I’ve worked on air pollution for a long time, going back to the days of my post-doctoral fellowship in the late 1970s, and I had worked on many, many pollutants. I came to Johns Hopkins in 1994, and one of the first projects I became involved in was reexamining some time-series studies that were coming out back then. At the time, there were a number of these new time-series studies, some of them from the Harvard group of Joel Schwartz and Doug Dockery, suggesting that there was a day-to-day effect of pollution on mortality. There was a lot of skepticism about these studies; people did not quite have confidence in these methods. So Scott Zeger and I became involved through the Health Effects Institute. Basically, we were looking at some of that evidence. We built some of the databases to evaluate them, and redid some of the analyses. Scott is a wonderful methodologist, and together we began to explore some of the issues related to these time-series analyses.

“…the idea was to analyze the data in the same way in each city, look at the day-to-day evidence of air pollution and mortality, and then join the evidence across cities.”

The basic idea is to take the number of deaths—say, day to day—or the number of hospital admissions or emergency room visits and look to see if there’s a relationship in that time series and variations in air pollution levels. You also take into account the other time varying factors, like temperature or seasonal disease outbreaks like influenza. So you use mathematical models to look at the relationship between air pollution and mortality. I had done some time-series work way back in a paper I published around 1981. In the 1990s there was an increasing sophistication of methods, new analytical tools, and the development of computer capacity that made things possible that we couldn’t do before.

The 2000 study grew out of the work Scott and I had been doing. Up until then, people would use a particular lengthy series of data, or perhaps the data for wherever they happened to live. Our "big idea" was to use cities without any selection; take the largest cities or all those with pollution data, and analyze those data in exactly the same way within each city. Until that point, one group had done some multi-city studies in Europe. Their model was to go where the data and investigators were available. In the U.S., we could essentially take every city with data available and put it together. The power of the method was that we would be using large bodies of data, and we could try to optimize the signal-to-noise ratio. We could also look across the country and see whether the effect of the pollution varies. For example, in the Northeast, people are concerned with power plants. In California, they’re concerned about vehicle traffic. There is some variation across the country in what people breathe. We thought this method would allow us to better understand this heterogeneity.

There were a couple of things that made this possible. One was the availability of software and hardware that allowed us to do this. The kind of things we did in this New England Journal of Medicine paper were simply not possible even 10 years earlier. And my colleagues are just superb methodologists: Scott Zeger, as I’ve mentioned, and Francesca Dominici. They developed the regression models at the heart of this research. And so the idea was to analyze the data in the same way in each city, look at the day-to-day evidence of air pollution and mortality, and then join the evidence across cities. Is it the same? Or is it different? And, if it’s different, can we explain the evidence?

That paper used data from the 20 largest cities in the U.S., covering roughly 54 million people. It described a statistically significant and, I think, a fact of important public health magnitude in terms of an effect of particulate air pollution on mortality. And we showed that the effect persisted when we took account of other pollutants that might have been correlated with these particles.

We looked at all-cause mortality and at cardio-respiratory mortality, which included things like chronic obstructive pulmonary disease, which is what we used to call emphysema, and then pneumonia, heart attack, and congestive heart failure.

Not really. I think there are a couple of reasons the paper had such an impact. One is the demonstration that it was possible to carry out this kind of analysis. The second is that it showed that there was an effect of particles on mortality that could not be attributed to other pollutants. And I think, probably, another important thing with this paper is that it reduced huge amounts of data down to a couple of simple graphs and numbers, which is something that is useful for policy apparatus and policy making. By putting together data for such a large number of cities, it provided a very powerful piece of evidence for decision-making. I think that added to the significance of the paper.

Part of the story, which you might not be aware of is, that we actually had to correct the evidence because of a software issue. I remember saying to my colleagues, "Well, we found what everybody else has; maybe we’re all right, maybe we’re all wrong." We used the same software everybody else did. This was the standard statistical package. The code was written by a superb statistician, but the way these models are fit is that there is an iterative algorithm that narrows down whatever the fit method is, and then says, okay, the data are fit well enough. This particular model in the software had been set in a default mode a long time ago to not iterate many times. And probably it needed to be iterated more times than it did in our original paper. We identified this a couple of years later, and when we reran our data, the main message was the same, but the estimate of the overall effect dropped. And in fact, what we had done and the way we used the software was what everybody had done and did afterwards, as well, and a number of people ended up redoing their analyses. That’s one thing you always worry about, some unknown issue of methodology. We’re using very sophisticated tools and there’s always the possibility of some methodological glitch you don’t understand.

I think for those of us who model data, we’re always concerned that there’s some aspect of how we model the data that will mislead us, and this is particularly dangerous when you’re estimating these kinds of effects that are not huge. We’re not detecting the kind of effect you can see visually. Part of the strength of our approach is that we take all these cities and analyze all the data at once and apply our models systematically. Another concern with the literature before our study is that people might have been selective in their modeling. They might have tended to report the models that gave the strongest and most often statistically significant approach. You also worry about publication bias. We just had an interesting paper published in Epidemiology, in which we essentially compared our multi-city approach, this time for ozone, to what was in the literature, where people had taken some of the same cities individually that we had used for unified data. And we showed clearly that the published single-city estimates tended to be much higher than those we had arrived at.

We’ve taken the next step and we’re doing several things. One is that we joined with colleagues in Europe and Canada to put all the evidence together from around the world. The other data set we’ve turned to, which I think is extraordinarily powerful, is Medicare, which is basically an ongoing cohort study of 40 million people, age 65 and over. What we’ve done now is taken this Medicare data, which includes death and hospitalization, and we’re joining that with the air pollution data. We’re also looking at the effect of smaller particles that are being measured by the EPA. We have an idea that what we’re doing should be set up almost in a surveillance fashion. We have the Medicare data that’s ongoing; the air pollution data is ongoing. We’re trying to show that these can be used together.

The other thing we’ve done, credit for which goes to Scott and others, is to make our data and methods available. We have a web site where we can post those data, along with the code. The idea is that our findings should be robust to reanalysis.

Just to dwell on the epidemiology approaches: there are two ways to pursue it. One is to do more observational studies—more time series, for instance, like we’ve been doing—and the other line of approach that may help give evidence for causal influence, is to take advantage of circumstances in which there are sharp changes in exposure to air pollution. People are now studying several of these types of situations: the city of Dublin, for example, instituted a coal-burning ban and abruptly changed the nature of air pollution. Hong Kong removed sulfur from automobile fuels. So in these cases, exposure changes in time and decouples itself from changes in confounders. It gives us an opportunity to perhaps be stronger in inferring cause if the data supports the hypothesis.