An analysis of surveillance design and public interpretation, not advice on selecting an antibiotic for an individual patient. All numerical demonstrations are hypothetical. Published data years are separated from publication dates throughout.
1. A global report contains several kinds of observation
WHO's global antibiotic-resistance report published October 13, 2025 draws on more than 23 million bacteriologically confirmed cases and presents adjusted estimates for 2023. Its underlying reporting includes 104 countries for that year and 110 over the longer 2016–2023 period. Those are substantial contributions, but the dates and units matter: the publication is not a census of every infection occurring in 2025, and the reported cases are not a random sample of everyone in the world. WHO 2025 report
The distinction is central to reading antimicrobial-resistance statistics. A laboratory result can be technically reliable while the collection of people reaching that laboratory remains selective. An increase in resistant isolates can reflect a worsening problem, expanded testing, changed referral patterns or several processes together. A low reported count can reflect a favorable situation or a weak ability to detect the problem.
None of these cautions diminishes the seriousness of resistance. They define what the evidence can establish and where additional measurement is needed. Public-health systems cannot respond effectively if the same number is alternately treated as a laboratory percentage, a population infection rate and a measure of treatment failure.
This article follows the observation from the patient to the specimen, the isolate, the susceptibility category and the surveillance summary. It asks who enters the denominator, how repeat samples are handled, why laboratory standards matter and how better detection can initially make a system look worse. The goal is to make resistance statistics more actionable by preserving the distinctions that a headline often removes.
2. Begin with the organism, drug and specimen
Resistance is not a single property that can be summarized without context. A surveillance claim needs to identify the organism, the antimicrobial being considered and the source of the specimen. The relevant unit may be a particular organism-drug combination in a defined type of infection, rather than an undifferentiated category of superbugs.
A hypothetical hospital could report that a bacterial species is resistant to one drug in 30 percent of tested isolates. That does not mean 30 percent of all bacteria in the hospital resist every antibiotic. It does not establish that 30 percent of patients are infected. It also does not determine which treatment is appropriate for a particular person.
The specimen source helps define the clinical question. Bloodstream isolates describe a different observation than organisms obtained from a wound, urine or a screening swab. Combining them can change the mix of patients and conditions represented. A trend may move because that mix changed, even if each source-specific pattern stayed similar.
A clear chart title should therefore identify its target in full enough terms to prevent broad reinterpretation. The accompanying methods should explain whether results concern clinical specimens, screening specimens or both, and whether they include all eligible organisms or a selected subset.
This specificity is not merely technical etiquette. It changes what action the evidence can support. A signal from invasive infections may be particularly important for severe disease, while a colonization-screening program can inform infection-prevention work. The two can complement one another without being interchangeable. Reporting them as the same statistic makes it harder to understand both the burden of illness and the pathways through which resistant organisms are detected.
3. The first selection occurs before the specimen exists
A resistance dataset begins with decisions about who receives care and who is tested. People with mild illness, limited access, financial barriers or successful empirical treatment may never produce a specimen. People with severe illness or treatment failure may be more likely to do so. The resulting laboratory collection can therefore overrepresent particular clinical situations.
ECDC's explanation of EARS-Net explicitly warns that blood-culture frequency and diagnostic practices affect case ascertainment. It also notes that taking cultures mainly after empirical treatment fails can overestimate resistance percentages by excluding susceptible infections from the denominator. That is a direct example of a sampling mechanism capable of distorting an apparently straightforward proportion. ECDC surveillance methodology
Imagine a hypothetical clinic that tests only patients who remain unwell after initial treatment. Its resistant fraction may be high because the sampling rule selects harder-to-treat cases. Another clinic that cultures a broader group at the beginning of care could report a lower fraction without necessarily having a safer underlying population.
The comparison requires information about testing practice, not an assumption that both percentages represent the same population. A useful surveillance report would describe the clinical indications for sampling and major changes in those practices over time.
This also explains why expanding access to diagnostic testing can change a trend. The new results may reveal previously unseen susceptible infections, resistant infections or both. The direction of the percentage change depends on what enters the denominator. A program should not be judged solely by whether its resistance percentage immediately falls after it improves testing. Better observation can alter the picture before prevention changes the underlying burden.
4. An isolate is not automatically a new infected person
A laboratory isolate is material recovered and characterized from a specimen. One person can contribute multiple specimens, and the same infection can generate repeated isolates over time. If every isolate is counted equally, people who undergo more testing can contribute disproportionately to a surveillance summary.
The EARS-Net annual report for 2024, published in November 2025, states that its analysis de-duplicates to the first isolate per patient, year and bacterial species. That is a specific operational rule for that analysis. It should not be assumed to match every other surveillance system or every clinical question. ECDC 2024 annual-report methods
A hypothetical patient with five repeat resistant cultures and another patient with one susceptible culture would produce an isolate-level resistant fraction of five out of six, or about 83 percent. Counting one observation per patient would produce one out of two, or 50 percent. Neither arithmetic result is mysterious; they use different units.
De-duplication itself involves choices. A rule that counts only the first isolate in a long interval may miss a later distinct episode or a changed resistance profile. A rule that counts every episode requires a defensible episode definition and enough information to apply it. The appropriate choice depends on the purpose of the surveillance.
The report should state the rule clearly enough for comparison. If one dataset counts first isolates per year and another counts every positive specimen, their percentages should not be treated as equivalent without adjustment or qualification. The word cases is particularly risky when it obscures whether the count refers to people, episodes, specimens or isolates. A precise denominator begins by naming the unit being counted.
5. Colonization and infection require different interpretation
A resistant organism can be present without causing the illness under investigation. Surveillance based on laboratory positives alone may not distinguish colonization from infection, especially when screening programs deliberately look for carriage. That distinction affects how the numbers relate to disease burden.
CDC's discussion of multidrug-resistant-organism prevention notes that measures based solely on positive cultures without accompanying clinical information do not distinguish colonization from infection and may not fully demonstrate disease burden. The page presents material from the 2006 management guideline; its historical provenance should remain visible rather than being mistaken for a newly issued 2024 research finding. CDC MDRO surveillance discussion
A hypothetical hospital might expand screening at admission. The number of detected carriers could rise even if the number of clinical infections remains unchanged. That rise can be useful information for prevention, but it should not be reported as an equivalent rise in people made ill by the organism.
The opposite mistake is to dismiss colonization data as unimportant because it does not measure current illness. Carriage can be relevant to transmission and infection-control planning. The point is to describe what the surveillance was designed to detect and avoid converting one outcome into another.
A clear public report might separate screening detections, clinical infections and uncertain classifications. It would explain whether a change reflects a new screening policy. That allows readers to understand why a prevention program can detect more organisms while working toward fewer infections. A single undifferentiated count would obscure that relationship and could discourage the very testing needed to understand the problem.
6. Resistance categories depend on an interpretive framework
Susceptibility testing produces measurements that are interpreted using criteria appropriate to the organism, drug and clinical context. A categorical result is not simply a timeless label stamped on a bacterium. The standards and version used to interpret the measurement matter.
EUCAST's definitions distinguish susceptible at standard dosing, susceptible with increased exposure and resistant. Its page explains that the definitions of S, I and R changed in 2019. In particular, the I category means susceptible with increased exposure, not a generic halfway point between susceptibility and resistance. This is a description of the classification framework, not a recommendation that a reader change a dose. EUCAST category definitions
A hypothetical analyst who combines I and R into a single resistant category without checking the definition could misstate a published result. Another analyst comparing data across a change in standards might interpret a categorical shift as biological change when some of it reflects revised interpretation.
For surveillance, the useful record includes the testing method, interpretive standard and version. Where possible, retaining quantitative measurements can support later reanalysis, although the availability and quality of such measurements vary.
The practical lesson is to inspect the legend before interpreting the color. A resistance map is the end of a chain of measurement and classification. Two maps can use similar colors while applying different definitions. Good reporting explains those definitions and identifies discontinuities rather than assuming that every historical category means exactly what its current name suggests.
7. A breakpoint is linked to a clinical claim
A minimum inhibitory concentration is a laboratory measurement under specified conditions. A clinical breakpoint translates measurements into categories informed by evidence about achievable drug exposure and likely treatment effectiveness. The relationship is important because a laboratory value and a treatment outcome are not the same observation.
FDA's susceptibility-interpretive-criteria page describes categories in relation to concentrations achievable with recommended regimens and the likelihood of clinical efficacy. It also describes categories whose interpretation depends on exposure. The page should be read as a framework for professional interpretation, not as a self-treatment guide. FDA interpretive criteria
A hypothetical organism-drug result near a breakpoint can be sensitive to measurement variability or a revised standard. That does not mean the laboratory is unreliable. It means categorical summaries can make a continuous measurement appear more absolute than it is.
For a trend analysis, a change in the proportion classified resistant should be examined alongside any breakpoint revisions. If historical quantitative data are available and suitable, analysts may be able to apply a common definition. If not, the report should explain the limitation rather than draw an uninterrupted line that implies perfect comparability.
The same principle applies across jurisdictions using different standards. Harmonization can improve comparison, but a superficial recoding of labels may not reproduce the underlying interpretation. A reader should ask whether the methods genuinely align.
The clinical boundary remains clear: aggregate surveillance supports population understanding and professional decision-making, but it does not determine the correct treatment for an individual without the relevant clinical information. A public article should explain the measurement without turning it into a prescription.
8. Laboratory methods can change the observation process
Surveillance systems may use phenotypic testing, genomic information or combinations of methods. Each can provide valuable evidence, but they do not necessarily observe exactly the same property. A resistance-associated genetic feature and a measured susceptibility phenotype can require careful reconciliation.
CDC's Active Bacterial Core surveillance isolate page describes a historical transition in susceptibility information for its specified pathogens: testing was solely phenotypic before 2015, used both approaches in 2015 and has primarily inferred minimum inhibitory concentrations from whole-genome sequencing since 2016, with selected conventional testing. This is a program-specific description, not a claim about all US laboratories. CDC ABCs isolate methods
The reporting lesson is that a time series can contain a methodological transition. Whether the transition affects comparability depends on validation and the particular organism-drug combinations. The existence of a new method does not automatically invalidate the series, but it warrants documentation.
A hypothetical surveillance network adopting a genomic assay should explain what targets it detects, how results are interpreted and how discordant findings are investigated. A negative result for a limited genetic panel cannot automatically exclude every possible resistance mechanism. Conversely, detecting a gene does not by itself measure every clinical consequence associated with its presence.
Public summaries should avoid treating all laboratory technologies as interchangeable tests for a single abstract property. The appropriate question is what each method establishes and how it was validated for the surveillance purpose. This allows genuine improvements in speed or coverage to be recognized without erasing the boundaries of the measurement.
9. The tested-drug denominator can be selective too
Not every isolate is tested against every antimicrobial. Laboratories may use panels, staged testing or selective reporting based on organism identification and earlier results. The set tested against a particular drug can therefore differ from the set tested against another.
Imagine a hypothetical laboratory that tests a reserve drug mainly when an isolate resists several common alternatives. The resistance fraction for that reserve drug would describe a selected group of difficult isolates. Comparing it directly with the fraction for a drug tested routinely in all isolates could produce a misleading ranking.
A report should state how many eligible isolates were tested for each organism-drug combination and whether testing was conditional. Missing susceptibility results should not be silently treated as susceptible. They are missing observations, and their pattern can matter.
The distinction between testing and reporting also matters. A result may exist in the laboratory system but be suppressed from a clinical display or absent from the surveillance extract. The reason could be a stewardship policy, a technical interface or a data-quality rule. Those explanations have different implications for completeness.
For an analyst, a useful first step is a table of denominators by drug, specimen source and time period. Sudden changes can reveal panel updates or reporting changes before they are mistaken for biological trends. The table is not a complete solution to selection bias, but it makes the observation process visible.
A percentage without its tested denominator invites overinterpretation. Showing both does not burden readers unnecessarily; it gives them the information needed to distinguish a broad estimate from a result based on a small, selected subset.
10. Resistance proportion and infection incidence can move differently
The proportion of tested isolates that are resistant and the population incidence of resistant infection answer different questions. A proportion uses tested isolates as its denominator. Incidence uses a population or exposure-time denominator and a defined event count. Confusing them can reverse the apparent meaning of a trend.
Consider a hypothetical region with 100 resistant and 900 susceptible infections in one year. The resistant proportion is 10 percent. In the next year, resistant infections remain at 100 while susceptible infections fall to 400. The resistant proportion rises to 20 percent even though the number of resistant infections has not increased.
Now reverse the situation. If resistant infections rise from 100 to 150 while susceptible infections rise from 900 to 1,850, the resistant proportion falls from 10 percent to 7.5 percent. A lower fraction can coexist with a higher count of resistant infections. These examples assume complete comparable ascertainment solely to isolate the arithmetic.
The appropriate metric depends on the question. A resistance proportion can inform understanding of the organisms encountered among tested cases. An incidence measure can better address the burden of resistant infection in a population, provided the coverage and case ascertainment support it. Both may be necessary.
A good report names the metric in the headline or nearby explanation. It should not use the word rate ambiguously when the denominator is actually tested isolates. If a trend concerns a proportion, the article should ask whether the resistant numerator, susceptible denominator or both changed. That small discipline prevents a mathematically correct percentage from telling the wrong public-health story.
11. Population coverage is an estimate that needs a method
A surveillance network may include hospitals serving overlapping regions, referral centers receiving patients from distant areas and laboratories that process specimens for multiple facilities. Estimating the population covered can therefore be more complicated than drawing a circle around each participating hospital.
WHO's 2023 GLASS manual describes the objectives and methodology for surveillance of resistance in selected bacteria causing common human infections. Its role is to support standardized implementation rather than assume that every national system begins with the same coverage or infrastructure. WHO GLASS manual
A hypothetical national report might say that participating facilities serve half the population. That figure needs an explanation: does it refer to geographic catchments, healthcare utilization, laboratory service areas or another measure? If facilities overlap, simply adding their catchment populations can double-count people. If a referral hospital attracts severe cases from outside its area, geographic coverage alone may not describe its case mix.
Coverage estimates also change over time. A new laboratory may add a previously unobserved region or mainly duplicate an existing service area. The number of participating sites is therefore not a direct substitute for population coverage.
An incidence estimate that scales observed cases by coverage should communicate uncertainty in that scaling. If the unobserved population differs systematically from the observed population, proportional expansion can remain imperfect even when the arithmetic is correct.
The practical request is a coverage method and its limitations, not a demand for impossible precision. A transparent approximate estimate is more useful than an exact-looking denominator whose construction cannot be examined. The denominator is part of the evidence, and it deserves the same attention as the resistant isolates in the numerator.
12. Referral networks can concentrate difficult cases
Specialist laboratories and tertiary hospitals often receive cases selected for complexity or unusual findings. Their data can be excellent for detecting emerging mechanisms and supporting difficult investigations. That same selection can limit their use for estimating routine population prevalence.
A hypothetical reference laboratory might receive isolates because local testing already suggested an unusual resistance pattern. If 80 percent of those referrals confirm resistance, the fraction describes the referral stream. It does not establish that 80 percent of infections in the surrounding population are resistant.
This distinction is particularly important when a network expands its ability to receive unusual isolates. A rise in confirmations may reflect successful detection and referral rather than a proportional increase in population burden. It can still warrant action, especially if the finding is novel or consequential, but the action should be justified by the kind of evidence actually collected.
A public dashboard could separate routine sentinel surveillance from targeted referral testing. It could describe inclusion criteria and whether submissions are driven by local priorities, outbreaks or standardized sampling. Without that separation, readers may combine incompatible data sources into a single national trend.
The broader lesson is that usefulness does not require representativeness for every purpose. A targeted system can be highly valuable precisely because it concentrates on unusual events. The mistake is to ask it to provide a population estimate it was not designed to produce. Good communication recognizes the system's strength while stating the boundary of inference, rather than treating all laboratory counts as interchangeable measures of how common resistance is.
13. A changing network can create a changing average
Surveillance networks grow, contract and change composition. If new laboratories serve populations with different resistance patterns, a pooled percentage can move even when no participating laboratory experiences a change within its own population.
Suppose two hypothetical hospitals each test 1,000 eligible isolates. Hospital A reports 10 percent resistance and Hospital B reports 30 percent. The combined fraction is 20 percent. If Hospital B stops reporting, the network fraction becomes 10 percent even though Hospital A's underlying pattern is unchanged. A chart showing only the pooled number could appear to demonstrate a dramatic improvement.
A stable-site analysis can help examine this problem by comparing a consistent set of reporting laboratories. It has its own limitation: restricting to stable sites may exclude newly observed populations and reduce coverage. The full-network and stable-network views therefore answer complementary questions rather than competing for the title of correct result.
The report should also show changes in test volume. A laboratory contributing twice as many isolates gains more influence in a pooled isolate-level proportion. That may reflect greater disease burden, expanded sampling or a service reorganization.
A useful trend explanation separates changes within reporting units from changes in the composition of those units. It can identify major additions, losses and disruptions. This does not require exposing private patient data. It requires preserving the operational history needed to interpret the aggregate.
Readers should be wary of both celebratory and alarming narratives that ignore network composition. A trend may be real, but the evidence should demonstrate that it is not merely the consequence of observing a different collection of facilities.
14. Hospital case mix complicates league tables
Hospitals care for different populations. Some receive more critically ill patients, more transfers, more complex surgery or more people with prior antimicrobial exposure. A raw resistance percentage can reflect those differences as well as local infection-prevention performance.
A hypothetical referral center could have a higher resistant fraction than a community hospital because it treats patients whose infections were already difficult elsewhere. That does not prove its prevention program is poor. Conversely, a low fraction does not automatically establish excellent practice if testing is limited or high-risk patients are transferred out.
Risk adjustment can help with specified differences when suitable data exist, but it does not eliminate every source of confounding. The variables included, missing information and model assumptions should be clear. An adjusted estimate is an analytical result, not a direct observation free from uncertainty.
The comparison should also distinguish infections present on admission from those detected later under the surveillance definition. Timing categories can support investigation, but they are not perfect proof of where an organism was acquired. A patient may have undetected carriage before admission, and diagnostic timing can influence classification.
For public accountability, the useful question is what the measure was designed to assess. A hospital ranking should not imply causation unless its design supports that claim. More informative reporting may combine resistance patterns with testing practices, patient mix and prevention-process measures.
This approach does not excuse poor performance. It makes criticism more defensible by tying it to evidence that distinguishes a facility's actions from the population it serves. Fair comparison is necessary if institutions are to learn from the data rather than avoid testing or referrals that could worsen an oversimplified score.
15. Missing data can change the conclusion without changing a single result
A surveillance table may contain complete susceptibility results for some organisms and drugs but sparse information for others. Missing age, specimen source or facility details can further limit interpretation. These gaps are not neutral simply because they appear as blank cells.
Imagine a hypothetical analysis of 1,000 isolates in which 200 lack results for the drug of interest. If missingness is unrelated to resistance after accounting for the design, the observed fraction may be reasonably informative. If the missing results are concentrated in laboratories with limited testing capacity or in isolates requiring special work, the bias can be more substantial.
The first responsibility is to report completeness. The second is to investigate why information is missing. A statistical imputation may be appropriate for some purposes, but it introduces assumptions that should be stated. It cannot transform an unobserved process into direct evidence.
A useful sensitivity analysis asks how plausible alternative assumptions affect the conclusion. If the direction of a trend changes under modest assumptions about missing data, the report should be cautious. If the conclusion remains stable across reasonable scenarios, confidence can increase without pretending the gaps disappeared.
Readers should also distinguish no resistance detected from no usable data. A blank map region should not inherit the meaning of a low-resistance color. Institutions should make that distinction visually and verbally.
The public-health cost of missingness includes invisibility: places with the weakest measurement can appear least affected. Reporting completeness beside the resistance estimate helps prevent resources from being directed only toward locations capable of documenting their problem most thoroughly.
16. Antimicrobial use is a related measure with a different denominator
Resistance surveillance and antimicrobial-use surveillance inform each other, but they do not measure the same thing. A resistance percentage describes properties of observed organisms under a testing framework. A consumption measure describes quantities of medicines used or supplied under its own definitions and coverage.
WHO's Global Health Observatory explanation of GLASS describes the system as collecting official national data on antimicrobial resistance and use in humans. The combination supports a broader evidence base, but it does not mean that a simple correlation between national consumption and resistance proves the effect of a particular prescribing policy. WHO GLASS overview
A hypothetical country could reduce measured antibiotic consumption while resistance remains high because of transmission, historical selection, changes in the drugs used or other factors. Another could show rising consumption as access to necessary treatment improves. The interpretation requires more detail than a single total.
Consumption measures also vary in what they capture. Sales, dispensing and administration are different stages. Hospital and community data may have different completeness. A standardized unit can support comparison while still leaving differences in age, dosing and indication.
For evaluating a stewardship intervention, the question should be specific: which use changed, in which patients, over what period, and with what clinical and microbiological outcomes? Reducing unnecessary use is different from reducing all use regardless of need.
The article's role is to keep those distinctions visible, not to advise an individual to start or stop medication. Public-health analysis benefits from linked evidence, but linkage does not remove the need for a defensible causal design when attributing a resistance change to a policy.
17. Trends need enough time and a plausible comparison
Resistance can fluctuate because of outbreaks, seasonal case mix, changes in testing or random variation. A single annual increase may be important, but it does not automatically establish a sustained trend. The amount of evidence depends on counts, consistency and the structure of the surveillance.
A hypothetical small laboratory testing twenty isolates could move from two resistant results to four. The proportion doubles from 10 to 20 percent, but the estimate is based on few observations. A large network showing a similar proportional change across many stable sites would support a different degree of confidence.
Uncertainty intervals help communicate sampling variability, though they may not capture every source of bias. A narrow interval around a selected sample's proportion does not make that sample representative. Precision and validity remain separate properties.
Interrupted time-series or comparison-group designs can help evaluate interventions when their assumptions are appropriate. A simple before-and-after comparison may be confounded by simultaneous changes in infection prevention, diagnostics, patient mix or reporting. The analysis should explain which alternative explanations it can address and which remain.
A responsible news account should separate an observed rise, a statistically supported trend and a causal attribution. These are progressively stronger statements. It should also identify whether a report revises earlier data, because late submissions or corrected classifications can change historical estimates.
The strongest interpretation combines numerical analysis with operational knowledge of the network. An unexplained discontinuity deserves investigation before becoming a story about sudden biological change. At the same time, methodological caution should not become a reason to ignore a credible signal while waiting for impossible certainty.
18. Laboratory capacity is part of the public-health intervention
Reliable resistance surveillance requires more than a reporting website. It depends on specimen collection, transport, culture or other detection methods, identification, susceptibility testing, quality assurance and data management. Weakness in any stage can affect which infections become visible and how results are interpreted.
A hypothetical investment in laboratory capacity might initially increase reported resistant infections because previously missed cases are now identified. That increase can be a sign of better detection rather than immediate deterioration. Evaluation should therefore include measures of testing access, turnaround time and data completeness alongside the disease indicators.
Quality assurance matters because systematic measurement differences can distort comparisons. Laboratories need appropriate controls and procedures for unusual or discordant results. Public reports need not reproduce every technical record, but they should describe participation in relevant quality systems and major limitations affecting the data.
Workforce and supply continuity are also consequential. A laboratory without reagents for part of the year may contribute an uneven sample. A reporting backlog can make a recent period look artificially quiet. These are operational problems with statistical consequences.
The most useful funding question is therefore what additional decision-quality information the investment will produce. More tests can improve coverage, but only if the resulting data are interpretable and reach people able to act. Better surveillance should connect to prevention, appropriate treatment systems and investigation capacity.
A dashboard alone cannot accomplish those tasks. It can make the evidence accessible, but the evidence must first be produced through a maintained clinical and laboratory network. Treating that network as infrastructure helps explain why long-term capacity matters even when the immediate headline concerns a single alarming percentage.
19. A responsible public table should expose its boundaries
An antimicrobial-resistance table can remain readable while making its limitations explicit. At minimum, a reader should be able to identify the organism, drug, specimen type, data period, count tested and category definition. The report should also explain de-duplication, participating sites and major changes in testing or reporting.
An editorially proposed reading sequence begins with the numerator: what exactly was counted as resistant? It then examines the denominator: which observations were eligible and tested? Next comes the sampling frame: who could enter the dataset and through which facilities? Finally, it considers comparability: did methods, breakpoints or coverage change?
For a hypothetical regional release, a concise main table might present the proportion and tested count, with a linked methods sheet explaining selection and coverage. A separate completeness display could show missing drug results and site participation. This structure would let general readers understand the headline while allowing technical readers to assess the inference.
The report should avoid false precision. A percentage reported to several decimal places may look authoritative even when ascertainment is uncertain. Appropriate rounding and uncertainty communication are more honest than numerical ornament.
It should also resist causal labels unsupported by the design. A region with high resistance deserves attention, but the table alone may not identify why. A hospital with an increase deserves investigation, but the increase alone may not prove a particular lapse.
Clear boundaries make the data stronger, not weaker. They allow a reader to see what is established, what is estimated and what requires another source. That is the difference between a statistic that merely alarms and evidence that can guide a proportionate response.
20. Better surveillance changes what becomes visible
The central lesson is that resistance statistics are produced through a sequence of clinical and laboratory decisions. People reach care, specimens are collected, organisms are identified, susceptibility is assessed, duplicates are handled and results are combined. Each stage affects the population represented by the final number.
A responsible interpretation follows that sequence. It distinguishes resistance proportion from infection incidence, colonization from disease, referral testing from routine surveillance and a changed breakpoint from a changed organism population. It preserves the data year when citing a later publication. It asks whether a trend reflects a stable observation process or a network that is still changing.
These distinctions should not weaken the case for action against antimicrobial resistance. They improve it. A system that understands its blind spots can invest in the places where evidence is missing, compare institutions more fairly and assess whether interventions change the outcomes they are intended to change.
The alternative is a cycle of overclaim and dismissal. A dramatic percentage is presented as a complete population picture; critics identify a limitation; the entire dataset is then treated as worthless. Both steps are avoidable. Imperfect surveillance can be highly informative when its design is understood and its uncertainty is communicated honestly.
The number that matters is therefore not only the resistant fraction on the chart. It is also the number of people and infections the system could observe, the tests it performed and the definitions it applied. Following those denominators turns a headline about superbugs into a more useful account of where the evidence is strong, where the threat remains hidden and what better measurement could make possible.