Environmental Science & Technology 
pubs.acs.org/est 
EL 
25% Imp Imp > 1.5 - Ship Ref 
_ Kruskal-Wallis test I > IMPropze% >25 - Ship Ref 
>» (P<0.05) Ref+3 - SD z——— AND — 
N Imp > 1.5 - Low Ref 
= ANDI= = = OR = = -JAND ——— 
S Imp Impropzs% > 2.5 + Low Ref 
5 —_ 33% Imp | -  ———— AND) = = 
Ref > Imp > 1.5 + High Ref 
Ref+2 - SD Impropzs% > 2.5 + High Ref 
N 
> 
X 
U 
= 
S 
pe 
= 
U / 
” 
> 
SA 
& 
X 
wr 
Bm 
dur 
N 
25% Imp Imp|NbyYropzs% 
Kruskal-Wallis test > >23 -Ship Ref 
(P<0.05) Ref+3 - SD fo AND sr un 
— u m (JR mn Imp|Nbyrop25% 
Imp | Nby >3 Low Ref 
—_ 33% Imp ——— AN) — 
Ref — > Imp|Nbyropz5% 
Ref+2 + SD >23 -High Ref 
V Improp3> 3 + Ship Ref 
— A NIE) —U.—— 
V Improp3> 3 + Low Ref 
—— AND — = 
V Impsop3> 3 - High Ref 
Top-3 highest peak Intensities N 
AND 
measured at Imp 
! 
Figure 2. Overview of the score function used to select relevant features based on three selection strategies. The third criterion of strategies 1 and 2 and 
:he second criterion of strategy 3 use the log mean, whereas in the other cases, the average is used. 
with I, the measured intensity of that feature and f7oc the measured 
:otal organic carbon (TOC) value at that station.” 
2.5. Statistical Analysis 
For both the GC-MS and LC-HRMS data, features that showed higher 
'ntensities in the impact area compared to the reference areas were 
'dentified. Because of the heterogeneous pattern of most features, a 
score function using three different strategies was developed (Figure 2) 
*o quickly screen through thousands of features and select the ones that 
could potentially be emitted by the OWFs. All analyzed data were 
expressed as peak intensity for both GC and LC-HRMS. 
In the first strategy, a feature is selected if three criteria are met. First, 
:he median intensity at impact stations should be significantly higher 
*han at the reference stations, using a Kruskal— Wallis test with a p-value 
<0.05. Second, the intensity of a feature at the impact and reference 
10cations was compared. For each feature, 25% of the impact sites 
should have an intensity that is higher than three times the standard 
deviation plus the mean intensity of the reference sites, or 33% of the 
mpact sites should have an intensity that is higher than two times the 
standard deviation plus the mean intensity of the reference sites. Third, 
:he log mean intensity at impact sites should be 1.5 times higher than 
‘he mean log intensity at each reference area (ship ref, high ref, and low 
‚ef), and the log mean of the 25% ofthe highest intensities in the impact 
ocations should be 2.5 times higher than the mean log intensity at each 
‚eference area. 
The first strategy allows for capturing slightly higher feature 
'ntensities across the entire impact area or much higher feature 
ntensities in only a few stations in the impact area. In other words, it 
ıllows for the identification of continuous sources, as well as more local 
a1otspot contamination at only a few impact locations. The Kruskal— 
Wallis test is often used in NTS to identify differences between groups 
when the normality of the data cannot be assured.?”° However, after 
nanually checking the selected compounds, the statistical test alone 
>roved to be insufficient, as it selected compounds with only slightly 
different peak intensities. Therefore, the median intensity and standard 
deviation of the reference areas were also used to further narrow the 
compound selection. This procedure is often used to differentiate NTS 
compounds from the blanks.”*“ However, because a heterogeneous 
vattern was expected, only a percentage of the samples had to fulfill the 
criterion. A similar approach was used by Tisler et al.,”” where a more 
flexible filter was used to avoid compounds that were not detected in 
each sample being wrongly classified as transformation products. The 
log,o transformation in the third criterion was used to reduce the effects 
of extreme peak intensities and is often used for the same reason in 
other NTS studies before data processing. “7 The parameters of the 
score function (% of samples and k times the SD) were chosen by first 
selecting a subset of features and manually classifying them. Then, the 
parameters were optimized to mimic a similar selection based on fixed 
rules that can be applied to the whole data set. 
While being similar to the first strategy, the second strategy 
zompared not only the impact stations but also the nearby stations to 
the reference area. This also allows to account for the drift of certain 
leachates over a larger area due to hydrodynamic processes. Again, three 
zriteria were decisive: first, the median intensity at impact and nearby 
stations should be significantly higher than at the reference stations 
(Kruskal—Wallis; p < 0.05). Second, the intensity of a feature at the 
impact and reference locations was compared in the same way as in the 
first strategy. This ensures that the intensities are still elevated in the 
impact locations and not only at nearby stations. Third, the log mean of 
the 25% of the highest intensities in the impact and nearby locations 
should be three times higher than the mean log intensity at each 
reference area. 
The third strategy considers hotspots. A feature was labeled as a 
hotspot if the three highest intensities were measured in the impact area 
and each of these intensities was at least three times higher than the log 
mean intensity in each reference area. 
Features with a higher intensity in either the high or the ship ref area 
were also considered for this study. For the ship ref, a feature was 
selected if the mean intensity in the ship ref is at least twice as high 
zompared to the mean intensity of the three highest intensities 
measured in the low and high ref area and the mean intensity in the ship 
ref area is at least three times higher than the standard deviations plus 
the mean intensity measured in both the low and high ref area. A feature 
can also never be selected as a high ref or ship ref if it was already 
selected in the first three strategies, as this would still indicate that the 
intensity of the impact is still much higher than any reference area. 
Similarly, features were selected for the high ref area by comparing it 
against the low and ship ref areas. 
2.6. Tentative Identification 
Features that were selected to be relevant by the score function were 
identified at a level 2.°® For GC-MS data, identification was done using 
https://doi.org/10.1021/acs.est.5c17939 
Environ. Sci, Technol, XXXX, XXX, XXX XXX