Atmospheric Environment 256 (2021) 118432

Contents lists available at ScienceDirect

Atmospheric Environment
journal homepage: www.elsevier.com/locate/atmosenv

Calibration of low-cost PurpleAir outdoor monitors using an improved

method of calculating PM2.5
Lance Wallace a, *, Jianzhao Bi b, Wayne R. Ott c, Jeremy Sarnat d, Yang Liu d
a
Independent Researcher, 428 Woodley Way, Santa Rosa, CA, 95409, USA
b
University of Washington, USA
c
Stanford University, USA
d
Gangarosa Department of Environmental Health, Rollins School of Public Health, Emory University, Atlanta, GA, 30322, USA

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• 33 PurpleAir monitors were calibrated

against 27 EPA regulatory PM2.5
monitors.
• A reproducible method to estimate
PM2.5 from PurpleAir particle counts is
presented.
• The method is superior to the undis­
closed algorithm of the Plantower
sensor.
• The method has improved accuracy and
precision, with a lower limit of
detection.
• A different calibration factor for PM2.5
aerosols from wildfires was indicated.

A R T I C L E I N F O A B S T R A C T

Keywords: PM2.5 hourly average measurements from 33 outdoor PurpleAir particle monitors were compared with hourly
Low-cost monitors measurements from 27 nearby US EPA Air Quality System (AQS) stations employing Federal Equivalent Method
PurpleAir (FEM) monitors in California over an 18-month (77-week) period. A transparent and reproducible alternative
PM2.5
method (ALT) of calculating PM2.5 from the particle numbers in three size categories was used in place of the
Calibration factor (CF)
estimates provided by Plantower, the manufacturer of the sensors used in PurpleAir monitors. The ALT method
was superior in several ways (better precision, lower limit of detection, improved size distribution) compared to
Plantower’s CF1 or ATM data series. PurpleAir monitors were strongly correlated with the nearby US EPA Air
Quality System AQS stations. A calibration factor (CF) ranging between 2.9 and 3.1 was empirically derived for
the PurpleAir estimates using the ALT method. This value was based on comparing the average value of 177,329
PurpleAir measurements to the value calculated from the FEM stations. The monitoring period included about 13
weeks showing very high outdoor values due to several major fires covering several hundred thousand acres. The
CF during these 13 weeks averaged 2.39, whereas the CF for the remaining 64 weeks averaged about 3.21,
suggesting a different response to the smoke from wildfires compared with normal ambient fine particulate
matter (PM2.5). The standard Plantower CF1 data series overestimated the FEM values by about 40%, in
agreement with several other studies.

* Corresponding author.
E-mail address: [email protected] (L. Wallace).

https://doi.org/10.1016/j.atmosenv.2021.118432
Received 10 March 2021; Received in revised form 13 April 2021; Accepted 19 April 2021
Available online 4 May 2021
1352-2310/© 2021 Elsevier Ltd. All rights reserved.
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

1. Introduction monitors for possible use in their air quality monitoring programs,
including Lane County, OR and Salt Lake County, Utah (Kelly et al.,
Low-cost air quality monitors reporting particulate matter (PM) 2017). Sayahi et al. (2019) reported a limit of detection (LOD) of about
concentrations directly to the Internet are now being widely used in the 6 μg/m3.
US and in Asia (Williams et al., 2014; US EPA, 2017). Multiple manu­ A long-term comparison of four sensors, including the two Plantower
facturers are producing tiny electronic sensors for use in these monitors. models PMS 5003 and PMS 7003, took place at two schools near
The sensors often employ lasers to scatter light off the particles into Southhampton, England (Bulot et al., 2019) between March 13, 2018,
detectors. The scattered light is analyzed, generally using Mie scattering and February 28, 2019. The two Plantower models showed excellent
theory, to estimate particle size, number, and mass. agreement with each other (Spearman coefficient of 0.98). They were
Among the most widely used monitors, with >16,000 devices in also well correlated with the background reference station data, with a
operation, is PurpleAir (https://www2.purpleair.com/). PurpleAir bias ranging between 1.15 and 1.22 at one school, and similar values at
monitors use one or two identical sensors (Plantower Model PMS 5003) the second school.
manufactured by the company Plantower, (http://www.plantower. Zheng et al. (2018) compared three Plantower Model PMS 3003
com/en/). This sensor is constructed using a laser providing light at sensors to a reference monitor for 30 days at a US EPA facility in
an approximate 650 nm wavelength that is scattered throughout a 90◦ Research Triangle Park, NC. Correlation across the three sensors was
viewing area (He et al., 2020). The sensor measures particle numbers per very high (R2 = 97%). Correlation with the reference monitor was
deciliter for 6 overlapping size categories: >0.3 μm, >0.5 μm, >1 μm, moderate (R2 = 66%) for both 1-min and 1-h averaging periods, but a
>2.5 μm, >5 μm and >10 μm (Plantower User Manual, 2016). From semiempirical correction for RH improved the correlation to 93% for
these particle number measurements, Plantower uses a proprietary al­ both averaging times. The RH values at the site averaged 64% ± 22%.
gorithm to estimate various PM size fractions, including PM1, PM2.5, and The authors concluded that the 3003 sensors were highly precise and
PM10 concentrations. Two data series are produced (CF1 and ATM) for capable of measuring PM2.5 to within 10%.
these three particle mass concentration estimates. Plantower provides Levy Zamora et al., (2019) evaluated three Plantower PMS A003
no information on the calibration aerosol used to correct its sensor sensors when exposed to eight particulate matter (PM) sources (i.e.,
retrieval signal. incense, oleic acid, NaCl, talcum powder, cooking emissions, and mon­
Our aim in this study is to present the first calibration of PurpleAir odispersed polystyrene latex spheres under controlled laboratory con­
monitors using all monitors in the State of California within 500 m of ditions and also residential air and ambient outdoor air in Baltimore,
official US EPA Air Quality System (AQS) Federal Equivalent Method MD). Overall accuracy for PM2.5 ranged between 86% and 93% for in­
(FEM) stations. We also provide here an existing widely used transparent cense, cooking, residential air and ambient outdoor air, and precision
and reproducible independent alternative (ALT) approach to calculating ranged between 9% and 12% for these four sources. The authors remark
PM2.5 from the particle number estimates provided by the PurpleAir that performance was good for both PM1 and PM2.5 but poor for larger
monitors. This approach is widely used by manufacturers of optical particles.
particle monitors capable of collecting simultaneous data on multiple Three important studies appeared in early 2020. He et al. (2020)
particle size categories (e.g., the TSI Model 3330 (https://tsi.com/pr tested the response of three Plantower PMS5003 sensors to various
oducts/particle-sizers/particle-size-spectrometers/optical-particle-sizer aerosols of known composition and size. They developed the first esti­
-3330/)). Since this approach is central to our study, we devote an Ap­ mate of the transfer function governing the response of the Plantower
pendix to explaining the approach in detail and comparing it with the optical sensor to particles of different sizes. The authors found that
PM2.5 data provided by PurpleAir. In particular, we compare results on particles well outside the boundaries of the size channels contributed to
precision, bias, limit of detection, coefficient of variation, and size dis­ the sensor response. For example, 0.2 μm monodisperse particles were
tribution using multiple months-long datasets collected in two locations. readily detectable in the smallest channel (>0.3 μm). More concerning,
1.0 μm monodisperse particles provided the strongest signal in the
1.1. Previous studies employing PurpleAir monitors smallest size channel rather than the channels more directly related to
that size. This important finding adds considerable uncertainty to the
PurpleAir monitors have been evaluated in several studies. The size classifications employed by the Plantower sensors, and to the mass
South Coast Air Quality Management District, part of the California Air calculations based on those sizes, including our own PM2.5 estimates.
Resources Board (CARB), operates a program called AQ-SPEC that has Simultaneously, a major study used Plantower PMS A003 sensors in
issued reports on about a dozen low-cost monitors (http://www.aqmd. seven cities to predict the PM2.5 measurements made by Federal Refer­
gov/aq-spec). Their field evaluation of PurpleAir monitors took place ence Methods (Zusman et al., 2020). This study did not use the ATM or
from February to April 2016 (AQ-SPEC, 2016). This test found generally CF1 PM values, but rather the particle number counts. The authors chose
good agreement with two reference monitors for PM2.5 (R2 = 78% and to use the original overlapping definitions (i.e., >0.3 μm, >0.5 μm …)
90%), but poor agreement for PM10 (R2 = 34% and 45%). The PurpleAir after finding that they (slightly) outperformed the non-overlapping size
instrument overestimated PM2.5 concentrations by about 40% compared categories. The model used the first five of the six categories, choosing
to the two reference instruments. The report does not specify whether not to use the >10 μm category (even though all particles in that cate­
the CF1 or CF ATM series was used. Similarly, the Lawrence Berkeley gory also appear in the first five categories). Also entered were tem­
National Laboratory (LBNL) tested 7 monitors for ability to measure perature and humidity measurements by the low-cost monitors. Very
particles from a large number of common indoor sources such as cooking good results were obtained across the seven cities with 10-fold
and cleaning (Singer and Delp, 2018). Seven low-cost PM monitors were cross-validation (PM2.5 precision r = 0.99, accuracy R2 = 0.96, RMSE
tested using a variety of common indoor sources, including candles, 1.15 μg/m3).
incense, and cooking of various types. Four monitors, including Pur­ The third study (Bi et al., 2020) employed all outdoor PurpleAir
pleAir, were described as “quantitative” (within a factor of 2 compared monitors (2,090) operating in California in the year 2018 to provide
to two reference instruments) for most of the aerosol types measured. additional PM2.5 measurements in the geographic areas covered by 138
Indeed, the investigators commented that the low-cost monitors were official air monitoring stations. The density of sensors allowed creation
about as accurate as the research reference instruments. A presentation of a grid 1 km to a side (almost 500,000 cells). 5.84 million hourly
based on the same study concluded that the PurpleAir instruments were measurements were obtained. Geographic calibration of the instruments
“quantitatively much better than the others” Walker (2018). The CF1 with the reference monitors resulted in reducing the overall bias from
data series was used (Delp and Singer, 2020). 1.9 to ~0 μg/m3, and reducing the residual error by 36%. After
State and local air quality agencies have tested the PurpleAir correction for the 17% bias and down-weighting the PurpleAir values by

2
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

about a factor of 5 to account for lower-quality measurements, the au­ outdoor concentrations near 25–30 μg/m3.
thors found a useful improvement in estimating PM2.5 concentrations These and other studies dealing with the Plantower sensor are briefly
throughout the state. described in Table 1.
Wallace et al. (2020) co-located two PurpleAir monitors with two
research-grade monitor types (Piezobalance and SidePak) in 124 ex­ 2. Methods and materials
periments measuring PM2.5 from vaping marijuana liquid. The PurpleAir
instruments had a precision and coefficient of variation (COV) that was A database consisting of PM2.5 measurements from PurpleAir mon­
similar to the research-grade instruments, and both had been calibrated itors near EPA Air Quality System (AQS) monitors employing Federal
by comparison with gravimetric instruments. That study used the Equivalent Methods (FEM) to provide hourly estimates was prepared.
alternative (ALT) method presented in the Appendix below to estimate Across the entire state of California, 33 PurpleAir monitors were within
PM2.5. The calibration factor for PurpleAir was determined to be 3.0. 500 m of 27 EPA monitors. 500 m was used as a cutoff in an earlier study
Bi et al. (2021) examined 91 pairs of indoor/outdoor PurpleAir of PurpleAir monitors (Bi et al., 2020) based on low and slowly-changing
monitors within 500 m of each other over a 20-month period ending correlations between PurpleAir and AQS stations as a function of dis­
June 2020. Using LOWESS local regression, they determined infiltration tance. Data from November 2018 to April 2020 (18 months) was
factors for all indoor sites, with a mean value of 0.26 (IQR 0.15–0.34). downloaded from the PurpleAir website (PurpleAir.com) and the EPA
The resulting exposure error (difference between total indoor exposure website (https://aqs.epa.gov/aqsweb/airdata/download_files.html).
and exposure due to particles of ambient origin) was plotted as a func­ The 2-min average PurpleAir measurements were averaged over hourly
tion of outdoor concentration. The peak of the exposure error occurred periods to match the EPA measurements. The hourly PurpleAir averages
at low outdoor concentrations <5 μg/m3 and dropped to nearly zero at were required to have valid reading for at least 20 of the 30 possible

Table 1
Studies evaluating Plantower sensors.
Reference Year Data series Model N Location Time span Comment
PMS-

Chen 2017 CF1 or ATM 3003 N/A Taiwan N/A General outline of system with data archiving. Two short-
5003 term case studies.
Wang, K. 2018 CF1 or ATM 7003 17 Shanghai, China 7 days Compared with TEOM: Outdoor R2 0.72–0.78, mean RSD
21%; indoor R2 0.95–0.96, mean RSD 16%.
Zheng 2018 CF1 3003 7 Durham NC; NC: 90 days; 1-h mean errors of 200% in Durham, but 35%–46% in India,
Kanpur, India India 45 days indicating improved performance at high concentrations.
Following empirical RH correction, estimates are within
~10%.
Becnel 2019 CF1 or ATM 3003 50 Salt Lake County, 6 months Individual calibration improved RMSE by 1.8 times. Good
Utah agreement (88% R2) with reference monitors
Bulot 2019 CF1 or ATM 5003 6 Southhampton, UK 1 year Moderate to good correlation: 0.61 < r < 0.88.
7003
Francis 2019 ATM, Particle # 5003 N/A Sabah, Malaysia 1 week Use of particle number to estimate mass. PM1 & PM2.5 R2 =
0.82 & 0.88. PM10 unreliable.
He 2019 Particle # 5003 6 Clarkson Univ. NY Theoretical analysis of transfer function. All size channels
include response to sizes outside their boundaries.
Kaduwela 2019 Particle # 7003 1 Albany, CA 2 weeks School study including wildfire days (15X increase in particle
number indoors).
Levy 2019 CF1 or ATM A003 3 Baltimore, MD 1 month, 10 days Accuracy 87%–96%, precision 9%–10% for incense, cooking,
Zamora and residential indoor. Underestimates with precision
10–24% for NaCl, talcum powder, oleic acid. Overestimate of
ambient data by 1.6–2.4X, reduced to 0.9–1.4X after
correction for RH.
Magi 2019 ATM 5003 1 Charlotte, NC 16 months Multiple linear regression with T, RH, and BAM improved
accuracy by 27%–57%. 15% of data < LOD of 5 μg/m3.
Malings 2019 CF1 or ATM 5003 20 Pittsburgh, PA 17 months Multiple segmented linear regression with T and RH (10
parameters). Median correlation with BAM of 0.73; mean
absolute error 2.5 μg/m3; bias − 0.14 μg/m3
Masic 2019 CF1 or ATM 5003 2 Sarajevo, Bosnia 16 months R2 93% and 96%. Bias 37% and 31%
Tryner 2019 CF1 and ATM 5003 8 Fort Collins, CO 1 week Gravimetric correction reduced bias and RSD. CF1 values
50% higher than ATM values at high PM2.5 concentrations
Bi 2020 CF1 or ATM 5003 2090 California 1 year Overestimate of 13/11.1 = 17%, based on 137,000 1-h
measurements. Calibration reduced bias to ~0. Most
important parameter in estimating PM2.5 was the PM2.5/PM10
ratio.
Wang, Z. 2020 CF1 or ATM 1003 9 Berkeley, CA 1 month 3 different monitors incorporating the three models listed. All
3003 performed reasonably well (within a factor of two) with
5003 multiple different aerosols containing particles >0.25 μm.
Zusman 2020 Particle # A003 80 Seattle, 6 other 30–95 weeks PM2.5 precision r = 0.99, accuracy R2 = 0.96, RMSE 1.15 μg/
cities m3. Region-specific models in 6 other cities R2 0.74–0.95,
RMSE 0.84–2.46 μg/m3. Seattle model applied to other 6
cities: R2 0.67–0.84, RMSE 1.67–3.41 μg/m3.
Wallace 2020 Particle # 5003 4 Santa Rosa, CA 14 months (124 Indoor measurements of aerosol from vaping marijuana
Redwood city, CA experiments) liquids. PurpleAir monitors were collocated with research-
grade monitors. Precision and COV were comparable across
all monitor types. Calibration factor estimate was about 3.0.
Bi 2021 Particle # (ALT) 5003 91 indoor- California 20 months Infiltration factors and indoor-generated exposures were
outdoor estimated for pairs of PurpleAir indoor-outdoor monitors
pairs within 500 m distance. Exposure errors were calculated as a
function of outdoor concentrations

3
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

2-min measurements to be included in the analysis. Only PurpleAir sites the paper by Gupta et al. (2018), each of which found about a 35%–40%
with the PA-II monitors employing two independent Plantower PMS overestimate of PM2.5 for the PurpleAir monitors.
5003 sensors were included in the analysis. The PM2.5 estimates from the Individual calibration factors for each PurpleAir monitor can also be
two sensors within each monitor were required to agree within 30% of calculated as the ratio of the mean AQS to the mean ALT PM2.5 mea­
each other, corresponding to a precision [abs (A-B)/(A + B)] of 0.130. surement (Table S1). Four of the 33 monitors had fewer than 2 months of
About 90% of all PurpleAir 1-h averages met this condition. Only data and were not analyzed further. The 29 remaining monitors had at
monitors with at least 3 weeks of data were included. A small number of least 1800 h (75 days) to 10,770 h (>400 days) of hourly average
measurements with abnormal temperature and relative humidity read­ measurements. The median CF was 3.13 (Interquartile range (ICR) =
ings were also removed. The final database had 177,329 matched Pur­ 2.5–3.9). The range across the 29 individual sites was 2.2–6.0.
pleAir and EPA hourly PM2.5 estimates. A third approach to estimate a calibration factor used correlations
A crucial aspect of our approach is our use of a transparent and across each of the 33 sites. 31 of the sites had strong Spearman corre­
reproducible alternative method (ALT) to estimate the PM2.5 concen­ lation coefficients exceeding 0.50, and 16 of those sites had Spearman
trations. The ALT method is based on the number of particles per deci­ correlation coefficients >0.7. Averaging across these 16 “best” sites gave
liter reported by the PMS 5003 sensors in the PurpleAir instrument for a mean CF of 3.14 (SE 0.18).
the three size categories less than 2.5 μm in diameter. The method makes A fourth approach used regression of the ALT hourly values on the AQS
no use of either the CF1 or ATM data series calculated according to a values. Outliers more than a factor 5 away from the overall FEM/ALT ratio
proprietary and undisclosed algorithm by the Plantower manufacturers were deleted, leaving 154,900 hourly values in the regression. The
of the sensors. The method is fully explained in the Appendix to this resulting equation was ALT PM2.5 = − 0.10 (SE 0.0066) + 0.3409 (SE
article. The practical improvements allowed by the method in precision, 0.0006) FEM (adj. R2 = 0.679). This produced a CF estimate of 2.93 (0.01).
limit of detection, and distribution shape are also contained in the Ap­ The four approaches produced an average CF estimate of 3.05 (SE = 0.05).
pendix, using measurements by two co-located PurpleAir indoor moni­
tors in each of two homes, one in Santa Rosa, CA, and the other in
Redwood City, CA. 3.1. Effect of wildfires

3. Results and discussion The period of monitoring included three wildfires, each exceeding
100,000 acres burned. Each fire affected air quality readings for several
The means (standard errors) for the AQS measurements and our weeks up to two months. Over the 77 weeks of the study, 13 weeks
PM2.5 ALT estimates were 7.40 (SE 0.03) and 2.49 (0.01) μg/m3, affected by the wildfires had greater than 50% increases in the anoma­
respectively, leading to an overall average calibration factor (CF) of lies (departures from the 77-week mean) for both the PurpleAir and FEM
7.40/2.49 = 2.98 (0.01). An overall average calibration factor can also monitors (Fig. 1). There was a clear change in the FEM/ALT ratio during
be calculated for the Plantower CF1 data series. The same requirement these weeks. For those 13 weeks the CF was 2.39 (0.06) compared to
of the sensors to agree within 30% of their CF1 estimates was applied, 3.21 (0.07) in the remaining 64 weeks. The difference was significant (p
reducing the number of accepted measurements for this comparison only < 0.001) and likely due to the different composition, refractive index, or
to 154,900. The mean and standard error for the AQS measurements and reduced density of the wildfire smoke from the normal PM2.5 back­
the CF1 estimates were 7.98 (0.02) and 11.14 (0.03), respectively, ground level. Since not all FEM sites were affected by the fires, the
leading to a calibration factor for the CF1 series of 0.716 (0.003). This difference is likely to be even greater than the approximate 25% decline
finding of about a 40% (1/0.716) overestimate for the CF1 data series is observed. Delp and Singer (2020) studied the Camp wildfire of
in very good agreement with the finding of the AQ-SPEC program and November 8–21, 2018 in California using 53 PurpleAir monitors within
11 km of 12 Northern California AQS sites. They found a Smoke

Fig. 1. Variation of PM2.5 anomalies (ratios to 77-week mean values) associated with wildfires. The 13 weeks in which wildfires raised PM2.5 readings by 50% or
more had a lower average calibration factor (2.39) than the 64 weeks without wildfires (3.21).

4
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

Adjustment Factor of 0.42 for the PurpleAir monitors used. This were reaching their deliquescence limit at these higher RH values. The
reduction is in the same direction but larger than our value of 25% based curve shows a small effect of RH even at the lowest values, a finding also
on all monitors including those very distant and probably unaffected by of Zheng et al. (2018). A complicating factor is that the RH values
the wildfires. The overall CF from these 77 weekly averages was 3.09 (SE recorded by the PurpleAir monitor are considerably lower than the RH
0.09), within the values observed for the hourly and site averages. in the surrounding air. A mean decrease of about 15% was noted for two
Summarizing, five different ways of determining the overall CF using indoor PurpleAir monitors studied over a period of several months,
hourly, weekly, or site-specific approaches gave CF values in the range along with a corresponding increase of about 8 ◦ C due to the
of 2.9–3.1. near-infrared lasers.
Although it might be possible to use the curve shown in Fig. 2 above
to calculate the CF as a function of RH, the extreme variability of the 1-h
3.2. Effect of RH PM2.5 values suggests that not much gain would be expected from using
the exact relationship. Also, any extended data set (say a day or more)
Under conditions of increasing relative humidity (RH), inorganic will be likely to experience substantial diurnal swings in RH. The
atmospheric particles other than dust undergo deliquescence, a sudden average diurnal variation for the 33 PurpleAir sites is shown in Fig. 3.
increase in diameter and mass due to absorption of water (Seinfeld and Given the range of 30%–55% in expected diurnal variation (the
Pandis, 2nd edition). The RH at which this occurs varies according to diurnal variation curve would move upwards and downwards over the
chemical composition, but is often about 60%–85% RH for common seasons but continue to include a substantial range on most days), using
atmospheric salts such as KCl, Na2SO4, (NH4)2SO4, etc. For H2SO4, the overall calibration factor of about 3 would provide a reasonable
however, the molecule grows in diameter and mass with no threshold estimate of PM2.5 in many cases.
RH required. Most of these molecules also exhibit hysteresis, main­
taining the water in the particle as RH declines, even well below the RH
3.3. Comparison of alternate (ALT) PM2.5 estimates to plantower CF1
value at which the particle first took on water. For optical particle
system
monitors, then, an increase in RH will produce an increase in particle
size and a corresponding increase in the apparent mass. Since many of
The ALT and CF1 hourly PM2.5 estimates are very highly correlated
these increases will occur at fairly high RH values near 60%, we would
(Fig. 4). The CF1 estimates are about 4.5 times the ALT estimates (N =
not expect a strictly linear effect, but a convex upward shape. We can
159,400); R2 = 98%). This suggests that the unknown algorithm adop­
test this by observing the effect, if any, of RH on the FEM monitors. Many
ted by Plantower may be closely related to our approach using the
of these are beta attenuation monitors (BAM), in which the sampler inlet
particle numbers in the three smallest size categories. Based on this close
is heated to reduce moisture deposition in the pipeline, and thus these
agreement of CF1 and ALT measurements, we can calculate that for the
FEM monitors would not be expected to be influenced strongly by RH. In
2.93–3.15 range of the calculated CF for the ALT system, the calculated
fact, not much effect can be seen: a linear regression is PM2.5 (FEM) =
CF for the CF1 system would lie between 0.65 and 0.72. This corre­
7.186 (0.050) + 0.0037 × RH (R2 adj. = 0.000054), so the entire range is
sponds to the approximate 40% overestimate of the CF1 series as found
from 7.19 at RH = 0% to 7.56 at RH = 100%. We can then consider the
in other studies (AQ-SPEC (2016); Kelly et al., 2017; Sayahi et al., 2019).
ratio of the ALT to the FEM PM2.5 estimates (Fig. 2) using distance-
weighted least squares, since we expect nonlinear effects to appear at
relatively high RH. Indeed, the graph shows an inflection point at about 3.4. Precision and coefficient of variation
50% RH when the slope increases until about 80% RH and then stays
nearly constant until 100% RH. This would be expected if some particles As explained in the Appendix, the ALT system was compared to the

Fig. 2. Ratio of PurpleAir (CF 3) ALT PM2.5 estimates to FEM PM2.5 as a function of RH, using distance-weighted least squares regression. There is a slow nearly
linear increase from 0% to about 50%–60% but then a faster increase until leveling off at about 80%.

5
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

Fig. 3. Diurnal variation for PurpleAir (PA) RH values averaged across each hour of the day. (N = 174,169).

Fig. 4. Scatterplot of PM2.5 estimates for the CF1 and ALT systems.

CF1 system in experiments carried out using duplicate collocated Pur­ 3.5. Size distribution of PM2.5 concentrations
pleAir monitors in two homes in Santa Rosa and Redwood City. The data
were collected between Jan 10, 2019, and March 26, 2020, a total of The CF 1 (and CF ATM) series distort the distribution of the PM1 and
about 350,000 measurements over 433 days. During this time, mean PM2.5 mass concentrations due to cutting off low concentrations and
precision in the ALT series was excellent at 4%–6%, compared to 7%–14% replacing them with zero. In our example dataset from Santa Rosa, for
for the CF1 series. Over 358 days in Santa Rosa when only normal indoor one PurpleAir monitor there were 353,551 observations, but 73,266
activities were carried out, the COV for the ALT series improved by 26%; (20.7%) reported PM2.5 values of zero for the CF1 data series (see Ap­
the corresponding value over 85 days in Redwood City was 22%. pendix, Figures A1 and A2). This distorts the distribution to the point
that no fit could be determined by the software program (Statistica)
using the CF1 distribution, although a log-normal distribution gave a
good fit to the ALT distribution.

6
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

Table 2 4. Conclusions
Comparison of the CF1 and ALT series LODs for PM2.5 in two locations, with
fraction of observations exceeding the LOD. We have applied an alternative method (ALT) to calculate PM2.5
Santa Rosa Redwood City estimates in a transparent and reproducible way rather than relying on
CF1 ALT CF1 ALT
the unknown algorithm provided by Plantower for their PMS 5003
sensors used in PurpleAir monitors. We have tested the results using 433
LOD (μg/m3) 1.77 1.15 2.55 1.23
days of monitoring in two homes and demonstrate that the ALT method
fraction > LOD 0.13 0.48 0.51 0.90
has superior performance with respect to precision, coefficient of vari­
ation, limit of detection, and less distorted size distribution.
3.6. Limits of detection (LOD)
To calibrate PurpleAir outdoor monitors using the ALT method of
estimating PM2.5, we compared hourly averages of 33 PurpleAir moni­
To calculate an LOD for a continuous monitor such as the PurpleAir
tors to the hourly averages of 27 official FEM sites within 500 m distance
instrument, we use a method developed in Wallace et al. (2010). In brief, the
over 20 months (77 weeks). Several different ways of computing the
method searches for the lowest concentration above which mean concen­
calibration factor (CF) resulted in a range of estimates from 2.93 to 3.15,
trations exceed their standard deviations by a factor of 3 more than 95% of
with standard errors ranging from 0.01 to 0.18. This value near 3.0 for
the time. The LOD was improved by 30%–50% for the ALT system
27 outdoor PurpleAir monitors was also the estimated CF for 4 indoor
compared to the CF1 system (Table 2). (The ALT system LODs were
PurpleAir monitors collocated with research-grade monitors over a year-
multiplied by the CF1/ALT ratio of 4.5 to make the comparisons shown.)
long study (Wallace et al., 2020).
The time period included major wildfires affecting observations over
3.7. Limitations of the study 13 weeks, and we detected a difference in the CF during the 13 weeks
from a value of 2.78 to a value of 3.12 in the remaining 64 weeks. We
No information was available on the siting of the outdoor PurpleAir also observed a nonlinear effect of relative humidity (RH), with the
monitors. The manufacturer recommends mounting the instrument PurpleAir results compared to FEM values increasing slowly at levels
away from vents and foliage; mounting high enough off any nearby from 0 to 50%, more rapidly at levels from 50% to 75%, then remaining
surface to avoid splashing water or snow; and mounting it with the open constant or slightly declining up to 100%. Although this behavior could
surface pointing down. The finding by He et al. (2020) of imperfect conceivably be incorporated into a model, considering the wide range of
assignment of particles to the size categories affects our PM2.5 estimates uncertainty and particularly the fact that any period of monitoring
directly. The finding by Zusman et al. (2020) that using the overlapping exceeding a day will encounter swings in RH over ranges of about 25%–
observations gave slightly better results than using the non-overlapping 30%, employing the central estimate near 3 would be sufficient in many
size categories, is further evidence of problems in assigning particles to cases.
their proper size category. These findings add uncertainty of unknown We also found a CF ranging between 0.65 and 0.72 for the CF1 data
magnitude to our estimates. series provided by PurpleAir, corresponding to an approximate 40%
The FEM monitors are subject to occasional quality assurance pro­ overestimate of PM2.5 values using the CF1 series. Overestimates of
cedures carried out by EPA (https://www.epa.gov/outdoor-air-quali PurpleAir monitors in the 30%–40% range have been found in other
ty-data/pm25-continuous-monitor-comparability-assessments). The studies.
program provides gravimetric measurements (Federal Reference Our study is the first to provide a PM2.5 calibration factor for Pur­
Methods, or FRM) to compare to the FEM measurements. Ten of the 27 pleAir monitors based on 18 months of hourly observations by 33
FEM sites included in this study were visited in recent years (See Excel monitors near 27 FEM sites. We have also shown that an existing
file in Supplementary Information). Only 3 sites, all in Fresno, had biases alternative (ALT) method of calculating PM2.5 outperforms the CF1 and
<10%. Six of the seven remaining sites had positive biases exceeding ATM data series offered by PurpleAir. These results are presently being
20% (and up to 38%) during the years (2018–2020) overlapping our incorporated into the PurpleAir database as a correction factor “overlay”
study. The remaining site had both high positive and high negative bias available to all users.
in adjacent years. Overall, however, the assessment program has
roughly equal numbers of positive and negative bias of the FEM sites. Funding
We compared daily PM2.5 averages at FEM sites to nearby (500 m)
FRM sites. Only 16 of our 27 FEM sites were associated with FRM sites. Funding for the portion of this study concerned with calibration of the
There were 2290 days of data recorded. The mean (median) FEM/FRM Purpleair monitors was provided by the Multi-Angle Imager for Aerosols
ratios ranged from 0.74 to 1.55 (0.64–1.67). A linear regression of FEM (MAIA) science team and the Multi-angle Imaging SpectroRadiometer
on FRM had a slope of 0.89 and an intercept of 1.56 with an R2 of 85%. (MISR) science team, both at the Jet Propulsion Laboratory (JPL), Cali­
The FEM error was <20% for only 8 of 16 cases, and >30% for 4 cases. fornia Institute of Technology, led by D. Diner (Subcontract # 1588347
The errors had a slight tendency toward overestimates, with 10 of 16 and 1363692). No funding was provided for the study described in the
FEM/FRM ratios >1 and a mean (median) error of 1.11 (1.15). These Appendix. The work of J. Sarnat and Y. Liu was supported by the MAIA
errors clearly contribute to the “noise” observed across the comparisons science team at the JPL, California Institute of Technology (Subcontract
of PurpleAir monitors to FEM monitors. However, we are unable to #1588347), and the National Institute of Environmental Health Sciences
calculate an overall bias, if any, based on only 16 of the 27 FEM sites in of the National Institutes of Health (Award # R01-ES027892). The content
our database. is solely the responsibility of the authors and does not necessarily represent
The temperature values reported by the PurpleAir monitors are the official views of NASA or NIH.
influenced by the heated internal region due to the near-infrared lasers.
In a long period of comparison of an indoor monitor in a Santa Rosa CRediT authorship contribution statement
home, the increase in temperature averaged 8 ◦ F. A corresponding
average decrease of 15% in RH was also measured. The particles Lance Wallace: Conceptualization, Methodology, Investigation,
entering the monitors thus undergo a sharp change in both T and RH Formal analysis, Writing – original draft. Jianzhao Bi: Methodology,
occurring across a few cm. Depending on how rapidly the particles Software, Validation, Investigation, Formal analysis, Writing – review &
adjust to the changing environmental conditions, the effects associated editing. Wayne R. Ott: Methodology, Software, Validation, Investiga­
with the recorded RH may actually relate to some mixture of ambient tion, Formal analysis, Writing – review & editing. Jeremy Sarnat:
RH and the internal RH of the monitor. Methodology, Validation, Investigation, Formal analysis, Writing –

7
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

review & editing. Yang Liu: Writing – review & editing, Supervision, Acknowledgements
Project administration, Funding acquisition.
Adrian Dybwad of PurpleAir corrected a pair of misapprehensions in
Declaration of competing interest an earlier version of this paper. The authors appreciate the commitment
of PurpleAir to creating a completely open and accessible database of
The authors declare that they have no known competing financial thousands of these sensors.
interests or personal relationships that could have appeared to influence
the work reported in this paper.

Appendix

Since our study uses an alternative (ALT) method of calculating PM2.5, this Appendix presents an explanation of the method and a comparison with
the corresponding PM2.5 estimates supplied by PurpleAir. These concentrations are provided with no explanation from the sensor manufacturer
(http://www.plantower.com/en/) of how they are calculated. The alternative method presented here is a standard method employed in many optical
particle monitors with multiple particle size measurement capabilities. It is transparent and reproducible by any user. It depends on the particle
numbers in three size categories as reported by PurpleAir. The basic approach is to select an intermediate particle diameter within each size category,
calculate the number of particles and the associated particle volume (assuming sphericity) and resulting particle mass (assuming an arbitrary density).
The resulting estimates of particle mass (e.g., PM2.5) can then be compared to those supplied by Plantower.
The PurpleAir data include two data series: CF1 and ATM. The CF1 series is described as “for laboratory use” and the ATM for “atmospheric”
conditions. Many of the above studies (Table 1) employ the PM values reported in either the CF1 or ATM data series. However, we find serious
deficiencies in both of these data series. For example, both series report multiple values of zero for typical indoor and outdoor measurements; yet in
our studies of scores of sites with hundreds of thousands of observations there are never cases when the particle numbers in the 0.3–0.5 μm or 0.5–1 μm
size categories are zero. The values of zero are apparently applied by the unknown Plantower algorithm to concentrations below some arbitrary cutoff
number.
Another deficiency is the relation between the CF1 and ATM data series. A two-week period between 4/25/19 and 5/3/19 during which both high
and low concentrations of PM2.5 were produced provided data comparing the two series at a wide range of concentrations. For PM2.5, it is clear that for
low concentrations (less than about 28 μg/m3), CF1≡ ATM (Fig. S1). A breakpoint occurs near 28 μg/m3, after which the ATM concentrations increase
more slowly than the CF1 concentrations (CF1/ATM ratio increases). Another breakpoint appears at 78 μg/m3, at which point the CF1 value is close to
50% higher than the ATM value, where it remains for all values > 78 μg/m3. A comparison of the ALT series with the CF1 series shows a linear
relationship, but a comparison with the ATM series is nonlinear (Fig. S2).
Because of these deficiencies, we examined whether an alternative (ALT) method of estimating particle mass concentrations could perform better
than the CF1 or ATM series.

Outline of alternative (ALT) method

The approach adopted here is identical to that recommended by manufacturers of optical particle monitors capable of collecting simultaneous data
on multiple particle size categories (e.g., the TSI Model 3330 (https://tsi.com/products/particle-sizers/particle-size-spectrometers/optical-particle-si
zer-3330/)).

1. Calculate the number N of particles in each of the size categories 0.3–0.5 μm, 0.5–1 μm, 1–2.5 μm, 2.5–5 μm and 5–10 μm. Note: Since the
Plantower output provides the total number of particles >0.3 μm, the total number >0.5 μm, etc., one must subtract the total number >0.5 μm from
the total number >0.3 μm to arrive at the number of particles in the 0.3–0.5 μm size category. Similar subtractions will result in five size categories
under 10 μm: 0.3–0.5 μm, 0.5–1 μm, 1–2.5 μm, 2.5–5 μm, 5–10 μm.
2. Select an average diameter D for each of the five size categories up to 10 μm. For example, the average particle diameter D in the 0.3–0.5 μm size
category must be between 0.3 and 0.5 μm, so we can approximate it by the midpoint (0.4 μm) or by the geometric mean of the size boundaries
(0.387 μm). Several manufacturers use the geometric mean so we use it throughout this work.
3. Calculate the total particle volume V in each size category. This is given by multiplying the number of particles N times the volume per particle:
/
V = N πD3 6

4. Multiply by a density ρ to arrive at an estimate of the mass concentration M in each size category. We choose ρ to be the density of water, a choice
that is also made by the Model 3330 manufacturers:
M = ρV

5. Add the appropriate size categories to estimate PM1, PM2.5 and PM10. PM1 is the sum of the mass concentrations in the two smallest size categories:
0.3–0.5 μm and 0.5–1 μm. PM2.5 is the sum of the mass concentrations in the two smallest size categories plus the next largest size category 1–2.5
μm. PM10 adds the mass in the two largest size categories (2.5–5 μm and 5–10 μm) to the three smallest size categories.
6. Calculate a calibration factor (CF) for the aerosol mixture being monitored by comparison to gravimetric studies of the same mixture or to research-
grade monitors that have themselves been calibrated using gravimetric measurements.

8
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

Comparison of ALT method to plantower CF1 method

To compare results of the ALT method with the PM mass concentrations reported by Plantower, we consider measurements by two collocated
PurpleAir monitors in each of two homes, one in Santa Rosa, CA, and the other in Redwood City, CA.
In both homes, the PurpleAir monitors each contained two PA-II sensors, model PMS 5003, plus a sensor measuring temperature, relative humidity
(RH) and atmospheric pressure (Bosch Model BME 280) (https://www.bosch-sensortec.com/products/environmental-sensors/humidity-sensors-
bme280/). In the Santa Rosa home, the monitors were placed on the front of a dresser 33” (83 cm) high 0.5 m from the wall in a 30 m3 room in a
single-story 385 m3 private residence housing two people. The house employs forced air with a central fan on at all times. There is a return air filter
with electret fibers. Blower door tests have been performed on the home twice, and air exchange measurements have been made multiple times, with
results indicating a tight home with average outdoor air exchange rates of 0.15 (SD 0.05) h− 1.
The data were collected between Jan 10, 2019 and March 26, 2020, a total of about 350,000 measurements over 433 days. The PurpleAir in­
struments took samples every 80 s for the first 6 months and every 2 min thereafter. On most days (N = 379) the room’s door was open to the rest of the
house. On those days, normal household activities such as cooking were carried out. Neither resident was a smoker. On 55 days, the room was shut off
from the rest of the house, the floor register was sealed, the central fan was turned off, and experiments were run using various sources of fine and
ultrafine particles such as candles, laboratory hot plates, toaster ovens, and other typical particle sources. A full description of the experiments is
contained in Wallace et al. (2020). The experiments typically elevated the PM2.5 concentrations to levels above 1 mg/m3. These high concentrations
were permitted to decay over the next few hours before the room was opened up again.
In the Redwood City home, the two monitors were set up in a 43 m3 bedroom. The 2-story house has a volume of 430 m3. There is a gas furnace with
forced-air ventilation, but no air conditioner. There is one resident nonsmoker. The monitors collected mostly 2-min average data from May 21, 2019
through April 15, 2020. A 3-month dataset (July 23, 2019 through October 22, 2019) was chosen for analysis. This period included 7 days during
which experiments on various particle sources were carried out in the test room, which was closed off from the rest of the house during those days only.
A full description of the experiments at this second home is provided in a companion paper Ott et al., 2021. These periods of high concentrations were
analyzed separately from the remainder of the data (85 days).
Data Analysis. Raw data were examined for duplicates. Duplicates of every measurement occurred between March and November of each year
between 2017 and 2020, a period coinciding with daylight savings time. This error affects all downloaded reports of PurpleAir concentrations during
these months (up to April 10, 2020). The PurpleAir organization is working on correcting the error. Our study has eliminated all duplicates.
All data were examined for outliers, such as large negative values, and they were removed. These errors were rare; for one sensor, there were 3
negative values for PM2.5 out of 353,551 measurements, an error rate of about 10− 5.
The initial comparison of the ALT data series with the CF1 and ATM data series resulted in a strong linear relationship (R2 > 0.999) with the CF1
data series, but a nonlinear relationship with the ATM series. Therefore, the following analysis compares our ALT series with the CF1 series only.
Results are presented for PM1 and PM2.5. Although we have also considered PM10, the relationship between the ALT and CF1 series is not linear.
Also, the LOD is often high, such that the majority of PM10 concentrations as measured by the CF1 data series are below the LOD. We will not,
therefore, present results for PM10.
For both PM1 and PM2.5, we can compare our ALT values to the CF1 values considering bias, precision, coefficient of variation (COV), size dis­
tribution, and limit of detection (LOD).

Precision and bias

Bias across the four independent sensors was calculated with respect to the average of the four values. The bias was then corrected and the bias-
corrected precision was calculated using the equation.

Precision = abs(A-T)/(A + T)

where A is a measurement from one sensor and T is the average of the four sensors.
Over the 433 days measured at the Santa Rosa site, mean precision in the ALT series was excellent, ranging across the four sensors from 3% to 6%
for PM1 and from 4% to 6% for PM2.5 (Table A1). Corresponding values for the CF1 series are 9%–28% and 7%–14%. Bias relative to the mean of the
four sensors in the ALT series ranged from 0.95 to 1.06 for PM1 and from 0.93 to 1.06 for PM2.5. Corresponding values for the CF1 series are 0.86–1.12
and 0.96 to 1.03.

Table A1
Precision and bias for the ALT series compared to the CF1 series for PM1 and PM2.5 measured during 433 days in Santa Rosa.
Sensors A and B in monitors 1 and 2 are labeled as 1A, 1B, 2A, and 2B.

PM size Data series 1A 1B 2A 2B

Precision
PM1 ALT 0.03 0.05 0.04 0.06
CF1 0.11 0.10 0.09 0.28
PM2.5 ALT 0.05 0.06 0.04 0.06
CF1 0.07 0.09 0.12 0.14
Bias
PM1 ALT 1.06 0.99 0.95 1.00
CF1 1.12 1.01 1.01 0.86
PM2.5 ALT 1.01 1.00 1.06 0.93
CF1 1.02 0.96 1.03 0.99

For the 3-month period from July 23 to October 22, 2019 at the Redwood City location, two monitors were col-located. In this case, the mean
precision for the CF1 series was 9.6% compared to 2.86% for the ALT Series. Another way of looking at the difference is that 98% and 99% of values in

9
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

the ALT Series had precision better than 10% and 15%, respectively, compared with 73% and 83% in the CF1 series.
Coefficient of variation (COV)

This is a measure of the amount of variation within a dataset. It is calculated as the Relative Standard Deviation (RSD), the SD divided by the mean.
The smaller the COV, the more dependable will be the estimate of the mean.
Since the aerosol mixtures from the experiments at the two sites may vary considerably from the normal indoor aerosol, we compare the PM2.5
results for the experimental and non-experimental days separately (Table A2).

Table A2
Comparison of CF1 to ALT PM2.5 data series in Santa Rosa and Redwood City locations: mean, SD, and COV for days with and without major sources of indoor particles.

Santa Rosa (1/10/19 to 3/26/20) Redwood City (7/23/19 to 10/22/19)

Experimental days No experiments Experimental days No experiments

PM2.5 CF1 PM2.5 ALT PM2.5 CF1 PM2.5 ALT PM2.5 CF1 PM2.5 ALT PM2.5 CF1 PM2.5 ALT

N 54 54 378 378 7 7 85 85
mean 29.20 5.61 3.97 0.91 118.8 24.2 3.30 0.81
SD 19.57 3.68 7.96 1.35 10.4 2.42 2.58 0.49
RSD (COV) 0.67 0.66 2,00 1.48 0.09 0.10 0.78 0.61
COV improvement 2% 26% − 10% 22%
Ratio (CF1/ALT) 5.2 4.4 4.9 4.1

In both locations, the COVs of the CF1 and ALT series were not much different on the experimental days (higher concentrations), but for the lower
concentrations, the COV was substantially lower (better) for the ALT series, by 26% at the Santa Rosa location and 22% at the Redwood City location.

Calculation of the limit of detection (LOD)

To calculate an LOD for a continuous monitor such as the PurpleAir instrument, we use a method developed in Wallace et al. (2010). To be 95%
certain that a reported observation is > 0, we require that the mean value of multiple collocated instruments be > 3 times the standard deviation.
However, if the observations are in a series ordered by the mean of the instrument measurements, it can and typically does happen that an observation
will meet this criterion, whereas at a higher concentration it does not. Then the LOD is not this lower concentration. Therefore, we search for the lowest
concentration above which higher concentrations always have at least 95% of the calculated mean/SD ratios >3. For a large dataset, this can be done
by considering “batches” of, say, 100 observations at a time in the dataset ordered by mean concentrations, and counting the number of cases in which
the mean/SD ratio is < 3. As higher concentrations are examined, eventually a batch is found with fewer than 5 such cases. If testing more batches at
yet higher concentrations never shows 5 or more such cases, we assume that the earlier batch contains the limit of detection.
Two large datasets were employed to calculate the LOD for PM1 and PM2.5 in the two locations: a Santa Rosa dataset running from Jan 30, 2020, to
April 27, 2020, with 63,091 observations and a Redwood City dataset for the 3-month period from July 23 to October 22, 2019, with 65,319 ob­
servations. In all calculations, the bias of the four individual sensors compared to the mean was determined and the bias-corrected values were then
analyzed to determine the mean and standard deviation of each observation.
A Santa Rosa database running from Jan 30, 2020, to April 27, 2020, with 63,091 observations was used to calculate PM1 and PM2.5 LODs for the
CF1 and ALT series (Table A3). For the CF1 series, the LODs for PM1 and PM2.5 were 1.54 μg/m3 and 1.77 μg/m3, values that are comparable to those
reported in other studies (e.g., the PM2.5 LODs found for two PMS 5003 sensors (2.62–3.65 μg/m3) in two seasons by Sayahi, Butterfield, and Kelly;
Sayahi Supplementary Material, 2018). (Note: Since the LOD is determined by ordering all data by concentration and then determining the con­
centration above which the ratio of mean to standard deviation never exceeds 3 by as much as 5% for the remainder of the data, there is not an obvious
approach for determining the uncertainty of the estimate.) Despite the low CF1 LODs found here, only 6% of the PM1 concentrations and 13% of the
PM2.5 concentrations exceeded the LOD for the CF1 series. By comparison, the improved LODs for the ALT series resulted in 78% of PM1 concen­
trations and 48% of PM2.5 concentrations exceeding the LOD. (Note: The results for the ALT series have been multiplied by the factors relating the ALT
series to the CF1 series for proper comparison.)
In the Redwood City location, the 3-month database with 65,309 observations as described above was used. The PM2.5 LOD determined for the CF1
series was 2.55 μg/m3. The LOD for the ALT series was 1.23 μg/m3, about half that found for the CF1 series. For the CF1 data series, there were 32,100
(49.0%) values below the LOD. For the ALT data series, there were only 6900 (10.6%) values below the LOD. The improvement in the LOD was only by
a factor of 2, but it led to almost 90% detected vs. less than 50% for the CF1 series. For the PM1 CF1 series, the LOD was 1.98 μg/m3, and 47% of values
exceeded the LOD. For the ALT PM1 series, the LOD was 0.50, and 99.4% of values exceeded the LOD.

Table A3
Comparison of the CF1 and ALT series LODs for PM1 and PM2.5 in two locations, with fraction of observations
exceeding the LOD

Santa Rosa Redwood City

CF1 ALT CF1 ALT

PM1 LOD 1.54 0.52 1.98 0.50

fraction > LOD 0.06 0.78 0.47 0.99
PM2.5 LOD 1.77 1.15 2.55 1.23
fraction > LOD 0.13 0.48 0.51 0.90

These results are concerning, because outdoor values can be quite low in many parts of the US, and thus the PM1 and PM2.5 estimates in the CF1
data series will be below the LOD (i.e., indistinguishable from zero) for a possibly large fraction of all measurements. This problem can be largely

10
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

overcome by calculating the PM1 and PM2.5 values in the ALT method demonstrated here.
Size distribution of PM concentrations

The CF1 and CF ATM series distort the distribution of the PM1 and PM2.5 mass concentrations due to cutting off low concentrations and replacing
them with zero. In our example dataset, for one PurpleAir monitor there were 353,551 observations, but 73,266 (20.7%) reported PM2.5 values of zero
for the CF1 data series (Figure A1). This distorts the distribution to the point that a normal or lognormal fit could not be determined by the software
program (Statistica).

Fig. A1. Histogram of 353,511 reported PM2.5 concentrations in the CF1 data series provided by Plantower. “No of obs” = number of observations. The 73,266
reported values of zero distort the distribution so extensively that no fit could be determined.
For the ALT series, the PM2.5 size distribution is seen to approximate a lognormal distribution (Figure A2). No values of zero were reported.

Fig. A2. Histogram of PM2.5 concentrations reported by the ALT series. The shape is approximated by a lognormal fit.

11
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

Comparison of ALT and CF1 findings for the difference between PM2.5 and PM1

The PurpleAir estimates of PM1 depend on the smallest two size categories, and the estimates of PM2.5 depend on the smallest 3 size categories. This
means the difference between PM1 and PM2.5 depends only on a single size category (1–2.5 μm). This difference should therefore be a linear function of
the number of particles in that single size category. In our own estimate of the difference (PM2.5- PM1 ALT series), it is in fact a perfect (R2 = 1) multiple
of the number of particles in this size category. Yet this same difference in the CF1 series is far from a simple constant factor times the number of
particles, explaining only 8% of the variation (Fig. S3).
Many of the problems in using the CF1 or ATM data series provided by Plantower stem from the arbitrary decision to substitute zero for mea­
surements falling below some value. For example, this obviously leads to distortions of the mass distributions if a substantial number of measurements
fall below this limit. But it also leads to worse precision and higher coefficients of variation, since the clustering of otherwise positive values at zero
serve to widen the standard deviation. Statisticians generally oppose the practice of substituting a single value, such as 0, the LOD/2, or the LOD itself
because it leads to exactly these problems. A review of many of these studies is found in Helsel (2010). His first conclusion reads: “In general, do not
use substitution. Journals should consider it a flawed method compared to the others that are available and reject papers that use it …” (emphasis
added).
However, this is not the only problem with the CF1 and ATM data series. The example discussed above of the difference between PM2.5 and PM1,
which should rest entirely on the mass M of particles in the 1–2.5 μm size category, achieves only an 8% R2 when plotted against M. The ALT method
achieves an R2 of precisely 1 in this case, as expected from theory. Apparently, the hidden algorithms of the Plantower approach assign values to
measurements that in some way depart from using the mass calculated from the numbers of particles in the size categories.
The ALT method works well when calculating PM1 or PM2.5 from the two or three smallest size categories. However, when adding in the two largest
size categories (2.5–5 μm and 5–10 μm) to estimate PM10, it finds a very large LOD in both of our selected data series, such that <10% of the data
exceeds the LOD. Other studies have also found problems with the Plantower PM10 estimates. The SPEC study (AQ-SPEC, 2016) found that the PM10
estimates showed a poorer relationship to the reference monitor than the PM1 or PM2.5 estimates. Therefore, we do not attempt to apply the ALT
method to estimate PM10.

Comparison of number estimates with research-grade monitors

Plantower number estimates in the three size categories from 0.3 to 2.5 μm were compared to those produced by a collocated reference monitor,
the TSI Optical Sizer Model 3330 (Wallace, unpublished data). For the two smallest size categories, the Plantower number estimates were smaller by
factors on the order of 8–10. For the sum of numbers in all three categories, the Plantower estimate was also smaller, by about a factor of 4, than the
Model 3330 estimate. This observation may partially explain the factor of 4.5 by which the CF1 data series exceeds the ALT data series (using a density
of 1).
A calibration factor for indoor PM2.5 estimates has been recently estimated (Wallace et al., 2020). This was a study of an indoor aerosol using
PurpleAir monitors collocated with research-grade SidePak monitors which themselves had been calibrated using gravimetric techniques (Zhao et al.,
2020). This study arrived at a calibration factor of 3 for the indoor PM2.5 estimates, in good agreement with our present estimate of 2.9–3.1 (SE < 0.2)
for the outdoor PM2.5 data calibrated using FEM monitors.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.atmosenv.2021.118432.

Conflicts of interest/competing interests

No conflict of interest exists.

Availability of data and material

All data available from author LW on request.

Code availability

Some R scripts were written for this study. Contact author JB. Software applications include Excel and Statistica.

References of multiple low-cost particulate matter sensors in an outdoor urban environment.

Sci. Rep. 9, 7497. https://doi.org/10.1038/s41598-019-43716-3.
Chen, L.J., Ho, Y.-H., Lee, H.-C., Wu, H.-C., Liu, H.-M., Hsieh, H.-H., Huang, Y.-T.,
AQ-SPEC, 2016. http://www.aqmd.gov/docs/default-source/aq-spec/field-evalua
Lung, S.-C.C., 2017. An open framework for participatory PM2.5 monitoring in smart
tions/purpleair—field-evaluation.pdf. (Accessed 19 December 2020).
cities. IEEE Access 5, 14441–14454. https://doi.org/10.1109/
Becnel, T., Tingey, K., Whitaker, J., Sayahi, T., Le, K., Goffin, P., Butterfield, A., Kelly, K.,
ACCESS.2017.2723919. Accessed. (Accessed 18 December 2020).
Gaillardon, P.E., 2019. A distributed low-cost pollution monitoring platform. IEEE
Delp, W.W., Singer, B.C., 2020. Wildfire smoke adjustment factors for low-cost and
Internet of Things Journal 6 (6), 10738–10748.
professional PM2.5 monitors with optical sensors. Sensors 20, 3685. https://doi.org/
Bi, J., Wildani, A., Chang, A.H.H., H, H., Liu, Y., 2020. Incorporating low-cost sensor
10.3390/s20133683.
measurements into high-resolution PM2.5 modeling at a large spatial scale. Environ.
Francis, A.S., Chee, F.P., Chang, J.H.W., Sentian, J., Dayou, J., Payus, C.M., 2019.
Sci. Technol. (Article ASAP).
Parametric model for estimation of mass concentration based on particle count
Bi, J., Wallace, L.A., Sarnat, J.A., Liu, Y., 2021. Characterizing outdoor infiltration and
distribution for ambient air monitoring. J. Phys. Conf. 1358 https://doi.org/
indoor contribution of PM2.5 with citizen-based low-cost monitoring data. Environ.
10.1088/1742-6596/1358/1/012042, 012042.
Pollut. 276, 116763.
Gupta, P., Doraiswamy, P., Levy, R., Pikelnaya, O., Maibach, J., Feenstra, B., et al., 2018.
Bulot, F.M.J., Johnston, S.J., Basford, P.J., Easton, N.H.C., Apetroaie-Cristea, M.,
Impact of California fires on local and regional air quality: the role of a low-cost
Foster, G.L., Morris, A.K.R., Cox, S.J., Loxham, M., 2019. Long-term field comparison

12
L. Wallace et al. Atmospheric Environment 256 (2021) 118432

sensor network and satellite observations. GeoHealth 2, 172–181. https://doi.org/ Tryner, J., Quinn, C., Windom, B.C., Volckens, J., 2019. Design and evaluation of a
10.1029/2018GH000136. Accessed. (Accessed 18 December 2020). portable PM2.5 monitor featuring a low-cost sensor in line with an active filter
He, M., Kuerbanjiang, N., Dhaniyala, S., 2020. Performance characteristics of the low- sampler. Environmental Science Processes and Impacts 21, 1403–1415.
cost Plantower PMS optical sensor. Aerosol. Sci. Technol. 54 (2), 232–241. https:// US EPA, 2017. https://www.epa.gov/air-sensor-toolbox/how-use-air-sensors-air-sensor-
doi.org/10.1080/02786826.2019.1696015. guidebook.
Helsel, D., 2010. Much ado about next to nothing: incorporating nondetects in science. Walker, 2018. Accessed. http://conference2018.resnet.us/data/energymeetings/present
Ann. Occup. Hyg. 54 (3), 257–262. ations/RESNET2018_LBL_LowCostMonitors_walker.pdf. (Accessed 19 December
Kaduwela, A.P., Kaduwela, A.P., Jrade, E., Brusseau, M., Morris, S., Morris, J., Risk, V., 2020).
2019. Development of a low-cost air sensor package and indoor air quality Wallace, L.A., Wheeler, A., Kearney, J., Van Ryswyk, K., You, H., Kulka, R.,
monitoring in a California middle school: detection of a distant wildfire. J. Air Waste Rasmussen, P., Brook, J., Xu, X., 2010. Validation of continuous particle monitors for
Manag. Assoc. 69 (9), 1015–1022. https://doi.org/10.1080/ personal, indoor, and outdoor exposures. J. Expo. Sci. Environ. Epidemiol. 21,
10962247.2019.1629362. 49–64.
Kelly, K.E., Whitaker, J., Petty, A., Widmer, C., Dybwad, A., Sleeth, D., Martin, R., Wallace, L.A., Ott, W.R., Zhao, T., Cheng, K.-C., Hildemann, L., 2020. Secondhand
Butterfield, A., 2017. Ambient and laboratory evaluation of a low-cost particulate exposure from vaping marijuana: concentrations, emissions, and exposures
matter sensor. Environ. Pollut. 221, 491–500. determined using both research-grade and low-cost monitors. Atmos. Environ. X.
Levy Zamora, M., Xiong, F., Gentner, D., Kerkez, B., Kohrman-Glaser, J., Koehler, K., https://doi.org/10.1016/j.aeaoa.2020.100093.
2019. Field and laboratory evaluations of the low-cost plantower particulate matter Wang, K., Chen, F.E., Au, W., Zhao, Z., Xia, Z.-L., 2019. Evaluating the feasibility of a
sensor. Environ. Sci. Technol. 53 (2), 838–849. personal particle exposure monitor in outdoor and indoor microenvironments in
Magi, B.I., Cupini, C., Francis, J., Green, M., Hauser, C., 2019. Evaluation of PM2.5 Shanghai, China. Int. J. Environ. Health Res. 29, 209–220. https://doi.org/10.1080/
measured in an urban setting using a low-cost optical particle counter and a Federal 09603123.2018.1533531.
Equivalent Method Beta Attenuation Monitor. Aerosol. Sci. Technol. 54, 147–159. Wang, Z., Delp, W.W., Singer, B.C., 2020. Performance of Low-Cost Indoor Air Quality
https://doi.org/10.1080/02786826.2019.1619915. Monitors for PM2.5 and PM10 from Residential Sources. Building and Environment.
Malings, C., anzer, R., Hauryliuk, A., Saha, P.K., Robinson, A.L., Presto, A.A., https://doi.org/10.1016/j.buildenv.2020.106654. Accessed. (Accessed 19
Subramanian, R., 2019. Fine particle mass monitoring with low-cost sensors: December 2020).
corrections and long-term performance evaluation. Aerosol. Sci. Technol. 54, Williams, R., Vasu, K., Snyder, E., Kaufman, A., Dye, T., Rutter, A., Russell, A.,
160–174. https://doi.org/10.1080/02786826.2019.1623863. Hafner, H., 2014. Air Sensor Guidebook. U.S. Environmental Protection Agency,
Masic, A., Bibic, D., Pikula, B., 2019. On the Applicability of Low-Cost Sensors for Washington, DC. EPA/600/R-14/159 (NTIS PB2015-100610).
Measurements of Aerosol concentrations, 0452-0456. In: Katalinic, B. (Ed.), Zhao, T., Cheng, K.-C., Ott, W.R., Wallace, L.A., Hildemann, L.M., 2020. Characteristics
Proceedings of the 30th DAAAM International Symposium. Published by DAAAM of Secondhand Cannabis Smoke from Common Smoking Methods: Calibration
International, ISBN 978-3-902734-22-8. https://doi.org/10.2507/30th.daaam. Factor, Emission Rate, and Particle Removal Rate. Atmospheric Environment.
proceedings.060, 1726-9679, Vienna, Austria. https://doi.org/10.1016/j.atmosenv.2020.117731.
Ott, W.R., Zhao, T., Cheng, K-C, Wallace, L.A., Hildemann, L.M., 2021. Measuring indoor Zheng, T., Bergin, M.H., Johnson, K.K., Tripathi, S.N., Shirodkar, S., Landis, M.S.,
fine particle concentrations, emission rates, and decay rates from cannabis use in a Sutaria, R., Carlson, D.E., 2018. Field evaluation of low-cost particulate matter
residence. Atmos. Environ. 10, 100106. https://doi.org/10.1016/j.aeaoa.2021.100 sensors in high-and low-concentration environments. Atmospheric Measurement
106. Techniques 11 (8), 4823–4846. https://doi.org/10.5194/amt-11-4823-2018.
Plantower, 2016. Accessed. https://www.aqmd.gov/docs/default-source/aq-spec/res Zusman, M., Schumacher, C.S., Gassett, A.J., Spalt, E.W., Austin, E., Larson, T.V.,
ources-page/plantower-pms5003-manual_v2-3.pdf. (Accessed 17 January 2020). Arvlin, G.C., Seto, E., Kaufman, J.D., Sheppard, L., 2020. Calibration of low-cost
Sayahi, T., Butterfield, A., Kelly, K.E., 2019. Long-term field evaluation of the Plantower particulate matter sensors: model development for a multi-city epidemiological
PMS low-cost particulate matter sensors. Environ. Pollut. 245, 932–940. study. Environ. Int. 134, 105329.
Singer, B.C., Delp, W.W., 2018. Response of consumer and research grade indoor air
quality monitors to residential sources of fine particles. Indoor Air 28, 624–639.
https://doi.org/10.1111/ina.12463. Accessed. (Accessed 19 December 2020).

13