Counting people properly, and what the count is allowed to mean
Clinical work treats the patient in front of you. Epidemiology asks what is happening to everybody, which is the only way to know whether a problem is growing, whether an intervention worked, whether an outbreak is occurring, and which of the many things that could be done would save the most lives for the money available. Nurses generate most of the raw material this subject runs on — every notification, every cause of death recorded, every immunisation counted, every fluid chart totalled — and they also have to read its conclusions, because screening programmes, vaccination schedules and public health advice are all built from it. A nurse who cannot tell a rate from a count, or an association from a cause, is at the mercy of whoever is claiming something.
Clinical reasoning asks what is wrong with this person and what will help them. Epidemiological reasoning asks how much of this exists, who gets it, whether it is increasing, and what would reduce it across everybody. Both are necessary and they frequently give different answers, because what is best for an individual patient and what is best for a population are not the same question and are not always compatible.
This subject runs on data that nurses generate: notifications of infectious disease, causes recorded on death certificates, immunisations entered, births and weights recorded, incidents reported, and charts totalled. When those are recorded carelessly under pressure, the error does not stay local — it becomes a district figure, then a national one, and eventually a policy built on something that was never true.
Nurses are also consumers of this subject. Screening programmes, vaccination schedules, dietary advice and public health campaigns are all built from population data, and staff are expected to explain and defend them to patients who ask. Understanding how the underlying claim was reached is what makes that explanation honest rather than a recitation.
Almost all descriptive epidemiology answers three questions: who, where and when. Who is affected by age, sex, occupation and circumstance; where the cases are; and how the pattern has changed over time. Those three, applied to almost any health problem, produce most of the useful hypotheses about what is causing it.
Forty cases in one district and forty in another are equivalent only if the districts are the same size. Without a denominator — the population at risk — a count cannot be compared with anything, and comparing raw counts between places or between years is one of the commonest errors in reporting. The first question about any number is always: out of how many.
The denominator should include only people who could actually have the outcome. Cervical cancer rates calculated over a whole population including men, or rates of a condition of old age calculated over everybody, produce figures that are arithmetically correct and clinically meaningless. Choosing the denominator is where most honest error and most dishonest manipulation happen.
A rate expresses events per population per period — cases per thousand people per year. Without the time element, two figures cannot be compared if they cover different lengths of time, and a figure quoted without its period is uninterpretable. This sounds pedantic until you compare a monthly figure with an annual one and conclude that something has improved twelvefold.
A population with more older people will have more cancer and more heart disease regardless of anything else, so comparing two populations of different age structure directly is misleading. Age-adjusted figures remove that effect so that a real difference can be seen, and when a report compares places or periods without adjustment it is usually worth asking why.
Incidence counts new cases occurring in a period; prevalence counts all cases existing at a point. They answer different questions: incidence tells you about risk and about whether something is spreading, prevalence tells you about burden and about how much service is needed. Confusing them is the single most examined error in this subject.
Prevalence depends on how many people get a condition and on how long they live with it. A treatment that prevents death without curing the disease increases prevalence, which looks like deterioration and is the opposite. This is precisely what has happened with several chronic infections and cancers, and reading the rise as failure is a serious misinterpretation.
For planning services, staffing and supplies, prevalence is what matters — how many people need care now. For investigating a cause, evaluating prevention or detecting an outbreak, incidence is what matters, because it reflects what is happening currently rather than the accumulated past.
Neither measure counts the disease; both count the disease that has been found. Introducing a screening programme, a new test or better access will raise measured incidence and prevalence without any change in the underlying disease, and this artefact has been mistaken for an epidemic more than once.
A mortality rate is deaths in the whole population; a case fatality proportion is deaths among those who had the disease. A condition can have a very high case fatality and a low mortality rate because it is rare, and a condition with modest case fatality can dominate mortality because it is common. Mixing them up produces wildly wrong impressions of danger.
Infant mortality, under-five mortality and maternal mortality are used as indicators of a health system's condition rather than only as measures of those particular deaths, because they are sensitive to nutrition, sanitation, access and the quality of care around birth. This is why they appear in every international comparison.
What appears on a death certificate is a judgement, made under pressure, by somebody who may not have known the patient, using a classification that forces a single chain of causation onto a person who had several conditions. In much of the world most deaths are never certified at all. National figures built from this are estimates, and treating them as measurements overstates what is known.
Comparing total deaths in a period with the number expected from previous years gives a figure that does not depend on anybody attributing a cause. It is the most robust way of measuring the impact of an epidemic, a heatwave or a disaster, precisely because it sidesteps the weakest part of the data.
Surveys measure a population at a point and answer how common something is, what people do, and what they believe. They cannot establish sequence, because exposure and outcome are recorded together, and a survey showing that people with a condition behave differently cannot say which came first.
A cohort study identifies people, records their exposures, and waits. Because exposure is recorded before outcome, sequence is established, and because outcomes are counted in both exposed and unexposed groups, risk can be compared directly. The cost is time, expense and losing people along the way, and the people lost are rarely a random sample.
A case-control study begins with people who have the condition and a comparison group who do not, then looks back at exposures. It is efficient for rare conditions and for long delays between cause and effect, and it is vulnerable to people recalling their past differently once they are ill, and to the comparison group being poorly chosen.
Ecological studies compare aggregate figures between places or periods — average intake against average disease rate. They are cheap, use existing data, and cannot support conclusions about individuals, because a correlation between country averages says nothing about which individuals within them were affected. Making that leap has a name and is a standard examination question.
Trials allocate an intervention, and community trials allocate it to whole villages, schools or clinics rather than to individuals, which is necessary when the intervention affects everybody nearby. They are the strongest design available for establishing effect, and they are frequently impossible for the exposures that matter most.
Comparing the risk in an exposed group with the risk in an unexposed group gives a ratio: twice as likely, half as likely, no different. This is the standard way an association is reported, and it says how strong the relationship is without saying anything about how common the outcome is or how many people are affected.
Subtracting rather than dividing gives the additional cases produced by the exposure, which is what tells you how much would be prevented by removing it. A doubling of a very rare risk produces almost nothing; a small increase in a very common one produces a great deal. Both figures are needed and only one is usually quoted.
Combining the strength of an association with how common the exposure is gives the share of cases attributable to it across the whole population. This is the figure that decides where public health effort goes, and it explains why a weak risk factor that almost everybody has can matter more than a strong one that is rare.
A measure applied to a whole population usually prevents more cases than one aimed at the highest-risk individuals, because most cases arise in the large number of people at moderate risk rather than the small number at high risk. It also means most people who accept the measure receive no benefit themselves, which is why population prevention is a hard argument to make.
Screening offers a test to people with no symptoms, which changes the ethical position entirely: these are people who were fine until the service approached them. The obligation to be sure of benefit, and honest about harm, is therefore higher than for a test requested by a patient with a complaint.
A sensitive test rarely misses those who have the condition, so a negative is reassuring. A specific test rarely flags those who do not, so a positive is meaningful. Screening tests are usually built to be sensitive, which necessarily produces false positives, and a positive screen is therefore the start of investigation rather than a diagnosis.
When a condition is uncommon, even an accurate test produces more false positives than true ones, simply because there are so many more people without the condition. This is counterintuitive, it is the reason mass screening for rare conditions causes net harm, and it is the single most useful statistical idea for explaining a worrying result to a frightened patient.
Screening produces anxiety in people who turn out to be well, investigation with its own risks, and overdiagnosis — finding disease that would never have caused harm, and then treating it. Overdiagnosis is invisible to the individual, who believes their life was saved, which makes it particularly hard to discuss and particularly important to measure.
Finding a disease earlier lengthens the measured time from diagnosis to death even if the date of death does not move, and including slow-growing cases that would never have progressed improves the figures further. Both effects make screening look effective in survival statistics while changing nothing, which is why screening is judged on mortality rather than survival.
Surveillance is the continuous collection and use of health data to guide action, and the last part is what distinguishes it from record keeping. A system that gathers notifications nobody reads is not surveillance. The purpose is to detect change early enough to respond, which means timeliness matters more than completeness.
Every country requires certain diseases to be reported, and the list reflects what that country considers capable of spreading or signalling a failure. Notification is frequently a legal duty falling on whoever identifies the case, which in practice is often a nurse, and under-notification is the normal state of affairs almost everywhere.
Passive surveillance waits for reports to arrive and is cheap and incomplete. Active surveillance goes and looks — contacting clinics, checking registers, visiting — and is expensive and far more sensitive. Systems switch from passive to active when something is suspected, which is why an outbreak appears to grow rapidly at the moment somebody starts looking properly.
A rise in reported cases may mean more disease, more testing, better reporting, a changed definition, or a single enthusiastic clinician. Before concluding that an epidemic is occurring, the question is always whether anything changed about the counting, and that question is skipped constantly.
An outbreak is more cases than expected, linked in place and time, and it is usually recognised by somebody at ward or clinic level noticing that this is the third similar case this week and saying so. Detection depends far more on an individual joining the cases together than on any system, particularly outside large institutions.
Confirm the cases are real, define precisely what counts as a case, find them all, describe them by person, place and time, form a hypothesis about the source, test it, and act. Control measures usually start before the investigation finishes, because waiting for certainty costs cases and the early measures are often broadly protective anyway.
A definition that is too narrow misses cases and hides the outbreak; one that is too broad includes unrelated illness and points at the wrong source. Definitions frequently change during an investigation as understanding improves, and any figure quoted must be read against the definition in use when it was counted.
Plotting cases against the date of onset shows the shape of the outbreak, and the shape suggests the mechanism: a sharp single peak indicates a common source over a short period, a prolonged plateau indicates continuing exposure, and successive waves indicate spread from person to person. This is a genuinely useful piece of reasoning available from a simple chart.
Cohorting patients, maintaining precautions consistently when they are inconvenient, and keeping meticulous records of who was where and when. The movement record is frequently what identifies the source, and it exists only if somebody kept it accurately at the time rather than reconstructing it afterwards.
A hospital outbreak has features a community one does not: the population is unusually susceptible, they are concentrated in small spaces, staff move between them constantly, and the organisms circulating have been selected for resistance by the antimicrobials used there. Detection depends on somebody comparing today's isolates with last month's, which in many hospitals nobody is doing, and the movement records that identify the source exist only if they were kept accurately at the time rather than reconstructed afterwards.
The average number of further cases produced by one case in a fully susceptible population determines whether an infection grows or dies out. Above one it grows; below one it fades. Everything done to control an infection — vaccination, isolation, distance, treatment, hygiene — works by pushing that number below one, which is a useful way to understand why several partial measures together can succeed.
When enough people are immune, the chain of transmission breaks and those who cannot be vaccinated are protected indirectly. The proportion required rises with how transmissible the infection is, which is why some diseases need very high coverage and others do not, and why a small fall in coverage can produce a sudden outbreak.
The interval between exposure and illness determines how long contacts must be watched, how quickly an outbreak becomes visible, and whether intervention after exposure can work at all. It is the first thing established about an unfamiliar infection, because the practical response cannot be designed without it.
Many infections produce mild or symptomless cases that never reach a clinic, so reported cases are a fraction of real ones and the fraction changes with testing. This is why case fatality estimates fall as an epidemic progresses and testing widens, without the disease becoming any less dangerous.
A figure given to two decimal places carries an impression of exactness that the underlying data frequently cannot support, particularly where deaths are uncertified and births unregistered. Modelled estimates are legitimate and useful, and presenting them without their uncertainty is where the misleading happens.
Almost any trend can be made to look like an improvement or a disaster by selecting the starting year, the comparison area or the age group. When a claim rests on a comparison, the question is who chose it and what a different reasonable choice would have shown.
A national average conceals that one group has improved substantially and another has got worse, and this is the normal pattern rather than an exception. A programme can improve the average while widening inequality, and reports that give only the average are unable to show it.
Anything measured when it is unusually high will tend to be lower next time simply by chance, which makes any intervention applied at a peak look effective. This accounts for a large number of apparently successful initiatives launched after a bad year, and it is why comparison groups exist.
Civil registration is the foundation of health statistics and is incomplete across much of the world, with a large share of deaths never recorded and many births unregistered. Where registration is weak, national figures come from surveys and models rather than counts, and every figure built on them inherits that uncertainty.
Attendances, admissions, immunisations and notifications are collected as a by-product of care, which makes them cheap and biased towards people who reached a service. They describe the served population rather than the whole one, and using them as if they described everybody systematically hides those who never arrived.
Household surveys reach people who never attend a service and are the main source of reliable population data in many countries. They are expensive, occasional, and dependent on sampling properly, and they miss people without a fixed household — the homeless, the displaced, those in institutions, and mobile populations.
Every one of these sources begins with somebody writing something down during a busy shift. Ethnicity recorded as assumed rather than as stated, a cause of death chosen for convenience, a chart not totalled, a notification not sent — each becomes a national figure eventually, and nobody downstream can tell that it was wrong.
Every additional field on a form is time taken from care, and a system that demands a great deal of recording from people who never see any result from it will be completed carelessly. Data quality is therefore an organisational question rather than a matter of individual diligence, and the single most effective way to improve it is to show the people collecting it what their own figures said and what changed because of them.
A programme is not evaluated by whether it was delivered but by whether the thing it was meant to change actually changed, and by whether that change can reasonably be attributed to it rather than to something happening anyway. Most health programmes are reported on activity — sessions held, leaflets distributed, people reached — because activity is easy to count and outcome is not, and the substitution is rarely stated openly.
Without a comparison, an improvement means nothing, because almost everything changes over time for reasons unrelated to any intervention. A comparable area that did not receive the programme, the same area before and after with a long enough run of prior data, or a staged roll-out where later areas act as controls for earlier ones, all provide the comparison that a simple before-and-after does not.
A programme can fail because it did not reach people, because what was delivered was of poor quality, or because the intervention does not work even when delivered well. These require different remedies and are frequently confused, and distinguishing them means measuring reach and delivery separately from outcome rather than concluding from the outcome alone.
Interventions produce effects nobody planned: a screening programme that frightens people away from a clinic, an incentive that shifts effort from unmeasured work, a campaign that increases stigma, a free service that displaces an existing one. An evaluation that measures only the intended outcome will report success while these accumulate, and asking what else might have changed is a routine part of honest evaluation.
A confounder is associated with both the exposure and the outcome and produces an apparent relationship between them. Studies finding that coffee drinkers had more lung disease were describing smoking; studies finding that people who take vitamins live longer are frequently describing income, education and the sort of person who takes vitamins. Recognising this pattern is the most transferable skill in the subject.
People in employment are healthier than the general population because illness removes people from work, so comparing workers with everybody makes any workplace exposure look protective. Similarly, people who adhere to a treatment differ systematically from those who do not, in ways that affect outcome independently of the treatment. Both effects are named because both are common and both mislead in a predictable direction.
Data collected in services describe the people who arrived. Those who could not travel, could not pay, were turned away, died before arriving, or avoided the service entirely are absent, and their absence usually makes outcomes look better than they are. The sicker and poorer the excluded group, the larger the distortion, which means the error runs in the same direction almost every time.
Confounding can be reduced by restricting a study to one group, by matching participants, by adjusting statistically for known factors, or by randomising, which handles unknown factors too. Statistical adjustment can only correct for things that were measured, so a paper reporting adjustment is telling you which confounders it knew about and nothing about the rest.
Epidemiology describes; it does not decide. Turning a figure into an action requires judgements about how much benefit is worth how much cost, whose freedom may be restricted, and what a society will accept, and none of those are settled by the data. Presenting a decision as though the numbers compelled it is a way of avoiding an argument that should be had openly.
Primary prevention stops the problem arising — vaccination, clean water, safe roads. Secondary prevention finds it early, which is what screening is. Tertiary prevention limits the damage once it exists, which is most of clinical care. Programmes are frequently argued about because the parties are talking about different levels without saying so.
Measures acting on the conditions themselves — price, regulation, availability, infrastructure — protect everybody without requiring individual effort, and therefore reach the people least able to act on advice. Measures relying on individual behaviour are taken up most by those with the most resources, which is why they can improve the average and widen the gap at the same time.
Public health sometimes restricts individuals for collective benefit — isolation, compulsory notification, restrictions in an epidemic. These are justified by necessity and proportion, they require review and a route of appeal, and they carry real costs to the people restricted. Health workers are frequently the ones enforcing them and should know that they are exceptions requiring justification rather than ordinary practice.
Telling somebody that a risk is zero point four per cent conveys very little; telling them that four people in a thousand are affected conveys more, and showing a thousand figures with four marked conveys most. Using the same denominator throughout a conversation matters as well, because switching between one in fifty and two per cent forces the listener to convert mid-sentence.
Doubling a very small risk is still a very small risk, and a fifty per cent increase in something that affects two people in a hundred thousand is not news. Reporting risk only in relative terms is the standard way of making a finding sound alarming or a treatment sound powerful, and asking for the absolute figures is the standard defence.
Risks that are unfamiliar, involuntary, dreaded, catastrophic or caused by somebody else provoke far more fear than familiar, voluntary, everyday ones that kill many more people. Dismissing a patient's concern as irrational misses this entirely, and acknowledging why something feels worse is more persuasive than restating the number.
Saying that the evidence is limited, that estimates may change, and that reasonable people disagree, is more durable than projecting confidence that later has to be withdrawn. Public trust is damaged far more by a certainty that turns out false than by an acknowledged uncertainty, and health workers are usually the ones who absorb the consequence.
A patient told their screening test is positive usually hears that they have the disease. Explaining that the test is deliberately built to over-call, that most positives turn out to be nothing, and what the next test will establish, converts a diagnosis into a step in an investigation. Giving the actual frequency — of a hundred people with this result, about this many turn out to have anything wrong — does more to settle somebody than any amount of general reassurance, and it has the advantage of being true.
Health data are collected from individuals, frequently without a meaningful choice, and used for purposes they were never told about. Surveillance, research and service planning all depend on this, and the justification is collective benefit rather than individual consent, which means the obligation to hold the data securely and use it only as described is correspondingly stronger.
A figure describing a rare condition in a small district identifies the individual to anybody local, even without a name. This is why statistical outputs suppress small cells, and it is a live problem in any setting where a nurse writes a report about a community they work in.
Deciding which groups a system counts separately determines which inequalities can ever be demonstrated. A country that does not record ethnicity cannot show ethnic disparity; one that combines many communities into a single category cannot show differences between them. These are political decisions with statistical consequences and they are rarely presented as such.
What gets measured gets managed, and what is not measured disappears from the plan. Maternal deaths, deaths at home, violence against women, illness among undocumented people and conditions affecting populations without services are all under-counted, and the under-counting reflects who has been considered worth counting rather than any technical difficulty.
The mean is the arithmetic average and is dragged by extreme values; the median is the middle value and is not. For anything skewed — length of stay, income, waiting time, number of admissions — the median describes a typical person far better, and a report giving a mean for such data is either careless or flattering itself. The spread matters as much as the centre, since two groups with identical averages can be entirely different populations.
A confidence interval gives the range within which the true value plausibly lies. A narrow interval means a precise estimate; a wide one means the study or the data cannot say much. An interval that includes the possibility of no difference means no difference has been established, whatever the headline figure sitting inside it appears to show, and this is the single most useful thing to look at in any results table.
A statistically significant result means the finding would be unlikely to arise by chance alone if there were truly no effect. It says nothing about whether the effect is large enough to matter to a patient or to a service. A significant but trivial difference and a non-significant but substantial one are both common, and both are routinely reported as though the arithmetic had settled the clinical question.
If twenty things are compared, one will typically appear significant by chance. This is why the main outcome should be declared before a study or an evaluation begins, and why a striking result found in a subgroup nobody specified in advance is a suggestion for further work rather than a conclusion. Reports that present the one comparison that worked are extremely common.
A statement that attendance improved by thirty per cent means something entirely different if it rose from ten people to thirteen than from a thousand to thirteen hundred. Percentages calculated on small numbers swing wildly with one extra case, which is why good reporting gives both the proportion and the count, and why a proportion quoted alone should prompt the question.
International estimates of how many people die of what are built from civil registration where it exists, surveys where it does not, verbal accounts from families in many countries, and statistical models that fill the remaining gaps. They are the best available and they are estimates, sometimes revised substantially, and treating them as counts overstates what anybody knows.
Because many conditions disable without killing, population health is also measured in years of life lost combined with years lived with impairment. This allows mental illness, back pain and blindness to appear in comparisons that would otherwise be dominated by fatal disease, and it involves value judgements about how much a year with a given condition is worth, which are made explicitly and are contested.
Differences between countries reflect the completeness of their data as much as their health, and a country with good registration can appear to have more disease than one that records almost nothing. Ranking tables are read as performance and are frequently measuring how well something was counted, which is worth remembering before quoting one.
These figures determine which programmes are funded, where staff are posted, which diseases get attention, and what is taught in curricula. A condition that is common where you work but under-represented in global estimates will be under-resourced, and the under-representation frequently begins with records nobody completed at a bedside.
Resistance is not a property of a patient; it is a property of a population of organisms shaped by how antimicrobials have been used across a whole community, a whole hospital and a whole country, including in agriculture. This is why an individual prescriber's decision feels inconsequential and collectively is not, and why the measures that work are surveillance, stewardship and regulation rather than better judgement in any single consultation.
Which organisms are common and which antimicrobials still work differ between countries, between hospitals and between wards in the same hospital, and they change over time. Empirical treatment guidelines are built from local laboratory surveillance, which means the laboratory reports nurses help generate are directly what the next guideline will be written from.
Measuring deaths from resistant infection is genuinely difficult, because the counterfactual — what would have happened had the organism been susceptible — cannot be observed. Estimates therefore vary widely and are model-based, which is a reason to read specific figures cautiously and not a reason to doubt the direction, which is consistent across every method used.
Taking the specimen before the first dose, giving doses at the times prescribed, noticing treatment that has continued without review, asking whether an intravenous route is still needed, and reporting resistant organisms into whatever surveillance exists. Each is small, each is countable, and together they are a substantial part of what any stewardship programme actually consists of.
Incidence against prevalence, rates against counts, sensitivity against specificity, matching a design to a question, and the steps of outbreak investigation. Almost all of it is interpretation in short scenarios rather than calculation, and the wording of the question usually contains the distinction being tested.
Reading rising prevalence as worsening disease when better survival explains it; comparing counts between populations of different size; treating a positive screen as a diagnosis; concluding about individuals from a comparison of country averages; assuming a rise in reported cases means a rise in disease.
Take any health figure in a newspaper and ask what the denominator was, over what period, compared with what, and whether the comparison was chosen or given. Ten minutes on a real claim teaches more than an hour of definitions, and the same four questions work on almost everything.
Ask out of how many. Notify what you are required to notify. Record ethnicity as the person states it. Total the chart. Say aloud when you notice the third similar case this week. None of these require statistical training, and together they are most of what nursing contributes to this subject.