Preliminary results

The results are in, and drum roll please…. There’s reason to investigate further.

The air quality beside the Weston—Mount Dennis tracks is, if my calculations are correct (and there are lots of ifs), ever-so-slightly worse than the air away from the tracks.

Proving so was tricky business. A recap:

First, I calibrated the sensors by running them all at my house. I found that, indeed, they are all measuring well, and that roughly 95% of the variability in the sensors can be attributed to variability in the atmosphere.

Each sensor, though, measures ever-so-slightly differently. So while Pi0 might show an average over 24 hours of 10.2μ/m³, Pi1 would measure 11.4, and Pi2 measure 12.1. Over time, however, these differences are more-or-less constant.

Figure 1: Correlation for a 24-hour period

In Figure 1, Pi1 (green) consistently measures higher than Pi0 (blue), which tends to measure lower than the average of all four. But, importantly, all four Pis move in sync–so we can be reasonably sure that they’re actually measuring something.

The problem with using Pi0 and Pi1 to measure the air quality beside the tracks is obvious: if I put Pi0 there, the air would seem better than it is. If I put Pi1 there, it would seem worse. The trick, then, is to adjust the Pis.

Unfortunately, there is no calibration screw on the sensors, so we need to calibrate them mathematically.

The difference between each Pi’s measurement and the average of all Pis can be used as a constant to adjust each individual Pi. Pi0 can be adjusted up (or Pi1 down) by adjusting by the difference between its average readings over time, and the average of all the average readings over time.

To calibrate the machines, I took 100 measurements over a several days outside my house. The results were as follows:

PM10 adjustmentPi0Pi1Pi2Pi3
This Pis average8.7910.228.958.35
Average of Pis 0-39.089.089.089.08
Adjustment to average0.29-1.140.130.73
Adjustment to Pi00-1.43-0.160.44
PM2.5 adjustmentPi0Pi1Pi2Pi3
This Pis average5.977.616.656.20
Average of Pis 0-36.616.616.616.61
Adjustment to average0.64-1.005-0.040.41
Adjustment to Pi00-1.64-0.68-0.23

So, in order to adjust, say, the Pi1 so that it reads the same as Pi0, we subtract 1.64 from its PM2.5 readings, and 1.43 from its PM10 readings. (Pi0 is an mathematically arbitrary reference, but it’s the one farthest from the tracks—it’s at my house—so it makes sense as a comparison. And we could just as easily adjust to the average of the Pis instead of a particular Pi; again, it is just an arbitrary choice.)

The results:

Knowing, then, how to make a comparison, I did so.

Locations next to the tracks appear to have slightly higher concentrations of PM2.5. My house is approximately 425m from the tracks, and I have an average concentration of 2.9μ/m³. The house nearest the tracks (Pi2) has 6% more, and the house slightly farther away (Pi1) has 8.4% more.

PM2.5Pi0Pi1Pi2
Raw average2.884.793.75
Adjusted average relative to Pi02.883.153.07
Measured difference  to Pi0 (abs)00.260.19
Percent difference08.416.14

For PM10, the results are similar. My house (Pi0),  has 5.3μ/m³. The house nearest the tracks has a 15% higher concentration. The house slightly farther away from the tracks has a 16% higher concentration.

PM10Pi0Pi1Pi2
Raw average5.307.766.39
Adjusted average (ref Pi0)5.306.336.23
Measured difference to Pi0 (abs)01.020.93
Percent difference016.1814.89

So, there’s good news and bad news–and much news in between.

The bad news first.

  • It does look like the homes next to the tracks have higher pollution, though there are many caveats. First of all, this is a short-term reading (one week). Second, the sensors, though I’ve done my best, are cheap. Third, and most importantly, I’m very new to this. I might have totally screwed it up.
  • Third, and most importantly, I did screw this up. For PM2.5 the statistical certainty of my findings is quite low: p=.36 (for Pi0:Pi1 PM2.5) and p=.28 (Pi0:Pi2 PM2.5. That means there is a roughly 30% chance that these effects were due to random variations, and not physical differences.
  • For PM10, my results are much more certain. It seems like the PM10 pollution near the tracks is about 20% worse, with a p<.03.

The good news:

The sensors appear to work. We can (and will) continue to monitor the pollution levels, and we can do so cheaply.

The in-between news:

  • It is very far from clear that these differences in concentration are meaningful in terms of human health. The air quality, even when it is slightly more polluted, appears to be very good. The differences are small.
  • It is not at all certain that the tracks are the source of the pollution. Mount Dennis is a few kilometers away (and closer to downtown) than my house. The differences could be attributable to that distance, or something else entirely.
  • Indeed, it is possible that the tracks are not the source of pollution. The house 72m from the tracks had slightly higher levels than the house 39m away. This defies my predictions.
  • We should study the pollution more; the PM2.5 results could be due to chance.

You can check my work here.