Two animal surveys record many of the same identifiable individuals. A simple capture–recapture calculation may interpret that overlap as evidence that the population is fairly small. But what if the repeatedly observed animals were also the easiest animals to encounter?
The short answer is that overlap depends on the observation process as well as population size. When some individuals are consistently easier to encounter, a pooled two-occasion estimate can miss much of the less observable group. A large overlap is not automatically a representative overlap.
The specialist question here is how a stable observation advantage differs from one that switches between occasions, and why their overlap can push the same simple calculation in opposite directions.
This guide extends The Gold Standard of Zoology through original fictional encounter histories and probability models. It teaches interpretation rather than a field protocol. It provides no instructions for capturing, marking, handling or disturbing animals; actual animal research requires appropriate expertise and authorisation.
eduKateSG · Zoology
Find your next learning step
Compare stable and switching encounter probabilities through original population tables, overlap covariance and explicit limits on capture–recapture inference.
Chapter contents
Understand the question · Chapters 1–4
Work the first cases · Chapters 5–8
Test the boundaries · Chapters 9–13
Transfer and repair · Chapters 14–18
Report and continue · Chapters 19–23
CHAPTER 1 OF 23 · Understand the question
1. Define the population before estimating it
A population estimate needs a target: which animals, in which area and over which period? A method cannot supply a meaningful total if the membership boundary changes without being represented in the model.
The simplest teaching model here assumes a closed population during two encounter occasions. Closure means no births, deaths, immigration or emigration affecting membership over that defined comparison. It is an assumption for the example, not a claim that real animal populations are permanently fixed.
Individuals must also be recognisable across occasions under the stated method. The word encounter can represent an appropriately identified observation, not necessarily physical capture. A collection of photographs without reliable individual identities does not automatically supply the overlap count this method requires.
Begin with the intended claim in plain language: “Estimate the number of members of this defined closed population.” Keep that distinct from counting files, observation events, calls or sites occupied by a species.
The zoology root and its Zoologists guide explain the wider evidence and professional context. Here we isolate what unequal encounter probabilities can do even after the population boundary and individual identities have been stipulated.
CHAPTER 2 OF 23 · Understand the question
2. Distinguish three counts in the simple estimator
Let M be the number of distinct individuals encountered on occasion one. Let C be the number encountered on occasion two. Let R be the number encountered on both occasions.
The familiar simple estimate is Nhat = MC/R when R is positive and the relevant assumptions justify the model. It is often called the Lincoln–Petersen estimate. Its numerical value need not be an integer even though the actual population size is an integer.
Suppose M = 30, C = 48 and R = 12. The estimate is 30 × 48/12 = 120. The calculation does not mean 120 animals were observed. The observed number of distinct individuals is M + C − R = 66.
The difference between those quantities matters. One is a model-based population estimate; the other is a count of identified individuals in the combined record. Treating the estimate as a direct census would misstate the evidence.
The USGS Mark-Recapture Science programme provides the wider research setting for modelling encounter processes. The small datasets below are original teaching constructions, not reported findings from a real population.
CHAPTER 3 OF 23 · Understand the question
3. Reconstruct the ratio argument behind the formula
In an idealised well-mixed situation with suitable equal encounter probabilities, the proportion previously encountered among the second occasion’s individuals can represent the proportion previously encountered in the whole population.
That suggests R/C ≈ M/N. Rearranging gives N ≈ MC/R. The approximation is statistical, not an exact identity for every finite sample. Sampling variation can make the observed overlap differ from its expected value.
The crucial question is why R/C should represent M/N. It is not enough that the two fractions can be written down. The observation process and model assumptions must connect them.
If the first occasion mainly includes easy-to-encounter animals, and those same animals are especially likely to appear again, the second occasion may contain an unusually large proportion of previously encountered individuals. Substituting that proportion as though it represented the population can make the estimated total too small.
This is a representation problem, not an arithmetic error. Multiplying M and C and dividing by R can be performed perfectly while the interpretation fails. A good explanation therefore gives both the formula and the reason its overlap fraction is, or is not, appropriate for the intended population claim.
CHAPTER 4 OF 23 · Understand the question
4. Keep encounter histories separate from the unseen group
With two occasions, an individual has one of four binary histories: 11 means encountered on both, 10 only on the first, 01 only on the second and 00 on neither.
The observed data contain the first three categories. The fourth is the category the population inference must account for, despite its members not appearing in the encounter record.
For M = 30, C = 48 and R = 12, there are twelve 11 histories, eighteen 10 histories and thirty-six 01 histories. Their sum is sixty-six observed distinct individuals. The number of 00 histories remains unobserved.
A model-based estimate of 120 implies a corresponding unseen total of 54 in this arithmetic description. That does not mean a hidden list of 120 identified animals was found. It is the model’s implication about the missing category.
Writing the four-category table is useful because it shows exactly where inference begins. The observed categories can be counted and checked. The unseen category depends on assumptions linking the observation process to population membership. A software output should not erase that distinction simply because all four categories appear together in a completed table.
CHAPTER 5 OF 23 · Work the first cases
5. Work a closed population with two observation groups
Imagine a fictional population of one hundred animals divided into two groups of fifty. Their group labels describe observation probabilities, not species, health, intelligence or value. For teaching, the complete population is known so we can inspect what the estimator misses.
Group H has thirty-two 11 histories, eight 10 histories, eight 01 histories and two 00 histories. Group L has two 11 histories, eight 10 histories, eight 01 histories and thirty-two 00 histories.
| Fictional group | Both occasions, 11 | First only, 10 | Second only, 01 | Neither, 00 | True total |
|---|---|---|---|---|---|
| H | 32 | 8 | 8 | 2 | 50 |
| L | 2 | 8 | 8 | 32 | 50 |
| Combined | 34 | 16 | 16 | 34 | 100 |
The pooled first count is 50, the pooled second count is 50 and the overlap is 34. The simple estimate is 2500/34, approximately 73.53, while the stipulated true population is 100.
The observed distinct total is 34 + 16 + 16 = 66. Most of the unseen individuals belong to Group L. The dataset demonstrates a possible underestimate under structured encounter differences; one constructed dataset alone is not a proof of every finite-sample bias property of the estimator.
CHAPTER 6 OF 23 · Work the first cases
6. See why the encountered sample is unrepresentative
The true population contains equal numbers in H and L. But on the first occasion, forty H individuals and only ten L individuals are encountered. The encountered sample is therefore 80% H even though the population is only 50% H.
On the second occasion, the same group imbalance occurs. Among the previously encountered H individuals, thirty-two appear again. Among the previously encountered L individuals, only two appear again. The overlap is strongly concentrated in the easier group.
The pooled overlap fraction is 34/50 = 68% of the second occasion’s sample. The fraction of the true population encountered on the first occasion is 50/100 = 50%. Those are not equal in this fictional record.
Replacing the unknown population fraction with the larger overlap fraction makes the simple estimate smaller. The numerical mechanism can be seen directly in Nhat = M/(R/C).
Do not describe Group L as absent or irrelevant merely because it contributes few records. That is exactly the inference problem. A dataset can contain many high-quality repeated identities from an accessible subset while saying much less about other members of the target population. Repetition is useful evidence, but its representation must be examined.
CHAPTER 7 OF 23 · Work the first cases
7. Build a probability model that explains the pattern
Suppose each H individual has probability 0.8 of encounter on each occasion, and each L individual probability 0.2. Assume occasions are independent conditional on the individual’s group and that individuals’ outcomes follow the stated model.
For H, the probabilities of 11, 10, 01 and 00 are 0.64, 0.16, 0.16 and 0.04. Multiplying by fifty gives expected category counts 32, 8, 8 and 2. For L, the expected counts reverse to 2, 8, 8 and 32.
Thus the earlier table equals this model’s expected category counts, while also being a possible integer-valued realised table. A real random sample need not equal those expectations exactly.
Conditional independence does not imply that all individuals share one common encounter probability. It also does not imply independence after the group information is discarded. Group membership creates a relationship between being encountered once and being encountered again.
The distinction is worth stating carefully. A model can contain no direct behavioural effect of the first encounter on the second and still produce excess pooled overlap because the same individuals are consistently more observable. That mechanism differs from an animal changing its behaviour after an encounter, even though both can affect a simple pooled estimate.
CHAPTER 8 OF 23 · Work the first cases
8. Derive the ratio-of-expectations diagnostic
Let p vary across individuals but remain the same for each individual across the two occasions. In the ideal independent-occasion model, the expected first and second totals are N times the mean of p. The expected overlap is N times the mean of p².
Substituting those expected counts into MC/R gives the diagnostic expression N × mean(p)²/mean(p²). Because mean(p²) equals mean(p)² plus the variance of p, this expression is no greater than N when the probabilities have positive variation.
In the two-group example, mean(p) = 0.5 and mean(p²) = 0.34. The diagnostic is 100 × 0.25/0.34, approximately 73.53, matching the table calculation.
This is explicitly a ratio of expected counts. It is not a claim that the expectation of the random estimator MC/R equals that ratio in every finite sample. Random ratios, zero-overlap outcomes and finite-sample corrections require separate treatment.
The diagnostic explains a mechanism: consistently high-probability individuals contribute disproportionately to overlap because their probabilities enter as squares. It is a useful reasoning tool when kept at that level. A complete statistical analysis must use an appropriate estimator and uncertainty treatment for the actual study design.
CHAPTER 9 OF 23 · Test the boundaries
9. Explain why more repeats do not automatically repair the model
Seeing the same easily encountered animals more often can improve information about those animals. It does not automatically establish how many difficult-to-encounter individuals remain unseen.
In the fictional table, Group H supplies most repeated histories. Increasing observations in conditions that continue to favour H may generate an even richer record of that group while leaving the target population’s less observable portion weakly represented.
Additional occasions can be valuable because their encounter histories contain more information for suitable models. But “more data” is not a complete answer unless the observation design and model address the relevant heterogeneity.
The primary research paper Capture–Recapture Abundance Estimation Using a Semi-Complete Data Likelihood Approach examines models that explicitly handle unobserved individual heterogeneity. Its methods are substantially more advanced than the classroom diagnostic developed here.
For a learner, the practical question is what the added observations distinguish. Do they help separate population size from differences in encounter probability, or do they mostly repeat an already accessible subset? A record can become longer and more precise about its observed members without making every broader population claim equally well supported.
CHAPTER 10 OF 23 · Test the boundaries
10. Compare pooled and group-specific calculations
If the two groups and their encounter histories were validly known, the simple calculations could be examined separately. For H, M = 40, C = 40 and R = 32, giving 40 × 40/32 = 50. For L, M = 10, C = 10 and R = 2, giving 10 × 10/2 = 50.
The sum is 100, matching the stipulated truth in this constructed table. The pooled estimate was approximately 73.53. Pooling concealed the differences in observation process that the group-specific arithmetic retained.
This result does not prove that stratification always solves the problem. Real group membership may be unknown for unseen individuals, groups may retain internal variation, and small overlaps can make separate estimates unstable.
The expected-count construction was chosen to make the mechanism exact. In realised random data, separate simple estimators do not necessarily reproduce the true group totals. A proper analysis must consider sampling variation and the conditions supporting each group model.
The useful lesson is narrower: a single pooled overlap fraction can hide structure relevant to abundance inference. Before combining groups, ask whether their observation probabilities and study roles justify the pooling. Do not add a group label after seeing the results merely to force agreement with a preferred population size.
CHAPTER 11 OF 23 · Test the boundaries
11. Do not infer the hidden groups from observed frequency alone
An analyst cannot generally reconstruct all unseen low-probability individuals simply by labelling repeatedly encountered animals easy and rarely encountered animals difficult. Those labels are themselves affected by random sampling and by the fact that unseen individuals have no observed histories.
In a two-occasion record, an individual with history 10 might have a low encounter probability, or it might be a high-probability individual that happened to be missed once. One history does not reveal its exact p.
Similarly, the complete absence of an individual from the record does not reveal whether it belongs to a large low-probability group or a smaller group with different conditions. The unseen category is precisely where model assumptions matter.
Additional covariates, design information or richer encounter histories can support more appropriate models when their relevance and measurement are established. But a flexible model is not automatically correct merely because it can reproduce the observed table.
For a classroom exercise, ask students to describe two different hidden populations compatible with the same observed histories. The task demonstrates that fitting the observed categories and identifying the unseen total are different achievements. A method needs assumptions or further information that connect those two parts of the problem.
CHAPTER 12 OF 23 · Test the boundaries
12. Show how a different probability pattern can reverse the direction
Consistent easy-versus-difficult probabilities are one structure, not every possible structure. Consider two equal groups of fifty. Group A has encounter probabilities 0.8 on the first occasion and 0.2 on the second; Group B has 0.2 first and 0.8 second.
The expected first and second totals are again fifty each. Each group has expected overlap fifty × 0.8 × 0.2 = 8, so pooled expected overlap is sixteen. Substituting those counts gives 2500/16 = 156.25, above the true total of one hundred.
The observation advantage switched between occasions, reducing overlap. In the corresponding expected-category table, there are sixteen 11 histories, thirty-four 10 histories, thirty-four 01 histories and sixteen 00 histories.
This example does not contradict the previous diagnostic. That diagnostic assumed an individual’s probability stayed the same across occasions. Here the first- and second-occasion probabilities differ systematically.
The broader lesson is to avoid the rule “unequal detectability always causes an underestimate.” The direction depends on the relationship between encounter probabilities across occasions and on other model features. State the pattern being considered and derive its implication rather than turning one important mechanism into a universal slogan.
CHAPTER 13 OF 23 · Test the boundaries
13. Use covariance to compare the two patterns
Let p1 and p2 be an individual’s encounter probabilities on the first and second occasions. Under the conditional independent-occasion model, the ratio-of-expected-counts diagnostic is N × mean(p1)mean(p2)/mean(p1p2).
The denominator can be written mean(p1)mean(p2) plus the covariance between p1 and p2 across individuals. Positive covariance increases the overlap relative to the product of the means, pushing this diagnostic below N. Negative covariance has the opposite effect.
For the stable high/low example, the means are 0.5 and 0.5, mean(p1p2) = 0.34 and covariance is 0.09. For the switching example, mean(p1p2) = 0.16 and covariance is −0.09.
Those values explain why the two diagnostic results are 73.53 and 156.25 despite identical expected first and second totals. The missing information lies in who is likely to appear on both occasions.
Again, this is a model diagnostic based on expected counts, not a complete finite-sample theorem for every estimator. Its value is explanatory: marginal totals alone cannot reveal the joint observation structure. The same M and C can accompany very different overlaps and therefore very different simple population estimates.
CHAPTER 14 OF 23 · Transfer and repair
14. Distinguish heterogeneity from behavioural response
Stable heterogeneity means individuals differ in their encounter probabilities even without the first occasion changing their subsequent behaviour. A behavioural response means the first encounter itself changes the chance of a later encounter.
Both can affect overlap, but they are different explanations. An individual may be repeatedly observed because its habitat use makes it consistently accessible. Another observation process might change after an encounter, creating a history-dependent probability.
The observed two-occasion table alone may not identify which mechanism produced the overlap. A suitable study design and model must distinguish the explanations relevant to the actual setting.
Do not attach a psychological label such as clever, shy or cooperative merely because an animal was encountered repeatedly or not at all. Those descriptions require their own behavioural evidence and can conceal the actual observation conditions.
For students, write two causal stories that could produce a high overlap: one with stable individual differences and one with an encounter-related change. Then list what additional evidence would distinguish them. This makes clear that a successful numerical fit is not automatically a complete biological explanation. The mathematics identifies a dependence worth understanding; it does not supply the animal’s motive.
CHAPTER 15 OF 23 · Transfer and repair
15. Check identity errors and changing membership separately
The worked cases assume correct individual identities and a closed population. If identities are mismatched, the overlap count itself can be wrong. If membership changes, an absent individual may have left the target population rather than merely escaped observation.
A false match can inflate R, tending to reduce MC/R when M and C are held fixed. A missed match can reduce R, tending to increase it. Those are arithmetic sensitivities under the stated fixed-count comparison, not complete predictions for every real identification error process.
Likewise, births, deaths or movement across the boundary may invalidate the simple closed-population interpretation. Different models address different population dynamics; they cannot be selected solely because one produces a convenient total.
Keep these issues distinct from heterogeneity. Correcting encounter-probability differences does not automatically correct identity errors or changing population membership. Conversely, perfect identities do not guarantee representative overlap.
A useful evidence checklist has separate lines for population definition, closure, identity quality, observation probabilities and uncertainty. Combining every concern under the word bias makes it harder to identify the additional information needed. Specific model assumptions lead to specific checks and clearer limits on the resulting claim.
CHAPTER 16 OF 23 · Transfer and repair
16. Examine the zero-overlap case without inventing infinity
If R = 0, the simple expression MC/R is undefined. A spreadsheet error or a displayed infinite value is not a direct census result and does not establish an infinitely large population.
Zero overlap can arise under several circumstances, including limited observation, changing membership, incompatible identities or an observation process that reaches different subsets. The appropriate interpretation depends on the study design and model.
For a teaching example with M = 10 and C = 12, the combined observed distinct count is twenty-two when identities confirm no overlap. That observed count remains meaningful. The simple estimator’s denominator does not support a finite value through this formula.
Do not repair the problem by silently replacing zero with one. That changes the data and can produce a dramatically different estimate without justification. Finite-sample methods or alternative models need to be chosen and explained on their actual statistical basis.
The learner’s useful next step is to identify which conclusion remains directly supported: at least the observed distinct individuals under the stated identity and membership assumptions. Any broader population inference requires an appropriate method and uncertainty treatment, not an improvised arithmetic patch.
CHAPTER 17 OF 23 · Transfer and repair
17. Keep precision separate from model suitability
An estimate reported as 73.52941176 may look more convincing than approximately seventy-four. The extra decimal places do not establish that the observation model is suitable or that the population lies close to that number.
The fictional hundred-animal example deliberately produces a simple estimate far below the stipulated truth. Calculating it with greater numerical precision does not repair its representation problem.
Uncertainty also has several sources: random encounter outcomes, estimated observation parameters, identity quality, boundary assumptions and model choice. A standard error conditional on one model does not automatically account for every structural uncertainty outside that model.
When reading a study, ask what its interval or uncertainty summary includes. Do not invent a confidence level for a range that was constructed only from alternative assumptions. Likewise, do not treat one constructed counterexample as a confidence interval for a real population.
The distinction is constructive. Good statistical work makes its model and assumptions visible so that uncertainty can be interpreted appropriately. Students can practise by labelling a number as an observed count, an estimate conditional on a model or a diagnostic calculation from expected counts. The label is part of its meaning.
Imagine two fictional groups of one hundred animals. Group A has encounter probability 0.6 on each occasion, while Group B has probability 0.3. Assume independent occasions conditional on group and use expected category counts for this diagnostic exercise.
The expected first and second totals are sixty plus thirty, or ninety. Expected overlap is one hundred × 0.6² plus one hundred × 0.3² = 36 + 9 = 45.
The pooled ratio-of-expected-counts result is 90 × 90/45 = 180, while the stipulated population is 200. The expected observed distinct total is 90 + 90 − 45 = 135, leaving sixty-five expected unseen individuals.
Within the groups, the same diagnostic calculations give 60 × 60/36 = 100 and 30 × 30/9 = 100. Their sum recovers the stipulated total because the construction supplies the group structure and expected counts exactly.
Explain why this does not mean every random realised pair of samples would produce those answers. Expected counts summarise a probability model; actual histories vary. A complete answer should state which calculations use the synthetic expectation table and which conclusions concern an observed dataset. That distinction prevents a useful teaching model from becoming a false promise about estimator performance.
CHAPTER 19 OF 23 · Report and continue
19. Read a population claim through five checks
First identify the target population and period. Second identify what constitutes a reliable individual identity. Third separate observed distinct counts from inferred unseen counts. Fourth ask how encounter probabilities and their dependence across occasions were modelled. Fifth identify what the reported uncertainty actually covers.
Apply those checks to the sentence “Most of the second sample had been seen before, so almost every animal must already be known.” The overlap may be concentrated in a repeatedly accessible subset. Without a suitable observation model, the sentence extends beyond the record.
Apply them again to “The model included heterogeneity, so the estimate is correct.” Including a relevant mechanism is useful, but the model still needs appropriate data, assumptions, fitting and evaluation. A model name is not an independent guarantee of truth.
For a classroom discussion, ask which additional evidence would be most informative for the actual unresolved question. It might concern identity quality, sampling coverage, covariates, more encounter occasions or population movement. The answer should follow the specific problem rather than default to collecting more of the same observations.
CHAPTER 20 OF 23 · Report and continue
20. Use observed counts as a lower-bound check
Under the stipulated correct identities and closed membership, the combined observed distinct count M + C − R cannot exceed the true population. It is a direct bookkeeping lower bound, even when a particular population model is uncertain.
For the high/low example, that count is sixty-six. A claimed population estimate of fifty would conflict with the record before any sophisticated encounter model was considered. The record already contains more than fifty identified members.
This check does not turn sixty-six into the best estimate of the full population. It simply identifies a constraint every compatible population total must respect under the stated identity and membership assumptions.
The assumptions matter here too. Duplicate identities or a changing membership boundary can alter what the combined record represents. Do not preserve a numerical lower bound while discarding the conditions that made its interpretation valid.
For a useful student task, reconstruct the observed union from the three visible history categories, then compare it with any proposed total. Basic set bookkeeping is often the first and clearest way to catch an impossible inference.
CHAPTER 21 OF 23 · Report and continue
21. Explain a model result without turning it into a census
A concise report can separate three statements. First, state how many distinct individuals were identified. Second, state the estimate produced by the specified model. Third, state the observation-process assumptions and uncertainty relevant to extending from the first number to the second.
In the constructed high/low dataset, the report would distinguish sixty-six observed individuals, the simple pooled estimate of approximately 73.53, and the stipulated true total of one hundred available only because this is a teaching construction. A real analyst would not have that final truth supplied by the exercise.
This distinction is especially useful when evaluating a simulation. Knowing the simulated population lets a researcher assess model behaviour. It does not mean the same truth is directly observable in a field study.
Ask which number came from counting and which came from assumptions about what was missed. If the reader can answer that question, the report is preserving the evidence chain. If all numbers are described as animals counted, a central part of the inference has been lost.
CHAPTER 22 OF 23 · Report and continue
22. Questions about recapture probabilities and abundance
Does high overlap always mean a small population?
No. Its interpretation depends on the observation process. Consistently accessible individuals can create high overlap while a less observable group remains poorly represented.
Does unequal encounter probability always cause underestimation?
No. The stable high/low diagnostic and the switching-probability diagnostic produce different directions. The relationship between probabilities across occasions matters.
Is the ratio of expected counts the expected estimator?
Not generally. Expectation does not commute with division. The diagnostic explains a mechanism; it is not a substitute for finite-sample statistical analysis.
Can the article be used as an animal-handling protocol?
No. It supplies fictional records for reasoning. Real animal studies require appropriate expertise, welfare safeguards, authorisation and study-specific methods.
CHAPTER 23 OF 23 · Report and continue
23. Continue through animal observation and statistical modelling
Return to Zoology and the Zoologists guide for the wider research and evidence roles. The existing all-zero survey-history guide addresses a different question about non-detection and presence.
For the broad abundance and field-method setting, read Counting Wild Animals Without Seeing Every Animal. This specialist article adds the explicit stable-versus-switching probability comparison, covariance diagnostic and distinction between a ratio of expectations and an expected estimator. For the basic abundance setting, read Biodiversity Surveys, Species Richness and Capture–Recapture. For primary research context, consult USGS Mark-Recapture Science, King and colleagues’ semi-complete likelihood approach and Schofield and Barker’s closed-population heterogeneity modelling. Their methods are not replaced by the simple diagnostic examples here.
The wider route is Learning That Transfers into Life within How X Works. Reconstruct one four-history table and ask which part was observed and which part was inferred. That question is a concrete next step toward a population claim whose meaning can be checked.