\(\hbox {PM}_{2.5}\) is a critical air pollutant associated with many health effects. In the USA, \(\hbox {PM}_{2.5}\) monitoring relies on reference monitors, and a limited number of monitoring sites restrict the availability of high-resolution data. Low-cost air sensors, such as PurpleAir (PA), offer a promising alternative due to their potential for dense deployment. However, these sensors often produce biased data, typically overestimating \(\hbox {PM}_{2.5}\) measurements. To address these biases, correction models have been developed, though the influence of co-location distance and the number of sensors used in these correction models remains underexplored. This work evaluates the impact of co-location distance and sensor numbers on PA sensor correction models using EPA reference monitors as the ground truth. Multiple linear regression models were created incorporating \(\hbox {PM}_{2.5}\) readings from single and multiple PA sensors, along with relative humidity and temperature data from PA sensors. Constrained inference with ANOVA was applied to analyze the relationship between co-location distance and model accuracy. Our results show that using data from multiple PA sensors improves predictive accuracy ( \(R^2\) = 0.50–0.70, RMSE = 2.5–3.0 \(\mu g/ m^3\) ) compared to single-sensor models ( \(R^2\) = 0.30–0.60, RMSE = 1.9–3.5 \(\mu g/ m^3\) ). Model performance declines with increasing co-location distance, with a significant drop in accuracy beyond 30 km. This study suggests that considering distance and incorporating data from multiple sensors in correction models improves the accuracy of PA sensor data. Findings from this study will enhance the understanding of low-cost sensor networks for optimizing corrections and obtaining accurate \(\hbox {PM}_{2.5}\) estimates at a higher spatiotemporal resolution, which can significantly improve air quality monitoring and policy-making efforts in regions where only low-cost sensors are available.
