HealthCAir

• Aim 1: Issues of air quality can affect all children. However, some may be disproportionately impacted by having high concentrations of PM2.5 at or near their school. To understand this issue better, we built unique high-quality data of PM2.5 pollution exposure at every school in CA.


• Aim 2: We established an empirical relationship between PM2.5 levels recorded at schools in CA and being upwind from key point sources of pollution.

We model each of our health outcomes using a two-stage least squares instrumental variables (2SLS IV) regression. The two modeling stages are as follows:

1. Model PM2.5 using our instrument and fixed effects:

\[ \widehat{PM}_{2.5}=\beta_0 + \beta_1 \kern 0.05em I + \beta_2 \kern 0.05em year + \beta_3 \kern 0.05em month + \beta_4 \kern 0.05em county + \beta_5 \kern 0.05em temperature \]

\[ \kern 1.5em + \beta_6 \kern 0.05em elevation + \beta_7 \kern 0.05em income + \beta_8 \kern 0.05em population + \beta_9 \kern 0.05em wspd_{avg} + \epsilon \]

2. Model the 12-month change in medical diagnoses using this predicted PM2.5 and same fixed effects:

\[ Y = \Delta \frac{\sum Diagnoses}{Pop_{zipcode, 0-19}} = \delta_0 + \delta_1 \kern 0.05em \widehat{PM}_{2.5} + \delta_2 \kern 0.05em year + \delta_3 \kern 0.05em month + \delta_4 \kern 0.05em county + \delta_5 \kern 0.05em temperature \]

\[ \kern 6em + \delta_6 \kern 0.05em elevation + \delta_7 \kern 0.05em income + \delta_8 \kern 0.05em population + \delta_9 \kern 0.05em wspd_{avg} + \epsilon \]

• \( \widehat{PM}_{2.5} \) = the predicted PM2.5 9-month average for every zip code and month
• \( \beta_i \) = the first stage regression coefficient for variable i
• \( \delta_i \) = the second stage regression coefficient for variable i

The medical diagnoses in our study are likely to develop over long term exposures to PM2.5. This is why we used a rolling 9 month average of PM2.5 to predict each medical diagnosis. The medical outcomes are measured as the change in the number of diagnoses per 1,000 children over a 12-month period, where the medical counts in a given month are measured as the sum of total diagnoses over a 3-month period. We applied this regression framework to each of our ten medical outcomes of interest.

The instrument, \(I\), is calculated using several different variables involving wind, pollution, and distance variables. We considered the nearest 15 pollution point sources to each school, and for each, we used their PM2.5 emissions measured in tons per year (TPY), distance to the school, wind speed, and wind direction to compute this instrument:

\[ I_{zmy} = \sum_{ps^{*}=1}^{15} \sum_{d_{m}=1}^{D_{m}}\sum_{H_{d}=1}^{24}\theta_{downstream_{zd_{m}}} \times \widetilde{TPY}\kern -0.2em_{ps} \times S_{zd_{m}} \]

• \( zmy \) = zip code, month, and year combination
• \( ps^{*} \) = pollution point source prefiltered to only include those with average TPY levels greater than 4
• \( D_m \) = number of days in month
• \( d_m \) = day of month
• \( H_d \) = hour of day
• \( \theta_{downstream_{zd_{m}}} \) = wind bearing in relation to each school
• \( \widetilde{TPY}\kern -0.2em_{ps} \) = point source pollution output (TPY), scaled using standard scaler, followed by min-max scaler
• \( S_{zd_m} \) = wind speed

Instrument Demonstration

We have included a video which helps to illustrate this instrument calculation and how it was used to predict PM2.5 levels at each school.