This manuscript presents a novel approach to improve calibration of low-cost PM_2.5 sensors using reference NO₂ measurements from distant reference monitors (<20km).  Publically available US EPA and PurpleAir PA-II outdoor sensor data were used in this study. The authors first located and quality controlled several PA-II sensors located near a reference monitor site (<100m). After sucessfully locating a sensor-monitor pair in Rubidoux, CA the authors made daily comparisons between the sensor and FRM instrument and hourly comparisons between the sensor and a BAM instrument. They concluded that the BAM instrument was sufficiently accurate to make the hourly comparisons. Over a two year study period they observed a lower correlation in summer months compared to the winter and that temperature and relative humidity had less of an impact in winter compared to summer. They tested two calibration methods, Multiple Linear Regression and Random Forest, considering additive and multiplicative terms including PM_2.5, temperature, relative humidity, and NO₂. The inclusion of NO₂ in these methods did result in improved sensor performance, even when distant NO₂ was included. While this study did test many variables/cominations of variables on sensor performance the results are only representative of a single PA-II sensor. Additionally, a whole year was used to train the calibration models. These two factors ultimately limit the applicability of their conclusions without major revisions.

Major Comments:

Section 3.1: 14 sensors were originally idenitfied in this study, however only 5 were selected based on their months of valid measurements data. Of these 5, 2 were explicitly eliminated based on correlation analysis between the sensors' and their A and B units. Based on Figure 1 it seems like both PA-II 7 and 8 would be suitable for this study while PA-II 2, 3, 5, & 6 were not (Sensor 5 included in Figure 1 but not on line 188). Your final results will be more applicable if you are able to demonstrate improvements in more than 1 sensor, even if the study period is less than 2 years. 

Line 298: What is the reasoning behind this 1:1 data split, specifically using the whole year of 2018 to train the models and apply to 2019. This implies that in practice you have to wait a whole year before collecting valid/corrected data with this method which hinders the use of low-cost sensors. And assuming minimal sensor drift from 2018 to 2019 and similar environmental conditions.

Minor Comments:

Figure 1: Please include info about PA sensors A and B in the caption as you did on line 193.

Figure 2: Include a 1:1 line for comparison.

Figure 3: Ensure x-axes are the same for the PM_2.5 graph and temperature+RH graph.

Figure sizes could be increased to improve readability. 

Line 36: Please clarify that FRM and FEM are US EPA designations and may not be applicable to every county. 

Line 61: "good a correlation" Please correct to "a good correlation".

Line 74: More discussion needed on how NO₂ contributes to PM_2.5 formation.

Line 127: Typo for US EPA

Line 131: What is the purpose of the 2-minute vs 80 sec interval?

Line 178: Please clarify the difference between the FRM instrument and the BAM instrument. Does the FRM only report daily values?

Line 206 + 236: You list 6 significant figures/3 decimal points for several of the PA-II sensors, yet these sensors are not that accurate. As per the manufacturer +/-10 ug/m³ for 0-100 ug/m³ and +/-10% for 100-500 ug/m³. Please correct.

Line 219: How are you defining the r correlation of 0.928 as "good"?

Line 220: You say performance of FRM and BAM did not correlate favorably, yet in line 203 you state that the non-FEM method compared well to FRM? Why do you conclude that the BAM is less favorably correlated to the FRM when its statistics are better than the PAs?

Line 230: Please clarify why the FRM instrument was not used to evaluate hourly performance? Were hourly FRM measurements not available?

Line 272: The referenced article does not actually consider NO₂ in their PM_2.5 calibration. They only used PM_2.5, Temperature, RH, CO, and wind speed in their models.

Line 293: "because month has a different slope..." Do you mean " because each month..."?

Lines 311 + 355: Can these lists be included as Tables rather than in-text to improve readability and when readers look at Tables 3-5.

Line 395: "Corresponding R² values did not differ meaningfully" Based on what statistics, do you have a p-value?

Line 408: How are you defining moderate and high correlations?

Line 412: "We used NO₂ for training a calibration model" Which NO₂ data to train from, from Rubidoux? Please clarify.

Line 430: "but not significantly" Based on what statistics, do you have a p-value?

Line 447: Please re-word sentence as the point is unclear.

Line 448: Please re-word to clarify that the inclusion of NO₂ as an environmental factor in the calibration has potential to improve...

Section 2.2 Please include more information about the monitoring instrumentation used, especially the NO₂ monitoring sites. 

Section 3.2 + 3.3: At various points you include or drop units for your RMSE, MSE, MAE and r stats. Please be consistent. Shouldn't r (R²) be unitless? Please be consistent in using r vs R².

Section 3.6.3: Please check units of ug/m³ as you often have "ugm³" in this section.

Equations 3, 4, & 5 could be included in the methods section rather than results.

The article's concept is very interesting and could benefit the scientific community. Having a calibration methodology that allows scientists to leverage openly available data to increase the spatial resolution of currently available official measurements at a low cost could provide novel insight into the field. Unfortunately, the study structure and data do not support the authors' claims due to the lack of a robust dataset and an unclear strategy between model training and model evaluation data groups.

- The article is confusing and hard to follow. Too much detail is given for non-relevant information but not enough for evaluation.

- The authors argue against multivariable linear regression analyses but use MLR without offering a reasonable justification for its use nor explain why its results from RF and MLR are comparable.

I recommend a major revision of the dataset, hypothesis setup, and result analysis. 

Other Notes:

- Title misspelled "Collocated".

- Line 47: "however" seems to be misplaced.

- Line 121: This sentence is poorly constructed and confusing.