One More Air-Pollution Study

In the last post I confused this study (PDF) with an earlier one by the same group of researchers; I wrote about the earlier one, but linked to a post on Paul Whiteley's blog about the more recent one, which was published just last November.

(I also started describing this one, and then switched to describing the earlier one; in my last post, only one of the studies I mentioned used air-pollution data from the EPA's air-monitoring stations. The earlier study by these authors only used proximity to high-traffic roadways as their variable indicating pollution exposure.)

This paper combined the methods of the two studies I wrote about in the last post; it used the same pool of children born in California between 1997 and 2006 and drew on two sources of data on air pollution at the time and place those children were born: the EPA's air-quality data that I wrote about yesterday, and a computer model of average traffic flow, and exhaust emissions, along California's major roadways.

To some extent, you could see it as a more geographically dispersed version of the study I described yesterday that looked at prenatal exposure to air pollution in just Los Angeles County. 

(Weirdly, the LA-County-only study, though restricted to a smaller geographic area, involved way more people than the two traffic-related studies: 7,603 autistic and 75,782 control subjects, as opposed to this study's 279 autistic and 245 control subjects.)

But the relatively straightforward EPA data, which are direct measurements of the concentration of various pollutants at regular intervals, and which are only abstracted from each child's actual prenatal exposure in that 1) they do not measure what concentration of those pollutants actually got into the mothers' bodies, much less the fetuses', and 2) they were taken at sites some distance away from where the children actually live. 

So it's not a perfect data source, but it's still a lot more directly reflective of reality than this computer model seems to be:

The principal model inputs are roadway geometry, link-based traffic volumes, period-specific meteorological conditions (wind speed and direction, atmospheric stability, and mixing heights), and vehicle emission rates. Detailed roadway geometry data and annual average daily traffic counts were obtained from Tele Atlas/Geographic Data Technology in 2005. These data represent an integration of state-, county-, and city-level traffic counts collected between 1995 and 2000. Because our period of interest was from 1997 to 2008, the counts were scaled to represent individual years based on estimated growth in county average vehicle-miles-traveled data. Traffic counts were assigned to roadways based on location and street names. Traffic volumes on roadways without count data (mostly small roads) were estimated based on median volumes for similar class roads in small geographic regions. Meteorological data from 56 local monitoring stations were matched to the dates and locations of interest. Vehicle fleet average emission factors were based on the California Air Resource Board's EMFAC2007 (version 2.3) model. Annual average emission factors were calculated by year (1997-2008) for travel on freeways (65 mph), state highways (50 mph), arterials (35 mph), and collector roads (30 mph) (to convert to kilometers, multiply by 1.6). We used the CALINE4 model to estimate locally varying ambient concentrations of nitrogen oxides contributed by freeways, nonfreeways, and all roads located within 5 km of each child's home. Previously, we have used the CALINE4 model to estimate concentrations of other traffic-related pollutants, including elemental carbon and carbon monoxide, and found that they were almost perfectly correlated (around 0.99) with estimates for nitrogen oxides. Thus, our model-based concentrations should be viewed as an indicator of the traffic-related pollutant mixture rather than of any pollutant specifically.

So, to arrive at an estimate of how much of a certain category of air pollution (traffic-related air pollution) each mother and child in their study had been exposed to, they used a computer model to come up with average emissions for vehicles all over the state, traveling at various average speeds corresponding with their various categories of roads, for each year in their study. Then they entered that, along with all the other types of data mentioned above (winds, atmospheric conditions, traffic volume, road layout) into another computer model to arrive at the final answer.

I'm not criticizing their model; it actually seems like a pretty good one to my untrained eye. But my point is that there's a lot of extrapolating, averaging, assuming that what's true for location x will also be true for location y, and other things that make the model work but aren't grounded in direct observation and thus might not actually be true. 

That will be the case for any model, and this one has a few serious gaps in its data pool. They're missing eight years of traffic data from their eleven-year "period of interest," so they have to guess at what those numbers might be based on expected growth in traffic volumes. They're also missing traffic counts for some roads, so they estimate them based on the counts for other, similarly-sized roads.

It bears repeating that this model was not their only source of data on pollution exposure; they also used direct measurements taken by the EPA air-monitoring station(s) nearest to study participants' houses throughout the study period.

For traffic-related air pollution --- the type of pollution exposure they modeled rather than measured directly --- they found a difference between the highest- and lowest-exposure groups (with the former three times as likely to develop autism as the latter), but no difference between the lowest-exposure group and the two groups in the middle.

For the specific pollutants measured at EPA air-monitoring stations --- coarse and fine particulate matter, nitrogen oxides, ozone --- they found an increased likelihood of autism associated with greater exposure to particulate matter and nitrogen oxides, but not ozone. This effect was strongest during the third trimester of pregnancy. 

Unlike the other study I described that used the EPA air-quality data, this one did not find any change in the pattern when they adjusted for sociodemographic variables like child's sex, race/ethnicity, parents' educational level, mother's age, or mother's smoking during pregnancy.

Volk, H., Lurmann, F., Penfold, B., Hertz-Picciotto, I., & McConnell R. (2013). Traffic-Related Air Pollution, Particulate Matter, and AutismAir Pollution, Particulate Matter, and Autism JAMA Psychiatry, 70 (1) DOI: 10.1001/jamapsychiatry.2013.266