 It's also great to have spent the past year at TSE. Thank you for having me as a visitor. In the spirit of it being a keynote talk, I'm mostly going to focus on big picture questions and on issues that I think are important to resolve that remain unresolved. But I will touch on a couple of papers I've written on these issues as well. So we all know that drastically reducing fossil fuel emissions of greenhouse gases is necessary for combating climate change. So that hopefully we have fewer 38, 39, 40 degree days like the one outside. But of course, when you reduce fossil fuel use, you also reduce emissions of air pollutants that are associated with burning fossil fuels like sulfur dioxide and nitric matter. We know that these pollutants affect human health. And this has been long recognized as a co-benefit of climate change policy. Of course, ideally we regulate these things separately, wisely, but in practice, don't always. And so we can ask ourselves the question, if we address climate change, what will that mean for human health in terms of the air pollution reductions? Now, of course, the answer depends on how harmful air pollution is to human health and for other outcomes. And you'd think that after a calmness jumped on the health effects of air pollution bang-wagon more than two decades ago, we'd have a pretty good answer. But I'm going to argue that we actually don't accept that we know pollution is bad for you, which is pretty clear. But as far as the magnitude of this harm, I think we still have a ways to go, especially in places where air pollution levels are already low. The evidence on how much beneficial further reductions would be is not that clear. To motivate why this question is important, I've graphed air pollution levels in the United States from the time that widespread monitoring began to today. I normalized each pollutant to 100 in the first year of widespread monitoring. And you can see that air pollution levels in the US have fallen drastically since the early 1970s. And there are similar trends for Europe as well. In particular, SO2, which is emitted largely by coal-fired power plants, has fallen by more than 90% over this time period. Nitrous oxide, which is nitrogen dioxide, sorry, has also fallen by over 70%. And then total suspended particulates, another pollutant that is not dust from fossil fuel, but fossil fuels are also responsible for a non-trivial share of it, have fallen by less. So part of this is changes in monitoring compositions, because we've realized that total suspended particulates are not as bad as smaller particles. So we stopped monitoring PM10, which is any particle 10 micrometers or smaller in diameter. Then we realized even smaller particles are bad. So we started monitoring PM2.5. So there are some compositional changes here that kind of explain this flattening out. But even if you look at PM10, which we started monitoring in the late, in the early 1990s, and then PM2.5, which we started monitoring in the United States in 1999, since widespread monitoring, there's been a reduction of 40%. So air pollution levels have fallen quite a bit in the United States. If you consider places like China and India that are burning a lot of coal, their SO2 levels are really high. We want these places to decarbonize. That will mean that air pollution will go down as well. And so what kind of benefits will this have? That, I think, is a really important question. Now, here's why it is difficult. First, air pollution is not as good as randomly assigned. It's associated with driving with electricity use with burning fossil fuels, and so on. As economists, we know pretty well how to deal with that conceptually. But finding good quasi-experimental variation in practice can be difficult. Something that the literature has realized lately is the problem, is that ambient air pollution levels are also measured with air. So typically, you have monitoring stations that are measuring pollution in a specific point. And then you have the population, which moves nearby. But of course, the measurements at the monitor might be a noisy proxy for what people are actually experiencing. And that kind of measurement error is its classical degree attenuation bias. Even with products where you have satellite-sensed, pollution can have this kind of error as well. Potentially not as bad as a single monitor, but nonetheless could cause your estimates to be attenuated. That less econometrically, more big picture issues is that air pollution might have delayed health effects. So ideally, you want to measure the total effect of a given unit of exposure up through infinity that is very difficult to do. But if you're not measuring delayed health effects, you're not capturing the full effects of air pollution. Or in addition to, increases in air pollution might kill people who would have died soon anyway, regardless of this shock. This is known as mortality displacement. And soon is in quotes because it's not clear how soon is. If we knew that, we could obviously account for it. It could be that air pollution kills people who would have lived for another year, another three years, or just another month. So it's really important to know if you believe that kind of value of statistical life should be adjusted for how long you would have lived and that somebody who would have lived for 10 more years, absent a shock. This should be valued at more than somebody who would have lived for 10 more days, which I think most economists at least do. Then this is really important to account for. And a fifth major problem is that people might take defensive actions like buy medication or relocate somewhere else, counteract the negative health effects of air pollution. Now, this is a good thing as far as minimizing the health effects, but as far as understanding the cost of air pollution, this makes it harder. Because at the extreme, if there's really expensive medication that perfectly offsets any health effects, and I take it and you don't see any measured health effects on me, you might erroneously conclude that air pollution is costless. But in fact, I had to undertake a very costly investment in order to avoid these negative health effects. So just for the rest of the talk, have some taxonomy of what I'm talking about, there's a couple of effects of air pollution that we can measure. So we can have a short run exposure to air pollution, let's say one day of higher pollution levels, and then a measured effect that's over the same timeframe or maybe a bit longer, like three to five days a month. And I would consider that the short run effect of short run exposure, so the exposure period is short, and the timeframe over which you're measuring the outcome is short, that is undesirable because there could be delayed effects, and because there could be mortality displacement, right? So this doesn't give you a full picture of the cost of air pollution. So what can you do? Well, you can extend the time period over which you measure the effect, and that will get you the long run effect of short run exposure, and that is something that you can use for policyful undertaking any defensive investments or other mitigative behaviors. And then what you can also do is find a good exogenous source of longer run exposure. Why is this useful? Because a series of short run exposures might not be additive. They might not have additive health effects, and so then you actually need to measure what would be the long run effect of long run exposure. It could be concave, it could be convex. It might be additive. That is something that we don't know very well. So that brings me to kind of the state of the economics literature today. And that is that the best identified studies focus on outcomes that are measured over fairly short time frames following the pollution exposure. And I know I'm missing a lot of studies, including by some of the people in this room. This is just to give you an idea of where the literature is. So there's a large literature on infant mortality and air pollution exposure. The reason that infants are somewhat attractive to studies because they haven't been alive very long, so you can measure their lifetime exposure to pollution fairly well. So there's a study that looks at birth weight and prematurity. So immediately after the infants are born, and then makes conclusions about long run outcomes based on an association between low birth weight and prematurity. But it doesn't empirically measure what happens to these infants over a longer period of time. Then there's a paper that looks at weekly infant mortality. Again, it doesn't measure what happens to kind of longer run infant mortality. And if these infants are infants that would have died soon anyway, or maybe they're infants that would have lived for 80 more years. We have a similar picture on for adults. So these are all well identified studies. And we have outcomes that are daily, sometimes going out to monthly, like Hollingsworth et al. 2021, sometimes going as far out as 90 days. And the longest ones, there's two studies that look at annual mortality. But even there, you might ask, okay, annual mortality goes up, but what would have happened over two years? Are these people that would have lived for 10 more years, 20 more years, two more years, without looking over a longer time frame, it's very hard to determine this empirically. Okay, so are there studies that look over longer periods of time? There are very few. One example is Boreka et al. that studies reductions in sulfur dioxide pollution as a result of the acid rain program, and finds that their larger cross-sectional mortality decreases in counties that experience larger reductions in SO2, and that these mortality reductions also grow over time, so they're smaller immediately after the beginning of this program, and then they become larger. Now, one way to interpret this is that this is, this is totally not stuff, can just add up shorter run effects to get longer run effects, but this is cross-sectional data, and so what the studies challenge that longer run outcomes face is whether these are the same people that are dying, right? So when you have an air pollution shock, and there are studies showing all of this, right? People might respond by moving, right? So there's a big hedonic literature about how much people value clean air, and so if air pollution goes down or goes up, you might have resorting of people into neighborhoods. That in itself can have health effects. Air pollution can also affect labor markets, which could have consequences for your health. The health effects themselves can affect your labor market outcomes. So there's basically a lot of complex mechanisms tied to air pollution exposure, which makes it really hard to be sure that you're really picking up this link between air pollution, health, and mortality. That's the outcome that you're interested in. Now, why don't you just want this reduced-form relationship between air pollution and mortality? Well, as I mentioned, because all of these things could moderate the health impact, and at the extreme, you might find that even in a well-identified study that there's no act of pollution on mortality, but that's a medicine or other defensive investments in response. Alternatively, you can also overestimate the health effects it goes. Okay, so where do we go from here? So I think assessing the counterfactual survival of those who are killed by air pollution is really important for understanding the welfare consequences of air pollution reduction. And I see two options that are not mutually exclusive. I think we should be trying both of them. One is to do more work considering mortality impacts over different time windows following air pollution shocks. So now I think papers are starting to do this, and that's really great, because previously people would kind of look at contemporaneous, continuous change in outcomes, and that doesn't give you enough to understand the full welfare impacts. Now, of course, if you extend the time period T over which you look at the effects of air pollution, eventually the effects of air pollution on mortality will go to zero. In the long run, we're all dead. Like in 200 years, it doesn't matter how much pollution we were all exposed to, right? So in some sense, the goal of the research should be to measure what is this time period capital T over which the effect of pollution on mortality disappears. And that informs the counterfactual survival of those affected, right? Like is air pollution kind of like a car accident in that if you escape being killed in a car accident, you're probably in good shape for a really long time, or is it more like somebody who is very unhealthy and has a heart attack, even if they survived that heart attack, they're at a very high risk of another one that will kill them sooner, right? And we don't really know. Oh, well, we're getting better understanding of what kind of shock pollution is, but I don't think we're fully there. So why doesn't everybody just do option one? Because statistical power may quickly become an issue, right? So especially if you're looking at a daily or a weekly shock and you're trying to see what happens over 10 years, that's just statistically not feasible. Even with an annual shock after 10 years, you might not have the right data to track the mortality of the people, or you might have against statistical power issues. The second option that I am actually personally excited about and actively working on is to use data to model counterfactual survival of people. So kind of combined, what I think have a lot of good papers have done lately, not just in this literature, but in economics in general, is combine well-identified estimates with a structural approach to get the best of both worlds. And the advantage of this method is that it's not limited by the length of the follow-up window. So if you model counterfactual survival well in principle, you could look at a daily shock and you could look at daily counterfactual survival and you could get the long run effect of short run exposure at least, which we still do not have. The downside is that this is data intensive as I'll demonstrate with a few examples and it also requires some assumptions because it partly involves structural work, but I think it's a promising way forward. So here's an example of the second approach from my recent paper with Garth Hugh Till, Nolan Miller, David Molliter and Julian Reif, and you're gonna kind of have to trust me that the identification here is good, right? We have an instrument for a one-day increase in fine particulate matter and then we see what happens to mortality over the following and the subsequent two days. And in terms of mortality, we find, I say all age, this is actually data on the US elderly administrative records from Medicare, the health insurance for the elderly, so this is all elderly. We find that all age mortality over the day of the shock and the subsequent two days go by 0.85 deaths per million. And then to understand what this means in terms of how many life years are lost as opposed to just how many lives, we model the counterfactual survival of the people who died using Medicare records. So what we're able to observe is essentially every single plane that they have with Medicare and we use machine learning to create basically a survival model and we assume that people who died had they not died would have followed that survival model going forward. And then it makes a big difference for valuing the counterfactual life expectancy of people killed by air pollution. So this column shows what happens to life years lost. So here we have lives and then basically what we do is for every person who died, we assign them the different counterfactual life expectancy. So here we just take the average life expectancy for the Medicare population as a whole, which is slightly over 11 years and we assume every person who dies would have lived for 11 years otherwise. So this is what would happen if the person killed by air pollution was just a random individual plucked from this population. And then you get that a one unit increase in PM 2.5, which is about 10% of the mean during this time period in our sample loses about 9.7 life years per million individuals. And of course you can multiply whatever or preferred value for a life year is, you can multiply this to get the smaller amount. And this is obviously just this estimate scaled by this average life years lost per decedent. Then you can do kind of a bit better and this is something that some studies do just to count for the age and sex of the people that died. We know that on average and we showed that on average pollution kills older people. So you can adjust for that and even just that adjustment will take three life years lost. So we'll take about a third of the estimate from your counterfactual. And then here we also add chronic conditions. So Medicare has flags for a variety of chronic conditions. That reduces the estimate further. And then when you do machine learning with thousands of predictors of life expectancy thrown in, the estimate was reduced to less than a third of what kind of the naive approach would be. So the counterfactual life expectancy of the typical, here we calculate what it would be per complier is three and a half years. Now that's actually still not small, right? So it might make you think, well, then these short run shops don't matter. But even though this is smaller than average, when added up for millions of individuals, you still get annual benefits of air pollution reduction for about 24 billion dollars, right? So I think this is a promising approach going forward. If you have this kind of data, regardless of what is the timeframe of the shock that you're looking at, whether it's one day or one year, is explicitly modeling counterfactual life expectancy of those who are killed by air pollution. Now I'm gonna use a more recent work of mine to talk about the other approach and that is the structural model. And here we're really able to demonstrate that there's both delayed effects and short run mortality displacement going on and they're important to account for. So here we're taking us back to the 88, the time when SO2 concentrations in the United States were quite high. So you can kind of think about India and China today to approximate approximation, possibly India and China are even worse. And we use daily changes in wind direction as an instrument for changes in sulfur dioxide concentrations, lots of controls. Again, I kind of wanna focus on bigger picture things here so you just have to trust me that it's all well identified. And then our dependent variable is county level, cumulative mortality over X days following the shocks where X is on the X axis up here. So basically we're doing option one and empirically tracing out how long people would have lived in absence of the shock peak and also delayed effects. So this is gonna pick up the net of the two. And when you look at all cause mortality, you see that on net there are delayed effects. As you extend the time window further out, the shock is still the same. It's always a one day increase in sulfur dioxide. The estimated mortality impacts grow. And when you look break it down by cause, it looks about a third due to cardiovascular causes, a third due to other and a third due to cancer. And of course the cancer, because you can't develop cancer and die from it as a result of an air pollution shock in one day. So this is already evidence that air pollution is killing individuals who probably were already pretty sick. Anyway, how this looks at the end of the 28 day window, the cancer mortality has gone to zero. So, and you know, kind of goes to zero after about two weeks. So what this is saying is that when it comes to cancer that about a third of the individuals who died on the day of pollution would have lived less than 14 days, depending on if you wanna see if when it goes to zero and when it ceases to be statistically significant. So as we might expect, air pollution is definitely killing people who are in very bad health. But when it comes to cardiovascular and these other causes of which chronic obstructive pulmonary disease, so respiratory related disease is a big one, this does not go to zero. So it suggests that we can't just assume, you know, it's all mortality displacement. We can't assume that none of it is mortality displacement. We really have to think hard about counterfactual life expectancy. So that leads us to propose a new approach and that is leveraging a structural model which we've used in many other settings. So why not when it comes to health effects of air pollution? And the reason we propose this is because estimating the long run effects of air pollution with quasi-experimental methods has proven difficult. So I'm not aware of any studies that sort of have a definitive answer to this. And this could be a nice complimentary approach because we still don't have a good answer to how much life expectancy in the U.S. improved as a result of pollution declines, for example. And organizations such as the World Health Organization, the Environmental Protection Agencies, while some of them are using some of the well identified estimates are still relying on epidemiological estimates. And one reason is because when it comes to epidemiological estimates, if you're not restricting yourself to be well identified, then you can look over any time periods that you want, right? So they can deliver the answers that policy makers want in terms of the units, but they're not necessarily the right answers. And so if we want to sort of be competitive when it comes to estimates being used for policy, we need to get better estimates. And this is where I think well identified short run estimates combined with a dynamic model of health to predict longer run impacts can be very useful. So I'll talk about the proposed model and this is actually not our model. It's a general model of mortality developed by Yaris Mooney and Marot, which I think in some sense gives it nice external validity because we're not cherry picking it for our purpose. We're not just picking it up for our purpose. It's a good model in general. So this model posits that you have some health capital at time T, where T can, that depends in the last period. Minus some depreciation parameter, which we can conceptualize as some constant delta times T to the alpha. Plus some investment, you have some initial at birth. And so this is a model that depends on five parameters, mu, delta, alpha, I am squared. In this model, death occurs. So this is latent health, right? We don't observe this. When latent health falls below a threshold eight to zero without loss of generality, you die. When it falls below a slower bar for the first time, this model was developed to capture a variety of real world mortality dynamics. So by parameterizing it, you can capture rectangular return. More people are making it to advanced ages like 80, 85, and so on. But there's a sharp drop off, you know, around 80 or making it past that. So kind of survival curves are becoming more correct angle. It can also capture socioeconomic gradients with this parameter I, where some people have more of an endowment than others, and that explains why they live longer. Importantly for thinking about the health of air pollution, it can capture scarring effects. So if I depreciate your health once for a period of time, let's say through a pollution shock, you might wanna die whether or not you're healthy. Like if you're a pretty healthy individual, if your age is high, and I shock you for one period of time, you will probably be fine. But because your health tomorrow depends on your health today, your health will be permanently lower than somebody who didn't receive that shock, and that will produce a scarring effect or a delayed effect if you will on mortality. This model can also capture mortality displacement. If you increase age lower bar temporarily, what will happen is the individuals who are the sickest will die. But then because you now have this gap in the distribution, if age lower bar refers to where it was before, you're gonna have a decrease in mortality, and that's the mortality displacement effect. So what we do in this paper is we calibrate this model using a period life table from the 1970s, kind of the time period of our study. It doesn't capture infant mortality very well, but we actually don't detect increases in infant mortality and we're still working on calibrating this. Here are the model parameters. The units don't really matter. Except to the extent that they fit the actual survival curve pretty well. We simulated with a million agents. So here are some example health trajectories that can come out of this model. This is simulated using those parameters, so there's four people here. At birth, they have the standard deviation of health is pretty good. They follow fairly different trajectories in life. On average, health initially increases because there's this curve on it and when your young depreciation is low because it's a T alpha, so T is low. So your health stock is growing and then eventually as the depreciation parameter starts becoming important, people die. So this person made it to over 80 years old. This person dies at 70, this person dies at 60. And this person gets unlucky and dies at 50. Notice that this person comes pretty close to death a couple of times, but manages with lucky shocks, manages to survive for another decade. So this also kind of hits that this model can be used to think about counterfactual life expectancy if you think it's well calibrated. How long would these people have lived? Who did not get killed by air pollution? So once you have a baseline fit for this model, you need to think about how air pollution changes the underlying health parameters, right? So the basic idea is that of course, air pollution will kill some people, but then the people that it doesn't kill, it might weaken so that they live less long. Maybe they die three days later, maybe five days later. Maybe it just makes a tiny bit more likely to die five years from now. And so basically what you can do is you can take your regret estimates on daily mortality and you can use those to calibrate the changes and model parameters. And there's really two ways that you can approach this that are not mutually exclusive. You can either assume that air pollution changes health capital, delta, alpha, or I. Delta and alpha behave pretty similarly in this model. I is a little bit different, but I don't think I have time to go into more details. And then this will produce both short and long run effects on mortality. The short run effects will be calibrated to whatever the regression estimates calibrate the model. And then the long run effects is something that you can simulate by basically assuming that future pollution, well, you can turn off the pollution shock or you can turn on a permanent shock and then potentially have non-additive interactive effects over time because this model is highly non-linear. Alternatively, or in addition to, you can model this as a change in the death threshold H-short bar and that will generate short run mortality displacement. Here to give you an example, these are the estimates, the ID estimates for the effects of sulfur dioxide on mortality following a one day change in sulfur dioxide over various days since exposure. And here we sort of see that the delayed effects are there on average, but we assume that 50% is gonna be due to the mortality displacement or an increase in H-lower bar. So we calibrate the model, the change in parameters to match the regression estimate for same day exposure. And I guess maybe something closer to 30% is more appropriate, but we're going with half for now. And then we run the model forward for 28 days. And so here's what the model predicts. If you calibrate, so this is just calibrated using one day exposure, everything going forward completely, separate, independent from the actual IV estimates. And overall it fits pretty well. So if I were to put confidence intervals over here, the red line model predictions would be within the confidence intervals of the actual IV estimates. So this model is flexible enough and realistic enough to at least fit the estimated pattern as well within the time window over which we have good statistical power to measure the delayed effects ourselves. And then if we're willing to be very brave, we can actually run this model on an entire cohort and increase SO2 by one unit to see what happens to long run survival. And here we have the choice of whether to go with alpha, delta, or I. We're keeping the H-lower bar at 50%. As we said, now alpha and delta deliver very similar predictions. So our baseline life expectancy based on the 1972 table calibration is 74.8 years that a one unit increase in SO2 would lower survival. So suppose air pollution levels had gone down by one unit fewer would have lowered survival by 0.6 to 0.7 years. Now it gives you a less realistic estimate. So it does matter what parameter you assume is changing. Our argument in the paper is not a good parameter to talk because it doesn't match the age that we see in the data. So we can also use regression estimates to inform ourselves which parameter should be changing. In the case of I, if the effects of pollution were kind of operating in this uniform fashion, affecting the latent health of everybody equally, then we'd expect to see more younger people dying. And we don't think these estimates are both reasonable and they're consistent with the patterns that we see in the data. And notice that, so here basically air pollution changes starting from birth, but really the two curves deviate from each other at older ages. And this is because of course, we're capturing these scarring effects. So our model doesn't predict that more infants will die as a result of air pollution, partly because alpha and delta most affect old people, but more infants become scarred. So we basically estimate that compared to what you would were to do if you were to take our naive estimates based on short run and multiply them, you think this model about doubles the estimated mortality effect of air pollution compared to just using the shorter run estimates. Okay, so this is what I think a promising way forward is. Definitely we need to grapple with the fact that while our studies have been very well identified, they haven't delivered the estimates that policy makers really want, which is what happens to life expectancy, right? What happens over the longer run? And if we want to compete with and supersede the epidemiological estimates, then we need to be able to see these kind of impacts. Now I'm not saying that studies that exploit longer outcome windows or longer lasting pollution shocks are not valuable, but as I've mentioned, there's other issues that come up with these studies. So I think we need to kind of attack on both fronts and pursue good long run variation in air pollution if we can find it. So far, we haven't found that $20 bill on the sidewalk, at least not many good ones, but that doesn't mean it's not worth looking more and thinking about issues of offsetting investments and relocation and so on. But I also think that we should take more seriously this idea of combining well-identified short run estimates of health so we can better speak to welfare costs of air pollution arc. And I think I will stop here to not prevent you from enjoying the beautiful outdoor weather and see if there's any questions.