A variant is not found among sixty usable sequence records. That is a real observation about those records. It does not establish that the variant is absent from the population they were intended to represent.
The short answer is conditional: under independent random sampling, a fixed population proportion and perfect classification, the probability of zero detections is (1 − p)ⁿ. Inverting that relationship gives an upper confidence bound after a zero count. The result changes when records are duplicated, samples are selected unevenly or classification can miss the target.
This guide extends Virology into the interpretation of supplied fictional surveillance records. Every proportion, sample size and performance value below is invented for mathematical learning. There are no instructions for obtaining, propagating, modifying or experimentally testing a virus. The examples do not interpret a person’s medical result or establish present disease conditions.
eduKateSG · Virology
Find your next learning step
Interpret a zero count through its population, independent observation unit and explicit sampling model.
ROUTE 1 · CHAPTERS 1–5
Define the sampled evidence
Define the quantity and reconstruct the first worked relationship.
ROUTE 2 · CHAPTERS 6–10
Interpret finite non-detection
Follow the original examples and check why the operation is valid.
ROUTE 3 · CHAPTERS 11–16
Preserve independence and selection
Change the evidence and identify which assumptions still hold.
ROUTE 4 · CHAPTERS 17–21
Change detection and population models
Check uncertainty, scope and independent model alternatives.
ROUTE 5 · CHAPTERS 22–26
Practise and continue
Repair common errors, attempt transfer and choose connected reading.
Chapter contents
Define the sampled evidence · Chapters 1–5
1. State the population before counting records
3. Derive the probability of missing the category
Interpret finite non-detection · Chapters 6–10
6. Invert the zero-count relationship
7. Interpret confidence without changing the question
8. Explain the approximate rule of three
Preserve independence and selection · Chapters 11–16
11. Recognise clustering without inventing an effective count
12. Keep selection separate from classification
13. Compare stratified non-detection probabilities
14. Add a known chance of missing the category
Change detection and population models · Chapters 17–21
17. Preserve the reporting period
18. Avoid converting a bound into a posterior probability
19. Attempt an independent changed-sample case
Practise and continue · Chapters 22–26
22. Distinguish a census statement from a sampling statement
23. Preserve simultaneous coverage when asking several questions
24. Separate statistical exclusion from logical impossibility
CHAPTER 1 OF 26 · Define the sampled evidence
1. State the population before counting records
Imagine a fictional reporting period and a defined collection of eligible, already classified cases. We are interested in the proportion p belonging to a symbolic variant category A within that collection. The labels describe an abstract category; they carry no assigned biological function or clinical importance.
The population is not automatically all residents, all infections, all particles or all sequence reads. Those are different units. An analysis of eligible cases cannot silently become an analysis of every infection, particularly when entry into the eligible collection depends on observation and reporting.
For the first model, assume each retained record represents a different independently sampled eligible case. Every case has the same probability p of category A, and the classification is exact. These assumptions deliberately isolate the sampling question.
Write the target in a sentence before using a formula: the proportion of category A among eligible cases in this specified collection and period. If the target changes to a different place, time or eligibility rule, the original calculation needs review. A denominator supplied by a database does not decide the population question on its own.
A usable record classified as a different category supplies a non-detection of A under the stipulated exact classification. A record with no usable classification supplies something else: missing information. Treating both as the same outcome changes the evidence.
Suppose seventy records enter an exercise, but only sixty have usable classifications. All sixty are non-A. The initial binomial model uses n = 60, provided its other assumptions hold for those retained records. It does not use seventy merely because the file contains seventy rows.
The ten unresolved classifications might be missing for reasons unrelated to category. They might also be systematically different. The observed records alone cannot choose between those possibilities. Conditioning on successful classification can change which population the retained records represent.
For a paper task, give each row one of three statuses: usable A, usable non-A or unresolved. Ask students to identify which statuses enter the stated count and which require further explanation. The objective is to preserve an observation’s meaning, rather than to manufacture a larger denominator by converting every blank into a reassuring zero.
CHAPTER 3 OF 26 · Define the sampled evidence
3. Derive the probability of missing the category
In one independent draw, the probability of a non-A case is 1 − p. For two independent draws with the same p, the probability that both are non-A is (1 − p)². Repeating that multiplication gives P(zero A records) = (1 − p)ⁿ.
The multiplication uses independence. The repeated factor uses a fixed p. Exact classification connects an observed A record with the target category. Each condition has a specific role; none can be replaced merely by saying that the dataset is large.
The probability of at least one detection is the complement, 1 − (1 − p)ⁿ. These statements describe possible results under a specified population proportion. They do not, by themselves, assign a probability distribution to the unknown proportion after the records have been observed.
The NIST binomial distribution reference provides the fixed-probability count framework. Here the zero-count expression is derived directly as an original teaching calculation. Begin with the multiplication and complement before introducing confidence language; learners should know which event the formula actually describes.
Set n = 60. If the population proportion were p = 0.05, the probability of no A records would be 0.95⁶⁰, approximately 0.04607. A zero count would be relatively unusual under that particular model value, but it would remain possible.
If instead p = 0.01, the zero-count probability is 0.99⁶⁰, approximately 0.54716. More than half of repeated sixty-record collections would contain no A under these assumptions. The same observed zero therefore has a very different compatibility with the two proposed proportions.
This is more informative than saying either that zero proves absence or that zero tells us nothing. The observation can constrain a model without settling every population claim. Its strength depends on the sampling conditions and on which proposed value is being examined.
Ask a learner to state the direction: as p increases, missing A in every independent draw becomes less probable. As n increases at fixed positive p, it also becomes less probable. If their interpretation runs in the opposite direction, return to the single-draw non-detection probability before adding more complicated formulas.
| Independent usable records n | Miss probability if p = 0.05 | Miss probability if p = 0.01 |
|---|---|---|
| 20 | About 0.35849 | About 0.81791 |
| 60 | About 0.04607 | About 0.54716 |
| 100 | About 0.00592 | About 0.36603 |
| 300 | About 0.000000208 | About 0.04904 |
Each row is a separate fictional sampling model. The table does not prescribe a real surveillance programme, establish a biological threshold or describe actual variant prevalence. Its purpose is to make the relationship between rarity, independent information and non-detection visible.
A large number of records can still have a substantial chance of missing a sufficiently rare category. Conversely, a modest sample can be quite informative against a relatively common proposed proportion under the stipulated conditions. The relevant question is not whether the sample sounds large, but what detection probability it provides for the specific model and target.
CHAPTER 6 OF 26 · Interpret finite non-detection
6. Invert the zero-count relationship
Choose a one-sided confidence procedure with tail level α = 0.05. After observing zero A records in n independent usable draws, solve (1 − pᵤ)ⁿ = α. The resulting endpoint is pᵤ = 1 − α^(1/n).
For n = 60, pᵤ is approximately 0.04870, or 4.87%. Under the binomial model, this is the zero-count endpoint of the exact one-sided 95% upper confidence procedure. A larger proposed p makes the observed zero even less likely than the chosen 0.05 boundary.
The estimate zero and the upper endpoint answer different questions. The sample fraction is zero divided by sixty. The confidence endpoint describes the uncertainty procedure under repeated sampling and its model assumptions. It is not a measured proportion of 4.87%.
Keep the one-sided convention explicit. A two-sided interval constructed by splitting the error allowance between two tails has a different upper endpoint. Reports that simply say 95% interval can therefore hide a methodological difference. State the procedure and tail convention rather than comparing endpoints as if every interval used the same rule.
CHAPTER 7 OF 26 · Interpret finite non-detection
7. Interpret confidence without changing the question
A frequentist confidence procedure has a coverage property across repeated samples under its assumptions. After one particular sample, the population proportion is not treated as randomly moving inside and outside the calculated interval in that interpretation.
Consequently the endpoint should not be translated casually into a 95% probability that the unknown p is below 4.87%. Such a probability statement would require an explicit Bayesian model, including a prior, or another justified interpretation. The same printed endpoints do not erase that distinction.
A practical sentence can remain clear: no A records were observed among sixty stipulated independent eligible cases; under the fixed-proportion binomial model with exact classification, the one-sided 95% upper confidence endpoint is about 4.87%.
That sentence preserves the zero observation, the target model and the type of uncertainty statement. It also leaves space for the more consequential limitations: selection, independence and classification. Statistical precision cannot compensate for a population definition that does not match the claim. The interval describes the modelled question, rather than granting permission to generalise to every place or time.
CHAPTER 8 OF 26 · Interpret finite non-detection
8. Explain the approximate rule of three
For small p, log(1 − p) is approximately −p. Thus (1 − p)ⁿ is approximately exp(−np). Setting that approximation equal to 0.05 gives p approximately −log(0.05)/n, or about 2.996/n.
Rounding the numerator to three produces the familiar approximate zero-count upper bound 3/n. With n = 60 it gives 0.05, close to the exact one-sided endpoint 0.04870. The approximation is convenient because its derivation and limitations are visible.
It is not exact, and it does not apply by magic to arbitrary dependent records or a poorly defined denominator. At n = 5, 3/n gives 0.6 while the exact endpoint is about 0.45072. A short rule can be noticeably loose when the sample is small.
Ask students to compare exact and approximate endpoints for several n values. They should explain why the approximation approaches the exact value as n grows. Then ask what remains unchanged: even an exact formula still depends on the sampling model. Improving the algebraic approximation addresses a different issue from improving the evidence’s representativeness.
CHAPTER 9 OF 26 · Interpret finite non-detection
9. Distinguish two kinds of detection question
One question asks how likely a sample is to contain at least one A if the population proportion is specified. Another asks how precisely a sample can estimate an unknown proportion. Detecting one example and estimating a frequency with useful precision are different objectives.
For an abstract calculation, demand a 95% chance of at least one A when p = 0.01. Solve 1 − 0.99ⁿ ≥ 0.95, giving n ≥ log(0.05)/log(0.99). The smallest integer is 299 under this exact independent model.
That number does not establish that 299 observations estimate a 1% proportion accurately. It addresses only the event of at least one detection at a stipulated p. Nor is it a professional sampling recommendation; eligibility, clustering, time, resources and classification are absent from the calculation.
ECDC’s guidance on representative and targeted genomic monitoring, dated 3 May 2021, distinguishes surveillance objectives and sampling approaches. It provides historical institutional context. The fictional arithmetic here should not be presented as a replacement for that guidance or as a statement of current operational requirements.
CHAPTER 10 OF 26 · Interpret finite non-detection
10. Count independent cases rather than copied rows
Suppose ten independent case records are each copied ten times during preparation. The stored file now has one hundred rows. It still contains only ten independently sampled cases in this stipulated example.
If p = 0.05, the probability that all original cases are non-A is 0.95¹⁰, approximately 0.59874. Using the copied row count gives 0.95¹⁰⁰, approximately 0.00592. The latter calculation pretends that copying generated ninety new independent opportunities to observe A.
The difference is substantial. Both files contain a zero count of A, but the evidence has not become more informative through duplication. An exact case identifier and a clear observation unit matter more than the apparent length of the spreadsheet.
The cluster-precision guide develops how a shared group structure changes information about a mean. It complements the non-detection problem without supplying its binomial assumptions. The existing consensus-sequence guide examines another loss of meaning in sequence summaries, including repeated support and linkage. Here the question is population non-detection. A record’s identity must remain clear before the probability model can describe how much independent information it contributes.
CHAPTER 11 OF 26 · Preserve independence and selection
11. Recognise clustering without inventing an effective count
Records can be related without being literal duplicates. Several cases from one shared setting may have correlated category membership. The independent binomial model then need not describe their joint non-detection probability.
Use an extreme fictional construction: ten independent groups are selected, each containing ten perfectly matching case categories. All cases within a group are A together or non-A together. Each group’s category is A with probability 0.05. The zero-count probability is again 0.95¹⁰, despite one hundred case records.
This construction isolates perfect within-group dependence. It does not establish that every real cluster behaves that way, or that dividing a row count by a convenient group size gives a correct effective sample size. Partial dependence needs its own joint model.
A useful evidence note identifies the grouping structure and which independence assumptions are supported. If the information is insufficient, report that the simple binomial endpoint is conditional on an unverified assumption. Replacing the sample size with an improvised smaller number may look cautious, but it does not automatically create a validated confidence procedure. The repair should match the source of dependence, not merely reduce a total until the answer feels plausible.
CHAPTER 12 OF 26 · Preserve independence and selection
12. Keep selection separate from classification
Even exact classifications of independent selected cases do not ensure population representativeness. Suppose a fictional eligible population has two groups, each containing half its cases. A occurs in 10% of the first group and none of the second. The whole-population proportion is 5%.
If every retained record comes from the second group, zero A records are expected regardless of how many such records are collected. Applying the homogeneous model with p = 0.05 to those selected records misrepresents the selection process.
This example distinguishes sampling bias from classification error. The category calls are perfect in the model. The problem is that the selected cases do not represent the full target collection. More accurate calling of the same selected cases would not fix the missing group.
The useful next question is how cases entered the dataset. Which groups, places, dates or reporting routes were included? Were eligibility and inclusion probabilities stable? A proportion estimate or confidence endpoint must remain attached to the population its observations can support. A large denominator from a narrow selected group cannot be converted into whole-population evidence solely by using a smaller error bar.
CHAPTER 13 OF 26 · Preserve independence and selection
13. Compare stratified non-detection probabilities
If independent sampling is conducted within defined groups, the non-detection calculation can preserve those groups. For two strata with proportions p₁ and p₂ and independent counts n₁ and n₂, the probability of no A is (1 − p₁)^n₁(1 − p₂)^n₂.
Take the previous fictional half-and-half population, with p₁ = 0.10 and p₂ = 0. Sample thirty independent cases from each stratum. The probability of zero A is 0.9³⁰, approximately 0.04239. Treating every draw as identical with p = 0.05 instead gives 0.95⁶⁰, approximately 0.04607.
The models are similar numerically here but not identical. The stratified expression uses specified sampling allocations and group proportions. It does not follow merely from knowing the overall mean proportion.
For a learner, change the allocation to ten cases from the first group and fifty from the second. The miss probability becomes 0.9¹⁰, approximately 0.34868. The total remains sixty records, but their information about the heterogeneous population changes. This is why sample composition belongs in the record: an unchanged total can accompany a substantially changed observation process.
CHAPTER 14 OF 26 · Preserve independence and selection
14. Add a known chance of missing the category
Relax exact classification while retaining a very simple model. Let q be the known probability that a true A case is correctly detected as A, independently across sampled cases. Assume there are no false A detections and that other conditions remain as before.
The probability of an observed A in one draw is then θ = pq. The probability of no observed A is (1 − pq)ⁿ. A zero count directly constrains this observed-detection probability, rather than separating p and q without further information.
For n = 60, the earlier one-sided endpoint for θ is approximately 0.04870. If q is independently known to equal 0.8, the corresponding model-based upper endpoint for p is 0.04870/0.8, approximately 0.06088, capped at one where necessary.
The numbers do not describe a real sequencing system. Their purpose is to expose an assumption. Imperfect detection can weaken the inference from zero observed categories to a bound on the underlying proportion. If q is uncertain, inserting one convenient estimate and presenting the result as exact would hide another layer of uncertainty. Classification performance requires its own appropriate evidence.
CHAPTER 15 OF 26 · Preserve independence and selection
15. See why unknown detection prevents separation
Suppose only the product pq is constrained by the observed record. A small p with high q and a larger p with low q can produce the same observed-detection probability. The zero count cannot uniquely identify which explanation applies.
For a fictional illustration, p = 0.02 with q = 1 and p = 0.10 with q = 0.2 both give θ = 0.02. Under the stipulated independent model, every possible observed-detection count has the same distribution in those two cases.
This is a structural limitation, rather than a shortage repaired simply by adding more records under the same unknown performance relationship. More records can constrain θ more precisely while leaving the p-versus-q distinction unresolved.
If independent evidence only establishes q ≥ 0.8, then p ≤ θᵤ/0.8 provides a conservative conditional upper bound within this model. Without any positive lower bound on q, zero detections may remain compatible with very large p. An honest report identifies the unresolved detection component. The correct question is what additional evidence connects the observation process with the underlying category, not which wording makes a zero count sound most reassuring.
CHAPTER 16 OF 26 · Preserve independence and selection
16. Change to sampling without replacement
A finite collection requires another model when a substantial fraction is sampled without replacement. Suppose a fictional archive contains exactly N = 100 cases, of which K = 5 belong to A. Twenty distinct cases are selected uniformly without replacement and classified exactly.
The probability of zero A is the number of twenty-case subsets containing only non-A cases divided by the number of all twenty-case subsets: choose(95,20)/choose(100,20). Equivalently, multiply 95/100, 94/99 and the following factors until twenty distinct selections are accounted for.
This probability is about 0.31931. The independent with-replacement approximation 0.95²⁰ is about 0.35849. The difference arises because each observed non-A case slightly increases the remaining fraction of A cases in the finite archive.
The hypergeometric model refers to this finite collection with fixed K. It is not automatically a model for a continually changing wider population. Label which population parameter is being constrained. The exercise shows why a model should follow the stated selection process. The convenient binomial expression is valuable when appropriate, rather than a universal formula attached to every zero in a dataset.
CHAPTER 17 OF 26 · Change detection and population models
17. Preserve the reporting period
A zero count from one week constrains the population defined for that week under its observation model. It does not establish absence in an earlier unsampled week or a later population whose composition changes.
Imagine two fictional periods, each with thirty usable independent records. All are non-A. If both periods share the same fixed p and classification model, pooling gives sixty draws and the sixty-record calculation. That is an additional shared-proportion assumption.
If the scientific question permits p to differ between periods, each period has its own zero-count information. Each thirty-record upper endpoint is approximately 0.09503 under the same one-sided procedure. Pooling would constrain an assumed shared parameter, not automatically each changing period’s parameter.
This distinction becomes important when a larger cumulative database is presented beside a statement about the most recent date. Old records can be useful historical evidence without supplying current independent draws. Keep event dates, eligibility periods and reporting cut-offs visible. A stable cumulative total can hide a thin or incomplete recent record, especially when delayed classifications have not yet arrived.
CHAPTER 18 OF 26 · Change detection and population models
18. Avoid converting a bound into a posterior probability
For comparison only, suppose a Bayesian teaching model assigns p a uniform prior on zero through one and uses the same independent exact-classification likelihood. After zero A among n cases, the posterior density is proportional to (1 − p)ⁿ.
Its 95% upper credible endpoint is 1 − 0.05^(1/(n + 1)). For n = 60 this is approximately 0.04792. It differs from the one-sided frequentist endpoint because it answers a different probabilistic question and incorporates the stipulated prior.
The purpose is not to rank the approaches by which endpoint looks more desirable. A different prior can produce another posterior endpoint. A different confidence procedure can also change the reported frequentist endpoint. The method and interpretation must travel with the number.
Ask the learner to write two sentences, one for each procedure, without borrowing the other’s meaning. This exercise repairs a common communication error: reporting a confidence endpoint and then explaining it as though it were a posterior probability. Statistical language is useful when it preserves the reasoning that generated the result, rather than merely attaching a reassuring percentage.
CHAPTER 19 OF 26 · Change detection and population models
19. Attempt an independent changed-sample case
A fictional record has one hundred independent, randomly selected eligible cases, all classified exactly as non-A. Find the one-sided 95% upper confidence endpoint. Then repeat the calculation if the independent detection probability for an A case is known to be 0.75, with no false A calls.
The first endpoint is 1 − 0.05^(1/100), approximately 0.02951. In the second model, this constrains θ = 0.75p. Dividing by 0.75 gives an upper endpoint for p of approximately 0.03935.
Now change the evidence rather than the formula: the one hundred rows are twenty independent cases, each copied five times. Under exact classification, use n = 20, giving an endpoint approximately 0.13911. The copied-row calculation would be unjustifiably narrow.
Explain the three changes in words. The first gains information from more independent cases. The second acknowledges imperfect detection. The third restores the correct independent observation unit. A student who can distinguish those operations has understood why no single formula can repair all weaknesses in a surveillance claim.
CHAPTER 20 OF 26 · Change detection and population models
20. Build a practical evidence statement
Record the target collection, category definition, observation unit, dates, selection method, usable count, unresolved count and duplicate handling. Add the classification assumptions and any supported information about performance. These details should precede a strong population statement.
Then report the observation separately from the inference. No A records were observed is a count statement. A model-based upper endpoint adds statistical assumptions. No A exists anywhere is a much larger claim that this exercise does not support.
Preserve a clear route back to the source records and analytical choices. If the result changes after a revision, identify whether new cases arrived, duplicate rows were removed, category calls changed or the statistical method changed. These explanations have different meanings.
For research reading, a useful question is whether targeted and representative collections have been mixed without accounting for their different selection goals. Targeted evidence may be valuable for finding unusual categories, while a representative account serves another population question. Neither should be dismissed merely because it answers a different task. The record should explain which task the reported number is intended to serve.
CHAPTER 21 OF 26 · Change detection and population models
21. Repair, stabilise and extend the learning
If a learner thinks zero proves absence, begin with a small token model: draw a few symbols from a much larger imagined collection. Show that missing a rare label is compatible with its presence. Derive the single-draw probability before repeating it.
If the learner confuses probability of zero given p with probability of p given zero, write the conditioned quantity in words. Use the sixty-record comparison to show that the calculation begins with a proposed p. Introduce inference only after that direction is secure.
If they count copied rows as new evidence, use identifiers and reconstruct the original cases. If they treat unresolved classifications as non-A, restore the three record statuses. These errors need different repairs, even when both produce an inflated denominator.
For stabilisation, vary p or n one at a time. For extension, change selection, clustering, time or detection. Ask which formula remains justified and which assumption must be rebuilt. Return later with a new dataset description and require the learner to choose the relevant model independently. Transfer is visible when they recognise the evidence structure without relying on the original labels.
CHAPTER 22 OF 26 · Practise and continue
22. Distinguish a census statement from a sampling statement
Suppose a fictional archive contains exactly one hundred eligible records. Every record has a usable exact classification, and none is A. The observed absence is now established for that complete archive under the stated classification assumption. No sampling interval is needed to describe a census of those one hundred records.
The conclusion still does not extend automatically beyond the archive. Cases that were never eligible, never reported or occurred outside its time period are outside the counted collection. A complete archive census can coexist with incomplete coverage of a wider biological population.
Now return to sixty classified non-A records with forty archive records unresolved. Without a random-sampling or missingness model, a deterministic account permits anywhere from zero to forty A cases among the unresolved records. The archive proportion could therefore range from zero to 40%. This is a hard information bound from the supplied finite record, not the earlier 4.87% confidence endpoint.
There is no contradiction between those numbers. They answer different questions using different assumptions. One assumes random independent sampling from a fixed-proportion model. The other asks what an incomplete finite table alone logically permits. A learner should state which evidence justifies moving from the broad information bound to a statistical statement. The confidence formula cannot create a random selection mechanism that the records do not supply.
CHAPTER 23 OF 26 · Practise and continue
23. Preserve simultaneous coverage when asking several questions
An analyst may inspect several pre-specified variant categories and report a zero-count upper endpoint for each. A 95% procedure for each category separately does not automatically provide 95% simultaneous coverage for the complete set of statements.
For a fictional set of five pre-specified category questions, one conservative construction assigns tail allowance 0.05/5 = 0.01 to each one-sided confidence procedure. The probability of any failure is then at most the sum of the five allowances under their respective valid models. This union-bound reasoning does not require the five category results to be independent.
If a particular category has zero detections among sixty independent exactly classified cases, its adjusted endpoint is 1 − 0.01^(1/60), approximately 0.07388. It is wider than the single-question 0.04870 endpoint because the procedure now protects a larger family of claims.
The calculation is a teaching example, not a general recipe for every surveillance analysis. The questions, models, category definitions and complete interval procedures must be specified. Choosing a category after inspecting the data can introduce another selection issue. The point is to keep the inferential task visible: a statement about one pre-specified proportion and a statement covering several proportions jointly require different accounts, even when both begin with an observed zero.
CHAPTER 24 OF 26 · Practise and continue
24. Separate statistical exclusion from logical impossibility
For any fixed p strictly between zero and one and a finite positive n, the probability (1 − p)ⁿ is positive. A zero count is therefore logically possible even for a proposed proportion above the reported confidence endpoint. The confidence procedure marks some results as sufficiently unusual at a specified error allowance; it does not declare them mathematical impossibilities.
In the exact-classification model, p = 1 is different: every eligible case is A, so observing a non-A record is impossible. That sharp conclusion belongs to the model’s perfect classification and fixed-population assumptions. Allowing classification misses changes the observation relationship.
There is also a useful empty-record boundary. If n = 0, no usable classification has been observed. The formula involving the exponent 1/n cannot be applied. The record supplies no sampling constraint on p beyond its ordinary zero-to-one domain. Do not replace an unavailable count with a tiny positive number to force an endpoint.
These boundary checks help learners read a confidence statement with appropriate care. Finite evidence can be informative without proving absence, and a declared statistical threshold can support a disciplined decision rule without changing a rare event into an impossible one.
Does zero detections mean the population proportion is zero?
No. It is compatible with positive proportions under finite sampling. The strength of the evidence depends on the selection and observation model.
Is the upper confidence endpoint an estimated prevalence?
No. The observed sample fraction and the confidence endpoint have different roles. Report the method and target alongside both.
Can more sequence reads replace more independent cases?
Not automatically. The relevant unit is defined by the population question. Repeated information from one case does not create additional independently sampled cases.
Does a negative category call establish no infectivity?
No. Our classification concerns a symbolic sequence category. Infectivity, disease and transmission are separate biological questions requiring suitable evidence.
Can the same zero-count result be used for every period?
No. Pooling or temporal transfer requires assumptions about stable proportions, sampling and classification. Keep the reporting period attached to the conclusion.
CHAPTER 26 OF 26 · Practise and continue
26. Continue through biology, information and inference
Keep the biological question precise
Return to Virology for genomes, particles, hosts and measurement targets. The consensus-sequence guide develops linked information lost by a positional summary. Bioinformatics connects sequence records with analysis and interpretation.
Strengthen the statistical foundation
Use Probability for conditioned events and dependence, and Statistics for inference from observations. The NIST binomial reference supplies the basic distribution. The ECDC 3 May 2021 genomic-monitoring report supplies historical context on sampling objectives; its recommendations are not recast here as current local instructions. Sources checked on 23 October 2026.
Test independent transfer
Continue through the Gold Standard education parent and How X Works. Choose a supplied non-detection claim, identify the population and observation unit, and state one assumption whose failure would change the conclusion. Preserve what the zero genuinely establishes while leaving the unresolved population question visible.