms. Location tracking apps are attractive, but they require fast COVID testing, and lots more human contact tracers. A simple rule of thumb is that 1% of infected people die (the CDC and many studies put the rate closer to half that, which would double the number of contact tracers needed). So the number of people infected each day is 100 times the number that die, and each infected person requires roughly four contact tracers. So, at 500 deaths/day in the U.S., we have 50,000 new cases/day, and need roughly 200,000 contact tracers. And we need contact tracers who come from the communities being called, as they are more likely to be trusted. In New York City, less than half the people contacted admit to having been in contact with anyone they might have exposed to the virus.

Technology sounds like an attractive solution to contact tracing, but apps are at best a minor supplement to a large effort. We should be taking best practices from call centers, where human callers are supported by chatbots and information systems, supplemented with privacy-respecting apps on people’s phones that allow them to share information more easily and accurately. In the end, contact tracing is not an app, but a combined effort between technology, human tracers, and the general population. Without good faith and trust, contact tracing will fail.