Chances are, if you are listening to this around the time it was released, you’re listening alone. Right now the human species is conducting one of the most sweeping synchronized experiments of all time: physical isolation, restricted travel, shuttered businesses, our social lives moved online. Many people wonder whether all of this is truly necessary to halt the spread of COVID-19—or do not understand what differences there are between closed borders and closed schools and businesses, how epidemiologists derive the interventions they advise, and why it matters that we all stay home right now.
This week’s guest is Laurent Hébert-Dufresne, Assistant Professor of Computer Science at The University of Vermont’s Complex Systems Center, former SFI James S. McDonnell Foundation Postdoc and Research Fellow, and Editor of PLOS Complexity Channel. In this episode we discuss how network epidemiology studies contagions as they unfold across multiple scales, how co-infections (both biological and informational) change disease transmissibility, and how the best available research supports drastic containment measures.
Note that this episode was recorded on March 17th and we’d like to issue a blanket disclaimer that our understanding of the novel coronavirus pandemic evolves by the hour. We believe this information to be up to date at the time of publication but the findings discussed in this episode could soon be refined by more research.
Due to the pace at which the news is changing, we’ll ignore our normal schedule for the next few weeks and publish new episodes as quickly as we can. Please take a moment to subscribe wherever you listen to podcasts, and feel free to suggest questions for upcoming guests on Twitter or in our Facebook group.
Laurent’s Website & Twitter Page.
Read the papers we discuss in this episode at Laurent’s Google Scholar Page.
Visit our website for more information or to support our science and communication efforts.
Join our Facebook discussion group to meet like minds and talk about each episode.
Podcast Theme Music by Mitch Mignano.
Follow us on social media:
Twitter • YouTube • Facebook • Instagram • LinkedIn
Michael: Laurent, it is a pleasure to have you here on Complexity—in spite of the fact that you're on the show because we're in a pandemic. It's one of those mixed blessings that you get to step forward and speak with some expertise and authority here. This is our first remote recording, listeners, and it's appropriate to the social isolation, but there might be some latency issues in the recording for which we apologize. You've been looking at epidemic transmission, social reinforcement, contagion networks, and so on for years. You've given us a couple papers to look at. I'm not sure whether this is the right place to start, but let's start here. You did a paper with Sam Scarpino and Antoine Allard, “The Effect of a Prudent Adaptive Behavior on Disease Transmission,” from November 2016. That seems like the simplest place to start. Why don't you talk a little bit about the gross outline of this paper?
Laurent: Right. So this paper started during the Ebola epidemic in West Africa, in 2014. If you remember, back then, the news was much more focused on how different the Western African societies were than our own. There were a lot of conversations and new stories about traditional burial practice in West Africa and how that contributed to the spread of Ebola: funeral practices that involve touching the body a lot and that, of course, would spread Ebola, and those people are obviously going to be at risk. That's one of the first insights of network science and epidemiology is that you have individuals with just a lot more social connections, whether that's school teachers or funeral workers in West Africa, and those are going to be high risk individuals. If the disease is going to reach someone first, it's probably going to be them. They have so many connections.
Now, healthcare workers and funeral workers are in this weird position where they're really at risk, but they're also essential. If a funeral worker gets sick from Ebola because of their job, chances are they get sick, they stay home for the next few days, someone else is going to take care of the funerals. We're replacing high risk individuals with new healthy individuals.
The same thing happens in any outbreak, really. It's just we can't afford ... Right now, we're talking about the capacity of our healthcare system. If nurses and doctors who are at risk get sick, we’ve got to bring other people in, we’ve got to ask other people to put themselves at higher risk. It's the idea that you have those high risk essential individuals that you keep replacing, and it seems like a good idea because you want to treat people and you want to take care of the burials, but that can accelerate the spread of the disease because you have this turnover of hubs in your social network.
Michael: Talk a little bit about the model that you created for this, and then you compared it to empirical evidence from a number of different outbreaks.
Laurent: The model is fairly simple. We use the standard network models for disease transmission, which is simply you track individuals based on their current state, whether they're healthy, infectious, or dead or recovered, but you also track interaction: how many interactions occur between healthy people or one healthy person and one infectious individual, and just having this netrwork perspective just gives you the opportunity to model a social mechanism. In our case, it's the idea that people don't like being in contact with infectious individuals. The classic example I always give in talks is you go to the butcher to buy a steak for dinner, and you see the person behind the counter sneezing all over the meat. You're just going to eat something else that day. You're just rewiring that potential edge that you had with the butcher to whoever, the cheese maker because you're going to buy cheese instead.
Having a network perspective allows us to account for local rewiring mechanisms that we all do, like that social distancing. That's us cutting links. It's all very local network mechanisms. We had a simple model of that with the idea that some connections have to be preserved because they're essential. If my nurse gets sick, I'm just going to get rewired to a different nurse that's healthy, and now that nurse is being put at risk. That was a simple model, and it gave us this idea of super exponential spread. The classic picture is I infect two people, they each infect two people, these four infect two people, to get this exponential transmission tree for a disease. With this mechanism, because you keep re-feeding new, highly connected, high risk individuals into your population, it can get this super exponential spread, so faster than exponential.
We got the theoretical results, and we're like, "Well, has there ever been a disease where we think that there is a faster than exponential spread?" It turns out if you go in almost any flu seasons in the U.S. in almost any state, you're going to get this regime just before the peak of accelerating spread. That can be social patterns, what I'm describing now. It could be related to weather. The flu virus likes dry weather. That could be driving that. It could be it's flu season, we think about it more, we go to the doctor more. It could be a lot of things, but it seems like one of the most potent drivers of disease dynamics has always been social mechanisms and social behavior. So it seems likely, at least, that this could be one of them: this idea that when we think we're being prudent sometimes we make things worse for the collective by having a locally prudent behavior.
Michael: There are some implications that you discuss in this paper, and I'm curious how you see these results fitting in with COVID-19 and how we can make sense of this because, obviously, one of the issues is that it's a lot more difficult to identify who has been infected, that there's a latency with symptoms. How does that complicate this situation?
Laurent: Well, when we published that paper in 2016, we got some media attention, not all of which understood the subtlety of the model and of the results, I guess. There were a few headlines that read, "Don't Stay Home Sick," as if we were arguing against people staying home when they're sick. Of course, that's not the message. The message is stay home sick, and if your job needs someone to take over, then they should take some preventive measures. Right now, with coronavirus, the idea is if you are sick, absolutely stay home sick. If you can do your job remotely, do that. If you can't, or maybe your job can go without an employee for awhile, if you have to bring someone in, it seems like it's probably a high risk job because you got sick, and so that person should take additional measures to prevent their infection.
The idea is just, if the teacher gets sick, we're going to bring a substitute in if the school is still open, but that substitute should be aware that the virus might be going around the school, and, therefore, the way they teach and just the way they interact with people should change accordingly, which is just something we're not used to doing in normal seasonal flu outbreaks.
Michael: There is another fold in this. I want to talk about “Beyond R0: the importance of contact tracing when predicting epidemics,” this other paper that you did with Benjamin Althouse, Sam Scarpino, and Antoine Allard. One of the problems that we're in now is, because of this latency, it's been extremely difficult to trace the contact network.
Laurent: Right.
Michael: Nonetheless, I think it's worth getting a little bit into the details. Talk a little bit, for those unfamiliar with this term, although I'm sure it's getting a lot more widely known, what is R0? How do we measure that? And then why is R0 not enough? Where did this paper fill in an important gap in trying to think about tracing epidemics with that value alone?
Laurent: R0 or “R naught” is the basic reproduction number of a disease. The best definition is the simplest one. It's essentially the average or expected number of cases that you're going to create if you get infected early in an outbreak. If R naught is two, that means that if I get infected early in an outbreak, you can expect that I'll infect, on average, two people. That's a powerful number because, obviously, if that number is below one for every new case that you get, you expect less than one more case, so you don't even get a chain of transmission. That means that you'd expect a disease to die out quickly. If it's above one, you get branching trees of transmission, and then you'd expect it to grow and to get an outbreak.
It's a powerful number because it gives you, in one number, sort of an aggregated metric for both how transmissible the disease is, how dense the population is, the behavior that people are putting into place, the intervention that is being put in place, and all the complexity of disease dynamics and epidemics are put in one number, which at least below one, above one gives us a quantitative idea of how transmissible it might be. It's really popular because of that. It's one number, so it's tempting to try and use it to compare different outbreaks.
The classic example is the movie Contagion, which has that five-minute scene of them trying to figure out that one number. But it doesn't tell you the whole picture. There, my favorite example is the 1918 influenza epidemic and, again, the Western Africa Ebola epidemic of 2014/2016. Our best estimates for R naught in both cases are around 1.5, meaning that everyone that gets it, you transmit to one person, maybe two, but really the distributions are much more complex than that. For flu, maybe it's true that it was maybe one, maybe two, maybe three, average of 1.5. For Ebola, it's much more likely that you infect no one…and then, like I said, you have some super spreading events, like funerals, that end up infecting dozens of individuals. You can have the same average, but a very, very different underlying distribution. The reason why the average doesn't describe that distribution well is because whether you have this fat-tailed distribution or your average is driven by super spreading events will dictate how robust your epidemic is to intervention and to just stochastic chance. If an epidemic is driven not by the average individual, but by super spreading events, then an intervention that cuts an outbreak before you reach one of these events will be super effective. It's much harder to intervene against diseases that spread very steadily. Everyone infects two people: that’s way harder to fight against.
Michael: One of the points in this paper is about how little we actually know about secondary infections with the 2019 novel coronavirus. What insights does this paper offer in terms of how to address outbreaks when we have such a profound unknown—when the uncertainty in outbreak size is modeled everywhere from, you say, 5-40% of susceptible individuals?
Laurent: Well, that's the very tricky question, and that's why I guess I ignored part two of your previous question. It's very hard to measure the basic reproduction number because, most diseases, we don't have a clear idea of who infects whom. We have something like that for hemorrhagic fevers, like Ebola, where people will put into place a contact tracing protocol. So contact tracing is in the title, and it's this idea that when someone shows up sick at a treatment center or at a hospital, you might want to stay ahead of the disease. If you have the resources, you go to their workplace, you go to their home, and you test people for early symptoms, and then you might be able to stay ahead of the disease and catch cases early. If you do that, you also get the great benefit of treating them well, but also getting the data of who infected whom and have an idea of this secondary number, this number of secondary infections or of R0. So for Ebola, we have some contact tracing data and we do it well, but it's hard to do in real time. It's often during chaotic outbreaks and we often don't have the resources to do that. The data that comes there is noisy and hard to play with, but it's incredibly valuable. For respiratory infections, it's almost impossible. You get the flu, maybe it’s from someone on the bus, maybe someone at work, someone that touches the same door handle 10 minutes ago. It's just much more tricky.
And then we rely on a lot of simulations. So we look at the time series and how much noise we see in the time series of cases. We simulate, to the best of our abilities, diseases with different distribution of secondary infections and we try to get a posterior estimate for what these distributions might look like. So we have different ways. There's also a way through, if we have an idea about the mutation rate and if we see sequence of virus enough, we might have good ideas about who infected whom. So there are different ways to get to it, but none of them are super reliable early in an outbreak, which is one reason why the classic models that only use one number to parameterize everything about an epidemic are still successful because we still struggle to go beyond that, even if…obviously, my paper is saying we should and people have been saying that for decades, but it's just very, very, very hard.
Michael: That seems like it tails us into another paper that you coauthored on “School Closures, Event Cancellations, and the Mesoscopic Localization of Epidemics and Networks With Higher Order Structure.”
Laurent: I love that title so much! Because the title was meant to reflect sort of that's sometimes very technical results like the mesoscopic localization of epidemics in higher order networks can inform us on why and how useful it is to do very concrete things like close school and cancel events. So that's why I love the title.
Michael: Yeah. So, as you've already mentioned, one of the main challenges here is—as you and your co-authors put it—the failures of the surveillance system in this particular case. We’re pretty deep into the outbreak already. I want to spend a little bit more time in a moment on some of the questions that have been coming up from folks in our social media audience…one of those questions was about, how do we know from the models how containment strategies at different scales work? When is this about the distance that you're taking from another person? When is it about a school or event closure? When is about border crossings and closing those? How do we understand how those different skills relate to one another? How did you model this and this in this particular paper?
Laurent: Oh, this is such a good question. So I won't get into the model right away. I want to first answer what do we know about the different scales and why I felt like we needed this mesoscopic scale. So when I say mesoscopic, the idea is that microscopic is the individual, so that's me choosing to avoid a friend for a night. Mesoscopic is anything that involves multiple individuals, so that's school closures or events. And then the macroscopic would be mass quarantines at the country level. So let's say from the top down, I think we know from history that mass quarantines don't work or at least they don't work well, especially late in an outbreak. People find a way to avoid them. People need to travel and it's easier to make it easy for people to travel but actually test them and track them as they do. So I checked my visa status and requirements recently because I've been asked to comment about some of the decisions in the intervention from the U.S. and apparently I'm allowed to criticize my host country. I'm Canadian.
But that's why the intervention has been shit: because it's focused a little bit too much on the macro scale and hasn't given much guideline to what I think in this outbreak is the critical scale, which is the meso scale. So I've been organizing conferences and it's up to me at the end to decide whether I cancel my conferences or not. It was up to SFI to decide whether to shut down or not. And some places can't make those decisions because they have to weigh the collective good when facing the outbreak with the fact that their business maybe can't afford to close down for a month or two. And that's not a decision that anyone should be asked to make, really to choose between their own business and livelihood or the health of their community.
And we don't have enough information or even knowledge really at the individual level to make those decisions. So the intervention has been messy in that regard I think, and then the kind of information that we would need is also not available. So a lot of people, we cancel a conference and we get emails from speakers who are like, "Oh, I think you're overreacting." At the time of the cancellation there were only nine cases in Quebec, which is where I'm from, and while that's true, there's a lot that goes into that. I'm sure other people who will be talking too will bring up the fact that this is an exponential curve. So when you're early on, you don't want to use the number at T = 0 to try and predict what is going to be in two months, because it's an exponential spread and might be gigantic.
But also it's not true that those cases are uniformly distributed in the population. We tend to think of, “Oh, it's nine cases, every person has nine over n chances of being infected in this population,” which is just not the case. Some people are more at risk. That's what network epidemiology has been telling us for decades and some structures are more at risk, and that's why the meso scale is really important. So you're much more likely to find cases in schools or hospitals or big events just because that's where most of the social connections are out there, and therefore that's where most of the infections are going to occur. So if the disease lives at this meso scale, which not all diseases do, but some do, and I think this might be the case for Coronavirus, our intervention is much more powerful if it operates on that same scale.
And what it does mean is that you, by closing a school, you delete, you remove, a lot of those social connections that might spread the disease, but you also reduce the coupling across different schools that might still be open or across other structures. So it's really like an intervention that works in two ways by both protecting the individuals in that school, but also reducing the overall mesoscopic spread of the outbreak. So that's what it means, really. The important thing is try and think about, what are the key mechanisms for the spread of this disease? And individual interventions are super useful, washing your hands, social distancing, but you can spread to someone very inadvertently by just touching something you didn't even realized you touched and 10 minutes later they touch the same surface, touch their face. So unless you do it extremely well, those microscopic interventions are not going to be super efficient.
They'll work well for sexually transmitted infections, not so well for something as unknown. You don't even know if you're infected. We don't clearly know that the mechanisms with Coronavirus. So, in that case, mesoscopic works much better in my opinion. So that's how I'd try and think about it.
Michael: And so, in the model, this is just about pruning at different levels. I guess the question that I see coming up from people is why 500 people going to a church service is the number that people are picking. It seems to a lay person relatively arbitrary, but you talk about the efficacy of interventions at different levels. So, yeah, maybe a little bit more detail about how you actually model these interventions, and how it is that you and other epidemiologists are coming to these numbers and making these specific suggestions?
Laurent: Right. Yeah, so when we wrote the paper, I remember friends had a ban on all gatherings of above 1000 people. California at another number, maybe 5500, and now my university, University of Vermont, we can still have seminars on campus, but nothing with a crowd of more than 25 people. So the question is, how do you pick 25? How do you pick 1000? I think we have a good idea of just broad distribution. So we know those distributions are heterogeneous. We know there's hockey games in Montreal. That's what, 21,000 people? There's not that many big events that occur weekly in Quebec and there are a lot of events with 50 people. We know how many classes occur and all that. So we have an idea that those distributions are very heterogeneous, and we have an idea of how heterogeneous they are. Similarly, we have an idea of how many social groups and events a given individual will participate in.
So we have broad ways to parameterize these types of networks that are created not by simple pairwise interactions but really by higher order structure events and groups and classes. And once you have that, you can calculate the coupling between different structures. So if I start an outbreak in this hospital, maybe it's going to stay in the hospital. But if the coupling is strong enough, maybe it's going to be able to jump to other hospitals or schools. So then the idea is, how much do I have to play with this distributions of participants per event to decrease the coupling? It's almost like bringing the reproduction number down below one, not on the individual level but at the level of events and structures. So it's just a coupling that you decrease by making A less and less big event and eventually you get this critical threshold, which is one in the case of R0, something else in the case of this mesoscopic transmission. But then you stop the spread at the mesoscopic level. And that's where those interventions can be super beneficial.
And actually you end up cutting less social interactions by doing it at that level than if you were just randomly cutting pairwise interactions or social interactions through social distancing. So it is a powerful metric, but it is very hard to calculate or to estimate the critical size that you need because it's specific to different populations. We still don't know just how transmissible the virus even is. So it is tricky. In the paper, 27…but that's for one set of parameters and a bunch of assumptions. But the idea is that you want to be safe, so this idea of 1000 might be too much. That's already a really big event and 25 might be too low because then it's hard to operate. Not everything can move online. So in my mind it's this intermediate thing of numbers like 50 and 100 sort of make sense, they're in the right ballpark. 25 is very safe. 1000 is too much. It's hard to explain what it is, but we have a good understanding of what they should be.
And of course, the safer we are now, the better it's going to be in the long run. And my new philosophy is that almost no event is really critical. Nothing is that important. I could cancel my class right now. My students might miss some of my good jokes and a little bit of good material, but it's not that important. At the end of the day, it's not that critical. So canceling events is really not that big of a deal.
Michael: So there's another fold to this, which you talk about your 2015 paper in PNAS with Ben Althouse, “Complex Dynamics of Synergistic Co-Infections on Realistically Clustered Networks.” And I think it's important to stress that Coronavirus isn't happening in isolation here and that for a lot of people, this has been a particularly vicious flu season. In my own family we're coming off of three weeks of head colds just in time for campus closure. All of this.
Laurent: Right!
Michael: You've done some research into how modeling networks of different structures changes our expectations when we're looking at diseases, not in isolation, but in co-infected individuals. I'd love to hear you talk a little bit more about that.
Laurent: Yeah, this is an interesting line of work because we often or almost always model disease in a vacuum. If you look at any models, it's rare that they're going to take more than one disease into account. And yet we know that those things interact, right? The classic examples are more sexually transmitted infection, so if you have syphilis and HIV, your transmission rate of HIV goes up by maybe a factor of 40, right? Syphilis gives you surface wounds which are going to just create more bodily fluid exchange and you're going to transmit HIV a lot more than you would if you only had HIV and not syphilis.
We know that for some diseases it's incredibly important to take interactions into account. Interactions can be mechanistic, like the one I just described. It could just be that you have the flu. It makes you weaker. It makes your immune system weaker and then you get colonized by pneumonia or something else. That's almost an indirect interaction, but it can also be just you declare a state of emergency to fight Ebola in Congo. Well then you're sending resources to fight Ebola and those resources, whether it's healthcare workers, or hospital beds, or actual dollars, they have to come from somewhere else. They often come from somewhere else in public health interventions. That's why Congo is now seeing more deaths because of malaria or cholera than they did before Ebola. That's an indirect interaction that goes through political and social systems.
We know those things: they all interact in one way or another. We were just curious about how different would those models be? Of course they become very hard to track because you need to keep track of my state regarding flu, my state regarding pneumonia, my state regarding something else. The number of variables and mechanisms just blows up. It is harder, but it's doable. When you do it well, you realize that they also have completely different dynamical patterns. The big one for me that was a big surprise back then, I guess in 2015 I think, was the fact that classic models almost always give us this monotonous relationship between the expected epidemic size, so how many people are going to get sick, and the transmission rate or R0 of a disease. If R0 is below one, nothing happens. When R0 goes above one, there's this monotonous relationship: the larger it is, the bigger the outbreak.
When disease interact, it's not quite like that because the idea is maybe Syphilis can spread alone and HIV can't. Then if Syphilis just gets a little stronger, then you can get this discontinuous jump where you go from almost no HIV epidemic to huge HIV epidemic. I don't want to use HIV and Syphilis as an example too much because I don't know that much about sexually transmitted infections, but the idea is that when things interact, you can get this latent heat where one disease would be like, "I would really blow up. If you spread a little more I'll blow up." The other one could say the same thing, so if one of them gets a random mutation, just gets a little more transmissible, then they both blow up. Right? You get these discontinuous jumps in epidemic size, which are just unheard of in classic models.
We know that dynamics can be very different. The super exponential spread that we get with some social behavior we can get as well with interacting diseases so that the range of possible behavior is just incredible. It means that if we want to do robust forecasts, we need to take more than one thing into account and we can't model them in a vacuum, really. I've been talking about different diseases, but what the coronavirus outbreak is showing us is that those contagions can also, or maybe more easily, be social and not so much other infectious diseases, right? You get spread of news information of social media…that's going to help coronavirus spread. If you want to get a good model to forecast the spread of coronavirus or to forecast the effectiveness of interventions, you have to take those contagions of misinformation into account because they're just as critical to the public health questions you're trying to ask as the virus itself.
Michael: I do want to get to the work that you've done on informational contagions in relation to biological contagions, but first I want to explore this just a little bit more because one of the questions that came up in the Facebook group was about how this relates to the paper we were just discussing a moment ago. We know that it's not just the size of the events, but it's the network structure within them. When we're talking about limiting events to, say, 50 to 100, but some of those events are school closures, some of them are sporting events and the contact networks in those different kinds of events look different.
Laurent: Right.
Michael: Could you talk a little bit more about specifically how randomizing a network affects the spread of co-infections and then what that means for people who might have to make difficult choices about the priority of events of similar sizes and how those events actually are structured?
Laurent: Right. When you have multiple things that are spreading on a network, but interact together in a synergistic way, so news information and coronavirus, or syphilis and HIV, they help each other spread. Well, these synergistic contagions, as I like to call them, also benefit from this mesoscopic clustering of many people being part of the same structure. That's also a little surprising because the classic models tell us that clustering your social connections, social isolation, is great, right? If the disease spread from me, to my sister and my mom, then it can go from my mom to my sister. There's a triangle, so that triangles ends up being a wasted link for the disease. Clustering of connections is good in classic models.
For synergistic interactions, not so much. The diseases or the different contagions like to be kept together, right? They want to be kept together because these spread better if they're together. Then having these dense groups actually help their spread and makes those group a hotspot for the spread of contagions. That's just another way that the dynamics of interacting contagions is just very different from classic models. They interact differently with the mesoscopic scale. I think if we went back to that paper where we use a really simple model to show this phenomenon of mesoscopic localization, if we did it with some form of interacting contagions that that benefit from mesoscopic clustering, I think the results would be even stronger and would make even a better case for closing certain structures.
The problem is we don't have good data about co-infection almost ever, right? For privacy reasons, you don't share the identity of the cases. So when we do studies on flu, we have flu incidence data. When we do a study for coronavirus with coronavirus incidence data. We rarely know who had both. The one thing we know for coronavirus now is that it seems like most cases of deaths are due to people with preexisting conditions. That can be asthma. That could be another respiratory infection. It could be cancer. That's one way that there seems to be an interaction, but just in terms of spread or incidence, we rarely have coincidence data so it's very, very hard to be able to do these models in the first place.
One thing that I'd like to make a call on is, coronavirus is showing us how important the social contagion aspect around diseases can be. That's one place where we could have access to data, right? If you wanted to parameterize a good model for coronavirus, you would want to know what fraction of that population thinks coronavirus is just like another flu, or is that hoax, or that social distancing doesn't work, or that everything is dumb, right? You want to know what kind of misinformation is spreading and where, so that your model can take that into account. We have access to that data, or some people do. I don't. That data exists out there. I think the idea that public health messaging, whether from officials or just individuals on social media, should probably be treated as public health data, and therefore should be public in some anonymized way, but that would be incredibly useful for modelers and public health officials alike.
Michael: You've led right into this last paper to discuss macroscopic patterns of interacting contagions are indistinguishable from social reinforcement. This is one you wrote with Sam Scarpino and Jean-Gabriel Young. It seems as if what this paper and what you've just been talking about suggests is that if we treat beliefs as something that we can be infected by, then this is an instance where it becomes especially obvious that there are benefits to associating online with people who have different beliefs than you do. That's sort of a speculation. We can just table that for a moment. [Laughs.]
Laurent: Yeah, yeah.
Michael: But that does seem like it would sort of help randomize a contact network in the meme structure of our society.
Laurent: Right.
Michael: Anyway, I'd love to hear you talk a little bit more deeply how you modeled and what insights you drew from those models about co-infection of a biological contagion and a memetic contagion and how that might change behaviors.
Laurent: Right. Well, before I get into paper, that's again a very interesting point you bring up. It's one thing that we haven't studied enough, but we tend to assume one or multiple network structures, but the connection between the physical, social or contact network on which a virus might spread, to the sort of virtual nature or de-localized nature of information networks on social media, is something that we don't study enough. We know they overlap, so we assume that to some extent they're very much co-related, but that would be interesting to look at the importance of how different are your physical networks and your information networks that mostly live online. The reason why that'd be important is in part because of this paper.
The way this started was in 2015 when Ben Althouse and I published our first paper on interacting contagions, a lot of the dynamical features that I talked about, so the idea that expected outbreak size is non-monotonous with respect to R0, or that you can have super exponential spread, or that these contagions can interact and can benefit from social isolation in clustering…all of that was brand new in public health modeling, but sort of old news in social modeling.
In social sciences they've looked at contagions for a long time as well. There, contagions tend to be different. One way that I like to talk about this difference is that the number of exposures and the identity of people who expose you to a contagion matters a lot in the social realm, not so much in the infectious disease realm. The dumb example that I use all the time is if you have 10 friends telling you to go see the new Wonder Woman movie, that has more impact than one friend telling you 10 times in a row to go see the movie, right? It matters who signals to you, who sends you an exposure to that idea. In the realm of infectious disease, if you're healthy, whether it's one friend sneezing on you 10 times or 10 friends sneezing on you once each, well, really it's like roughly the same exposure. And maybe there's a genetic diversity in the different sneezes, but really, you're going to get sick or you're not, and you can sum up those exposures in a linear way. Right? So, models of infectious disease tend to be linear in terms of exposure, and that's not quite true in the social sciences.
So social sciences that had all those features, so for exponential spread, discontinuity, like importance of clustering, so I was wondering (and that was in 2015) if you have data about the incidents of a very infectious disease through time, how can you tell whether it's a classic case of a classic model from public health—interacting contagions like the ones I described—or this type of social contagions that spread through social reinforcement? How can you tell these things apart?
It took years to get a negative results, which is, well actually, you’re probably never going to be able to tell a social contagion time series apart from interacting contagion time series. Especially if you don't know what it might be interacting. So that was a negative result. The answer to my question was, no, you can't do it. But that turned out to be just as interesting, because I did mention that models of interacting contagions get really, really complicated, right? It's just, there're so many states I can be in. So if I'm looking at n contagions I need to follow like, "Do I have this one? Do I have this one? That one?" And I have so many assumptions to make about how they interact with each other.
But if in some way I can tell what comes out of those models or that dynamical system apart from social models, then why not use social models in the public health space, right? They're much simpler, and while they don't make sense, like this idea of social reinforcement doesn't quite make sense from a disease perspective, but if it's useful as a forecasting tool or as a modeling tool, then maybe we should embrace that. Right? So the idea is that this thing seems to spread like you'd expect social contagion to spread, and that helps us understand like who can leverage the work that has been done for decades in the social sciences to hopefully better understand interacts with contagions in the future.
Michael: Kind of a stray question for me after listening to you talk about this is whether you would have gotten the same results 50 years ago? Or whether what we're really looking at here is a transition in the structure of society into this sort of more virtualized network structure?
Laurent: Yeah, I think we just have a lot more data. We've been able to track social contagions like never before because of social media. There's always been people refusing to take social distancing measures but we just never knew that, and we didn't quite know what messaging they were using to talk to each other. But now, it's all like mostly out in the open and on social media so we do know a lot more and we're able to look at the interaction of social contagions with infectious disease like never before.
Before coronavirus…I think this might be my new favorite example, although it's a terrible situation, but from a dynamical system perspective, it's just so rich. My previous example that I was using before to talk about this was the measles outbreak in the Philippines. You might have heard me talk about this before because I do talk about it a lot, but it's just a fascinating system. So, the Philippines also have quite a bit of dengue, right? And dengue has four strains. So dengue 1, 2, 3, 4.
And it can interact in some ways because you can get a phenomenon by which like if you've had dengue 2 and then six months later you see dengue 3, your immune response might still be good enough so that your antibodies are typed to the virus, but not strong enough to neutralize the virus. And really, what you end up doing is that you're just providing sort of a genetic material disguise to the virus and helping it invade you. So by having had dengue 2 six months ago, you're making yourself more susceptible now to other strains of dengue, potentially.
So what this meant was that in the Philippines, they introduced... I think they were one of the first country to introduce a tetravalent dengue vaccine, meaning of vaccine that works for all four strains. But it didn't work equally well for all four strains. So a lot of especially young kids that got the vaccine ended up getting a reaction by which they were more susceptible to certain strains of dengue than if they didn't get the vaccine at all. So the government, as far as I know, did the right thing of pulling back on the vaccine and messaging publicly saying, "Okay, this vaccine puts certain populations at risk so you're only going to get this vaccine if you're above a certain age or if you've had experienced dengue before."
Messaging around infectious diseases as we're seeing now in the United States and pretty much everywhere in the world is just very, very tricky. So what people heard was not, "Take the vaccine only under these conditions." Really people heard, "See the vaccine is dangerous. We knew all along," and it did spark a huge anti-vaccination movement in the Philippines, which is why there's now a big measles outbreak. And I like this story because it's a terrible story whose origins are very well understood interactions between dengue strains. Everyone involved was meaning well. And then this interaction between dengue strains interacted with the existing anti-vaccination movement, which then fed back into measles, which would spread in many countries now.
So it used to be like my favorite example, but now especially in China and the U.S., we're seeing so many patterns of how social viral messaging online and online media are interacting with the spread of coronavirus that I think all of this together is going to be a big wake-up call that we need to treat messaging around public health data as public health in and of itself. Right? So how people talk about diseases should be considered public health and should be included in our models.
Michael: I want to get you back to your working group, but before I let you go, there were just a couple of quick questions from the social media audience that I want to give you an opportunity to respond to. One is about what kind of tools, software and capabilities you find you and your colleagues are currently lacking in your daily work, where the bottlenecks are for you in research. And then possibly, how amateurs could support this research, and if there's anything that the citizen science audience can do to assist right now.
Laurent: Yes, there's been different efforts throughout the last decades to try and get more people involved. So we've been in the United States and Canada, which is the two country I know best, we’ve been okay but not incredible at infectious disease surveillance. And that's one way that I think the general public could really, really help. So there are different initiatives, like Flu Near You is one, where basically it's self-reported flu cases, or flu-like symptoms. And that gives us a better idea than a lot of official data that comes out, because really what better way than to have a distributed surveillance network that just relies on individuals?
There are different initiatives like that to try and get citizen science involved in disease surveillance. And really I think going forward that's going to be critical both for existing outbreaks, existing pandemics that we just live with, like the seasonal flu, or emerging outbreaks like coronavirus right now.
The second part of the question I think is what resources we're lacking. It's been really interesting. So at the University of Vermont and the Vermont Complex Systems Center where I'm based, we're lucky to be partnered with Google. And it was very interesting to see companies like that that have their own problems to deal with, but they did reach out and offered cloud services and computing powers and all that. So, technology is a big issue because public health science is well-funded but not to the extent that I think it should, and it's been interesting to see a lot of people offering help and resources in that way. Either computing resources or just help sharing the data.
I think being critical of the messaging that you see online, just reading to make sure what you're reading is correct and validated by other sources, and then making sure that you spread the validated messages far and wide and try and make sure that misinformation doesn't spread as much is probably more important than the two previous things I mentioned.
Michael: Last question for you. If you are going to advise people to track this at home with limited time and attention…if you can only offer one or a handful of data points to keep an eye on, what would those be? What do you consider the most important factors here to be staying on top of?
Laurent: You're talking to Carl Bergstrom, too, right? Are you talking to him later?
Michael: I will be, yes.
Laurent: Can you ask him that question? [Laughs.]
Michael: I will definitely ask him that question.
Laurent: As now, I'm just passing it along.
Michael: All right, sounds good. All right, thank you.
Laurent: He’s the misinformation expert, so you know.
Michael: Yeah, definitely. Thank you so much for your insights today.
Laurent: This was great. Thank you for inviting me.