The Gold Standard of Zoology | False-Positive Species Records and Weighted Validation of Site Presence

For this lesson, a site is either occupied by the target species during a defined period or unoccupied under the stipulated model. The observation method produces a positive or negative target label for one specified survey occasion.

Those definitions are deliberately limited. A site can contain several species, an occupied site can yield no target record and a positive label can be wrong. The binary model does not imply that the natural world contains only two kinds of animal.

Write the observation unit precisely: one site-survey label, rather than one animal. A thousand positive files do not automatically represent a thousand occupied sites or a thousand individuals. Grouping repeated files into visits is another decision that needs a documented rule.

The first evidence question is therefore, “What event does this positive describe?” If the answer is only “the software displayed a species name”, retain that as the observation. A stronger biological conclusion needs support beyond the appearance of a confident label.

The Bioacoustics guide already develops rare-call detector precision in labelled recording windows. This article extends the evidence question to site presence, unequal validation sampling, pooled environments and false-positive occupancy corrections. A recording window and a surveyed site require different denominators.

Contents · Next chapter

Let ψ be the proportion of sites occupied by the target in the model population. Let s be the probability of a positive label conditional on occupancy. Let a be the probability of a positive label conditional on non-occupancy.

The probability s includes the whole defined observation process for an occupied site. It should not be silently equated with the accuracy of classifying a perfect specimen photograph. The false-positive probability a describes target labels at sites stipulated to lack the target.

The overall positive-label fraction is q = ψs + (1 − ψ)a. The first term represents positives from occupied sites; the second represents positives from unoccupied sites. The terms have different biological meanings even though both enter the positive total.

The USGS discussion of false-positive models explains why species misclassification and confirmation information matter in occupancy inference. The numerical population below is an original construction that makes the mixture visible without claiming to reproduce a field study.

Contents · Previous chapter · Next chapter

Suppose a fictional study population contains 10,000 sites. Exactly 1,000 are stipulated to be occupied, giving ψ = 0.10. Let s = 0.80 and a = 0.02 for the defined survey occasion.

Among occupied sites, the expected positive count is 1,000 × 0.80 = 800 and the expected negative count is 200. Among 9,000 unoccupied sites, the expected false-positive count is 9,000 × 0.02 = 180 and the expected negative count is 8,820.

The entries are model expectations, not reported survey results. Their integer values were chosen for clarity. A realised random survey can differ from these counts even if the stipulated probabilities describe its process.

Check the margins before interpreting the table. Each row must reproduce its stated site total, and the positive column must include both true and false positives. A table with tidy percentages can still conceal a mismatched denominator.

Contents · Previous chapter · Next chapter

Consider a separate fictional dataset with 1,000 positive-labelled sites and 9,000 negative-labelled sites. A validation exercise samples 100 of the positives and ninety of the negatives, using a stipulated reliable site-state reference.

Suppose eighty of the 100 reviewed positives are confirmed occupied, while two of the ninety reviewed negatives are confirmed occupied. If the samples are representative within their respective label strata, the positive stratum suggests 1,000 × 80/100 = 800 occupied sites. The negative stratum suggests 9,000 × 2/90 = 200.

The resulting model-based population estimate is 1,000 occupied sites out of 10,000. Reviewing positives alone would reveal information about positive reliability but miss the occupied sites hidden among negatives.

Pooling the reviewed records without weights would give 82 occupied out of 190 reviewed sites, approximately 43.16%. That is the occupied fraction in this deliberately unequal review sample, not the estimated population occupancy of ten percent.

The example assumes appropriate sampling within each stratum and a reliable reference for the supplied exercise. It does not provide a complete uncertainty calculation. Real validation must address sampling variability, reference errors and which records were actually eligible for review.

Ask learners to explain why two verified occupied negatives can represent more population sites than twenty verified false positives in the other stratum. The difference comes from the sampling weights, not from assigning greater biological importance to one record. This is another place where the denominator carries the scientific meaning.

Contents · Previous chapter · Next chapter

Among the 980 expected positive labels, 800 come from occupied sites. The model probability of occupancy conditional on a positive is 800/980, approximately 0.8163, or 81.63%.

This quantity is not s. The detection probability s = 80% asks about positive labels among occupied sites. The 81.63% asks about occupancy among positive labels. Reversing the conditioning changes the denominator.

In symbols, P(occupied given positive) = ψs/[ψs + (1 − ψ)a], provided the denominator is positive. The population table supplies an accessible route to the same calculation before the conditional-probability notation is introduced.

Ask the learner to point to the relevant row or column before dividing. If they use 800/1,000 when asked about the reliability of a positive, the arithmetic is fine but the conditional question is reversed. Repair the question-to-denominator connection rather than assigning more multiplication practice.

Contents · Previous chapter · Next chapter

Now stipulate only 100 occupied sites in another 10,000-site population, so ψ = 0.01. Keep s = 0.80 and a = 0.02 unchanged.

The expected true positives are eighty. The expected false positives are 9,900 × 0.02 = 198. Positive labels total 278, of which eighty arise from occupied sites. The positive’s model reliability is 80/278, approximately 28.78%.

The method’s conditional detection and false-positive probabilities did not change. The population did. When target presence is rarer, even a relatively small false-positive probability applied to many unoccupied sites can contribute a large share of positive labels.

This does not mean rare records should be dismissed. It means a positive label needs an evidence and confirmation strategy appropriate to the question. The model explains why a classifier’s advertised detection percentage cannot simply be read as the probability that every positive label in a new setting is correct.

Contents · Previous chapter · Next chapter

In the first population, there are 1,000 occupied sites and 980 expected positive labels. The totals differ by only twenty. A superficial comparison might suggest that the method is almost perfectly identifying site occupancy.

The table tells a different story. Two hundred occupied sites receive negative labels while 180 unoccupied sites receive positive labels. Large classification errors partly cancel in the total.

Agreement between an aggregate count and the true total does not establish correct identification of the individual sites. The distinction matters if the next decision concerns which sites need review, protection or further investigation.

For a transfer exercise, compare two methods with the same total number of positives but different row-level errors. Ask which evidence is needed to know whether the same sites were correctly identified. A total alone cannot answer that question, just as a zero average signed measurement error does not establish accurate individual readings.

Contents · Previous chapter · Next chapter

From q = a + ψ(s − a), solve ψ = (q − a)/(s − a) when s differs from a. For q = 0.098, s = 0.80 and a = 0.02, the result is (0.098 − 0.02)/0.78 = 0.10.

The equation recovers the stipulated occupancy proportion because its inputs match the exact model. It does not establish that s and a are known in an actual study. Estimating or validating those quantities is part of the scientific work.

If s equals a, positive labels occur at the same probability whether sites are occupied or not. The denominator is zero, expressing a lack of information about occupancy in that observation channel under this simple mixture.

If estimated inputs produce ψ below zero or above one, do not hide the conflict by clipping the answer to the probability range and declaring success. Sampling variation, parameter uncertainty or an inadequate model may need treatment. The algebraic contradiction is evidence to investigate.

Contents · Previous chapter · Next chapter

Keep the observed positive fraction q = 0.098. If s = 0.80 and a = zero, the implied ψ is 0.1225. If a = 0.02, the implied ψ is 0.10. If a = 0.08, it becomes 0.025.

All three parameter combinations reproduce the same q under their respective assumptions. The positive fraction alone cannot identify detection, false-positive probability and occupancy simultaneously.

More sites can estimate q more precisely without resolving this structural ambiguity. A very narrow uncertainty interval around the positive-label fraction is not automatically a narrow interval around biological occupancy when the observation-error parameters remain uncertain.

Ask the learner which extra evidence could distinguish the three explanations. Suitable answers might involve independently reviewed records, additional observation methods or repeated histories within an appropriate model. “Collect more of exactly the same unvalidated label” is incomplete unless the additional structure actually separates the unknowns.

Contents · Previous chapter · Next chapter

A confirmation process can provide information about labels, but its definition matters. Is it an independent expert review, another evidence source, a repeated automated pass or a label supplied by someone who saw the first result?

A review of only positive labels may estimate how many reviewed positives are supported under that review standard. It does not directly measure how many occupied sites received negative labels, because those negatives were not examined.

Likewise, confirming a clearly visible specimen is different from confirming site occupancy during a defined survey period. The reference procedure must match the question and acknowledge its own limits.

The USGS research on acoustic-survey confirmation design connects misidentification, confirmation effort and study objectives. Its study-specific findings should not be converted into a universal instruction to review a fixed percentage of every dataset. Here the transferable point is to design validation around the intended inference.

Contents · Previous chapter · Next chapter

Return to the rare-target population with ψ = 0.01, s = 0.80 and a = 0.02. Stipulate two observation methods that are conditionally independent given occupancy state, each with those same probabilities.

Requiring both methods to be positive gives probability s² = 0.64 at occupied sites and a² = 0.0004 at unoccupied sites. In 10,000 sites, expected supported positives are 100 × 0.64 = 64 and expected false positives are 9,900 × 0.0004 = 3.96.

The model reliability among joint positives is 64/(64 + 3.96), approximately 94.17%. This is higher than the single-method 28.78%, but the probability of a joint positive at an occupied site has fallen from 0.80 to 0.64.

The tradeoff is visible: stricter joint confirmation can reduce false positives while also missing more occupied sites. The calculation is conditional on the declared independence and method definitions. It is not evidence that any two real reviewers or algorithms provide those joint probabilities.

Contents · Previous chapter · Next chapter

Suppose the “second method” is simply the same classifier run again on the same unchanged recording with deterministic output. Its answer repeats the first. Two positive outputs do not provide two independent observations.

In that limiting construction, joint positivity can have the same false-positive probability a rather than a². Squaring 0.02 would therefore overstate the reduction in false positives.

Less extreme dependencies can arise when reviewers share the same misleading feature or when two systems were trained on similar errors. The question is not whether there are two names on the method list, but how their joint errors behave under the relevant conditions.

Ask learners to compare two diagrams: independent conditional checks and a copied label. Both display two positive boxes, yet they carry different evidence. Counting boxes is not a substitute for analysing dependence. A carefully documented independent source can add more information than many repeated transformations of the same ambiguous record.

Contents · Previous chapter · Next chapter

If two independent methods are used and a site is called positive when either is positive, the occupied-site positive probability is 1 − (1 − s)². With s = 0.80, that is 0.96.

The unoccupied-site false-positive probability becomes 1 − (1 − a)². With a = 0.02, that is 0.0396. The rule catches more occupied sites while also admitting more unoccupied sites into the positive group.

The both-positive and either-positive rules therefore answer different decision objectives. Neither is automatically the best biological method. Their suitability depends on the use of the label, review resources and the consequences of missed or false detections.

For a classroom task, state the rule before calculating the table. Changing the rule after seeing which total looks more appealing obscures the evidence. A useful report preserves how positive labels were aggregated and why that aggregation fits the study question.

Contents · Previous chapter · Next chapter

Suppose an unoccupied site has independent false-positive probability a = 0.02 on each of five visits. The chance of at least one positive is 1 − 0.98⁵, approximately 0.09608.

Repeated opportunities can increase the chance of at least one false positive as well as improve detection of truly occupied sites. Calling a site occupied whenever any visit is positive needs an account of both processes.

The independence and stable-occupancy assumptions are explicit teaching conditions here. Real visits can share observers, environmental noise or identification problems. Occupancy may also change during the period if the chosen definition does not support closure.

The earlier all-zero survey-history guide develops non-detection. This article supplies the complementary question: what if some positive records are wrong? A complete study should not borrow a no-false-positive conclusion from a lesson that deliberately assumed it.

Contents · Previous chapter · Next chapter

A classifier may display a score beside a species label. That number’s meaning depends on the model, calibration and conditions. A score of 0.90 should not automatically be interpreted as a demonstrated 90% probability that the target occupies the site.

The score may concern a recording segment rather than site occupancy. It may be uncalibrated in the new environment, or the site-level aggregation may combine many segments. Each step changes the interpretation that needs validation.

For a supplied exercise, keep a model score column separate from the reviewed label and the biological state stipulated for checking. Ask what evidence would connect the score to long-run reliability for comparable records.

Increasing a score threshold can change both s and a. Without a validation table or other suitable evidence, the direction and size of those changes should not be invented. A more confident-looking display is not itself a new observation of the animal.

Contents · Previous chapter · Next chapter

A false species label concerns which species the record supports. A missing individual identifier concerns whether a later record can be linked to the same animal. Both involve identification, but they affect different questions.

The earlier tag-retention guide explains how an animal can remain alive while its original identity becomes unavailable. Its simplified chain explicitly set aside false-positive identity assignment.

Here a record may be assigned to a target species that was not present under the stipulated site model. Correcting that process does not automatically solve tag loss, repeated individuals or abundance estimation.

Use a three-column diagnostic: species label, individual identity and site state. A record may establish one while leaving another uncertain. This prevents a correct calculation about classification reliability from being presented as a population-size or survival estimate that the data cannot support.

Contents · Previous chapter · Next chapter

Suppose an identification method was checked on clear recordings from one environment. Applying its estimated error probabilities to noisier recordings elsewhere assumes that the relevant conditional performance transfers.

That assumption may fail if background sounds, similar species, recording equipment or observer conditions differ. The point is not that transfer is impossible; it is that the scope of validation belongs in the inference.

For an original sensitivity case, hold ψ = 0.01 and s = 0.80, but let a increase from 0.02 to 0.05. In 10,000 sites the true-positive expectation remains eighty while false positives increase from 198 to 495. Reliability among positives becomes 80/575, approximately 13.91%.

The example isolates one changed condition. It does not claim that any real environment produces that particular change. A good learner explains which assumption moved and how the conclusion depends on it, rather than treating an old validation percentage as a permanent property of the method’s name.

Contents · Previous chapter · Next chapter

Suppose q = 0.098 and s = 0.80 are stipulated, while a lies between 0.01 and 0.03. The correction is ψ = (0.098 − a)/(0.80 − a).

At a = 0.01, ψ is 0.088/0.79, approximately 0.11139. At a = 0.03, it is 0.068/0.77, approximately 0.08831. Under these conditions, increasing a lowers the inferred occupancy because more positives can be attributed to unoccupied sites.

The resulting range is about 8.83% to 11.14%. It reflects the supplied deterministic uncertainty in a while treating q and s as fixed. It is not a confidence interval that includes all sampling and model uncertainty.

This is a useful alternative to reporting a single unqualified correction. Partial information can support a qualified range. The calculation should name which uncertainties it includes and which remain outside the bound.

Contents · Previous chapter · Next chapter

Retain the original recording or observation reference, initial label, method version, score where relevant, review status, reviewer or reference method, final accepted label and reason for any revision. Preserve timestamps and the site-survey grouping rule.

Do not overwrite every uncertain record with a convenient negative. Uncertain, unreviewed and confirmed negative describe different evidence states. A missing review is not proof of target absence.

Keep sensitive wildlife-location details within the appropriate review arrangements. A teaching table can use fictional site identifiers without exposing a real animal location. Scientific traceability does not require publishing every coordinate indiscriminately.

Ask another learner to trace one positive in the summary table back through its review status and original record. If they cannot tell whether it was machine-labelled, independently confirmed or merely repeated in another file, the evidence categories need clearer definitions.

Contents · Previous chapter · Next chapter

Mistake one: “The method finds eighty percent of occupied sites, so eighty percent of its positive labels are correct.” Repair by comparing the occupied row with the positive column. The conditional questions differ.

Mistake two: “Only two percent of unoccupied sites are falsely labelled, so only two percent of positive labels are false.” Repair by calculating how many unoccupied sites are available to contribute false positives.

Mistake three: “The positive total nearly matches the true occupied total, so the labels are accurate.” Repair by inspecting false negatives and false positives separately. They can cancel in an aggregate.

Mistake four: “Two positives square the error probability.” Repair by asking whether the checks are conditionally independent and whether they refer to the same defined observation event. Repetition alone does not establish independence.

For each repair, use one small table that the learner can explain aloud. More elaborate formulas are useful after the conditioning and denominator are stable, not as a substitute for them.

Contents · Previous chapter · Next chapter

Task A: 2,000 model sites contain 200 occupied sites. Let s = 0.90 and a = 0.05. Find expected positives. Answer: 180 from occupied sites and ninety from unoccupied sites, totalling 270.

Task B: among those positives, what proportion arise from occupied sites? Answer: 180/270 = two thirds. Explain why the answer is neither ninety percent nor ninety-five percent.

Task C: keep s = 0.90 and a = 0.05 but make only twenty of the 2,000 sites occupied. Answer: eighteen true positives and ninety-nine false positives, so reliability is 18/117, approximately 15.38%.

Task D: q = 0.14, s = 0.90 and a = 0.05. Recover ψ in the exact mixture. Answer: (0.14 − 0.05)/0.85, approximately 0.10588. Check by substituting into q = a + ψ(s − a).

Task E: an unchanged deterministic classifier produces the same positive twice from one file. Can its false-positive probability automatically be squared? Answer: no. The repeated output is not an independent conditional check.

Contents · Previous chapter · Next chapter

For repair, begin with a hundred fictional records divided into known occupied and unoccupied groups. Sort the positive and negative labels, then ask the learner to identify the denominator for each question.

For stabilisation, change prevalence while keeping conditional performance fixed. Require a predicted direction of change before calculating positive reliability. Then reverse the task and recover occupancy under supplied s and a.

For extension, introduce dependence between methods, uncertain error probabilities or a new environment. Ask which previous conclusion survives and which needs additional evidence. The learner should recognise that a precise calculation can still rest on a fragile transfer assumption.

A parent can ask, “Among which group are we calculating this percentage?” A teacher can ask, “What extra evidence would separate these explanations?” Those questions preserve the first-principles clarity of the learning guide and help numerical reasoning travel beyond animal surveys.

Contents · Previous chapter · Next chapter

Suppose two fictional environments each contain 5,000 sites. In environment A, ψ = 0.20, s = 0.90 and a = 0.01. In environment B, ψ = 0.02, s = 0.60 and a = 0.05.

Environment A contains 1,000 occupied and 4,000 unoccupied sites. Its expected true positives are 900 and false positives forty. Environment B contains 100 occupied and 4,900 unoccupied sites, giving sixty true positives and 245 false positives.

Across both environments, there are 1,100 occupied sites, 960 true positives and 285 false positives. The positive fraction is 1,245/10,000 = 0.1245, while occupancy is 0.11.

The correctly pooled detection probability is 960/1,100, approximately 0.8727. It is weighted by occupied sites, not by the equal sizes of the two environments. The pooled false-positive probability is 285/8,900, approximately 0.03202, weighted by unoccupied sites.

Simply averaging s to 0.75 and a to 0.03 would discard those conditional denominators. Substituting those averages with ψ = 0.11 predicts q = 0.1092, which differs from the actual constructed mixture’s 0.1245.

The probability identity has not failed. The unweighted averages did not represent the conditional quantities required by it. This is a useful extension because a large pooled validation summary can look authoritative while concealing how species prevalence differs across environments. Preserve subgroup counts when they matter to the inference.

Contents · Previous chapter · Next chapter

A positive-label threshold can be used to select records for further review rather than to make a final biological declaration. Those jobs need different performance descriptions. A broad candidate queue may tolerate more false positives if an appropriate confirmation stage follows; a final published species claim needs evidence suitable for that stronger conclusion.

In the rare-target example, the expected positive queue contains eighty occupied sites and 198 unoccupied sites. Reviewing a simple random hundred from that queue would yield approximately 100 × 80/278 = 28.78 occupied-site records in expectation under the constructed model. The actual reviewed number can vary.

This expectation describes review yield, not the number of animals discovered or a guarantee about any particular hundred records. A targeted review order can produce a different yield if it uses additional informative evidence, but that evidence and selection rule must be described.

Do not choose a threshold solely because it maximises one familiar score without identifying the intended decision. A threshold that reduces false positives may also miss relevant records. A different rule may be appropriate for screening, confirmation or estimating a population quantity.

For a classroom task, ask learners to propose one metric for the candidate queue and another for the final accepted record set. They should explain why the two stages need different denominators and validation evidence. This keeps the analysis connected with a useful action while preserving the limits of the supplied probabilities.

Contents · Previous chapter · Next chapter

A strong conclusion separates the observed label fraction from the biological inference. For the main construction, write: “The stipulated survey process produces an expected positive fraction of 9.8%. With detection probability 0.80 and false-positive probability 0.02, the exact mixture corresponds to occupancy 10%.”

If the error probabilities have not been validated for the population, stop at the observed fraction and explain what is missing. If they are bounded, report the compatible range with those bounds. If the validation sample was stratified, retain the weights that connect it to the target population.

This approach leaves room for useful uncertainty. A reviewed positive can strengthen a species record, while a population estimate still requires an account of missed and false detections across the wider survey. The report becomes clearer when it names the inference step rather than presenting every output as the same kind of fact.

Contents · Previous chapter · Next chapter

Does a positive label prove occupancy? Not automatically. Its interpretation depends on the observation method, error probabilities, validation and the site-state definition.

Does a rare target make every positive false? No. It changes the balance of possible true and false positives in the model. Appropriate confirmation can provide stronger evidence for particular records.

Can a negative label prove absence? Not generally when detection is imperfect. Use the all-zero history guide for that separate question.

Can these calculations estimate animal abundance? Not from site labels alone. Occupancy, detections and individuals are different quantities. The encounter-probability guide develops another part of that evidence chain.

For foundations, return to Zoology. For primary scientific context, read USGS: modeling false positives and acoustic-survey confirmation design. Continue through The Gold Standard of Education and How X Works with the next useful question: what does this label establish, and which part still needs independent evidence?

Contents · Previous chapter · Return to Zoology