This paper examines air-quality trends in Wake County, North Carolina, from 2017 to 2025 using the U.S. Environmental Protection Agency (EPA) Air Quality Annual Summary dataset. Annual measurements of PM2.5, PM10, and ozone recorded at two monitoring sites, Millbrook School and Triple Oak, were cleaned, deduplicated, and aggregated into a dataset of 33 site-year-pollutant observations. Descriptive statistics, time-series plots, boxplots, and geographic maps were used to summarize the data, and multiple linear regression was implemented in R with the tidyverse and tidymodels frameworks to quantify temporal trends and site differences. The results show that PM2.5 and PM10 fluctuated across the study period while ozone remained stable, and that Millbrook School recorded PM2.5 concentrations approximately 1.10 micrograms per cubic meter higher than Triple Oak after adjusting for year (p = 0.0125). Observation completeness averaged above 90 percent in most years, indicating the dataset is reliable for long-term analysis. A multiple linear regression model with year, monitoring site, and pollutant as predictors explained 98 percent of the variability in annual means (R-squared = 0.98), and a train/test split evaluation produced a root mean squared error of 0.88. The project demonstrates a fully reproducible workflow for environmental monitoring analysis and identifies directions for seasonal and health-focused follow-up research.
This project addresses four research questions:
The analysis uses the EPA Air Quality Annual Summary dataset, which reports annual summary statistics for pollutants measured at monitoring sites across the United States. This project focuses on PM2.5, PM10, and ozone at two sites in Wake County, North Carolina, from 2017 through 2025.
Data were filtered to the target pollutants and years, deduplicated by selecting a single representative measurement duration for each pollutant, and collapsed across monitoring instruments within each site-year combination. Descriptive statistics and four visualizations were produced, and multiple linear regression models were fit and evaluated using the tidymodels framework.
PM2.5 and PM10 varied over the study period, ozone remained stable, Millbrook School recorded higher PM2.5 than Triple Oak, and observation completeness was high in most years. The regression models quantify these patterns and achieve strong predictive accuracy on held-out data.
Air pollution is one of the most important environmental risk factors for public health. Exposure to fine particulate matter (PM2.5) has been linked to cardiovascular and respiratory hospital admissions (Dominici et al., 2006), and the World Health Organization has identified PM2.5, PM10, and ozone among the pollutants of greatest health concern (World Health Organization, 2021). In the United States, the Clean Air Act requires the Environmental Protection Agency (EPA) to set National Ambient Air Quality Standards (NAAQS) for these criteria pollutants, and a nationwide monitoring network maintained under the Air Quality System (AQS) provides the data needed to evaluate whether those standards are being met (U.S. Environmental Protection Agency, 2024).
North Carolina monitors air quality through a network of state and federal sites. Wake County, the most populous county in the state and home to Raleigh, contains multiple monitoring stations operated in cooperation with the North Carolina Department of Environmental Quality (2025). Understanding how pollutant levels change over time and vary across locations within a county is essential for public-health surveillance, land-use planning, and the evaluation of air-quality policy.
Annual summary data are well suited to long-term trend analysis because they aggregate thousands of hourly or daily measurements into comparable yearly statistics. However, the raw AQS files contain considerable redundancy: the same underlying measurement can appear under multiple monitoring durations, parameter occurrence codes, and pollutant standards. A reproducible analysis therefore requires careful, documented data preparation. This project was motivated by the opportunity to demonstrate a transparent end-to-end workflow, from raw data to statistical inference, using free public data and open-source tools.
The objectives of this study are (1) to describe and quantify trends in PM2.5, PM10, and ozone from 2017 to 2025; (2) to test whether the two monitoring sites differ after controlling for time; (3) to identify which pollutant varies most; and (4) to evaluate data completeness as a measure of reliability.
The data were obtained from the Air Quality Annual Summary dataset
published by the U.S. Environmental Protection Agency and distributed
through the catalog.data.gov open-data portal (U.S. Environmental
Protection Agency, n.d.). The file
annual_37_183_0021_2017.csv covers monitoring site
37-183-0021 in Wake County, North Carolina (state code 37, county code
183, site code 0021), beginning with the 2017 calendar year.
Each row of the dataset corresponds to an annual summary record for one pollutant, one monitoring duration, and one point of operation at a site. The underlying data were collected by automated continuous monitors and filter-based samplers operated according to EPA quality-assurance protocols, then submitted to the AQS Data Mart. For this project the raw rows are treated as annual summary statistics: each record includes the arithmetic mean concentration, the number of valid observations, the percentage of expected observations that were collected, and extreme values such as the first maximum concentration (U.S. Environmental Protection Agency, n.d.).
After filtering, the analysis uses the following variables:
local_site_name: the monitoring site (Millbrook School
or Triple Oak).latitude and longitude: geographic
coordinates of each site (WGS84 datum).parameter_name: the pollutant measured.year: the calendar year of the annual summary
(2017–2025).units_of_measure: the units in which concentrations are
reported.arithmetic_mean: the annual arithmetic-mean
concentration (the response variable of the statistical models).observation_count: the number of valid observations in
the year.observation_percent: the percentage of expected
observations that were collected.minimum_value and first_maximum_value: the
minimum and first maximum concentrations recorded.Several limitations are important. First, the two sites are not fully comparable: Triple Oak monitors only PM2.5, whereas Millbrook School monitors PM2.5, PM10, and ozone, so the site comparison is restricted to PM2.5 and pollutant comparisons within Triple Oak are impossible. Second, the dataset contains annual summaries rather than daily or hourly measurements, which obscures seasonal patterns and short-term pollution events. Third, only nine years of data are available for most series, limiting statistical power for detecting slow trends. Fourth, the raw file contains duplicate records arising from multiple monitoring durations and pollutant standards; these were resolved through documented data-preparation decisions described below. Finally, the annual means include only the observations that were actually collected, so years with low completeness (for example, PM2.5 in 2021–2022) may under-represent true conditions.
Data preparation was performed in R version 4.5.0 using the
tidyverse package family (Wickham et al., 2019). The raw
file was imported with read_delim() using the semicolon
delimiter. Records were then filtered to the three pollutants of
interest and to the years 2017 through 2025; the 2026 records were
excluded because 2026 was still in progress at the time of analysis and
its observation percentages were far below the completed years.
To resolve the redundancy described above, a single representative
measurement duration was selected for each pollutant: hourly averages
(“1 HOUR”) for ozone, and daily block averages (“24-HR BLK AVG”) for
PM2.5 and PM10. This choice ensures that all pollutants are compared on
a comparable averaging basis and that the annual mean reflects a full
day of sampling. Exact duplicate rows produced by multiple pollutant
standards were removed with distinct(), and the remaining
observations from different monitors (parameter occurrence codes) at the
same site and year were averaged to yield one observation per
site-year-pollutant.
## Rows: 1880 Columns: 52
## Pollutants monitored: 222
## Monitoring sites: Millbrook School, Triple Oak
## Year range: 2017 - 2026## Rows in the cleaned dataset: 33| Monitoring Site | Pollutant | Site-Year Observations |
|---|---|---|
| Millbrook School | Ozone | 8 |
| Millbrook School | PM10 Total 0-10um STP | 8 |
| Millbrook School | PM2.5 - Local Conditions | 8 |
| Triple Oak | PM2.5 - Local Conditions | 9 |
Two derived variables were created to support modeling and
visualization. A short pollutant factor (PM2.5, PM10,
Ozone) was created from parameter_name so that model output
and figure facets are readable, and local_site_name was
converted to a factor with Millbrook School as the reference level, so
that regression coefficients are interpreted relative to Millbrook
School. The year variable was kept as a continuous
predictor so that each coefficient can be interpreted as the average
change in concentration per additional calendar year.
Descriptive statistics (mean, median, standard deviation, minimum, maximum) were computed for each pollutant. Data completeness was summarized by the average percentage of expected observations per year. Because the response variable, the annual arithmetic-mean concentration, is continuous, linear regression is the appropriate modeling framework (Kuhn & Silge, 2022). Three regression analyses were performed:
Models were specified with the parsnip package from
tidymodels using a linear_reg() specification with the
lm engine. Parameter estimates, standard errors, test
statistics, and p-values were extracted with tidy(), and
the fitted model was summarized with
extract_fit_engine() |> summary() and
coef(), following the workflow introduced in the course
notes (Kuhn & Silge, 2022). To evaluate predictive performance, the
cleaned data were split into training (80%) and testing (20%) sets with
initial_split() using stratification by pollutant, the
model was refit on the training set, and predictions on the testing set
were evaluated with the mean squared error (MSE), mean absolute error
(MAE), and root mean squared error (RMSE). A fixed random seed was used
so that the split is reproducible.
| Pollutant | Term | Estimate | Std. Error | t Statistic | p-value |
|---|---|---|---|---|---|
| Ozone | (Intercept) | 0.1632 | 0.7266 | 0.2246 | 0.8298 |
| Ozone | year | -0.0001 | 0.0004 | -0.1640 | 0.8751 |
| PM10 | (Intercept) | -46.9812 | 357.6437 | -0.1314 | 0.8998 |
| PM10 | year | 0.0304 | 0.1769 | 0.1719 | 0.8692 |
| PM2.5 | (Intercept) | 405.3891 | 269.3991 | 1.5048 | 0.1831 |
| PM2.5 | year | -0.1963 | 0.1333 | -1.4728 | 0.1912 |
pm25_data <- air_clean |>
filter(pollutant == "PM2.5")
site_fit <- linear_reg() |>
set_engine("lm") |>
fit(arithmetic_mean ~ year + local_site_name, data = pm25_data)
site_fit_results <- site_fit |>
extract_fit_engine() |>
summary()
coef(site_fit_results)
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 264.3150620 157.45574296 1.678663 0.11538930
## year -0.1264823 0.07789043 -1.623849 0.12670181
## local_site_nameTriple Oak -1.1010461 0.38422435 -2.865633 0.01245963
tidy(site_fit)
## # A tibble: 3 × 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 264. 157. 1.68 0.115
## 2 year -0.126 0.0779 -1.62 0.127
## 3 local_site_nameTriple Oak -1.10 0.384 -2.87 0.0125| Term | Estimate | Std. Error | t Statistic | p-value |
|---|---|---|---|---|
| (Intercept) | 134.993 | 117.165 | 1.152 | 0.259 |
| year | -0.063 | 0.058 | -1.078 | 0.290 |
| local_site_nameTriple Oak | -1.069 | 0.385 | -2.775 | 0.010 |
| pollutantPM10 | 5.851 | 0.395 | 14.803 | 0.000 |
| pollutantOzone | -8.587 | 0.395 | -21.727 | 0.000 |
| R-squared | Adj. R-squared | F Statistic | p-value | Residual df |
|---|---|---|---|---|
| 0.98 | 0.977 | 337.936 | 0 | 28 |
## # A tibble: 1 × 5
## arithmetic_mean .pred mean_squared_error mean_absolute_error
## <dbl> <dbl> <dbl> <dbl>
## 1 0.0437 0.518 0.773 0.668
## # ℹ 1 more variable: root_mean_squared_error <dbl>## # A tibble: 3 × 8
## pollutant n_years mean_value sd_value median_value minimum maximum
## <fct> <int> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 PM2.5 17 8.08 0.964 7.83 6.51 9.94
## 2 PM10 8 14.5 1.06 14.2 13.4 16.7
## 3 Ozone 8 0.0440 0.00216 0.0441 0.0402 0.047
## # ℹ 1 more variable: avg_observation_percent <dbl>Table 1 summarizes the distribution of the three pollutants. PM10 has the highest average concentration (14.5), followed by PM2.5 (8.08), while ozone is reported on a different scale in parts per million (0.044). PM10 also shows the greatest absolute variability, whereas ozone is the most stable. Average observation completeness exceeds 83 percent for every pollutant, with ozone the most complete at roughly 96 percent.
Table 2 reports the two monitoring sites and their coordinates. The sites are approximately 22 kilometers apart, with Millbrook School located east of downtown Raleigh and Triple Oak located near the western edge of Wake County.
Figure 1 maps the locations of the two monitoring sites. Millbrook School (35.856 N, -78.574 W) sits in the eastern part of the county, while Triple Oak (35.865 N, -78.820 W) sits to the west.
Figure 2 shows the annual trends for each pollutant. PM2.5 fluctuated between roughly 6.5 and 10.0 micrograms per cubic meter, with a visible dip in 2020 and an elevated period around 2023; PM10 fluctuated between about 13.4 and 16.7 micrograms per cubic meter; ozone remained essentially flat near 0.044 parts per million. Because Triple Oak monitors only PM2.5, the PM10 and ozone panels contain a single line for Millbrook School.
Figure 3 compares the distribution of annual concentrations between the two sites. For PM2.5, Millbrook School shows both a higher median and a wider spread than Triple Oak, suggesting systematically higher concentrations at the eastern site. Note: Triple Oak monitors only PM2.5, so the PM10 and ozone panels show data for Millbrook School alone, and the Triple Oak boxes are empty.
Figure 4 plots the percentage of expected observations that were collected. Completeness was above 90 percent in most years for all pollutants. The notable exception is PM2.5 and PM10 during 2021 and 2022, when completeness dropped to roughly 72 to 74 percent, coinciding with the period when Millbrook School reported fewer monitor records.
Figure 5 visualizes the site-comparison regression for PM2.5. After adjusting for year, the fitted lines show that Millbrook School is estimated to be about 1.10 micrograms per cubic meter higher than Triple Oak, a difference that is statistically significant (p = 0.0125). The slight negative slope of both lines reflects the weak, non-significant downward trend in PM2.5 over the study period.
The per-pollutant trend models for Millbrook School found no statistically significant linear trend for any pollutant over 2017–2025: PM2.5 declined by an estimated 0.196 micrograms per cubic meter per year (p = 0.19), PM10 increased by an estimated 0.030 micrograms per cubic meter per year (p = 0.87), and ozone changed by a negligible -0.000059 parts per million per year (p = 0.88). The small number of years (eight observations per pollutant) limits the power of these tests.
The site-comparison model for PM2.5 explained about 42 percent of the variability in annual means (R-squared = 0.417, adjusted R-squared = 0.334) and was statistically significant overall (p = 0.023). Triple Oak was estimated to be 1.10 micrograms per cubic meter lower than Millbrook School holding year constant (95% confidence interval approximately 0.28 to 1.92; p = 0.0125), indicating a meaningful spatial difference within the county.
The full model including year, site, and pollutant explained 98 percent of the variability in annual concentrations (R-squared = 0.980, adjusted R-squared = 0.977). Triple Oak remained significantly lower than Millbrook School (estimate = -1.07, p = 0.010), PM10 averaged 5.85 units higher than PM2.5 (p < 0.001), and ozone averaged 8.59 units lower (p < 0.001), reflecting the very different measurement scales of the pollutants. The year coefficient was not significant (estimate = -0.0625, p = 0.29). On a held-out testing set, the model achieved a mean squared error of 0.773, a mean absolute error of 0.668, and a root mean squared error of 0.879, indicating that the model generalizes well.
The results show that the two particulate pollutants varied considerably over the nine-year window while ozone remained stable, which is consistent with ozone being a regional secondary pollutant whose annual mean is buffered by photochemical and meteorological controls (U.S. Environmental Protection Agency, 2024). The dip in PM2.5 around 2020 and the elevated period around 2023 mirror nationwide and regional patterns associated with economic activity, wildfire smoke, and meteorology. The finding that Millbrook School records higher PM2.5 than Triple Oak suggests that local emissions sources, road traffic, or urban density differ across the county. The high observation completeness in most years supports the reliability of these comparisons.
For public-health surveillance, the stable ozone and moderate particulate levels observed here suggest that, while concentrations generally remained within a moderate range, the systematic PM2.5 difference between the two sites could be relevant for residents living near the Millbrook School area. For county planners, the results highlight the value of maintaining a geographically distributed monitoring network: the two-site comparison would be impossible if only one site existed. For students and analysts, the documented workflow demonstrates how free public data and open-source R packages can be combined into a fully reproducible analysis.
The principal limitations are the small number of years (nine), the incomparability of the two sites for PM10 and ozone, the use of annual summaries rather than daily or hourly data, and the incomplete observations for particulate pollutants in 2021 and 2022. Because only eight observations are available for the trend models of PM10 and ozone, the absence of a significant trend should not be interpreted as evidence that no change occurred.
This project analyzed air-quality trends in Wake County, North Carolina, from 2017 to 2025 using a reproducible R workflow. The main findings are: (1) PM2.5 and PM10 fluctuated over the study period, with no statistically significant long-term trend; (2) ozone remained stable; (3) Millbrook School recorded significantly higher PM2.5 than Triple Oak after adjusting for time; (4) PM10 varied most in absolute terms; and (5) observation completeness was high in most years, supporting data reliability.
Analysts and agencies working with the AQS Annual Summary should (a) always document the handling of duplicate duration, standard, and parameter-occurrence records; (b) restrict trend analysis to completed calendar years; and (c) verify that all sites being compared monitor the same pollutants. For residents and planners, the PM2.5 difference between the two sites suggests that localized monitoring remains important even within a single county.
Future research should (1) incorporate daily or hourly observations to study seasonal patterns and exceedance days; (2) add more years of data and additional sites across North Carolina to increase statistical power; (3) relate air quality to meteorological covariates such as temperature, wind, and precipitation; and (4) link concentration estimates to local health outcomes and population exposure, building on the framework demonstrated here.
Dominici, F., Peng, R. D., Bell, M. L., Pham, L., McDermott, A., Zeger, S. L., & Samet, J. M. (2006). Fine particulate air pollution and hospital admission for cardiovascular and respiratory diseases. JAMA, 295(10), 1127–1134. https://doi.org/10.1001/jama.295.10.1127
Kuhn, M., & Silge, J. (2022). Tidy modeling with R. O’Reilly Media. https://www.tmwr.org/
North Carolina Department of Environmental Quality. (2025). Air quality monitoring. https://www.deq.nc.gov/about/divisions/air-quality
U.S. Environmental Protection Agency. (n.d.). Air quality annual summary dataset. Data.gov. https://catalog.data.gov/dataset/air-quality-annual-summary
U.S. Environmental Protection Agency. (2024). NAAQS table. https://www.epa.gov/criteria-air-pollutants/naaqs-table
Wickham, H., Averick, M., Bryan, J., Chang, W., McGowan, L. D., François, R., Grolemund, G., Hayes, A., Henry, L., Hester, J., Kuhn, M., Pedersen, T. L., Miller, E., Bache, S. M., Müller, K., Ooms, J., Robinson, D., Seidel, D. P., Spinu, V., … Yutani, H. (2019). Welcome to the tidyverse. Journal of Open Source Software, 4(43), 1686. https://doi.org/10.21105/joss.01686
World Health Organization. (2021). WHO global air quality guidelines: Particulate matter (PM2.5 and PM10), ozone, nitrogen dioxide, sulfur dioxide and carbon monoxide. https://www.who.int/publications/i/item/9789240034228
## Analysis performed with R R version 4.6.1 (2026-06-24) on aarch64-apple-darwin23 ; tidyverse 2.0.0 ; tidymodels 1.5.0 ; knitr 1.51 ; rmarkdown 2.31