Abstract
Ecological surveys risk incurring false negative and false positive detections of the target species. With indirect survey methods, such as environmental DNA (eDNA), such error can occur at two stages: sample collection and laboratory analysis. Here we analyse a large qPCR based eDNA data set using two occupancy models, one of which accounts for false positive error (Griffin et al., 2020), and a second that assumes no false positive error (Stratton et al., 2020). Additionally, we apply the Griffin et al. (2020) model to simulated data to determine optimal levels of replication at both sampling stages. The Stratton et al. (2020) model consistently overestimated both overall and individual site occupancy compared to the Griffin et al. (2020) model and previous estimates. The inclusion of replication at both stages of eDNA analysis reduces both bias and credible interval width. Notably, collecting more than one field sample from a site can improve parameter estimates more effectively than increasing replicates solely within the laboratory analysis.
The Two-Stage Error Problem in eDNA
Environmental DNA—DNA shed by organisms into their environment—has revolutionized biodiversity monitoring. However, as an indirect method detecting molecules rather than organisms, it introduces two distinct stages where errors can occur.
- Stage 1 (Field Collection): DNA might be present in a water body but missed in a specific water sample (false negative), or a sample might be contaminated in the field (false positive).
- Stage 2 (Laboratory Analysis): DNA present in the sample bottle might fail to amplify during PCR (false negative), or lab contamination might create a "hit" where none exists (false positive).
Most standard occupancy models only account for false negatives. Furthermore, there is no consensus on the "golden ratio" of replication: should resources be spent collecting more water samples from a site (Stage 1) or running more PCR replicates on a single sample (Stage 2)?
Methods: Model Comparison and Simulation
The researchers compared two Bayesian multi-scale occupancy models:
- Griffin et al. (2020): Accounts for false positives at both stages.
- Stratton et al. (2020): Assumes false positives do not occur.
To optimize protocols, they simulated eDNA datasets across varying numbers of sites ($S$), field samples ($M$), and qPCR replicates ($K$). They also applied these models to a massive real-world dataset of 2,958 ponds surveyed for Great Crested Newts in England.
Key Insights
The Impact of False Positives
The most striking result was the degree of overestimation when false positives were ignored. In the Great Crested Newt case study, the Stratton model (no false positives) estimated occupancy at 47%. When the Griffin model accounted for false positives, the estimate dropped to 26%, aligning much more closely with historical ecological data. Samples with only 1 or 2 positive replicates out of 12 were identified as highly likely false positives.
Field Samples vs. Laboratory Replicates
A common practice is to increase laboratory replicates (qPCR runs) because it is cheaper than returning to the field. However, the simulation study revealed that this is not the most efficient path to accuracy.
As shown in the figure above, collecting just two field samples ($M=2$) reduced bias and uncertainty (PCI width) more effectively than running a high number of replicates on a single sample ($M=1$). For qPCR replication ($K$), the biggest gains in precision occurred between 2 and 6 replicates, with diminishing returns thereafter.
Conclusion and Recommendations
The study suggests that current "standard" protocols, which often rely on a single field sample with many lab replicates (e.g., $M=1, K=12$), may be suboptimal. A more robust and cost-effective protocol would involve $M=2$ field samples and $K=4–6$ qPCR replicates.
While the Griffin model provides a much more accurate picture of species distribution by acknowledging false positive errors, it requires careful implementation—sometimes needing informative priors or site covariates when replication is low. Future work should look to extend these optimization strategies to eDNA metabarcoding, where error profiles and costs differ.
Summary updated