A new approach to pandemic control by informing people of their social distance from exposure

Po-Shen Lo, a mathematician who works in graph theory, writes about a new approach he devised for pandemic control. He writes:

The significance of this new approach is potentially very high, because it not only can improve the current situation, but it would permanently add a new orthogonal tool to the toolbox for pandemic control, which works without vaccines or pharmaceutical treatments.

It’s a completely new type of phone app (www.novid.org), which resolves deep flaws in “contact tracing apps.” Functionally, it gives you an anonymous radar that tells you how “far” away COVID has just struck. “Far” is quantified by the graph-theoretic distance in the graph of physical relationships, automatically collected by the app via short-range communication with other people’s apps.

The simple idea flips the incentives. Previous approaches focused on controlling you, preemptively removing you from society if you were suspected of being infected. This new tool lets you see incoming disease to defend yourself just in time (e.g., by temporarily upgrading to a mask that protects you from others, as opposed to a cloth mask that mainly protects others from you). This uniquely aligns incentives so that even if everyone does what is best for themselves, they end up contributing to the whole. This solves the “tragedy of the commons.”

This could be important because (1) the COVID vaccines are being challenged by evasive new variants, and (2) at this rate, the next pandemic will paralyze the world again until new vaccine development and distribution.

Here’s the technical report, Flipping the Perspective in Contact Tracing.

It makes a lot of sense to me. You can read the report for details, but the key points are: (1) it’s measuring physical connection, not social connection (a link is established when two smartphones are within a specified distance of each other for a specified length of time), (2) it should work even if many people don’t enter their COVID-positive status themselves, and (3) Po-Shen tells me they’re trying it out now in two universities, so it’s a real thing. And here’s the general theme:

Instead of asking everyone within distance 1 of a positive case to quarantine, it tells everyone how far away the new positive cases have struck in their physical interaction network. This reversal changes the nature of the intervention, from one which “protects others from you” to one which “protects you from others.” Through that flip, the incentive structure also reverses, as users are given the opportunity to protect themselves before it is too late. Suddenly, users prefer false positives over false negatives (“better safe than sorry”), which is the opposite of the situation when they use apps that ask them to quarantine (the culturally unfamiliar “guilty until proven innocent”).

I find Po-Shen’s description persuasive, without my knowing enough about the details to say more than that. Speaking more generally, I like the combination of local data collection, centralized analysis, and individual decision making. This is an idea we’ve been interested in for a long time; see for example here.

There are also some interesting statistics problems here involving how the method performs as the data degrade in quality, how much this would be expected to improve outcomes compared to alternative procedures that don’t use the physical contact information, and how this procedure can be evaluated as it gets rolled out in different places.

P.S. Tyler Cowen beat me to this one by a few months. Po-Shen said that they’re doing it at Carnegie Mellon and Georgia Tech and that they have a critical mass of users. This would suggest that the way to go would be for a school or otther institution to get lots of people on it, and then the network effect works in a positive way, because if others are on it, there’s a benefit for individual users to join too.

P.P.S. Po-Shen responds to several concerns in commments.

30 thoughts on “A new approach to pandemic control by informing people of their social distance from exposure

  1. I see this as SUCH a good idea.

    I think it is almost universally accepted that people (largely) make decisions for emotional/psychological reasons, and that these often outweigh “data” or “scientific” based reasons.

    Here, the Po-Shen finds an extremely clever way to make the two compatible!

    I think the remaining barrier is convincing people that they aren’t being identifiably “tracked” through these apps. Aslo, we need to buy everyone a phone with NFC (e.g. I don’t have one :( )

    • Thanks for sharing your feedback. Actually, we don’t need NFC. Just Bluetooth is enough. This works on Android phones since “Marshmallow” (released 2015), and iPhones back to the 5s (released 2013).

      Regarding the “tracking” comment, we’ve discovered that the main issue for hesitation wasn’t tracking, but the fact that all of the other app attempts gave zero (or negative) direct selfish value to the user. The other apps would not tell you anything until they told you that you were already exposed to COVID, and that you should self-quarantine to protect other people from you.

      In our existing deployments, we discovered that once people learned that our new approach gave them a positive selfish value proposition (it lets you see the threat coming from afar so that you can defend yourself in time), they typically were curious about what data we collected, and most people were satisfied by the fact that the app doesn’t ask you for your name, phone number, or email address, and doesn’t use your GPS coordinates.

      At that point, the app falls into a similar category as any other app/website with a positive value proposition, where the user makes their decision on whether they think that the positive value proposition is worth the fact that we are constructing a pseudonymous interaction network. In many cases, user feedback was then that this was anyway much better than most of the other internet apps they used, as many of the other ones would even know their email address or name. For example, simply by using a smartphone, you already cede far more personal information to Apple and Google. Yet for better or for worse, people still use smartphones because they enjoy the positive value proposition. :)

      • This is very good on the privacy front. I have found that since apple’s app store started reporting what data was personally identifiable I was more likely to download an app that reported what data it used (even if it seemed like a lot) versus an app that provided no information.

        My only problem with tracking apps like this (and the general one in my state) is I hate having bt on because of battery impacts… though as you explained here the fact that this version of tracking provides the user a personal benefit also helps overcome hesitancy related to that. If I knew more about these kind of data I’d have interesting questions about how it works when people intermittently turn their BT on and off and how that affects the system you have in place. I’m sure that’s probably explained somewhere, but I am too swamped to look for it!

        Very interesting approach here!

      • Thanks for the reply Po-Shen.

        I realized the NFC thing after I read the full article. Unfortunately, one cannot edit comments on this site (I should have posted a reply to my own comment).

        FWIW – I have shared this with almost everyone I know as I really do see it as a brilliant idea.

        • Thank you! I am extremely interested in propagating the new concept. It is rare to discover a categorically new approach for fighting disease. Might you know other public intellectuals, bloggers, or podcasters who might be interested in discussing this with me? I’ve been analyzing this new approach from many different angles over the past months, and it appears to be very robust. It should change human civilization’s playbook for all future pandemics as well.

          Regarding the Bluetooth question, we have actually optimized our app for battery consumption. On iPhones, you can see how much battery is used by various apps, and NOVID uses about half as much battery as Apple’s own Exposure Notification system. That’s because we realized that it was relatively unimportant to detect phones that were around you for only 15 minutes, and therefore we do not need to scan as frequently. On my test devices (5 year old iPhones), I can leave the phone running NOVID and sitting on a bench for about 3 days without needing a charge.

        • > It is rare to discover a categorically new approach for fighting disease.

          I read most of the article, and the approach looks interesting, but I think you’re overstating your case. This is apparently a categorically new approach, but in the next paragraph you compare it directly to Apple’s Exposure Notification system. So what does categorically new mean there?

          And then statements like this (quoted from above):

          > This reversal changes the nature of the intervention, from one which “protects others from you” to one which “protects you from others.” Through that flip, the incentive structure also reverses, as users are given the opportunity to protect themselves before it is too late. Suddenly, users prefer false positives over false negatives (“better safe than sorry”), which is the opposite of the situation when they use apps that ask them to quarantine (the culturally unfamiliar “guilty until proven innocent”).

          Seem to be directly contradicted by different parts of the paper:

          > In order to model the impact of a signal delivered by our approach, one would need to estimate the probability p1 that the recipient of a signal of definitive nearby infection takes action to avoid infection, and the probability p2 that their action, if taken, interrupts a would-be transmission. It is an added bonus that there is also some probability p3 that they tell people nearby. Since the purpose of this article is to introduce our alternative approach, we leave it to future behavioral science research to precisely estimate the actual values of these parameters.

          Sure this approach gives people access to different data, but don’t say “Suddenly, users prefer false positives” as if this is some phenomena you have measured and characterized. There’s some serious story-time stuff happening here.

          I’m not even sure I really see the incentive change as clearly as you do. I like the graphs, but stuff like this:

          > And, with proper communication and education, our p1 can be increased because it is often directly in the user’s self-interest to protect themself; on the other hand, even with more education, it will be extremely difficult to convince people to substantially inconvenience themselves at low infection risk, unless there are no alternatives.

          What? So p1 in the hypothetical model of infection can be increased cause of user self-interest, but even with more education we can’t have users inconvenience themselves? Who said inconvenience themselves? Why can’t users of other apps be self-interested? This part is very unclear.

          Anyway, I like the graphs and this technique does seem interesting. The ultrasound technique is pretty cool too.

  2. I’ve read that in South Korea, they’ve been doing something similar to this utilizing phone cell data, except it was primarily focused on physical proximity. Apparently people were getting texts updating them if someone withing certain a certain distance (say who lived in their building two floors up} was identified as COVID positive and where that person had moved about.

    I can only imagine how much easier it would have been for me to navigate the pandemic if if had that kind if information. While I can get fairly rough estimates of the trend in infections in my county, and even rougher estimates of the trend in infections in my “hamlet,” I have no idea where, within my community, someone who was positive might have traveled.

    Imagine if I could have known if, say, no one who was identified as positive might have entered supermarket X in the past week, but someone who was identified as positive was shopping in supermarket Y yesterday.

    Of course, that kind of approach wouldn’t work nearly as well on a densely populated city with a particularly high positive rate – as the degree of differentiation would likely be almost nil. And I understand that people dying to put in their tri-cornered hats and see themselves as freedom fighters can’t swallow that level of “government tyranny” – so what worked in South Korea never had a shot if existing here.

    • > Imagine if I could have known if, say, no one who was identified as positive might have entered supermarket X in the past week, but someone who was identified as positive was shopping in supermarket Y yesterday.

      Would this information be useful? I would think this type of data would only be useful in real time. Like is there someone who tested positive in the supermarket right as I’m about to go in? But if you have even a few hours of lag then I don’t really see the use. Presumably by the time you go in that person will be gone unless they work there. But I’m not sure that’s a realistic scenario either because if the app knows you’re positive then presumably your employer would know too, right?

      • For months, piles of case reports have been accumulating; many of which, because of the details within, are potentially useful for inference — in the aggregate — to differential risk; that is: in a given region, activities and sites can be ranked in order of their apparent risk of transmission. The data may not tell a story any different from what we presume (stay away from church choirs and meat-packing factories). But then again it may. We do not know what story it would tell unless we attempt to elicit that story from the actual data; as it evolves; and wherever it evolves. I see great numbers of employee-cases at “big box stores” on the regularly updated county website here. Is this merely a reporting artifact (big box stores are more systematic in their reports to the county than ‘little box stores’)? Is this merely a phenomenon of the larger samples involved — so that counts drawn from those pools are closer to the overall prevalence in the county as a whole? Questions Questions Questions. Everyone has an “opinion” — but the *data* would tell a story, were it to be queried seriously. Certainly now that it’s mountainous.

        • Definitely! I wonder what the bottleneck is for this type of aggregate analysis. I would have thought that at this point – even considering early troubles with contact tracing, etc. – we would have more information on the relative and absolute risks of different activities. I mean we know general things like it’s safer outdoors than indoors but we’ve known that since the summer. Like is transmission driven by workplaces or small group gatherings? I mean it’s probably a bit of both but it would be useful to know the relative impacts and how they have changed over time.

          Is the bottleneck the data? I think contact tracing data is collected at the county-level so maybe it’s hard to aggregate across counties? Whether that be because of administrative/legal reasons or differences in data formatting. Or is it the analysis that’s difficult? This is not my field so although I presume it’s a difficult task, I don’t know if it’s intractably hard or a lot of work hard.

          Or maybe we do know some of these things (on an aggregate scale and not from county-specific reports) and I’ve just missed it? Very possible as I haven’t been paying as much attention the last few months.

      • Michael J –

        > I would think this type of data would only be useful in real time. Like is there someone who tested positive in the supermarket right as I’m about to go in?

        Yes, that’s a fair point. I kinda started thinking about this when surface spread was much more of a concern and haven’t updated my thinking.

        And yah, maybe information about something like the quality of a store’s ventilation system would be more useful than whether someone who was in there yesterday tested positive.

  3. Could be somewhat useful for phone manufacturers, not so much for infections. Using the distance b/w phones as a proxy for virulence is quite a stretch.

    Time spent in proximity of another phone tells us nothing about the major variables responsible for transmission (position, facing each other, talking or not, wind, air circulation indoors, viral load, viral susceptibility, shedding potential, etc.) The problem is not so much that all those other, crucial metrics, phones can’t capture, but behavioral changes they may introduce.

    Reminds me of Andrew’s good example of people measuring biceps circumference and reporting ‘upper body strength’ :-).

    There is so much unknown about this (or any other) virus, that is mind-boggling. The only metrics worth looking at are hospitalizations and deaths, but those are always after the fact.

    Instead, the focus could be on things we know for certain help a lot. For example, it’s been over a year into pandemic and public still can’t buy 3M real N95 masks. Why? Constant screaming from media how it should be reserved for medical workers meant that supply met demand and we got stuck. Just like with toilet paper. Ridiculous. A defense production act could have been used and everyone could be using the real thing, instead of a placebo gauze, cloth or other nonsense that simply doesn’t work.
    I would even go far and claim that the safest behavior we had was before the ‘masks'(cloth, surgical) were recommended, during the lock-downs, a year ago. Everyone kept distance and didn’t talk or linger around much. Now you see people yapping two inches apart from each other for half an hour while steaming through a piece of non-fitting cloth (surgical or not). But, hey, they are safe because they are following great CDC guidelines.

    Only annual vaccines are going to solve this, because people obviously misplaced brain manuals.

    • Your first point was my first thought on reading this. An outdoor contact is several orders of magnitude less risky than an indoors one, all else equal, and it’s not immediately obvious to me that this is adjusting for that. And even those non-fitting cloth masks do plenty to reduce transmission, but there’s no way to measure “were they appropriately masked” as a product of phone distance — and on and on it goes. I imagine this approach is very good at gathering data; I don’t know I’d believe that it’s particularly good at collecting information.

      • Thanks for your points. They actually reveal one of the key new ideas here.

        You are absolutely right that if you just know that two phones were around each other for 15 minutes (which is what all the other apps try to do), then that’s not a good way to estimate whether transmission occurred. Most people were focusing on measuring time and distance between phones, just as you say. However, we change the main “space” in question from physical space to network space. Specifically, we are most interested in anonymously detecting long and frequent interactions between people, and then telling you how many such long-and-frequent interactions separate you from the new positive cases. These correspond to relationships like living in the same house as someone (likely unmasked for hours), working in the same office area as someone else (likely only wearing cloth masks but in recirculated air day in and day out), or regularly having lunch with someone (hard to do with a mask on).

        By changing the space in question from physical space to network space, phones suddenly become much more useful, because the data recorded by phones is much more able to detect long-and-frequent interactions. Of course, this won’t protect you from every harm. Neither will your mask. On the other hand, this adds new power to detect the virus transmitting along the superhighway of long-and-frequent (and typically weakly protected) relationships.

        • This was an exceptionally clear and concise explanation of what makes this different. One can argue whether people will use the information in an effective way to protect themselves, but the fact that this is DIFFERENT information than is currently being considered is clear.

          Also thanks for the response RE bluetooth use!

  4. The impact of social measures on disease propagation is hard to tease out. Would government information alone have decreased tobacco use, or was it ever increasing taxes on tobacco? Will people do the smart thing based on information, or do we need some actual coercion? My network of friends and acquaintances is not composed of 100% rational actors. I suppose that the infectivity of the disease has a great influence on this question, but even a disease such as HIV had 36000+ cases in 2018 (many in especially vulnerable populations). The lowest rates for HPV vaccination are in Florida, Georgia, Texas, etc, which are also, perhaps not coincidentally, the states with the highest HIV rates. Will information alone improve this, or are more heavy handed measures needed?

    • >> Would government information alone have decreased tobacco use, or was it ever increasing taxes on tobacco?

      I kind of doubt either had the primary role. I think it was more a social shift.

      I’m too young to have directly observed it, but I wonder how much of it was in fact a “side effect” of broader social trends such as changing attitudes toward masculinity (IIRC smoking among women hasn’t dropped, and may even have risen).

      Social shifts are hard to predict ahead of time and even harder to cause *on purpose* (if it’s seen as purely “political” such an attempt will be opposed by ~half the population automatically, in polarized times like ours).

  5. Dallas County, TX is using a similar approach in Dallas using the location of reported cases, called MyPCI. Basically quantifies risk based on the proximity of your address and located cases close to you without reporting to you the close case location.

Leave a Reply to Joshua Cancel reply

Your email address will not be published. Required fields are marked *