https://orcid.org/0009-0003-1657-8050
Figures
Abstract
Analyzing surface-water PFAS datasets using EPA Method 1633 often yields many results below reporting limits and fewer qualified low-level detections, making post-processing choices important for site comparisons. In this pilot study, we used a single synoptic survey of 10 sites in Trinity River headwater tributaries near Fort Worth, Texas, to demonstrate a transparent workflow for summarizing such data. Samples were collected on 21 December 2025 and analyzed for 40 target PFAS using isotope-dilution LC-MS/MS at a commercial laboratory accredited to ISO/IEC 17025. Of the 400 site-analyte results, 82 (20.5%) were detected, and 318 (79.5%) were left-censored. We distinguished detect-only ΣPFAS40, defined as the sum of quantified detections, with non-detects set to zero, from median-imputed ΣPFAS40, a model-based total generated using 5,000 censoring-aware imputations per site. Median imputed ΣPFAS40 ranged from 55.5 to 685.2 ng/L. One high-total site (TW08; 95% conditional imputation-uncertainty interval: 673.2–696.8 ng/L) clearly exceeded the other nine sites. Imputation increased ΣPFAS40 by 32–38 ng/L at lower-total sites relative to detect-only sums, demonstrating that non-detect handling materially affects low-end site comparison. Site rankings were stable across distributional assumptions, substitution methods, and qualifier reclassification. River-network distance, catchment land-cover metrics, and multi-year national pollutant discharge records were retained only as screening-level context. Overall, this pilot case study shows that the main value of the dataset lies in a reproducible, qualifier-transparent, and censoring-aware workflow for PFAS reconnaissance, rather than in any definitive assessment of temporal conditions or causal source attribution.
Citation: Asare PT, Kharel G, Nice MS, Lavy BL, Birmingham M, Harvey OR (2026) Handling left-censored PFAS data in surface-water reconnaissance: A reproducible workflow applied to ten sites in the Trinity River headwaters, Texas. PLOS Water 5(6): e0000554. https://doi.org/10.1371/journal.pwat.0000554
Editor: Daniel Reddythota, Faculty of Water Supply & Environmental Engineering, ArbaMinch Water Technology Institute (AWTI), ETHIOPIA
Received: April 10, 2026; Accepted: May 20, 2026; Published: June 26, 2026
Copyright: © 2026 Asare et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All data necessary to replicate the findings of this study are deposited at Zenodo (https://doi.org/10.5281/zenodo.19500593) under a CC-BY 4.0 license. The deposited data package contains: the complete site-by-analyte concentration matrix with all measured values, reporting limits, censoring indicators, and qualifier flags; the Python imputation, sensitivity-analysis, table-generation, and figure-generation code; compiled NPDES/DMR summaries; catchment land-use and river-network distance data; and a machine-readable QA/QC appendix reproducing the method blank results, LCS/LCSD/LLCS recoveries, isotope-dilution recovery ranges, and qualifier decision rules extracted from the laboratory analytical report (Eurofins Job 810-176371-1). Supporting Information tables accompanying this article reproduce the 40-analyte detection results (S1 Table), site-by-analyte concentrations (S2 Table), catchment and NPDES indicators (S3 Table), per-site source-category counts (S4 Table), and exploratory Spearman screening correlations with Benjamini-Hochberg FDR-adjusted q-values (S5 Table). Table 1 in the main text presents the five-way sensitivity analysis.
Funding: This research was supported via Seed Funding from the TCU Natural Resource Research Initiative and TCU Invests in Scholarship (TCU-IS) grant (# 66071) to GK. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Per- and polyfluoroalkyl substances (PFAS) are a large family of synthetic chemicals that persist in the environment and have become ubiquitous contaminants in surface waters worldwide [1–3], including rivers receiving treated wastewater effluents [4] and urban stormwater runoff [5]. Reconnaissance datasets are often difficult to interpret because a large fraction of reported values fall below reporting limits. This is especially problematic for summed indicators such as ΣPFAS40, where site ranking can depend as much on the treatment of non-detects as on the observed detections themselves [6,7]. In such settings, deterministic substitution rules such as 0, RL/2, or RL can distort both absolute totals and between-site contrasts [8]. Established censored-data methods [8–10] therefore provide a more defensible basis for summarizing PFAS mixtures in screening studies.
A second challenge is the transparent handling of analytical qualifiers. Modern PFAS methods, such as EPA Method 1633, use isotope dilution, multiple transitions, and strict QC criteria [11], yet low-level environmental data can still include ion-ratio flags, elevated reporting limits, and other qualified results. In screening applications, these qualifiers must be retained explicitly and evaluated through sensitivity analysis so that downstream interpretation remains reproducible.
A third challenge is the use of spatial context. In river systems, along-network connectivity is often more meaningful than straight-line proximity [12,13], and public permitting datasets can help describe where regulated discharges occur [14]. However, in sparse-reconnaissance designs, river-network variables and NPDES records are best treated as contextual screening information rather than as source-attribution evidence, particularly when PFAS-specific monitoring is absent from public DMR and permit-limit records.
In this study, we use a single 10-site synoptic surface-water dataset from the Trinity River headwaters near Fort Worth, Texas, as a pilot application to demonstrate an integrated workflow for: (i) censoring-aware estimation of ΣPFAS40, (ii) transparent retention and sensitivity testing of laboratory qualifiers, (iii) screening with river-network and catchment covariates, and (iv) contextualization using multi-year NPDES records without causal claims. The purpose of the paper is methodological and applied: to show how a sparse EPA Method 1633 dataset can be summarized and interpreted transparently enough to support follow-up prioritization, while making the study’s inferential limits explicit.
Materials and methods
Study design and scope
Ten surface-water sites were sampled across major tributaries and sub-catchments in the Trinity River headwaters in and around Fort Worth, Texas (Fig 1). Site selection prioritized spatial coverage, safe access for synoptic grab sampling, and representation of gradients in drainage area, impervious cover, and developed land. Because only one sampling event was conducted, the study was designed and interpreted as a pilot spatial reconnaissance and workflow application, not as a temporal characterization of typical conditions or hydrologic variability.
(A) Locator maps showing Texas within the contiguous United States and the Dallas–Fort Worth study area. (B) Detailed study area showing the 10 sampling sites along five waterbodies in the Trinity River headwaters. Circle color indicates waterbody, and circle size is scaled to the site-level median ΣPFAS40 concentration, where ΣPFAS40 is the summed concentration of 40 target PFAS estimated using censoring-aware multiple imputation with 5,000 replicates. The gold square marks Naval Air Station Joint Reserve Base (NAS JRB) Fort Worth, approximately 1.4 km upstream of TW08 by river-network distance. Hydrography data are from USGS NHDPlus V2, and boundary data are from the U.S. Census Bureau TIGER/Line; all basemap data are public domain, and no copyrighted basemap was used.
Sampling was conducted from public access points or public rights-of-way, with no ground disturbance, equipment installation, entry onto private property, or collection of protected biological material. Therefore, no field-access or collection permits were required for the surface-water grab sampling reported here.
Field sampling and laboratory analysis
Synoptic grab samples were collected on 21 December 2025 and submitted to a commercial laboratory for analysis of 40 target PFAS (S1 Table) by isotope-dilution LC-MS/MS using EPA Method 1633 [11]. The laboratory received the samples on ice on 23 December 2025 at 3.4 °C, prepared them on 31 December 2025, analyzed them on 1 January 2026, and issued the report on 7 January 2026. The field campaign, therefore, took place in late 2025, while the analytical report was issued in early 2026.
QA/QC package and qualifier decision rules
The submitted laboratory analytical package documented a method blank, a laboratory control sample (LCS), a laboratory control sample duplicate (LCSD), a low-level laboratory control sample (LLCS), sample-specific isotope-dilution recoveries, and explicit qualifier definitions. The laboratory operated under A2LA ISO/IEC 17025 accreditation and Texas National Environmental Laboratory Accreditation Program (NELAP) accreditation. The laboratory report stated that analytical results met program requirements unless otherwise noted and that matrix-specific batch QC, such as MS/MSD, may not be reported when site-specific QC samples were not submitted. Field duplicates and site-specific MS/MSD results were not included in the submitted analytical report; this is acknowledged as a study limitation.
Two qualifiers affected the interpretation of the present dataset. The “I” qualifier denoted an ion ratio outside established control limits, and the “G” qualifier denoted an elevated quantitation/reporting limit caused by noise or matrix interference. Three PFOS results (TW12, TW20, and TW26) carried the “I” qualifier. These were retained in the primary analysis because the laboratory narrative reported positive identification, but they were reclassified as censored in a sensitivity analysis. One 8:2 FTS result at TW08 was reported as non-detect with an elevated RL (14 ng/L) and was treated as left-censored at that elevated threshold.
Data processing and ΣPFAS40 estimation
The data were harmonized into long format (S2 Table) with one record per site-analyte pair, including measured value when detected, site-specific RL, censoring status, and qualifier status. ΣPFAS40 was defined as the sum of the 40 targeted PFAS listed in EPA Method 1633 [11]. Detect-only ΣPFAS40 denotes the sum of quantified detections with non-detects set to zero, whereas median imputed ΣPFAS40 denotes the model-based site total after censoring-aware multiple imputation. ΣPFAS40 is used here as a simple screening surrogate for mixture magnitude; toxicity-weighted sums or health-risk indices were not calculated because the study was designed as a reconnaissance workflow application rather than a risk assessment. Subclass sums were also computed for perfluoroalkyl carboxylates (ΣPFCA), perfluoroalkyl sulfonates (ΣPFSA), and fluorotelomer sulfonates (ΣFTS) following established class nomenclature [1].
To avoid deterministic substitution bias, censored observations were handled using a distribution-based multiple-imputation workflow [8,9,15]. For analytes with three or more detections, lognormal parameters were estimated with a left-censored likelihood in which detected observations contributed log f(y_i) and censored observations contributed log F(RL_i), where RL_i is the observation-specific reporting limit. Censored values were then drawn from the fitted distribution conditional on being below the corresponding site-specific RL. For analytes with fewer than 3 detections, values were drawn from a Uniform(0, RL) distribution. This bounded weak-information distribution was chosen because fewer than three detections cannot support stable analyte-specific distributional fitting; it preserves the known interval constraint without imposing an unstable tail model. We generated 5,000 complete data matrices and summarized site-level ΣPFAS40 by the median and the 2.5th and 97.5th percentiles of the imputed distribution. All reported MI summaries were generated with a fixed random seed to ensure reproducibility; re-running with a different seed shifts site-level medians by less than 2 ng/L. These intervals are conditional imputation-uncertainty intervals reflecting uncertainty about censored values given the fitted distributions and reporting limits; they are not temporal confidence intervals and do not fully propagate uncertainty in the fitted distributional parameters. Because the fewer-than-three-detections rule is a pragmatic screening choice rather than a strongly data-rich model, low-total sites are interpreted primarily in terms of rank robustness and uncertainty propagation rather than precise absolute mixture burden. To evaluate sensitivity to distributional assumptions, qualifier decisions, and substitution methods, site-level ΣPFAS40 ranks were compared across five approaches (Table 1): (1) the primary MI workflow described above, (2) a Uniform-all MI variant in which all censored values were drawn from Uniform(0, RL) regardless of detection count, (3) RL/2 substitution, (4) a qualifier-sensitivity MI in which three I-qualified PFOS detections (TW12, TW20, TW26) were reclassified as censored before running the primary MI workflow, and (5) RL/4 substitution.
Catchment, river-network, and NPDES context
Catchment drainage area, developed land, and impervious cover were compiled from the National Land Cover Database [16] as screening covariates (S3 Table). Upstream NPDES facilities were assembled from ICIS-NPDES [14] and linked to sampling sites using hydrologically routed river-network distances rather than Euclidean distance alone [12,13]. Site-level contextual indicators included active upstream permit count, permit density (active permits per 100 km²), nearest along-network distance (km), and distance-weighted permit-flow proxies.
Five fiscal years of Texas DMR and permit-limit data (FY2021–FY2025) were summarized to characterize the upstream regulatory context. These records were used to describe permitted-feature density and monitoring history; no PFAS-related parameters were present in the compiled DMR or permit-limit records for these upstream permitted features. The DMR/permit-limit records were not used to attribute PFAS concentrations to individual permits or facilities. Each upstream permitted facility was assigned to one of six source categories (Stormwater/MS4, WWTP, Industrial, Airport, Landfill, Other/Unknown) based on SIC_Codes in the ICIS-NPDES record, which were mapped from 4-digit SIC entries to six screening groups; per-site category counts are reported in S4 Table.
Statistical treatment
PFAS occurrence was summarized using detection frequencies and site-level median imputed ΣPFAS40 distributions from the multiple-imputation procedure described in Section 2.4. Site-level ΣPFAS40 was reported as the median and 2.5th-97.5th percentiles of the imputed distribution. Associations between site-level median imputed ΣPFAS40 and contextual covariates were evaluated using Spearman's rank correlation. Two-sided raw p-values were calculated using scipy.stats.spearmanr, and Benjamini-Hochberg false-discovery-rate-adjusted q-values were calculated across the family of exploratory covariate-screening tests reported in S5 Table. Ninety-five percent intervals for Spearman ρ were estimated by nonparametric bootstrap resampling of site pairs using 10,000 resamples. Because the study included only 10 sites from a single synoptic event, and several ancillary DMR/flow covariates were available for only 4 sites, these statistics are reported only as descriptive screening metrics. No α-based significance threshold, formal hypothesis testing, causal inference, or source attribution was used. No formal sample size or power calculation was performed because the study was designed as a workflow-application reconnaissance rather than a hypothesis test. The complete imputation, bootstrap, and correlation code is deposited at Zenodo (https://doi.org/10.5281/zenodo.19500593) under a CC-BY 4.0 license, enabling independent reproduction of all reported statistics.
Results
Dataset structure and detection characteristics
Across the 10-site-by-40-analyte matrix, 82 of 400 results (20.5%) were detected, and 318 (79.5%) were reported as <RL. Twelve PFAS were detected at least once (S1 Table). PFBA, PFPeA, PFHxA, PFBS, and PFOS were detected at all 10 sites (100%), while PFHxS and PFOA were detected at 9 of 10 sites (90%). The remaining analytes occurred sporadically (detection frequencies of 20–40%). This distribution illustrates the type of dataset structure for which summed PFAS metrics become sensitive to censor-handling decisions: a small core of consistently detected compounds embedded within a much larger censored matrix.
Effect of censoring treatment on site-level ΣPFAS40
In this study, the principal result of the workflow is not the absolute value of any one site's total, but to demonstrate how censoring treatment materially changes low-end site comparisons. Median ΣPFAS40 ranged from 55.5 ng/L (TW25, Marine Creek) to 685.2 ng/L (TW08, West Fork Trinity River; 95% imputation-uncertainty interval: 673.2–696.8 ng/L) across the 10 sites (Fig 2). TW08 was the largest site total by a wide margin, followed by TW06 (269.7 ng/L) and TW04 (161.7 ng/L), while the remaining seven sites ranged from 55.5 to 93.6 ng/L in this single synoptic campaign.
Horizontal bar plot of median ΣPFAS40 (ng/L) for each sampling site, ordered from highest (TW08) to lowest (TW25). Error bars represent the 2.5th and 97.5th percentiles of the multiple-imputation distribution (5,000 replicates). Bar color indicates waterbody assignment. The dashed reference line marks the network median (75 ng/L).
At the lower-total sites, the imputed median exceeded the detect-only sum by 32–38 ng/L. The fraction of ΣPFAS40 supported by quantified detections was highly uneven across the network: detect-only results accounted for approximately 94% of the median total at TW08 and 89% at TW06, but only 33–47% at the lowest-total sites. The workflow, therefore, has modest consequences for obvious hotspots but large consequences for the ordering and interpretation of low-concentration sites. We want to emphasize this differential sensitivity to non-detect handling as the study's central applied result.
Site rankings were compared across all five censor-handling approaches in Table 1. The top three sites (TW08, TW06, TW04) and the mid-range cluster (TW11, TW20, TW09) maintained the same ranks across all five approaches. Approaches 1, 2, 3, and 5 (primary MI, Uniform-all MI, RL/2, and RL/4) produced perfectly identical full rankings (Spearman ρ = 1.00). The only rank perturbation occurred under Approach 4 (I-qualified PFOS censored), which swapped TW26 and TW07 at positions 7 and 8, separated by less than 4 ng/L. The hotspot decision–TW08 as the dominant follow-up priority–was insensitive to all five approaches.
PFAS class composition
PFSA dominated the mixture at TW08, TW06, and TW04, while short-chain PFCA contributed a larger relative share at sites with lower total (Fig 3). The heatmap of individual analyte concentrations (S1 Fig) confirmed that PFOS and PFHxS drove between-site differentiation in the PFSA class, and that PFHxA and PFBA were the most uniformly distributed PFCA across the network.
Stacked bar chart showing the contribution of four PFAS classes to site-level ΣPFAS40: perfluoroalkyl carboxylates (PFCA), perfluoroalkyl sulfonates (PFSA), fluorotelomer sulfonates (FTS), and all other detected PFAS (Other). Bar heights correspond to detect-only ΣPFAS40. Sites are ordered by decreasing ΣPFAS40 (TW08 on the left, TW25 on the right). Class-level sums are computed from detected concentrations only.
Qualifier sensitivity
Only four site–analyte results in the full matrix carried qualifiers. Three PFOS detections (TW12, TW20, TW26) were flagged with “I” (ion ratio outside established control limits), and one 8:2 FTS non-detect at TW08 was flagged with “G” (elevated RL of 14 ng/L). In the qualifier-sensitivity MI (Approach 4 in Table 1), reclassifying the “I”-qualified PFOS values as censored reduced site-level ΣPFAS40 at the three affected sites by 2.7 ng/L (TW12), 7.5 ng/L (TW20), and 4.0 ng/L (TW26). This caused TW26 and TW07 to swap rank positions (7 and 8, separated by 3.7 ng/L), but it did not change the top-three ranking or the identification of TW08 as the dominant follow-up priority. Qualifier handling, therefore, warrants transparent documentation and sensitivity analysis, but in this dataset, it has only a minor consequence: a reordering deep in the low-total cluster.
Application to spatial screening and follow-up prioritization
TW08 stood out not only in total burden but also in mixture composition, with elevated PFOS (180 ng/L), PFHxS (120 ng/L), and 6:2 FTS (81 ng/L) relative to the rest of the network. The upstream–downstream contrast between TW09 and TW08 on the West Fork Trinity River was pronounced: ΣPFAS40 increased approximately 10-fold, PFOS approximately 47-fold, and PFHxS approximately 23-fold across this transect in the single synoptic campaign. These concentration gradients are sufficient to identify TW08 as a priority reach for targeted resampling, upstream bracketing, and additional QA/QC in future campaigns, but they do not by themselves identify a specific source.
Contextual covariates were informative only at the screening level (Fig 4; S5 Table). ΣPFAS40 showed a moderate positive descriptive association with catchment imperviousness (Spearman ρ = 0.48; raw p = 0.162; BH FDR q = 0.374; 95% bootstrap CI: -0.33 to 0.89) and no consistent association with nearest upstream active-permit river-network distance (Spearman ρ = 0.15; raw p = 0.676; BH FDR q = 0.855; 95% bootstrap CI: -0.58 to 0.81). Active NPDES permit density also showed no interpretable inferential association with ΣPFAS40 after FDR adjustment (Spearman ρ = -0.45; raw p = 0.187; BH FDR q = 0.374; 95% bootstrap CI: -0.99 to 0.38). These correlations are reported as exploratory screening metrics only; with n = 10 sites from a single synoptic event, they do not support inferential conclusions or source attribution. The upstream permit inventory was dominated by stormwater-related coverage at all sites (S2 Fig; S4 Table), and the compiled Texas FY2021-FY2025 DMR and permit-limit datasets contained no PFAS-related parameters [14]. For TW08 specifically, the upstream inventory included 18 active permits, all classified as Stormwater/MS4, with no active upstream WWTP, Industrial, Airport, or Landfill permits (S4 Table). This inventory provides useful contextual structure but does not constitute evidence of source apportionment.
(A) Site-level median imputed ΣPFAS40 versus catchment imperviousness (%). (B) Site-level median imputed ΣPFAS40 versus hydrologically routed river-network distance (km) to the nearest upstream active NPDES permitted feature of any type. Spearman rank correlation coefficients and 95% bootstrap intervals (10,000 resamples) are shown in each panel. Raw p-values and Benjamini-Hochberg FDR-adjusted q-values are reported in S5 Table. Both associations are reported as exploratory screening metrics and do not support inferential conclusions given the n = 10 sites and single-event sampling design. Site labels are provided for reference.
Discussion
Multiple imputation for censored environmental data is well established [8,9,15]. The contribution of this study is the end-to-end application of those methods to a 40-analyte PFAS reconnaissance dataset generated under EPA Method 1633 [11], combined with explicit qualifier tracking and river-network/NPDES contextualization. Here, we demonstrate how to turn a sparse, qualifier-rich laboratory package into transparent site-level screening outputs that can be reproduced and audited. The five-way sensitivity analysis (Table 1) reinforces this point: the distributional assumption for low-detection analytes (lognormal versus Uniform), the substitution fraction (RL/2 versus RL/4), and the qualifier classification all change absolute totals but preserve the rankings. The only perturbation, a swap of two adjacent low-total sites under the qualifier-sensitivity scenario, occurs where the margin between sites is less than 4 ng/L, well within the imputation-uncertainty interval.
Our result supports three defensible claims. First, TW08 was the highest-priority site in this synoptic campaign. Second, the low-end site comparison was materially affected by censor treatment. Third, qualifier transparency and sensitivity analysis improved interpretability without changing the main hotspot decision. The five-way comparison showed that the top three rankings were invariant across all tested distributional, substitution, and qualifier assumptions.
TW08 remains an important result in this study, but strictly as a follow-up priority identified by the workflow. The mixture at TW08 was enriched in PFOS, PFHxS, and 6:2 FTS, and its median ΣPFAS40 was well above all other sites. Those are robust descriptive results. The correct interpretation within this workflow paper is that the procedure successfully isolates a high-priority site whose rank is stable to reasonable levels of censoring and qualifier decisions, as confirmed by its first-place ranking under all five sensitivity approaches (Table 1). The enrichment in long-chain PFSAs and 6:2 FTS at TW08 is consistent with hypotheses involving aqueous film-forming foam (AFFF) or similar fluorosurfactant releases [17,18], but this remains a hypothesis to be tested through targeted upstream bracketing, not a finding of this single-event reconnaissance. Because the compiled upstream permit inventory for TW08 is dominated by stormwater-related coverage and the public DMR and permit-limit records lack PFAS parameters, the present data do not support mechanistic source attribution.
From a water-management perspective, the TW08 result has immediate practical relevance. TW08 is located on the West Fork Trinity River downstream of Lake Worth, and the river continues south through the Fort Worth–Dallas metropolitan corridor before joining the mainstem Trinity River, which serves as a major water resource for downstream communities. A reproducible screening workflow that flags priority reaches with stable hotspot rankings, even under different censoring and qualifier assumptions, gives local water managers an actionable starting point for allocating follow-up monitoring resources. The workflow output at TW08 (median ΣPFAS40 of 685.2 ng/L, dominated by PFOS and PFHxS) is a structured, uncertainty-quantified result that can be incorporated directly into screening-level risk communication for source-water protection planning without overstating what a single synoptic event can demonstrate.
Implications for future reconnaissance studies
The practical lesson is that sparse PFAS reconnaissance can still produce actionable screening information when the inferential target is chosen correctly. Future campaigns at this study site should add temporal replication (baseflow and storm-event sampling), concurrent discharge or load estimates, field duplicates, site-specific matrix spikes where feasible, and fine-scale upstream bracketing at high-priority reaches such as TW08. However, the present paper does not need those additional data to justify its current, narrower contribution. Its value lies in serving as a transparent template for handling left-censored and qualified PFAS data in pilot river-network surveys.
Relevance to Texas water-quality management
The workflow demonstrated here addresses a gap that extends well beyond the Fort Worth study area. Texas does not currently appear to have numeric PFAS water-column criteria in the Texas Surface Water Quality Standards, and the ECHO-derived FY2021–FY2025 DMR and permit-limit records compiled for upstream permitted features in this study contained no PFAS-related parameters. The TCEQ Surface Water Quality Monitoring program coordinates sampling at over 1,800 sites statewide [19], but PFAS-specific monitoring was not present in the ECHO-derived upstream DMR and permit-limit records analyzed in this study. At the federal level, EPA’s 2024 National Primary Drinking Water Regulation [20] requires public water systems to monitor for six PFAS, but no analogous surface-water regulation exists under the Clean Water Act. Although the draft 2026 Texas Integrated Report includes PFAS-in-tissue impairment entries for selected assessment units elsewhere in the state [21], the absence of PFAS-specific discharge monitoring parameters in the upstream DMR records compiled for this study area is consistent with the broader regulatory context in which PFAS monitoring is not yet routinely required in Texas NPDES permits.
In this regulatory context, a transparent and reproducible screening workflow has practical value for three audiences. First, for municipal source-water managers evaluating whether PFAS monitoring of tributary inputs is warranted, the workflow provides a structured method for converting a pilot sampling campaign into defensible site-level rankings that can justify or deprioritize additional investment. Second, for TCEQ staff developing the state’s approach to emerging contaminant assessment, the explicit handling of left-censoring and qualifier sensitivity offers a model for summarizing sparse EPA Method 1633 datasets without overstating what single-event reconnaissance supports. Third, for watershed planning groups and stakeholders in the Trinity River basin, the combination of NPDES inventory context with censoring-aware site totals provides a screening framework that can be repeated as additional sampling campaigns are conducted, enabling consistent comparison across time and space even when analytical datasets are heavily censored.
Study limitations
The study dataset represents a single-day reconnaissance snapshot (n = 10 sites) and therefore does not resolve temporal variability associated with hydrologic conditions, seasonality, episodic storm events, or operational changes at potential sources. Although censoring-aware multiple imputation was used to propagate uncertainty from non-detects, the imputation intervals condition on the fitted distributional parameters and do not fully propagate uncertainty in those parameter estimates. Estimates for rarely detected analytes remain sensitive to distributional assumptions. The imputation procedure treats each site independently; in a nested river network, downstream sites partially integrate upstream areas, thereby limiting the independence of site-level predictors and constraining multivariable or spatial modeling. Field duplicates and site-specific MS/MSD results were not included in the submitted analytical report; therefore, field-level precision and matrix-effect estimates are not available for this campaign. Catchment-scale covariates and NPDES indicators provide useful screening context but do not directly quantify PFAS loadings or emissions [22]. Finally, SIC-based source categorization is used only as a screening inventory tool; SIC codes may not reflect PFAS use, treatment, or discharge at an individual facility.
Conclusions
This pilot case study demonstrates that transparent post-processing choices can materially affect site-level summed PFAS metrics in sparse EPA Method 1633 surface-water datasets. In the Fort Worth application, explicit treatment of left-censoring and qualifier sensitivity altered low-end ΣPFAS40 values by 32–38 ng/L relative to detect-only sums but did not change the main screening outcome: TW08 remained the highest-priority site for follow-up investigation, with a median ΣPFAS40 of 685.2 ng/L (95% imputation-uncertainty interval: 673.2–696.8 ng/L), and the top-three site ranking was perfectly stable across all five censor-handling approaches examined. River-network and NPDES information added a useful contextual screening structure, but they did not provide a basis for source attribution, particularly in the absence of PFAS-specific public DMR or permit-limit records. The main contribution of the study is a reproducible workflow for generating defensible screening summaries from heavily censored, qualifier-rich PFAS datasets, rather than a definitive characterization of PFAS occurrence across the Trinity River headwaters.
Supporting information
S1 Table. Target PFAS analytes (n = 40), detection frequencies, and concentration ranges.
Of 400 total site–analyte results, 82 (20.5%) were detected, and 318 (79.5%) were reported below the reporting limit. Twelve of 40 target analytes were detected at least once in the synoptic campaign. Detection frequency is the percentage of the 10 sampling sites with quantified (detected) results for each analyte.
https://doi.org/10.1371/journal.pwat.0000554.s001
(DOCX)
S2 Table. Site-by-analyte PFAS concentrations (ng/L) with qualifiers.
Values are reported as detected concentration (ng/L) or <RL (non-detect at site-specific reporting limit). Sites are ordered by decreasing ΣPFAS40 (TW08 highest, TW25 lowest). Qualifier flags: (I) = ion ratio outside established control limits; (G) = elevated reporting limit due to matrix interference. Default reporting limits: 1.3–1.4 ng/L for most analytes; 2.7–2.8 ng/L for PFBA, FTS compounds, and 3:3 FTCA; 6.7–6.9 ng/L for sulfonamide ethanols and 5:3/7:3 FTCAs. The 8:2 FTS reporting limit at TW08 was elevated to 14 ng/L due to matrix interference.
https://doi.org/10.1371/journal.pwat.0000554.s002
(DOCX)
S3 Table. Upstream catchment characteristics and NPDES permit indicators by sampling site.
Sites are ordered by decreasing ΣPFAS40. Land-cover data were compiled from the National Land Cover Database (NLCD) 2021. NPDES permit data were compiled from EPA ECHO ICIS-NPDES records for fiscal years 2021–2025. Permit density is reported as active permits per 100 km² of upstream drainage area. Nearest Permit, Any Type = minimum hydrologically routed river-network distance (km) to an upstream active permit across all source categories, including Stormwater/MS4. Nearest Individual Permit = river-network distance (km) to the nearest upstream NPDES permit classified as WWTP, Industrial, or Other/Unknown by SIC code; n/a indicates no permits in these categories are present upstream. The latter is a SIC-based proxy, not a formal NPD/GPC permit-type classification.
https://doi.org/10.1371/journal.pwat.0000554.s003
(DOCX)
S4 Table. Upstream NPDES permit counts by SIC-based source category.
Active permits are unique PERMIT_NMB records with active status (EFF, ADC, or EXP) from EPA ECHO ICIS-NPDES. Source categories were assigned by mapping 4-digit SIC_Codes from ICIS-NPDES facility records into six screening groups. Source-category definitions: Stormwater/MS4 comprises construction general permits, MS4 permits, and other general stormwater coverage; WWTP includes publicly owned treatment works and individual NPDES permits at sewerage/wastewater facilities (SIC 4952); Industrial includes individual and industrial stormwater permits at manufacturing and industrial facilities; Airport includes permits at air transportation facilities (SIC 45xx); Landfill includes permits at refuse and solid-waste facilities (SIC 4953); Other/Unknown includes permits not classifiable into the above groups or with missing SIC codes. Total unique active upstream permits across all sites: 855. Stormwater/MS4 accounts for 808 permits (94.5% of the total).
https://doi.org/10.1371/journal.pwat.0000554.s004
(DOCX)
S5 Table. Exploratory Spearman screening correlations with Benjamini–Hochberg FDR-adjusted q-values.
Raw two-sided Spearman p-values were adjusted using the Benjamini-Hochberg procedure across the family of exploratory covariate-screening tests shown in this table. Bootstrap intervals are nonparametric 95% intervals for Spearman ρ. Results are descriptive screening metrics only and are not interpreted as formal inferential tests because the study included only 10 sites from a single synoptic event. DMR- and permit-flow proxy rows are especially limited because values were available for only four sites.
https://doi.org/10.1371/journal.pwat.0000554.s005
(DOCX)
S1 Fig. Heatmap of log10-transformed individual PFAS analyte concentrations across the 10 sampling sites.
Rows represent sites ordered by decreasing ΣPFAS40; columns represent the 12 detected analytes grouped by class (PFCA, PFSA, FTS), separated by heavy vertical lines. The color scale represents log10(concentration, ng/L). Non-detect cells are imputed at the sample-specific reporting limit for visualization. TW08 exhibits distinctly elevated PFSA and FTS concentrations relative to all other sites, consistent with the PFSA/FTS-enriched fingerprint described in the main text.
https://doi.org/10.1371/journal.pwat.0000554.s006
(TIF)
S2 Fig. Upstream active NPDES permit counts by SIC-based source category for each sampling site.
Stacked bar chart showing the number of unique active upstream permits per site classified into six screening groups: Stormwater/MS4 (purple), WWTP (orange), Industrial (red), Airport (blue), Landfill (green), and Other/Unknown (gray). Sites are ordered by decreasing ΣPFAS40. TW07 has the highest total permit count (≈395), dominated by Stormwater/MS4, reflecting its large contributing drainage area (1,192 km²). TW08, despite having the highest ΣPFAS40, has only 18 upstream permits, all of which are Stormwater/MS4.
https://doi.org/10.1371/journal.pwat.0000554.s007
(TIF)
Acknowledgments
The authors thank the STREAM Lab (Sustainable Tools for Risk Evaluation and Climate-Water Modeling) in the Department of Environmental and Geological Sciences at Texas Christian University for laboratory and logistical support. The authors also thank Serene Aryal and Rishab Dhakal for their assistance with field sampling, Eurofins for analytical services, and the Texas Commission on Environmental Quality for providing public access to Surface Water Quality Monitoring, Discharge Monitoring Report, and permit-limit data. Publication charges for this article were supported by the TCU Library Open Access Fund.
References
- 1. Buck RC, Franklin J, Berger U, Conder JM, Cousins IT, de Voogt P, et al. Perfluoroalkyl and polyfluoroalkyl substances in the environment: terminology, classification, and origins. Integr Environ Assess Manag. 2011;7(4):513–41. pmid:21793199
- 2. Jarvis AL, Justice JR, Elias MC, Schnitker B, Gallagher K. Perfluorooctane Sulfonate in US Ambient Surface Waters: A Review of Occurrence in Aquatic Environments and Comparison to Global Concentrations. Environ Toxicol Chem. 2021;40(9):2425–42. pmid:34187091
- 3. Sims JL, Stroski KM, Kim S, Killeen G, Ehalt R, Simcik MF, et al. Global occurrence and probabilistic environmental health hazard assessment of per- and polyfluoroalkyl substances (PFASs) in groundwater and surface waters. Sci Total Environ. 2022;816:151535. pmid:34762945
- 4. Islam M, Thompson K, Dickenson E, Quiñones O, Steinle-Darling E, Westerhoff P. Sucralose and predicted de facto wastewater reuse levels correlate with PFAS levels in surface waters. Environ Sci Technol Lett. 2023;10(5):431–8.
- 5. Kali SE, Österlund H, Viklander M, Blecken G-T. Stormwater discharges affect PFAS occurrence, concentrations, and spatial distribution in water and bottom sediment of urban streams. Water Res. 2025;271:122973. pmid:39700609
- 6. George BJ, Gains-Germain L, Broms K, Black K, Furman M, Hays MD, et al. Censoring Trace-Level Environmental Data: Statistical Analysis Considerations to Limit Bias. Environ Sci Technol. 2021;55(6):3786–95. pmid:33625843
- 7. Antweiler RC, Taylor HE. Evaluation of statistical treatments of left-censored environmental data using coincident uncensored data sets: I. Summary statistics. Environ Sci Technol. 2008;42(10):3732–8. pmid:18546715
- 8.
Helsel DR. Statistics for censored environmental data using Minitab and R. 2nd ed. Hoboken, NJ: Wiley. 2012.
- 9.
Helsel DR, Hirsch RM. Statistical methods in water resources. Techniques of Water-Resources Investigations. Reston, VA: US Geological Survey. 2002.
- 10. Lubin JH, Colt JS, Camann D, Davis S, Cerhan JR, Severson RK, et al. Epidemiologic evaluation of measurement data in the presence of detection limits. Environ Health Perspect. 2004;112(17):1691–6. pmid:15579415
- 11.
US Environmental Protection Agency. Method 1633: Analysis of per- and polyfluoroalkyl substances (PFAS) in aqueous, solid, biosolids, and tissue samples by LC-MS/MS. EPA 821-R-24-001. Washington, DC: US EPA, Office of Water. 2024.
- 12. Dumelle M, Peterson EE, Ver Hoef JM, Pearse A, Isaak DJ. SSN2: the next generation of spatial statistical modeling on stream networks. J Open Source Softw. 2024;9(99):6389.
- 13. Ver Hoef JM, Peterson EE. A moving average approach for spatial statistical models of stream networks. J Am Stat Assoc. 2010;105(489):6–18.
- 14. US Environmental Protection Agency. ECHO: Enforcement and Compliance History Online. https://echo.epa.gov/ 2024. 2026 January.
- 15.
Rubin DB. Multiple imputation for nonresponse in surveys. New York: Wiley. 1987.
- 16. Dewitz J, US Geological Survey. National Land Cover Database (NLCD) 2021 Products. https://www.mrlc.gov/data/nlcd-2021-land-cover-conus 2023. 2026 January.
- 17. Place BJ, Field JA. Identification of novel fluorochemicals in aqueous film-forming foams used by the US military. Environ Sci Technol. 2012;46(13):7120–7. pmid:22681548
- 18. Barzen-Hanson KA, Roberts SC, Choyke S, Oetjen K, McAlees A, Riddell N, et al. Discovery of 40 Classes of Per- and Polyfluoroalkyl Substances in Historical Aqueous Film-Forming Foams (AFFFs) and AFFF-Impacted Groundwater. Environ Sci Technol. 2017;51(4):2047–57. pmid:28098989
- 19.
Texas Commission on Environmental Quality. Surface Water Quality Monitoring Program. Austin, TX: TCEQ. 2024. https://www.tceq.texas.gov/waterquality/monitoring
- 20. US Environmental Protection Agency. PFAS National Primary Drinking Water Regulation. Federal Register. 2024;89(82):32532–620.
- 21.
Texas Commission on Environmental Quality. Draft 2026 Texas Integrated Report of Surface Water Quality for Clean Water Act Sections 305(b) and 303(d), Including Draft Impairment Index. Austin, TX: TCEQ. 2025. https://www.tceq.texas.gov/waterquality/assessment
- 22. Byrne P, Mayes WM, James AL, Comber S, Biles E, Riley AL, et al. PFAS river export analysis highlights the urgent need for catchment-scale mass loading data. Environ Sci Technol Lett. 2024;11(3):266–72.