eDNA metabarcoding is not infallible. A species can be genuinely present at a site and fail to appear in a sampling event — and the failure is not random. Detection probability varies systematically with species biology, environmental conditions, sampling effort, and protocol design. For a biodiversity credit program where BHI scores are derived from detected species, the false-negative rate is not merely an academic concern: it affects credit valuations, potentially undervaluing ecologically rich sites, and creates an ethical obligation to account for it transparently.
The Sources of False Negative Detection
False negatives in eDNA detection arise at multiple stages of the sampling and analysis pipeline. Understanding which stage dominates under which conditions is necessary to design appropriate corrections.
Biological and Environmental Sources
The most fundamental source of false negatives is simply low eDNA concentration — the species is present but shedding insufficient DNA for reliable capture and detection at the sampling effort applied. eDNA shedding rates vary enormously across taxa: an active, mucus-producing fish in a small, contained pool sheds orders of magnitude more detectable DNA per hour than a cryptic, low-metabolic-rate reptile resting in a burrow adjacent to the sampled water body. Seasonal and behavioral factors further modulate shedding: amphibians shedding during active breeding are detectable at much lower densities than the same species in thermal torpor.
Hydrological conditions create additional detection uncertainty. In high-flow events, dilution reduces eDNA concentration below detection thresholds; rapid downstream transport means that a 1-liter sample at a given station reflects an unknown upstream source area rather than strictly local organisms. Temperature affects DNA degradation rate — in warm summer conditions, eDNA degrades faster in the water column, reducing detection windows relative to cold winter conditions.
Technical Sources
PCR inhibition is a significant and underappreciated source of false negatives in environmental samples. Riparian and wetland water often contains humic acids, tannins, and other compounds from plant decomposition that co-extract with DNA and inhibit PCR amplification. Samples from tannic, high-organic-load wetlands may show substantially reduced amplification efficiency compared to samples from clear, low-organic streams — meaning the same eDNA concentration in the extract can produce different detection rates depending on matrix effects.
We address PCR inhibition through two mechanisms: (1) BSA (bovine serum albumin) additive in our PCR reaction mix, which binds and sequesters many common PCR inhibitors; and (2) dilution series testing on extracts that fail our internal positive control threshold, to determine whether dilution improves amplification by reducing inhibitor concentration. Samples that remain below the positive control threshold after inhibitor-adjusted protocol are flagged as unreliable and excluded from BHI scoring, with a documented QC failure reason.
Reference database gaps represent a third technical false-negative source: ASVs that are genuine detections but cannot be assigned to a known taxon because the species is absent from the reference database. These become "unassigned" reads that are excluded from BHI scoring rather than attributed to species. For Pacific Northwest aquatic systems, coverage of fish and amphibian reference sequences in NCBI GenBank is reasonably good, but macroinvertebrate coverage (particularly Trichoptera at species level) has meaningful gaps for Oregon-endemic taxa.
Occupancy Modeling as a Correction Framework
The statistical tool most appropriate for handling imperfect detection in eDNA surveys is occupancy modeling. Standard occupancy models estimate the probability that a site is occupied by a species (ψ) separately from the probability that an occupied site is detected given a sampling event (p). With replicated samples per station and multiple stations per site, it is possible to estimate detection probability for each species and apply a site-occupancy correction to the raw detection data.
We apply occupancy modeling to our BHI species richness component using the replicate-level data from each monitoring event. The key assumption is that detection failure across replicates at a given station is independent — if the species is present and was detected in replicate 1 but not replicate 2, the non-detection in replicate 2 is a false negative, not evidence of absence. With three replicates per station, a species detected in at least one replicate is classified as present at that station; a species detected at no replicate at a given station but detected at adjacent stations is modeled under a spatially explicit occupancy framework.
This approach provides a statistically defensible species list that accounts for imperfect detection, rather than treating non-detection as confirmed absence. It is a more conservative approach (slower to claim presence) than treating any single-replicate positive as confirmed detection, but it is appropriate for credit scoring purposes where the risk of false-positive attribution (claiming species are present when they are not) exceeds the risk of false-negative undercount.
We are not saying every non-detection is a false negative that should be treated as presence. We are saying the probability of non-detection is non-trivial and must be explicitly estimated rather than ignored when species absence data is used to infer ecological community state.
Quantifying False Negative Rates Under Field Conditions
Our protocol validation work, conducted at sites with independently established species presence through long-term traditional monitoring, provides empirical estimates of detection probability under different conditions. These are internal validation results, not externally published data, but the ranges are consistent with the published eDNA literature:
- For high-density fish species in low-to-moderate flow conditions: per-replicate detection probability typically in the range of 0.85–0.95. With three replicates, the probability of at least one detection is ≥0.997 — effectively certain.
- For low-density or behaviorally cryptic amphibian species: per-replicate detection probability can fall to 0.30–0.50 depending on season and conditions. With three replicates, the probability of at least one detection ranges approximately 0.66–0.88 — meaningfully below 1.0, and below the threshold where we are comfortable treating non-detection as confirmed absence.
- For macroinvertebrate taxa at typical Oregon stream densities: per-replicate detection probability is intermediate (0.60–0.80 for EPT families under summer conditions) but drops substantially in high-turbidity post-flood conditions where inhibitor loads are elevated.
These estimates are used to set our minimum replicate requirements and to calibrate BHI scoring adjustments for conditions where detection probability is known to be reduced. A site sampled under post-flood high-turbidity conditions receives a reduced-confidence flag on its EPT richness component — the score is not inflated by the expected-but-undetected species, but it is also not penalized as if the non-detections were confirmed absences.
Implications for BHI Scoring Transparency
The honest disclosure posture for a biodiversity credit program is to state: this is the detected community, here is our confidence in the detection completeness, and here is the uncertainty interval on the BHI score that reflects that detection probability. We report BHI scores with an associated confidence tier (High / Moderate / Low) that reflects the combined effect of sampling conditions, replicate structure, and known detection probability for the site's target taxa.
Credits issued from High-confidence BHI scores carry essentially no detection-probability caveat — the evidence base is sound and the species list is reliable. Credits from Moderate-confidence scores carry a disclosure note that specific taxa in the detection profile have reduced detection certainty and that the BHI score should be interpreted accordingly. Low-confidence events trigger a protocol review and, where feasible, a targeted re-sampling visit before the BHI score is used for credit issuance purposes.
This transparency is not a weakness in the system — it is what distinguishes measurement-based credit programs from programs that present their ecological data without uncertainty quantification. A buyer who understands detection probability can make a more informed procurement decision than one who is presented with a species list as if it were a census. Scientific honesty about uncertainty is, in the long run, the more creditworthy position.