Three Per Cent or Ninety-One: What a Zero Actually Means, and Why Nobody Counts the Surveys They Did Not Do

1. Introduction: the number everybody misreads

This journal has spent several articles on how well detection methods find things. This one is about what happens when they do not, which is the more common result and the less examined one.

1.1 Why it matters here

A negative result is what clears a unit for letting, ends a treatment programme, closes a file and satisfies an auditor. If it carries less information than it appears to, every one of those decisions is weaker than it looks.

2. The statement the literature keeps making

The problem, as stated across fields.

Organisms are notoriously difficult to detect with certainty, and detectability varies strongly across space, seasons, habitat, sampling methods and species traits, making it difficult to distinguish true absences from non-detections.⁶

Explicitly modelling the probability of a true-absence versus false-absence reduces the bias that can result from assuming perfect detection.⁹

2.1 The phrase worth keeping

True absence against false absence. Those are the two things a zero can be, and ordinary practice treats them as one.

3. The two quantities

The conceptual move that the whole framework rests on.

Occupancy modelling works by explicitly estimating both the probability that a species occupies a site and the probability of detecting it when present, which allows a move beyond raw presence-absence data and reduces the bias introduced by imperfect detection.⁶

The two parameters are conventionally written as occupancy and detection probability.⁷

3.1 An everyday version of the idea

A smoke alarm that never sounds is consistent with no fires and with a dead battery, and the only way to tell is to test it against a known fire.

Monitoring devices are in the same position permanently, because no fire is ever deliberately set. The occupancy framework is a way of manufacturing the test from repetition instead.

3.2 Why separating them changes everything

Raw presence-absence data confounds two things: whether the organism is there, and whether your method would have found it if it were. A single survey cannot distinguish them, because both produce the same observation.

Separating them turns an unanswerable question into two answerable ones, provided the design supplies the right data.

4. How detection probability is estimated

The design requirement, which is the practical heart of this.

Repeated surveys of the same site produce a detection history. Where an organism is detected on some occasions and not others, we assume the site was occupied across all five sampling occasions, but not detected on sampling occasions 1, 3 and 5.⁷

4.1 The closure assumption

The inference in that sentence rests on something: that the site's true status did not change between occasions. If the organism arrived after occasion 1 and left before occasion 5, the misses are not misses.

That is the closure assumption, and the design has to make it plausible by keeping the occasions close enough together that the status is stable and far enough apart that the observations are independent. One camera study handled this by using two-week observation periods specifically to avoid autocorrelation between detection only a day apart and problems associated with zero-inflation.⁹

4.2 The logic

A site where something was found once is known to be occupied. The occasions on which it was not found at that same site are therefore known misses, and the proportion of misses estimates the detection probability.

That estimate then applies to sites where nothing was ever found, which is how a survey learns how much weight its own zeros deserve.

4.3 What this demands

Repetition. A single visit cannot produce a detection history, and therefore cannot support any statement about what a negative means.

That is the requirement most incompatible with how this trade operates, and §21 returns to it.

It also explains why the vector study needed three years. Detection probability is estimated from the pattern of hits and misses across occasions, so the number of occasions is the sample size for that parameter, and a handful of visits estimates it very poorly.

5. What varies detectability

The list, which is longer than intuition suggests.

Space, seasons, habitat, sampling methods and species traits.⁶ One study included day of year at sampling as a proxy for seasonal variation.⁶

5.1 The one that is easiest to forget

Sampling method. Detectability is not a fixed property of the organism that different devices measure more or less well; it is jointly produced by the organism and the device.

This journal's detection methods article found a passive interceptor over seven days matching an active trap over one night. Those are two detection probabilities for the same insect in the same room, and neither is the detectability of bed bugs.

5.2 The consequence

Detection probability is not a property of a device. It is a property of a device, in a place, at a time, against a species, in the hands of somebody.

This journal found the same structure in canine detection, where teams varied between themselves and between days, and in automated identification, where the colour of the card changed model accuracy. Detectability being context-dependent is the general case rather than the exception.

6. The vector surveillance case

The study that demonstrates the cost of ignoring this, which concerns mosquitoes rather than structural pests but makes the point better than anything we found in our own field.

Researchers used a modeling approach that explicitly accounts for imperfect detection and a 38-month, 55-site detection and non-detection dataset to quantify the effects of control interventions on site-occupancy dynamics, considering meteorological and dwelling-level covariates.¹

6.1 Why a mosquito study belongs in a structural pest journal

Because vector surveillance is the one field that has run this comparison at scale and published the answer. The organism differs; the logical structure does not.

A programme that inspects dwellings, records findings, targets intervention at the positives and evaluates itself on its own detections is the same machine whatever it is looking for.

6.2 The scale of the dataset

Three years, fifty-five sites, monthly. That is the repetition §4.2 requires, sustained long enough to estimate detection reliably.

7. The three numbers

What each method reported for the same quantity.

Site-occupancy estimates for one species were mean 0.91, range 0.79 to 0.97, which were much higher than reported by routine surveillance based on rapid larval surveys, at 0.03, range 0.02 to 0.11, and moderately higher than directly ascertained with oviposition traps, at 0.68, range 0.50 to 0.91.¹

7.1 Reading that gap

The operational surveillance method in routine use reported three dwellings in a hundred. The modelled estimate was ninety-one.

Even the better direct method, oviposition trapping, was at sixty-eight, so the gap between the best measurement and the estimate is real but modest. The gap between the routine programme and the estimate is a different order of thing.

7.2 What the ranges tell you

The model's range of 0.79 to 0.97 does not overlap the larval survey's 0.02 to 0.11 at any point.¹ These are not two noisy measurements of the same thing that happen to differ; they are incompatible.

The oviposition trap range of 0.50 to 0.91 does overlap the model's, which is what you would expect from a method that works but misses some.

7.3 Why the routine method fails so badly

Rapid larval surveys look for larvae in containers during a visit. That requires the surveyor to find the container, at a time when it holds detectable larvae, in a dwelling they have access to.

The failure modes compound, and none of them are visible in the output, which reports a clean number. That explanation is ours; the paper reports the discrepancy rather than decomposing it.

8. The finding that follows from them

What the authors conclude.

The marked contrast between the estimates of adult vector presence and the results from rapid larval surveys suggests, together with the lack of effect of local control campaigns, that many breeding sites were overlooked by vector control agents in our study setting, and that better sampling strategies are urgently needed, particularly for the reliable assessment of infestation rates in the context of control program management.¹

8.1 The phrase urgently needed

Researchers do not usually write that way. It appears here because the gap between what the programme believed and what the data showed was large enough to mean the programme was not doing what it thought it was doing.¹

8.2 The inference being drawn

The surveillance was not merely imprecise. It was systematically missing the thing it existed to find, and the programme acting on it was therefore aimed at a small fraction of the problem.

9. Why the control programme achieved nothing

The second finding, which is the one that should concern anyone running a programme.

Regular control campaigns based on breeding-site elimination had no measurable effects on the probabilities of dwelling infestation by dengue vectors.¹

9.1 The causal chain

If surveillance identifies three per cent of the infested sites, then elimination visits go to three per cent of the places that need them, and the other ninety-odd per cent continue undisturbed.

A programme cannot outperform its own detection. The intervention was not necessarily ineffective; it was aimed by an instrument that could not see most of the target.

9.2 The alternative explanation we should allow

Fairness requires it. Breeding-site elimination might be ineffective on its own terms, with sites refilling or being replaced faster than they are removed, regardless of how well they were found.

The study does not separate those two explanations, and the authors' reading is that the sites were overlooked.¹ We follow them, and note that an intervention that reached most sites and still failed would look identical in this data.

9.3 What this journal has said before

The rat counting article argued that a control programme without a measurement cannot know whether it worked. This is the stronger version: a programme with a bad measurement will conclude that it worked while the population is unchanged, because the same instrument reports both.

9.4 The covariates that did not explain it

Site-occupancy fluctuated seasonally, driven mainly by the negative effects of high maximum and minimum summer temperatures, while rainfall and dwelling-level covariates were poor predictors of occupancy.¹

10. The assumption being challenged

Stated plainly by the authors, and it is the assumption this whole trade operates on.

To date, studies have assumed that the vectors are truly absent from sites where they are not detected; since no perfect detection method exists, this assumption is questionable.¹

10.1 Why it took so long to be stated

Because a surveillance programme that assumes perfect detection produces coherent, internally consistent output. The numbers add up, trends can be plotted, and nothing in the data announces that the instrument is wrong.

Detecting the problem requires a second, better method to compare against, which is expensive and which nobody commissions while the first method appears to be working.

10.2 The strength of that objection

It does not require any particular method to be bad. It requires only that no method is perfect, which is not contested by anybody.

From that alone it follows that non-detection is evidence of absence rather than proof of it, and the strength of the evidence depends on a quantity nobody usually measures.

11. The effort question

How much sampling is enough, with a worked example from a different taxon.

An acoustic survey identified 351,771 passes of eight species or species groups in 5,856 detector-nights. On average approximately 30 sampling nights were needed to detect 90 per cent of the total species richness among locations and seasons, while relatively few nights, twelve or fewer, were needed to detect most species during summer.³

11.1 The denominator worth noticing

5,856 detector-nights to produce that guidance.³ Establishing how much effort is required is itself an enormous effort, which is why so few fields have the answer.

11.2 Eight species, thirty nights

Worth registering that the thirty-night figure is for detecting ninety per cent of richness among only eight species or species groups.³ A short list, and still a month of nightly sampling.

11.3 The shape of the curve

Most of what is present turns up quickly. The last portion takes disproportionately longer, which is the general form of a detection accumulation curve.

That shape is why short surveys feel adequate. They do find most of what is there, and the question is whether most is the standard.

12. The rare species problem

The part of the curve that matters for pest work.

Many more nights were needed to detect acoustically rare species, and the authors report that the sampling effort required to determine presence or probable absence of two federally endangered species was substantially greater.³

12.1 The asymmetry in the word rare

Rare in the ecological sense means uncommon across the landscape. Rare in the pest sense means few individuals in one building.

The detection consequence is the same either way, because what drives detection probability is how often an individual encounters the device, and that falls with density regardless of why the density is low.

12.2 Why this transfers directly

An early infestation is a rare species. A handful of individuals in a large structure is precisely the low-density case where detection probability per occasion is smallest and required effort is greatest.

Which is also the case where finding it matters most, because this journal has argued repeatedly that early intervention is cheaper. The effort requirement and the value of detection both peak in the same place, and the willingness to pay for either is lowest exactly then, because nobody has seen anything yet.

13. Declaring something free of a pest

The formal version of what a clearance inspection claims to do.

There is a critical need for more direction and guidance related to how many samples are enough to declare a unit of interest free of an organism, and work has been done to provide guidance on sample sizes required to be 95 per cent certain a target organism is absent from a site.²

13.1 Why this is the hardest claim to make

Establishing presence requires one positive. Establishing absence requires ruling out every place and moment the organism could have been, which no finite survey does.

The asymmetry is fundamental rather than a limitation of current methods, and it is why the literature converts the question into one about confidence at a given effort.

13.2 The form of the answer

Notice what the achievable claim looks like: ninety-five per cent certain, given a stated sampling effort. Not absent.

Certainty about absence is not available at any effort. What a survey can deliver is a confidence level attached to a stated design, and nothing stronger. A clearance statement that does not say how much looking was done has not made a claim that can be evaluated, because the same words would follow from a thorough search and from a cursory one.

13.3 The question this raises about our own reports

A pest control report saying no evidence of activity observed is accurate and nearly uninformative. It records an outcome without the effort that produced it, and §20 is about what would fix that.

14. The false positive complication

The literature has moved beyond misses alone.

Recent frameworks accommodate false negative, false positive and uncertain detections, and researchers found that the presence of uncertain detections increased the variability of resulting estimates.² Related methodological advances address incorporating false positives into the estimation and dealing with heterogeneity in detection probabilities.⁷

14.1 Why false positives matter for a programme

A false negative leaves an infestation untreated. A false positive produces treatment where none was needed, which this journal's credence goods article classified as overtreatment.

Both are errors in a monitoring system and only one of them is usually discussed, because only one of them generates a complaint.

14.2 Why uncertain detections are their own category

A result that might be the target is neither a detection nor a non-detection, and discarding it throws away information while counting it introduces error.

This journal's canine detection article described exactly this situation: an alert with nothing visible is an uncertain detection, and the industry resolves it by declaring it a hidden true positive. The statistical literature treats it as a third category requiring explicit handling.

15. The pheromone trap question

The version of this problem closest to ordinary practice.

When pheromone traps are used for detection of an invasive pest and then delimitation of its distribution, an unresolved issue is the interpretation of failure to capture any target insects. Is a population present but not detected, a so-called false negative?⁴

15.1 How it was approached

By simulation. Researchers modelled the probability of capture using a dynamic wind model generating a turbulent plume structure and varying wind direction, and a behavior model based on the documented maneuvers of the insect during plume acquisition and along-plume navigation.⁴

15.2 What the model had to include

Turbulent plume structure and varying wind direction on the physical side, and documented manoeuvres during plume acquisition and along-plume navigation on the behavioural side.⁴

That is the level of detail required to answer what looks like a simple question: does an empty trap mean anything. This journal's pheromone article found catch depending on trap design, colour, placement and airflow, and this is the same dependency expressed as a simulation requirement.

15.3 Why simulation rather than field data

Because the field cannot answer it. A site with no captures has no known status, so there is nothing to compare the zero against.

Simulation lets the true population be specified and the trap outcome computed, which is the same reason the credence goods experiments this journal described used a laboratory market: some questions are only answerable where the truth is known by construction.

16. Surveillance against delimitation

Two purposes requiring very different effort.

The modelled densities were 1 per 2.6 square kilometres for surveillance and up to 49 per 2.6 square kilometres for delimitation.⁴

16.1 Why surveillance can be so sparse

One trap across two and a half square kilometres sounds hopeless until you consider what it is being asked to do. It only has to catch one individual, ever, from a population that may be expanding, over a season.

Detection at the landscape scale trades spatial coverage for time and for the insect's own movement, which is why §17 finds daily dispersal mattering less than the number of devices: over a long enough period, movement integrates.

16.2 The ratio is the point

Roughly fifty times the density to map a population compared with detecting one.

Establishing that something is present somewhere in an area is a much cheaper question than establishing where it is, and the two get conflated whenever a monitoring programme is asked to do both with one deployment.

16.3 The structural analogue

A few monitors in a building can indicate that a pest is present. Locating which unit, which room and which harbourage is the delimitation problem, and this journal's article on detection methods found interceptors sampling only the furniture legs they sit under.

17. What determines detection probability

A modelling study of site-based surveillance gives the ranking.

For a given release size, time-dependent detection probability was primarily determined by trap density and lure attractiveness, whereas mean step size, being daily dispersal, had limited effect.⁵

17.1 Why dispersal mattering less is surprising

Intuition says a fast-moving insect is easier to catch, because it encounters more of the environment. The model found daily step size having limited effect.⁵

A plausible reading is that over the durations modelled, even slow movers cover enough ground to reach a trap if one is close enough, so density dominates. That is our interpretation rather than the authors', and it would predict dispersal mattering more for short deployments.

17.2 The useful implication

The two variables that matter are the two an operator controls. How far the insect walks each day is a property of the organism; how many devices are deployed and how attractive they are is a purchasing and placement decision.

17.3 The context of that study

Surveillance using lures is widely used to support market access requirements for traded articles that are hosts or carriers of quarantine pests, and site-based surveillance typically needs to detect pests that are already present in the site or that may be entering the site from surrounding areas.⁵

That is the same dual problem a food premises faces: resident population and continuing introduction, which this journal's stored product and small fly articles both identified as requiring different responses.

18. The timing finding

An example of this work changing an official programme.

Researchers modelled seasonal population dynamics to identify which days of the year are most appropriate for trapping exotic fruit flies, getting national authorities to change the seasonal fruit fly trapping calendar accordingly.⁸

18.1 What it required

Modelling seasonal population dynamics to identify the appropriate days.⁸ So the improvement came from knowing the organism's biology well enough to predict when it would be detectable, not from a better trap.

18.2 Why this is a satisfying result

Detection probability varies through the season, which means the same effort spent at a different time buys a different amount of information.

Moving a calendar costs nothing and improves the programme, which is a rare category of intervention. This journal's mosquito and yellowjacket articles both turned on seasonal timing determining what a given action achieves.

19. What reduced effort costs

A study that measured degradation directly by subsetting its own sampling.

A large ground beetle survey compared nested subsets of its pitfall traps, retaining usable data at 75, 61, 54, 52 and 48 per cent of sites as the number of traps per site was reduced.⁶

19.1 The design of that comparison

Subsetting an existing dataset is a clean way to answer the effort question, because the same sites, the same season and the same operators are held constant while only the number of devices changes.

It is also the only ethical way to ask it, since the alternative would be deliberately under-sampling some sites to see what gets missed.

19.2 What that shows

Halving effort does not halve information. It removes sites from analysis entirely, because below some threshold a site produces data that cannot support inference at all.

19.3 The selection problem underneath

Of 172 species detected, 42 species were retained for occupancy-based inference, including 26 with stable model fits and 16 with usable predictions despite moderate parameter instability.⁶

So even a large survey could support this analysis for under a quarter of what it found. Rare things are the ones that drop out, which returns to §12.

20. What this means for a service report

20.1 The principle

A finding is only interpretable alongside the search that produced it. Everything below follows from that.

Record the effort, not just the outcome. A clearance claim needs the design behind it.²

Repeat visits are what make a zero mean something. A detection history requires occasions.⁷

Say how many devices and for how long. Trap density and duration are the levers.⁵

Do not write absent. Not detected under this effort is the accurate statement.⁷

Treat an early infestation as the hard case. Low density is where detection is weakest.³

Ask what season it is. Timing changes what the same effort buys.⁸

21. Where this trade stands

An honest assessment of how far any of this is currently applied here.

It is not. We are not aware of any structural pest control programme, ours included, that estimates detection probability, reports sampling effort alongside a negative finding, or expresses a clearance as a confidence level.

21.1 What we do instead

A technician inspects, records what was observed, and the report states findings. Where monitors are placed, the number and location are sometimes recorded and the duration usually is not, because the next visit's date is treated as sufficient.

21.2 Why not

Because the design requires repetition at the same site with the possibility of both outcomes, and commercial inspection is usually a single visit whose purpose is to produce an answer rather than a distribution.

There is also no demand for it. A client asking whether their building is clear wants a yes, and a report offering ninety-five per cent confidence given a stated effort is a worse product in every respect except accuracy.

21.3 What could be done without any statistics

Recording the effort. How many devices, where, for how long, on what dates. That costs nothing, requires no modelling, and converts an uninterpretable zero into one somebody could evaluate later.

It would also accumulate. A company with years of effort-annotated inspections would have the raw material to estimate its own detection probabilities, which nobody in this trade currently has. That is our suggestion rather than an established practice.

22. Limitations and open questions

None of this literature concerns structural pest control. It is vector surveillance, wildlife monitoring, invasive species trapping and quarantine work. The application here is ours.

The vector comparison is one study in one setting. A 38-month dataset in one urban area, and the discrepancy it reports may not generalise to other surveillance programmes.¹

Occupancy and infestation are not identical. The model estimates the probability a site is occupied, which for a pest control purpose is not the same as whether treatment is warranted. A dwelling occupied at low density and one heavily infested both count as occupied.

We have read abstracts and summary sections. The methodological details of the models, including their assumptions about closure between sampling occasions, are not assessed here, and those assumptions matter for whether estimates are valid.¹²

The acoustic and beetle studies measure species richness. Detecting all species present is a different target from detecting one known species, and the effort figures should not be read across directly.³⁶

Sections 3.1, 7.2, 9.1, 12.1, 13.2, 16.2 and 21.2 are our reasoning. The confounding explanation, the decomposition of survey failure, the causal chain from detection to programme effect, the early-infestation analogy, the report criticism, the structural delimitation parallel and the effort-recording proposal are ours rather than sourced findings.

Our commercial position. This article argues that our own negative findings carry less information than our reports imply, and that we should record effort in a way that would make our results auditable. Both of those are costs to us.

23. Conclusion

Not detecting something does not imply it is absent.⁷ Detectability varies with space, season, habitat, method and species, so a zero is consistent with true absence and with a miss, and nothing in the observation itself distinguishes them.⁶ The framework that separates them estimates occupancy and detection probability as two quantities, which requires repeated occasions at the same site.⁶⁷

The demonstration is in the vector data. Routine larval surveys put infestation at three per cent of dwellings, oviposition traps at sixty-eight, and a model accounting for imperfect detection at ninety-one, while the control campaigns built on the first figure had no measurable effect on infestation.¹ The authors concluded that many breeding sites were simply being overlooked.¹ A programme cannot outperform its own detection, and one with a poor instrument will report success while the population sits where it was.

Nobody in structural pest control does any of this, including us, and the statistical version is probably out of reach for a commercial inspection. The part that is not out of reach is smaller and duller: write down how many devices went in, where, and for how long, so that the zero at the end has something attached to it. At present a clearance means a person looked and did not see anything, for an amount of time nobody recorded, which is a sentence about the inspector rather than about the building.

References

1. Modeling Dengue Vector Dynamics under Imperfect Detection: Three Years of Site-Occupancy by Aedes aegypti and Aedes albopictus in Urban Amazonia. PubMed Central PMC3589427. Principal demonstration source. Used for the statement that studies have assumed vectors are truly absent from sites where they are not detected and that since no perfect detection method exists this assumption is questionable; that imperfect detection may bias estimates of key surveillance and control parameters including site-occupancy rates and control intervention effects; for the study design using a modelling approach explicitly accounting for imperfect detection and a 38-month, 55-site detection and non-detection dataset with meteorological and dwelling-level covariates; for the results that site-occupancy estimates for one species were a mean of 0.91 with a range of 0.79 to 0.97, much higher than the 0.03 with range 0.02 to 0.11 reported by routine surveillance based on rapid larval surveys and moderately higher than the 0.68 with range 0.50 to 0.91 directly ascertained with oviposition traps; that regular control campaigns based on breeding-site elimination had no measurable effects on the probabilities of dwelling infestation; that site-occupancy fluctuated seasonally mainly due to the negative effects of high maximum and minimum summer temperatures while rainfall and dwelling-level covariates were poor predictors; and for the authors' conclusion that the marked contrast suggests many breeding sites were overlooked by vector control agents and that better sampling strategies are urgently needed. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3589427/
2. Inferring pathogen presence when sample misclassification and partial observation occur. United States Geological Survey publication record. Used for the statement that partial observations are either discarded or censored, disregarding information that could be used to make inference about the true state of the system; for the statement that there is a critical need for more direction and guidance related to how many samples are enough to declare a unit of interest free of a pathogen; for the development of a Bayesian hierarchical framework accommodating false negative, false positive and uncertain detections; for the provision of guidance on sample sizes required to be 95 per cent certain a target organism is absent from a site; and for the finding that the presence of uncertain detections increased the variability of resulting posterior estimates. https://www.usgs.gov/publications/inferring-pathogen-presence-when-sample-misclassification-and-partial-observation
3. Designing occupancy surveys and interpreting non-detection when observations are imperfect, together with an associated acoustic survey analysis. Research repository record. Used for the acoustic survey results that 351,771 passes of eight species or species groups were identified in 5,856 detector-nights; that seasonal patterns of activity varied among species; that on average approximately 30 sampling nights were needed to detect 90 per cent of total species richness among locations and seasons; that relatively few nights, twelve or fewer, were needed to detect most species during summer while many more nights were needed for acoustically rare species; that detection probability and occupancy were estimated using single-season occupancy models; and that the sampling effort required to determine presence or probable absence of two federally listed bat species was substantially greater. https://www.researchgate.net/publication/230538586_Designing_occupancy_surveys_and_interpreting_non-detection_when_observations_are_imperfect
4. Simulation Modeling to Interpret the Captures of Moths in Pheromone-Baited Traps Used for Surveillance of Invasive Species: the Gypsy Moth as a Model Case. PubMed record 27663859. Used for the statement that when pheromone traps are used for detection of an invasive pest and then delimitation of its distribution, an unresolved issue is the interpretation of failure to capture any target insects, and whether a population is present but not detected as a false negative; for the modelling of capture probability at densities typical for surveillance, being one trap per 2.6 square kilometres, and for delimitation, being up to 49 per 2.6 square kilometres; and for the simulation method using a dynamic wind model generating a turbulent plume structure with varying wind direction and a behaviour model based on documented manoeuvres during plume acquisition and along-plume navigation. https://pubmed.ncbi.nlm.nih.gov/27663859/
5. Simulation to investigate site-based monitoring of pest insect species for trade. PubMed Central PMC10413999. Used for the statements that pest insect surveillance using lures is widely used to support market access requirements for traded articles that are hosts or carriers of quarantine pests; that modelling has been used extensively to guide surveillance design for pest free area claims but is less commonly applied to provide confidence in pest freedom or low pest prevalence within registered sites; that site-based surveillance typically needs to detect pests already present in the site or entering from surrounding areas; and for the finding that for a given release size, time-dependent detection probability was primarily determined by trap density and lure attractiveness, whereas mean step size being daily dispersal had limited effect. https://pmc.ncbi.nlm.nih.gov/articles/PMC10413999/
6. Imperfect detection and sampling effort shape ground beetle (Carabidae) diversity indicators in large-scale monitoring. Ecological Indicators. Used for the statements that robust inference about these organisms remains challenging because they are notoriously difficult to detect with certainty; that species detectability varies strongly across space, seasons, habitat, sampling methods and species traits, making it difficult to distinguish true absences from non-detections; that occupancy modelling has emerged as a statistical framework addressing this by explicitly estimating both the probability that a species occupies a site and the probability of detecting it when present, allowing researchers to move beyond raw presence-absence data and reduce the bias introduced by imperfect detection; for the inclusion of day of year at sampling as a proxy for seasonal variation; for the nested pitfall subset results in which usable data were available for 75, 61, 54, 52 and 48 per cent of sites as trap numbers were reduced; and for the report that of 172 species detected, 42 were retained for occupancy-based inference including 26 with stable model fits and 16 with usable predictions despite moderate parameter instability. https://www.sciencedirect.com/science/article/pii/S1470160X26006217
7. Estimating percent-area-occupied and related dynamics from presence-absence data. Course notes on occupancy models, Colorado State University. Used for the statement that not detecting a species does not imply absence; for the worked example of a detection history in which a site detected on some occasions is assumed occupied across all five sampling occasions but not detected on occasions 1, 3 and 5; for the notation of occupancy probability and detection probability as the two estimated parameters; for the note that missing data are readily handled where a site was not visited on a given occasion; and for the summary that more recent methodological advances focus on incorporating false positives into the estimation, dealing with heterogeneity in detection probabilities, and applying these methods to estimating species richness. https://sites.warnercnr.colostate.edu/gwhite/wp-content/uploads/sites/73/2017/04/Occupancy-Models.pdf
8. Detecting target species: with how many samples? Royal Society Open Science, 9(8), 220046. Used for the statement that such models assume that if a species is present in a given sample its presence is not detected with some false negative probability, and that it is important to be aware of this detection error; for the statement that the procedure has been used in the early detection of pest invasions and diseases, which is of paramount importance for successful management of responses such as containment or eradication, implementing surveillance traps to maximise the probability of detection and minimise economic costs; and for the example in which authors modelled seasonal population dynamics to identify which days of the year are most appropriate for trapping exotic fruit flies, leading national authorities to change the seasonal trapping calendar accordingly. https://royalsocietypublishing.org/rsos/article/9/8/220046/96707/Detecting-target-species-with-how-many-samples
9. Survey techniques, detection probabilities, and the relative abundance of the carnivore guild on the Apostle Islands. Research preprint. Used for the statement that explicitly modelling the probability of a true-absence versus false-absence reduces the bias that can result from assuming perfect detection when calculating occupancy estimates; and for the design description in which image data were consolidated as detected or not detected for each species during two-week observation periods, providing discrete detection histories for repeated survey methods while avoiding autocorrelation between detections only a day apart and issues associated with zero-inflation. https://arxiv.org/pdf/1703.10726