12.4. Creating a Model to Correct PurpleAir Measurements

Now that we’ve matched together PM2.5 readings from AQS and PurpleAir sensors, we’re ready for the final step of the analysis: creating a model that calibrates PurpleAir measurements. Barkjohn’s original analysis fits many models to the data in order to find the most appropriate one. In this section, we’ll start by fitting several simpler models using the techniques from the Modeling with Summary Statistics chapter. Then, we’ll fit a few models from Barkjohn’s analysis, including the final model they chose for real-world use. Since the final model uses methods that we introduce later in the book, we won’t explain the technical details very deeply here—we instead encourage you to revisit this section after reading the Linear Models and ch:mult_intro chapters.

12.4.1. Loading in the Data

Let’s begin by loading in the cleaned dataset of matched AQS and PurpleAir PM2.5 readings. We do some data processing as we load in the CSV file. Since this section focuses on modeling, we’ve left it as an exercise to understand what the code does.

csv_file = 'data/cleaned_purpleair_aqs/Full24hrdataset.csv'
usecols = ['Date', 'ID', 'region', 'PM25FM', 'PM25cf1', 'TempC', 'RH', 'Dewpoint']
full = (pd.read_csv(csv_file, usecols=usecols, parse_dates=['Date'])
full.columns = ['date', 'id', 'region', 'pm25aqs', 'pm25pa', 'temp', 'rh', 'dew']
date id region pm25aqs pm25pa temp rh dew
0 2019-05-17 AK1 Alaska 6.7 8.62 18.03 38.56 3.63
1 2019-05-18 AK1 Alaska 3.8 3.49 16.12 49.40 5.44
2 2019-05-21 AK1 Alaska 4.0 3.80 19.90 29.97 1.73
... ... ... ... ... ... ... ... ...
12427 2019-02-20 WI6 North 15.6 25.30 1.71 65.78 -4.08
12428 2019-03-04 WI6 North 14.0 8.21 -14.38 48.21 -23.02
12429 2019-03-22 WI6 North 5.8 9.44 5.08 52.20 -4.02

12246 rows × 8 columns

We’ve included an explanation for each of the columns below.




Date of the observation


A unique label for a site, formatted as the US state abbreviation with a number. (We performed data cleaning for site ID CA1.)


The name of the region, which corresponds to a group of sites. (The CA1 site is located in the West region.)


The PM2.5 measurement from the AQS sensor.


The PM2.5 measurement from the PurpleAir sensor.


Temperature, in Celcius.


Relative humidity, ranging from 0 to 100%.


The dew point of the air. (Higher dew point means more moisture in the air.)

Before we start modeling, we want to understand the data. We’ll start by making some simple visualizations. First, we’ll make a plot of the weekly average AQS PM2.5 over time.

plt.figure(figsize=(8, 4))


plt.ylabel('Weekly avg PM2.5 (AQS)');

We see that most PM2.5 values range between 5.0 and 15.0 µg m⁻³.

Next, we’ll plot the distribution of both AQS and PurpleAir PM2.5 readings.

def plot_dist(col):
    # We need to make sure the x- and y-axes have the same limits for both plots,
    # or else the plots will be difficult to compare
    sns.displot(data=full, x=col, stat='density',
            kde=True, aspect=3, height=3)
    plt.xlim([0, 60])
    plt.ylim([0, 0.11])
# We need to make sure the x- and y-axes have the same limits for both plots,
# or else the plots will be difficult to compare
../../_images/pa_modeling_12_0.png ../../_images/pa_modeling_12_1.png

We see that the distributions of PM2.5 readings are skewed to the right.

Lastly, we’ll plot PurpleAir against AQS readings.

sns.relplot(data=full, x='pm25aqs', y='pm25pa', s=30, aspect=1, height=4);
plt.xlim([0, 80])
plt.ylim([0, 80]);

Before starting this analysis, we expected that PurpleAir measurements would generally overestimate the PM2.5. And indeed, this is reflected in the scatter plot.

12.4.2. Our Modeling Procedure

First, let’s go over our modeling goals. We want to create a model that predicts PM2.5 as accurately as possible. To do this, our model can make use of the PurpleAir measurements, as well as the other variables in the data, such as the ambient temperature and relative humidity.

Here, we treat the AQS measurements as the true PM2.5 values. The AQS measurements are taken from carefully calibrated instruments and are actively used by the US government for decision-making, so we have reason to trust that the AQS PM2.5 values are accurate.

We plan to fit and compare several models on the data. To do so, we first randomly split the data in full into a training and testing set. For all of our models, we’ll fit the model on the training set and report the root mean squared error (RMSE) on the test set. Intuitively, the held-out test set mimics new measurements that the model “doesn’t see” while we fit it. This means that we can treat the test set error as an estimator for the model’s real-world performance.

We’ll set aside 1000 observations aside for the test set.


n = len(full)
test_n = 1000

# Shuffle the row labels
shuffled = np.random.choice(n, size=n, replace=False)

# Split the data
test  = full.iloc[shuffled[:test_n]]
train = full.iloc[shuffled[test_n:]]
date id region pm25aqs pm25pa temp rh dew
2685 2019-04-25 CA2 West 7.17 8.06 22.99 48.33 11.46
10214 2018-08-30 NC4 Southeast 12.60 23.13 33.01 46.22 19.98
10516 2019-08-14 VT1 North 6.78 8.70 25.44 49.25 14.01
... ... ... ... ... ... ... ... ...
5402 2019-10-10 GA1 Southeast 6.70 11.99 23.52 60.24 15.35
861 2019-01-31 AZ1 Central Southwest 9.40 10.98 20.90 29.33 2.31
7282 2018-08-10 IA2 North 18.67 43.22 29.60 50.17 18.15

11246 rows × 8 columns

We also plan to evaluate each model’s predictions on the test set, so we’ll define a function to compute the RMSE of a set of predictions.

def rmse(predictions):
    return np.sqrt(np.mean((test['pm25aqs'] - predictions)**2))

12.4.3. Model 1: A Simple Constant Model

Now, let’s fit our first model—a constant model like the ones we worked with in the previous Modeling with Summary Statistics chapter. Constant models only predict one number. That is, our model \( f_{\theta}(x) \) is:

\[ \begin{aligned} f_{\theta}(x) = {\theta} \end{aligned} \]

For example, if we set \( \theta = 5.5 \), then the model would predict that the AQS PM2.5 reading is always 5.5.

In our previous discussion of the constant model, we found that we minimize the mean squared error when we set \( \hat{\theta} \) to the mean of the response variable. In this case, we set \( \hat{\theta} \) to the mean AQS PM2.5 value, and our model is simply:

\[ \begin{aligned} f_\hat{\theta}(x) = \text{mean(aqs)} \end{aligned} \]
def model_1(train):
    '''f(x) = θ'''
    mean = train['pm25aqs'].mean()
    def predict(data):
        return np.repeat(mean, len(data))
    return predict

Now, we can fit the model on the training set and check its accuracy on the test set.

predict = model_1(train)

We’ve found that this simple model has a loss of 5.45 µg m⁻³. Intuitively, this means that the model will usually be around 5.45 µg m⁻³ away from the actual AQS measurement. We can see this when we mark the mean on the distribution of AQS readings.

mean = train['pm25aqs'].mean()
plt.axvline(mean, c='red', linestyle='-.');
# Most points are within 5.36 µg m⁻³ away from the mean

This model is too simple to use in practice—it performs especially badly when PM2.5 values are high, which is exactly where we care about model accuracy the most! Still, this serves as a useful baseline for future models and demonstrates our modeling procedure.

12.4.4. Model 2: Adjusting PurpleAir by a Constant

Let’s fit a constant again, but with a twist: we’ll fit a model that adjusts PurpleAir measurements by a constant rather than simply outputting the average AQS measurement. Our model is:

\[ \begin{aligned} f_{\theta}(x_i) = \text{PA}_i + \theta \end{aligned} \]

In this model, \( \text{PA}_i \) represents the PurpleAir measurement for a row \( x_i \) in the data, and \( \theta \) is a constant we once again need to fit.

We want to minimize the mean squared loss. Let \( \text{AQS}_i \) denote the AQS reading for row \( x_i \) of the data. Then, the mean squared loss is:

\[\begin{split} \begin{aligned} L(\theta) &= \frac{1}{n} \sum_i (\text{AQS}_i - f_{\theta}(x_i))^2 \\ &= \frac{1}{n} \sum_i (\text{AQS}_i - (\text{PA}_i + \theta))^2 \end{aligned} \end{split}\]

To minimize the loss, we can derive that \( \hat{\theta} = \frac{1}{n} \sum_i(\text{AQS}_i - \text{PA}_i) \). In other words, \( \hat{\theta} \) is the mean difference between AQS and PurpleAir readings. We’ve left the derivation as an exercise for the reader. After fitting on the training data, we have \( \hat{\theta} = -4.86 \):

np.mean(train['pm25aqs'] - train['pm25pa'])

With this, we can implement our second model.

def model_2(train):
    '''f(x) = PA + θ'''
    mean_diff = np.mean(train['pm25aqs'] - train['pm25pa'])
    def predict(data):
        return data['pm25pa'] + mean_diff
    return predict

To compare our models, we’ve written a small helper function called model_results that outputs the models and their RMSE on the test set. We’ll call this function to compare the results of the two models we’ve defined thus far.

model_results([model_1, model_2])
f(x) = θ 5.45
f(x) = PA + θ 7.33 Why Did Model 2 Perform Worse?

Surprisingly, the second model has a worse RMSE than the first. Why did this happen? One way to diagnose these models is to subtract the model’s predictions from the actual observed AQS values. These differences are called the residuals.

def residuals(model):
    predict = model(train)
    return test['pm25aqs'] - predict(test)

Let’s plot the residuals for both models and compare.

def plot_residuals(models):
    model_names = pd.Index([model.__name__ for model in models], name='model')
    resids = (pd.concat(
        [test.assign(resid=residuals(model)) for model in models],
    g = sns.relplot(data=resids, x='pm25pa', y='resid',
                    col='model', col_wrap=3, aspect=0.8,
    g.map(lambda **k: plt.axhline(y=0, color='black', linestyle=":"))
plot_residuals([model_1, model_2])

Note that a residual of 0 means that the model correctly predicted the PM2.5. A positive residual means the model underestimated PM2.5, and a negative residual means the model overestimated PM2.5.

We see that both models perform poorly at higher PM2.5 measurements. Model 1, the simple constant model, tends to underestimate PM2.5. Likewise, Model 2, the model that adjusts PurpleAir by a constant, tends to overestimate PM2.5.

We saw earlier in this section that PurpleAir sensors tend to overestimate PM2.5. However, subtracting a constant value doesn’t seem to do enough to correct PurpleAir readings.

Next, we’ll fit linear models on the data. In this book, we cover linear models in later chapters (Chapters 15 and %s). However, we include linear models in this section in order to match Barkjohn’s analysis, and to produce the final model that is currently in real-world use.

12.4.5. Model 3: Simple Linear Regression

Making a scatter plot of PurpleAir and AQS measurements shows that a linear model can be appropriate:


By eyeballing this plot, we might guess that the PurpleAir PM2.5 is about twice as high as the actual PM2.5. This idea is encapsulated in the simple linear model:

\[ \begin{aligned} f_{\theta}(x_i) = \theta_0 + \theta_1 \text{PA}_i \end{aligned} \]

To predict the PM2.5, this model multiplies the PurpleAir measurements by \( \theta_1 \), then adds a constant \( \theta_0 \). Using a Linear Model for Calibration

This is a calibration scenario. We have PurpleAir sensors which we want to calibrate to match AQS sensors. Then, we can adjust the PurpleAir readings using the correction we get from calibration.

This is a two-step procedure:

  1. Fit a calibration model that uses AQS measurements to predict PurpleAir measurements.

  2. Invert the model to find \( \hat \theta_1 \) and \( \hat \theta_0 \).

In other words, we’ll fit this model first:

\[ \begin{aligned} \text{predicted } \text{PA}_i = b + m \cdot \text{AQS}_i \end{aligned} \]

Then, we invert this model:

\[\begin{split} \begin{aligned} \text{PA}_i &= \hat b + \hat m \cdot \text{AQS}_i \\ \text{PA}_i - \hat b &= \hat m \cdot \text{AQS}_i \\ \text{AQS}_i &= \frac{1}{\hat m} \text{PA}_i - \frac{\hat b}{\hat m} \end{aligned} \end{split}\]

Which gives \( \hat \theta_1 = \frac{1}{\hat m} \) and \( \hat \theta_0 = - \frac{\hat b}{\hat m} \)

Why invert the model?

This procedure might seem a bit roundabout. Why not fit a linear model for \( f_{\theta}(x) \) that uses PurpleAir to predict AQS directly? This would get \( \hat \theta_1 \) and \( \hat \theta_0 \) without needing to invert anything.

Intuitively, linear models assume that the explanatory variable, or the variable we put on the x-axis, has no error. Linear models also assume that all the error occurs when we measure the response variable, or the variable we put on the y-axis. Thus, during calibration we treat the AQS measurements as the x-axis variable since these measurements have relatively little error. This allows us to model the error in PurpleAir measurements. If we fit \( f_{\theta}(x) \) directly, we’ll say the opposite: that the measurement error happens in the AQS readings rather than PurpleAir readings.

As a simpler example, let’s say we’re calibrating a weight scale. We could do this by first placing known weights—say 1 kg, 5 kg, and 10 kg—onto the scale, and seeing what the scale reports. Analogously, the AQS measurements are the known quantities, and we check what the PurpleAir sensors report.

For a more rigorous treatment of statistical calibration, see [Osborne, 1991].

Now, let’s define the model.

from sklearn.linear_model import LinearRegression

def model_3(train):
    '''f(x) = θ₀ + θ₁PA'''
    # Fit calibration model using sklearn
    X, y = train[['pm25aqs']], train['pm25pa']
    model = LinearRegression().fit(X, y)
    m, b = model.coef_[0], model.intercept_
    # Invert model
    theta_1 = 1 / m
    theta_0 = - b / m
    def predict(data):
        return theta_0 + theta_1 * data['pm25pa']
    return predict
model_results([model_1, model_2, model_3])
f(x) = θ 5.45
f(x) = PA + θ 7.33
f(x) = θ₀ + θ₁PA 2.96

We see that the linear model performs considerably better than the other models we’ve done. This is reflected in the residuals:

plot_residuals([model_1, model_2, model_3])

Happily, the residuals of the linear model shows that it still performs relatively well even when PM2.5 is high.

Under this model, \( \hat \theta_1 = 0.52 \) and \( \hat \theta_0 = 1.54 \), so our fitted model predicts:

\[ \begin{aligned} f_{\hat \theta}(x) = 1.54 + 0.52 \text{PA} \end{aligned} \]

12.4.6. Model 4: Incorporating Relative Humidity

The final model that Barkjohn selected was a linear model that also incorporated the relative humidity:

\[ \begin{aligned} f_{\theta}(x_i) = \theta_0 + \theta_1 \text{PA}_i + \theta_2 \text{RH}_i \end{aligned} \]

Here, \( \text{PA}_i \) and \( \text{RH}_i \) refer to the PurpleAir PM2.5 and the relative humidity for row \( x_i \) of the data. This is an example of a multivariable linear regression model—it uses more than one variable to make predictions.

As before, we’ll fit the calibration model first:

\[ \begin{aligned} \text{predicted } \text{PA}_i = b + m_1 \cdot \text{AQS}_i + m_2 \cdot \text{RH}_i \end{aligned} \]

Then, we invert the calibration model to find the prediction model:

\[\begin{split} \begin{aligned} \text{PA}_i &= \hat b + \hat m_1 \cdot \text{AQS}_i + \hat m_2 \cdot \text{RH}_i \\ \text{AQS}_i &= - \frac{\hat b}{\hat m_1} + \frac{1}{\hat m_1} \text{PA}_i - \frac{\hat m_2}{\hat m_1} \text{RH}_i \end{aligned} \end{split}\]

Which gives \( \hat \theta_0 = -\frac{\hat b}{\hat m_1} \), \( \hat \theta_1 = \frac{1}{\hat m_1} \), and \( \hat \theta_2 = - \frac{\hat m_2}{\hat m_1} \).

def model_4(train):
    '''f(x) = θ₀ + θ₁PA + θ₂RH'''
    # Fit calibration model using sklearn
    X, y = train[['pm25aqs', 'rh']], train['pm25pa']
    model = LinearRegression().fit(X, y)
    [m1, m2], b = model.coef_, model.intercept_
    # Invert to find parameters
    theta_0 = - b / m1
    theta_1 = 1 / m1
    theta_2 = - m2 / m1
    def predict(data):
        return theta_0 + theta_1 * data['pm25pa'] + theta_2 * data['rh']
        return (data['pm25pa'] - data['rh'] * m2 - b) / m1
    return predict
model_results([model_1, model_2, model_3, model_4])
f(x) = θ 5.45
f(x) = PA + θ 7.33
f(x) = θ₀ + θ₁PA 2.96
f(x) = θ₀ + θ₁PA + θ₂RH 2.58

We see that out of the four models, the model that incorporates PurpleAir and relative humidity performs the best. This is also reflected in its residual plot.

plot_residuals([model_3, model_4])

From the residual plot, we can see that Model 4’s residuals are generally closer to 0. Model 4’s residuals trend downward less than Model 3’s, indicating a better model fit.

After fitting the model, we have: \( \hat \theta_0 = 5.77 \), \( \hat \theta_1 = 0.524 \), and \( \hat \theta_2 = -0.0860 \) so our fitted model predicts:

\[ \begin{aligned} f_{\hat \theta}(x) = 5.77 + 0.524 \cdot \text{PA} - 0.086 \cdot \text{RH} \end{aligned} \]

And thus, we’ve achieved our goal: we have a model to correct PurpleAir PM2.5 measurements! From our analysis, this model achieves a test set error of 2.58 µg m⁻³, which is a useful improvement over our baseline models. If the model can maintain this error rate for real-world data, it should be practically useful—for instance, a “moderate” PM2.5 level corresponds to 12-35 µg m⁻³, while a “unhealthy for sensitive groups” PM2.5 level corresponds to 35-55 µg m⁻³ [EPA, 2016]. Our residual plot shows that applying our correction to PurpleAir sensors makes most measurements within 5 µg m⁻³ away from the true PM2.5, so we can reasonably recommend using PurpleAir sensors to report air quality.