# Integrative Evidence Workshop

> A run of negative eDNA tests never proves a species is absent: each independent negative water sample multiplies the occupied sites still hidden among the negatives by the chance an occupied site is missed, and wells from one sample cannot push that sample's detection above the chance the water holds DNA. This workshop reads an environmental DNA (eDNA) dataset the way a careful analyst would: it separates the stages at which DNA can be missed, turns a string of negatives into a probability of occupancy, and makes a stop-or-continue decision with every assumption on the table. It then asks how firm that decision is, how to make the analysis reproducible and who has the authority to decide how the data are used.

Course: Ecology and Evolution (Global change and decisions). Level: core university (third year). About 20 minutes.
Lesson page (link this one when citing): https://stemhop.com/lessons/ecoevo-evidence-workshop/

Builds on: Movement and Spatial Populations · Conservation and Restoration · Global Change: Species on the Move

## What you'll learn

- Interpret eDNA detections and non-detections, naming the nested stages where DNA can be missed (sample, extraction, PCR run) and the sources of false positives and of downstream transport.
- Calculate the probability that an all-negative site is still occupied from stated detection probabilities and a prior, and defend a stop-or-continue decision against a stated threshold.
- Explain why field samples and PCR replicates are different kinds of replication, and plan a reproducible analysis with raw data, a data dictionary, scripts, versions and seeds.
- Apply the CARE principles alongside FAIR to decide how data about a place and its people may be accessed and reused.

The first film counted ponds with round numbers. Here we run it on data, decide, test how firm the decision is, and ask who decides how data are used. Background: [spatial populations](https://stemhop.com/lessons/spatial-populations/), [conservation](https://stemhop.com/lessons/conservation/) and [global change](https://stemhop.com/lessons/global-change/).

> **The one idea.** A run of negatives never proves a species is absent. Each independent negative water sample multiplies the occupied ponds still hiding among the negatives by the chance that an occupied pond is missed, while the empty ponds stay put. So the share still occupied falls with every sample but never reaches zero. (This needs a prior between 0 and 1 and a sample that can both detect and miss.) Wells and extractions from one water sample all test that sample’s water, so they cannot lift its detection above the chance the water holds DNA. Only new water is a new try at the scoop.

## What eDNA detects and what it misses

**Environmental DNA ([eDNA](https://stemhop.com/methods/occupancy-and-edna/))** is DNA collected from an environmental sample, such as water, soil or air, rather than from the organism itself; for animals, much of it is DNA they shed. A targeted survey filters water, extracts DNA and runs a quantitative [PCR](https://stemhop.com/methods/pcr/) (qPCR) assay for one species ([recombinant DNA and PCR](https://stemhop.com/lessons/recombinant-dna-and-pcr/), [sequencing and genomes](https://stemhop.com/lessons/sequencing-and-genomes/)). Each qPCR run, in one well of a plate, either amplifies the target or does not.

A site is **occupied** if the species uses it during the survey. A **false negative** is a miss: the organism or its DNA is present and nothing amplifies. A **false positive** is a report of the target when it is absent; contamination and a poorly specific assay are two routes.

The first film’s three nets get symbols, in its colours: **θ** (theta, blue), the chance that a water sample from an occupied pond holds the species’ DNA in a form the assay can detect; **γ** (gamma, green), the chance that an extraction captures DNA that is in the sample; and **δ** (delta, gold), the chance that one qPCR run amplifies captured DNA. One reason θ is below 1 is decay: it can leave too few intact copies of the target sequence in the scoop, even when shorter fragments remain. In upland Welsh stream mesocosms (outdoor experimental stream channels), DNA from three species fell by more than 90% within three hours and was nondetectable within two days, faster in acidic water (Seymour and colleagues, 2018): one experimental system, not a universal eDNA lifetime.

## Detection probability and an all-negative site

Made-up great crested newt data: four ponds, two water samples each, two extractions per sample, five [qPCR](https://stemhop.com/methods/pcr/) runs per extraction. Each string is one extraction’s five runs; 1 means amplification.

| Pond | Sample 1 (extraction 1 / 2) | Sample 2 (extraction 1 / 2) | Positive wells (of 20) |
| --- | --- | --- | --- |
| S1 | 10000 / 00000 | 00000 / 00000 | 1 |
| S2 | 00000 / 00000 | 00000 / 00000 | 0 |
| S3 | 00000 / 00000 | 00000 / 00000 | 0 |
| S4 | 11100 / 00000 | 11000 / 00000 | 5 |

Two of four ponds amplified at least once, so the **naive occupancy estimate** is 2/4 = 0.50. It ignores misses: S2 or S3 may be occupied. S1 counts as occupied only under the no-false-positive assumption below; its lone positive well is taken up after the decision.

**Worked problem: is S2 still occupied?** S2 gave 20 negative wells: 2 samples × 2 extractions × 5 wells.

| Measured elsewhere (borrowed, illustrative) | Our inputs |
| --- | --- |
| **γ = 0.49**, estimated without inhibitor removal | **θ = 0.60** (assumed) |
| **δ = 0.38** | **prior = 0.20**: our belief, before sampling, that S2 is occupied |
| Wild pigs, central Texas, one autumn (Amy Davis and colleagues, 2018); each has a standard error | **Stop rule** (a decision choice): stop when P(occupied) is 5% or below |

Assume no false positives (an empty pond always tests negative), S2 occupied or empty for the whole survey, and independence at every level: at an occupied pond, one sample missing does not change the chance the next one misses, and likewise for extractions from one sample and wells from one extraction. θ and the prior are assumed; a real analysis would estimate or inform them from relevant data, but four ponds are far too few to do so reliably. Work from the inside out with the product rule from [probability and crosses](https://stemhop.com/lessons/probability-and-crosses/): replicates at one level all miss with the product of their miss chances, and detection needs every step to succeed.

1. *Wells.* One run misses with 1 − 0.38 = 0.62; all five miss with 0.625 = 0.092. So five wells catch captured DNA with **A = 0.908**.
2. *Extractions.* One extraction finds the DNA only if it captures it *and* the wells catch it: 0.49 × 0.908 = 0.445, so it misses with 1 − 0.445 = 0.555. Both miss with 0.555² = 0.308, so the pair finds DNA in the water with **B = 0.692**.

A sample detects only if its water holds DNA *and* the lab finds it, so **q**, the first film’s chance that one water sample detects an occupied pond, is θ × B.

**Predict.** The water holds DNA only 60% of the time. Will q be closer to 0.69 or to 0.40?

> **Check your prediction.** 1. *Sample.* q = 0.60 × 0.692 = **0.415**, closer to 0.40: the scoop sets the ceiling.

1. *Two samples.* Each sample misses with 1 − 0.415 = 0.585, so two independent samples both miss an occupied pond with m = 0.585² = **0.342**.
2. *Weigh against the prior.* Count as the first film did: of 100 ponds like S2, 20 are occupied and 20 × 0.342 = 6.8 of them stay hidden among 6.8 + 80 = 86.8 all-negative ponds. That share is the **posterior** (the probability after the data):

> **P(occupied | all negative) = prior × m / [(1 − prior) + prior × m]** = 0.068 / 0.868 = **7.9%**

**Predict.** Stop, or sample again? If again, how many more negative samples will it take?

> **Check your prediction.** 1. *Decide.* 7.9% is above 5%, so sample again. A third negative sample gives m = 0.585³ = 0.200 and 0.040 / 0.840 = **4.8%**, just under the line.

*Interactive on the lesson page:* eDNA detection ladder — from negatives to a decision

A third sample means another field trip. Why not run more wells instead?

**Film transcript: Wells Share One Scoop**

### Film transcript: Wells Share One Scoop

More wells cost an afternoon in the lab, not a field trip. Pond S2 already gave twenty negative wells. Are those twenty tries, or only two? Draw S2 as a tree. The pond gives two blue water bottles. Each bottle gives two green extraction tubes. Each tube feeds five gold wells. So all twenty wells hang from just two scoops of water. Use the workshop's numbers. A scoop holds DNA 60% % of the time. That is theta, the blue line. The capture and well values are borrowed from a wild pig study in Texas, as illustrations only. Try adding wells. Pour more gold wells under the same two bottles. The gauge for one sample climbs at first. Then it levels off, still below the blue line.

It never crosses theta, however many wells you add. The reason is simple. If the scoop held no DNA, every tube and every well from it is testing empty water. No number of wells can find DNA that the sample does not contain. Watch the bar for the chance that S2 is still occupied. With unlimited wells on two samples, it stops at 7.2%. That is still above the 5% percent line. Now add a third blue bottle instead, a new scoop of water from the pond. With the usual five wells per extraction, the bar drops to 4.8%. That is below the line. Here is a tempting mistake.

Suppose we pretend the twenty wells were twenty separate trips to the pond. Will the pond look emptier, or fuller, than the evidence allows? Emptier. Redraw the tree as twenty bottles, each with one well. Each one now looks like its own try. The bar plunges to about two percent, and you would stop. But you would stop on an error in the model, not on evidence. Counting repeated measurements of one unit as if they were separate units has a name. It is called pseudoreplication. So wells and extractions from one sample all test the same water. They are second chances at the lab steps, not new tries at the scoop. Here, adding wells cannot cross the line.

A third water sample does.

From the film

- Wells and extractions from one water sample all test that sample's water: they are second chances at the lab steps, not new tries at the scoop, so they cannot detect DNA the sample does not contain.
- However many wells you add, one sample's chance of detecting an occupied pond cannot exceed θ (theta), the chance the water holds DNA.
- In this example, adding wells to two samples (two extractions each) leaves an all-negative pond above the 5% line; a third independent water sample brings it below.
- Counting wells as if they were separate field samples is pseudoreplication: it makes a pond look emptier than the evidence shows.

## Sampling design: field replicates versus PCR replicates

The second film called the twenty-well shortcut **pseudoreplication** (counting repeated measurements of one unit as separate units). It left out two things. First, the ceiling is on one sample, not on the decision: with enough extractions, even two samples can cross the line (Going deeper). And how firm is 4.8%? Davis’s γ has a standard error (SE, a measure of an estimate’s uncertainty) of 0.07.

**Predict.** Lower γ by one SE, to 0.42. After three negative samples, stop or continue?

> **Check your prediction.** Continue: three samples now give 5.9%.
>
> | P(occupied \| all negative) | 2 samples | 3 samples | 4 samples |
> | --- | --- | --- | --- |
> | γ = 0.42 (one SE lower) | 9.0% | 5.9% | 3.8% |
> | γ = 0.49 (worked problem) | 7.9% | 4.8% | 2.8% |
> | γ = 0.56 (one SE higher) | 6.9% | 3.9% | 2.2% |
> | γ = 0.49, unlimited wells | 7.2% | 4.1% | 2.3% |
>
> *New water samples move the posterior; unlimited wells barely do. Lowering γ by one standard error puts the third sample above 5%.*

Our choices matter too: across θ from 0.5 to 0.7 and priors from 0.1 to 0.3, the negative samples needed run from 2 to 5. Report a decision this close to its line with its sensitivity.

In the English newt surveys analysed by Andrew Buxton and colleagues (2021), twenty subsamples from around each pond were **pooled** into one water sample: wider coverage, but one field sample, leaving no field replicates to estimate θ. In the Texas pig trial, the authors suspected that pooling each site’s ten samples diluted the DNA, and pooled data underestimated occupancy at sites known to be occupied. Buxton’s simulations found that replicating both field and lab work reduced bias and narrowed credible intervals (Bayesian ranges of plausible values). Only new water is a new try at the scoop.

## What a positive can and cannot mean

So far we assumed no false positives. Clean blanks make contamination less likely but cannot exclude a rare, one-off event. S1’s single positive well out of twenty is the sparse pattern Buxton and colleagues judged more likely than others to be a false positive, and their two published models, one of which ignored false positives, gave overall occupancy of 0.26 and 0.47 for the same 2,958 English ponds surveyed in 2019 (one pooled sample each, twelve [qPCR](https://stemhop.com/methods/pcr/) replicates). With no ground truth, the authors judged that the model ignoring false positives overestimated; false-positive parameters can need calibration or informative priors to be identified.

Place matters too: in flowing water, a true DNA detection can falsely signal occupancy at that spot. In the river Glatt below Lake Greifensee, Switzerland, Kristy Deiner and Florian Altermatt (2014) detected the lake-dwelling water flea *Daphnia longispina* at every sampled river location, as far as 12.3 km downstream, but did not detect the mussel *Unio tumidus* beyond 9.1 km.

*Figure (AI-generated illustration).* In the river Glatt below Lake Greifensee, Deiner and Altermatt detected DNA of lake-dwelling animals in the river, so in flowing water a sample can carry DNA from upstream. A positive tells you about the water, not only about the spot where the bottle was filled. Image: a lake with a water flea and a mussel, and a river carrying gold dots of shed DNA downstream to a sampling bottle; an inset arrow points from the bottle back upstream to the lake.

### Design problem

A river survey detects an endangered mussel’s [eDNA](https://stemhop.com/methods/occupancy-and-edna/) at one site 5 km below a lake where the mussel lives. Does the mussel occur at the site? Design the follow-up.

> **One way to do it.** A positive can mean a local population, DNA carried down from the lake, or contamination. In the Glatt, mussel DNA was detected downstream though not beyond 9.1 km, so transport over 5 km is plausible in at least one river.
>
> 1. *Check for contamination.* Carry field blanks, extraction blanks and PCR negatives through every batch, and sequence positives to confirm the species.
> 2. *Test transport.* Sample from the lake outlet down past the site, plus each tributary between. A local population would tend to add signal at and below the site; transport alone would tend to fade with distance from the lake. DNA settling on the streambed and stirred back up, and patchy flow, blur both patterns, so use replicate samples and flow data.
> 3. *Replicate in the field.* Independent water samples on separate dates, not more wells; record flow and pH.
> 4. *Model and confirm.* DNA carried from the lake is a false positive for this site, so the no-false-positive model above fits only once transport is excluded; otherwise use a model that allows transport or other false positives, with stated assumptions. Either way, search the site directly with a method suited to mussels.
> 5. *Agree access first.* Settle with whoever governs the river and the data how results will be shared.

## Reproducible analysis and the decision

Keep the raw file untouched; work on copies. A **data dictionary** defines every column: site, date, sample, extraction, [PCR](https://stemhop.com/methods/pcr/) replicate, control, treatment, covariates. Record a non-detection differently from a missing, inhibited or inconclusive run; an inhibited well, where substances carried over from the sample blocked the PCR, is not a negative.

Sandve and colleagues (2013) set out rules that apply here: replace manual editing with scripts, archive the exact program versions, record every parameter and input, and record random seeds whenever an analysis uses randomness. FAIR (findable, accessible, interoperable, reusable) adds provenance and a clear licence. Then rerun from the raw file and check you get 7.9% and 4.8% again, with θ, the prior and the threshold beside them.

## Whose data? CARE and FAIR

Surveys happen on someone’s land and water. FAIR makes data usable by machines and people, and it permits authentication and authorization; it does not mean open to all. The **CARE principles** for Indigenous data governance, drafted in Gaborone, Botswana, on 8 November 2018 and published in 2020, add the question of who decides: **Collective Benefit** (data should benefit Indigenous Peoples), **Authority to Control** (Indigenous Peoples determine governance protocols for their data), **Responsibility** (respectful relationships with the people the data come from) and **Ethics** (their rights and wellbeing are central across the data’s life). A 2021 paper describes CARE as a complement to FAIR that covers scientific data as well as traditional knowledge.

For [eDNA](https://stemhop.com/methods/occupancy-and-edna/), a rare species’ location in a community’s water is itself data whose release that community may have the authority to control. This workshop uses no real community dataset and claims no permission. Merging eDNA results with a community’s observations would first mean identifying the rights holders, the benefit and the permitted purpose, and getting their protocols; a downloadable file is not permission.

## Going deeper

### The model in one line

With w wells per extraction and k extractions per sample, A = 1 − (1 − δ)w and q = θ × [1 − (1 − γA)k]; n independent samples miss an occupied site with m = (1 − q)n. In odds, each negative sample multiplies the odds of occupancy by 1 − q: the hidden-pond count in other words. These equations assume no false positives, closure (occupancy fixed during sampling), conditional independence at every replicate level (wells, extractions and samples, as in the worked problem), and parameters constant across sites and visits (or explained by covariates). A model that allows false positives needs extra false-positive parameters, which must be identifiable, and a different posterior.

### When lab work could cross the line

Perfect lab work makes q = θ = 0.60: one negative sample leaves 0.20 × 0.40 / (0.80 + 0.08) = 9.1%, two leave 3.8%. With unlimited wells an extraction still captures only with γ = 0.49, so q = 0.60 × [1 − 0.51k] for k extractions: two samples with four extractions each give 4.6% (three extractions, 5.4%).

### The pseudoreplication arithmetic

Treating each of S2’s 20 wells as an independent try detecting with 0.60 × 0.49 × 0.38 = 0.112 gives a miss probability of 0.88820 = 0.094 and the second film’s 2.3%.

### Carrying the uncertainty through

Borrowed point estimates hide their uncertainty and covariates (Davis and colleagues found DNA availability varied with month, pH and water-body type). The γ table is a one-at-a-time sensitivity check. For management use, compute the posterior across the joint uncertainty of γ, δ, θ and the other fitted parameters, with relevant covariates, and test whether the decision changes.

---

From STEMhop (https://stemhop.com), free lessons by Christopher Deppmann. Lesson text © the authors. Practice questions: https://stemhop.com/lessons/ecoevo-evidence-workshop/
