Measures of
Disease Frequency
Fundamental Epidemiological Concepts and Approaches
Learning objectives for this lesson:
- Distinguish counts, proportions, odds, and rates by their numerators, denominators, ranges, and units, and explain why a count alone is of limited use.
- Choose between an incidence risk, an incidence rate, and a prevalence for a described situation, using the study period, the risk period, and whether the population is closed or open.
- Calculate and interpret an incidence risk (subtracting half the withdrawals from the population at risk) and an incidence rate from person-time, stating the period or the units with each, and convert an incidence rate into a risk.
- Explain what prevalence measures and how it differs from incidence, calculate a point prevalence, and explain why prevalence depends on both incidence and duration and is therefore a poor measure for identifying causes.
- Calculate and interpret attack rates, secondary attack rates, case fatality rates, mortality rates, and proportional mortality.
- Calculate years of life lost (YLLs), years lived with disability (YLDs), and disability-adjusted life years (DALYs), and explain what DALYs add to mortality and morbidity measures.
- Explain why crude rates can mislead, distinguish direct from indirect standardisation, calculate and interpret a standardised morbidity ratio and a directly standardised rate, and calculate an approximate 95% confidence interval for a proportion or a rate.
This course was developed by Dr. Kiffer G. Card based on Dohoo, I. R., Martin, S. W., & Stryhn, H. (2012). Methods in Epidemiologic Research. VER Inc.
Glossary: Key Terms, People & Concepts
📚 Reference page, available throughout the lesson
This glossary collects the key concepts, people, and ideas you will meet in this lesson. Use it as a reference while you work through the material, or as a review before assessments. Type in the search box to filter entries.
Introduction & Foundational Concepts
⏱ Estimated reading time: 12 minutes
Overview
This section introduces the vocabulary on which the rest of the lesson is built. Measures of disease frequency turn the data produced by surveillance, sampling and questionnaires into the standard quantitative outputs of epidemiology, and the definitions of Rothman (2012) and Porta (2014) are the references used throughout. The section begins with why frequency has to be measured before disease can be explained, and why a count on its own says little. It then sets out the four mathematical forms that every measure takes, counts, proportions, odds and rates, and applies them to a single outbreak, before showing how the study period and the risk period guide the choice between a risk and a rate. Later sections build on this vocabulary, covering the two faces of incidence, then prevalence, mortality and burden-of-disease measures such as DALYs, and finally the standardisation techniques and confidence intervals that make fair comparisons between populations possible. Every measure of association, such as a risk ratio or a rate difference, is built from two of these frequency measures.
Learning Objectives
- 1.1 Explain why measuring disease frequency comes before explaining disease, and why a count alone is of limited use.
- 1.2 Distinguish counts, proportions, odds, and rates by their numerators, denominators, ranges, and units.
- 2.1 Use the study period and the risk period to decide whether a risk or a rate fits a described situation.
1.1 Why Measure Disease Frequency?
Measurement of disease (or event) frequency is the foundation of many epidemiological activities, including routine surveillance, observational research and outbreak investigations (Porta, 2014). Before an epidemiologist can explain why a disease occurs, they first need to describe how often it occurs, and in observational studies, measuring the frequency of a disease and linking it to an exposure are the first steps toward inferring causation (Rothman, 2012). Morbidity and mortality are the two main categories of events for which frequency measures are calculated, but the same approaches apply to other events of interest, such as vaccination, hospital admission or giving birth.
The simplest measure is a count of cases. Because a count does not take the size of the population into account, it has very limited use on its own: the same number of cases means something very different in a village than in a city. A count becomes informative only when it is set against the population it came from, which is what the proportions and rates in 1.2 do.
Frequency measures are also rarely reported for a whole population alone, for a reason worth remembering.
Stratification Matters
Because both morbidity and mortality are strongly associated with individual attributes, and because different diseases have different impacts, we often calculate frequency measures for specific host attributes (e.g., age, sex, race) and for specific diseases. This stratification allows us to detect patterns that would be hidden in overall population-level data.
Stratification returns in Section 4, where differences in the make-up of two populations are shown to distort a comparison of their overall rates. Turning a count into a useful measure starts with choosing a denominator, and the choice of denominator defines the forms of measure described next.
1.2 Counts, Proportions, Odds and Rates
Every measure of disease frequency in this lesson takes one of four mathematical forms. The simplest, a count, has no denominator at all; it is just the number of cases. The other three are ratios, and they all share the same numerator, the cases. What distinguishes them is the denominator, that is, what the cases are divided by. In a proportion, the numerator is part of the denominator. By this we mean that every person counted on top of the fraction is also counted underneath it, as when 30 ill residents are divided by all 100 residents, including those 30. In the odds, by contrast, the denominator is the non-cases, so the people on top are not counted underneath. In a rate, the denominator is not a number of people at all but the person-time at risk, the total amount of time that people spent being observed while they could still become a case. Because these three denominators answer different questions, the same cases can produce three different numbers, and learning to recognise the denominator is the first skill this lesson builds.
1.2.1 Numerators, Denominators, Ranges and Units
The four cards below give a fuller definition of each form with a short numerical example, including the nursing-home outbreak that Worked Example 4.1 returns to. Click each card to see how its denominator is built, and notice on the Odds card how the same stillbirth data give slightly different values as odds and as a proportion.
Table 4.1 then sets the four forms side by side, so that their numerators, denominators, ranges and units can be compared directly. A useful check follows from the ranges: a proportion can never be larger than 1, whereas odds and rates can, and only a rate carries units.
Table 4.1. The numerator, denominator, range and units of counts, proportions, odds and rates.
| Measure | Numerator | Denominator | Range | Units |
|---|---|---|---|---|
| Count | Cases | None | 0 to ∞ | None |
| Proportion | Subset of denominator | Total population | 0 to 1 | Dimensionless |
| Odds | Cases | Non-cases | 0 to ∞ | Dimensionless |
| Rate | Cases | Person-time at risk | 0 to ∞ | Per person-time |
1.2.2 One Set of Observations, Three Measures
Because the three ratios share a numerator, proportions, odds and rates can all be calculated from the same set of observations, and they give different numbers for the same events only because their denominators differ. Equation 4.1 divides the cases by the total population that contains them, Equation 4.2 divides them by the non-cases, and Equation 4.3 divides them by the person-time at risk.
Worked Example 4.1 applies Equations 4.1 to 4.3 to the nursing-home outbreak from the Rate card. As you work through it, watch how the numerator stays fixed while the denominator changes, and which of the three results carries units.
Worked Example 4.1: One Outbreak, Three Measures
The Rate card in 1.2.1 describes 30 cases of diarrhea among the 100 residents of a nursing home over 3 months. The same data yield all three measures.
- Proportion: p = 30 / 100 = 0.30, so 30% of the residents became ill during the 3 months.
- Odds: odds = 30 / (100 − 30) = 30 / 70 = 0.43, so there were 0.43 cases for every resident who stayed well.
- Rate: if each resident is counted as at risk for the full 3 months, as on the Rate card, the person-time is 100 × 3 = 300 person-months and the rate is 30 / 300 = 0.10 cases per person-month.
The proportion and the odds differ noticeably here because the outcome is common; with 3 cases instead of 30 they would be 0.030 and 0.031. Removing residents from the denominator once they fall ill would reduce the person-time and raise the rate, which is the refinement introduced in the next section. Only the rate carries units.
The worked example shows that the denominator, not the label, determines which measure has been calculated, so the first step with any reported figure is to write down what was divided by what. Labels are an unreliable guide, as the following caution explains.
Common Terminology Confusion
The term “rate” is often used loosely to refer to all types of measures of disease frequency. Similarly, people commonly say someone has a high “chance” or “risk” of disease when the underlying measure might actually be a rate. Being precise about these terms helps avoid misinterpretation of research findings.
The rate in Worked Example 4.1 counted every resident as at risk for the full 3 months. Whether a risk or a rate is the better summary of such data depends on how long people can develop the disease, which is the subject of 2.1.
2.1 Study Period and Risk Period
The choice among the measures in 1.2, and in particular between a risk and a rate, depends on two time-related concepts: the study period, during which people are observed, and the risk period, during which a person could develop the disease. Open each tab to read the definition of each period with its examples, and compare the very short risk period of post-partum eclampsia with the lifelong risk period of migraine.
Study Period
The study period is the time interval during which study subjects are observed for the outcome of interest. It is usually measured in calendar time, but sometimes represents a point in time. The study period could also be defined by a specific event, for example “at birth” or “congenital defects during 2008–2010.”
Risk Period
The risk period is the time during which an individual could develop the disease. For some conditions, the risk period is very short (e.g., post-partum eclampsia, usually less than 2 days), while for others it is essentially lifelong (e.g., migraine headaches).
The length of the risk period relative to the study period guides the choice between risk-based and rate-based measures. Diseases with a short risk period relative to the study period are good candidates for risk measures, because nearly all cases will appear while people are being observed, so the proportion who become ill summarises their experience well. Diseases with long risk periods are better suited to rate-based measures, which divide the cases by the person-time actually observed, and a rate is also the better choice when people are followed for different lengths of time. To decide in a given situation, ask how long a person could develop the outcome and whether everyone is observed for the whole of that time, and justify the choice in a sentence that names the risk period.
This section has defined counts, proportions, odds and rates, shown how the same observations yield each measure, and described the time periods that guide the choice among them. The key takeaways below summarise these points.
Key Takeaways
- 1.1 Before we can explain why a disease occurs, we first need to describe how often it occurs, which is why measures of disease frequency sit underneath surveillance, observational research and outbreak investigations alike. Measuring how often a disease occurs and then linking it to an exposure is the first step toward inferring a cause. A count on its own, such as 50 cases of flu, tells you very little, because you cannot judge whether 50 is a lot until you know how many people the cases came from. When a question or a news report compares raw counts from places of different sizes, turn each count into a proportion or a rate before you draw any conclusion. Because frequency often varies with age, sex and other host attributes, you will also often see these measures calculated separately for each group (subsection 1.1).
- 1.2 Counts, proportions, odds and rates are the four mathematical forms that the measures in this lesson take, and the easiest way to tell them apart is to look at the denominator. In a proportion the numerator is part of the denominator (cases divided by the whole group, ranging from 0 to 1 with no units); in the odds the denominator is the non-cases (ranging from 0 upwards with no units); and in a rate the denominator is person-time at risk, so only a rate carries units, such as cases per person-month. When you work through an example, write down what is being divided by what before you name the measure: in Worked Example 4.1, 30 cases among 100 nursing-home residents over 3 months give a proportion of 30 / 100 = 0.30, odds of 30 / 70 = 0.43 and a rate of 30 / 300 = 0.10 cases per person-month. Be careful with labels, because the word "rate" is often used loosely for proportions and odds; the denominator, not the name a source uses, tells you which measure you have. A quick check is that a proportion can never be larger than 1, whereas odds and rates can.
- 2.1 Two time periods guide the choice between a risk and a rate. The study period is the stretch of time during which people are observed, and the risk period is the time during which a person could actually develop the disease. When the risk period is short relative to the study period (post-partum eclampsia, for example, usually occurs within 2 days), nearly all cases will appear while people are being observed, so a risk (the proportion who become ill) works well. When the risk period is long or lifelong (as with migraine headaches), or when people are followed for different lengths of time, a rate with a person-time denominator fits better. In a question, look for how long people could develop the outcome and whether everyone was followed for the same time, and justify your choice in a sentence that names the risk period.
1. Over a 6-month winter, 25 of the 125 residents of a group home developed an influenza-like illness. A colleague reports the “rate” of illness as 25 / 100 = 0.25. Which statement is correct?
2. Two studies are planned in the same health region. Study A follows 600 patients for the 30 days after hip-replacement surgery to count surgical-site infections, which almost always appear within that window. Study B follows a cohort of adults for 10 years to count first diagnoses of type 2 diabetes; people can develop diabetes at any age, and many participants move away or join partway through. Which pairing of measures fits best?
3. In one year, Town A (population 4,000) recorded 40 cases of pertussis and City B (population 80,000) recorded 200. A local news report says pertussis is “five times worse” in City B. Which response is correct?
✦ Pass the knowledge check with 100% to continue
Incidence: Risk & Rate
⏱ Estimated reading time: 15 minutes
Overview
Section 1 set up the basic vocabulary of counts, proportions, odds and rates. This section takes the most consequential of these measures, incidence, and works through its two faces: risk, which is a probability, and rate, which counts events per unit of person-time. The distinction is central, because risk-based and rate-based designs lead to different measures of association: risk ratios and risk differences on one side, rate ratios and rate differences on the other. The section shows how the kind of population being followed decides between a risk and a rate, describes the four ways in which incidence can be expressed, calculates an incidence risk and an incidence rate from exact and approximate person-time, and ends by converting a rate into a risk.
Learning Objectives
- 2.2 Distinguish incidence risk from incidence rate, and closed from open populations, and choose the measure that fits a described study.
- 3.1 Define incidence as the occurrence of new cases and describe the four ways to express it.
- 3.2 Calculate an incidence risk, subtracting half the withdrawals from the population at risk, and an incidence rate from exact or approximate person-time, reporting the period or the units with each.
- 3.3 Convert an incidence rate into a risk, and explain when the shortcut R ≈ I × Δt is adequate.
2.2 Incidence Risk and Incidence Rate in Closed and Open Populations
Incidence relates to the number of new events (for example, new cases of a disease) in a defined population within a specific period. By this we mean the transitions from being well to becoming ill that happen while people are being watched, not the illness that was already present when observation began. Because it follows people through that transition, incidence is the measure used to identify factors associated with becoming ill: if smokers make the transition more often than non-smokers, that difference is a clue about cause. It also follows that anyone who already has the disease when observation begins cannot become a new case, so they are left out of the population at risk. For example, if 20 of 1,000 adults already have diabetes when a study starts, only the remaining 980 can contribute new cases and only they belong in the denominator. Measuring incidence in this way requires a clear case definition, so that everyone agrees on what counts as becoming ill, along with a surveillance programme capable of identifying all cases as they arise.
Incidence has two main faces. An incidence risk is the proportion of disease-free people who develop the disease during a stated period, so it ranges from 0 to 1. An incidence rate is the number of new cases per unit of person-time at risk, so it carries units of 1/time. Both need an estimate of the population at risk, and which of them can be calculated depends on whether the population being followed is closed or open. Open each tab below to compare the two kinds of population, their examples, and what each means for the population at risk.
Closed Population
A closed population has no additions and few to no losses during the study period. Examples include residents of a nursing home followed for a year, or women followed for one week post-partum.
Only disease-free individuals at the start are considered at risk. Even in a closed population, however, some people may be lost to follow-up, for example because they move away or stop attending. These people are called withdrawals, and they create a problem for the denominator: they were observed for part of the period but not all of it, so counting them as fully at risk would overstate the population at risk, while dropping them entirely would ignore the time they were watched. The simplest correction is to subtract half the number of withdrawals from the population at risk. The reasoning is that, if people leave at scattered times, they leave on average halfway through the study, so each withdrawal is treated as having been at risk for half of the period.
Open Population
An open population has individuals entering and leaving throughout the study period. For example, women served by a cancer centre who had mastectomies form an open population as new patients continually arrive.
An open population is considered stable (stationary) if the rate of additions and withdrawals, and the distribution of host attributes, remain relatively constant over time. By this we mean that, although the individual members keep changing, the population as a whole looks much the same from one month to the next: roughly as many people arrive as leave, and its age and sex make-up stays similar. This matters because several shortcuts later in the lesson, such as the link between prevalence, incidence and duration, assume a population of this kind.
Risk cannot be computed directly from an open population, because there is no fixed group whose members are all followed from the start to the end of the period. A risk asks what proportion of a defined group became ill, and in an open population the group itself keeps changing. Instead, risk can be estimated from the incidence rate (I) using the relationship between risk and rate described in 3.3.
The choice between a rate and a risk therefore follows from the design. In a closed population the fixed group can be followed from start to finish and its incidence reported as a risk over a stated period. When people join and leave at different times, as in an open cohort, the number of people says little about how long each one was observed, so the cases are divided by the person-time actually accumulated at risk rather than by a head count, and the result is a rate. A study description that says participants joined, left, moved away or were followed for different lengths of time therefore calls for a rate. Risk and rate are not the only ways to express incidence, however, as the next subsection shows.
3.1 Four Ways to Express Incidence
Incidence can be expressed as incident times, as an incidence count, as an incidence risk (R) or as an incidence rate (I). Expand each panel below to see how each form is defined and what it is used for, and pay particular attention to the last panel, which explains why a rate, unlike a risk, can be larger than 1.
Incident times are the specific times at which cases occur, measured as elapsed time since a reference event (e.g., days after exposure to a toxin, or days after parturition). Incident times form the basis of survival analysis and are discussed at length in more advanced topics.
The incidence count is the simple count of new cases observed. It is often used when a disease did not previously exist or was very rare in a population. Incidence counts are sometimes expressed as absolute rates, relating the number of cases to the time period of observation (e.g., “3 cases per year”). They have limited value unless combined with population-at-risk data.
Incidence risk (R) is the probability that an individual will contract or develop a disease in a defined time period. It is a proportion (dimensionless, ranges from 0 to 1) and the time period must be specified. Only the first occurrence in the period of interest counts. Sometimes called cumulative incidence.
R is used in studies focused on individual-level predictions (e.g., the probability of breast cancer recurrence within the next year is 14%).
Incidence rate (I) is the number of new cases per unit of person-time during a given time period. By this we mean that the cases are divided by the total time that people spent at risk, so the rate describes how fast new cases arise rather than what fraction of a group fell ill. Because time sits in the denominator, the rate has units of 1/time (for example, cases per person-year), and it is positive without an upper bound. It is also called incidence density, a name that reflects the idea of how densely cases are packed into the time at risk.
I is used to determine what factors are related to diseases and the effects of those diseases. It is the preferred measure for open populations where individuals enter and leave over time.
A rate has no upper limit of 1. A rate of 1.5 cases per person-year does not mean that 150% of people fall ill; it means 15 cases arose per 10 person-years of observation. Because the denominator is accumulated time rather than a head count, a rate can climb above 1, whereas a risk (which is a proportion) never can.
Incident times and incidence counts describe when and how many cases occurred, but only the risk and the rate relate the cases to the population that produced them, which is why the rest of this section concentrates on those two. Before either is calculated, one more decision has to be made about what counts as a new case.
First Cases vs. All Cases
Although incidence deals with “new cases,” this does not necessarily mean only the first-ever case within an individual. For some conditions (e.g., migraine headaches, recurrent infections), multiple episodes can occur. Researchers must decide whether to count only first cases or all new episodes, and clearly state this in their methodology.
With the population, the measure and the case definition settled, the next step is the calculation itself.
3.2 Calculating Incidence Risk and Incidence Rate
Both calculations divide the new cases by a denominator that describes the population at risk. For a risk the denominator is a number of people, and for a rate it is the person-time that those people spent at risk, which can be added up exactly or approximated.
3.2.1 Incidence Risk in a Closed Population
Incidence risk is a proportion, and its calculation requires a count of new cases in a defined period and an estimate of the population at risk. Equation 4.4 sets out the formula.
The population at risk is the number of individuals who are free of the disease when follow-up begins, less half of any withdrawals, as described on the Closed Population tab in 2.2. Worked Example 4.2 applies Equation 4.4 to a closed cohort of long-term care residents; watch where the withdrawals enter the denominator.
Worked Example 4.2: Incidence Risk in a Closed Cohort
A long-term care network admits 500 residents who are free of pressure ulcers and follows each of them for 90 days. During follow-up, 35 residents develop a pressure ulcer and 20 are transferred to other facilities before the 90 days are complete (withdrawals).
- Population at risk = 500 − ½ × 20 = 490, subtracting half of the withdrawals on the assumption that they left, on average, halfway through follow-up.
- R = 35 / 490 = 0.071, or 7.1% over 90 days.
Ignoring the withdrawals would give 35 / 500 = 0.070, a slight underestimate, because the residents who left early were not observed for the whole risk period. The result is a 90-day risk, and the period is reported with it.
The worked example shows that the withdrawal correction changes the risk only slightly, but that the period must always be reported with the result. Calculator 4.1 carries out the same calculation for other numbers of cases, starting population and withdrawals.
Risk can be computed directly only for a closed population like this one. Open populations need a denominator that reflects how long each person was observed, which is the incidence rate.
3.2.2 Incidence Rate from Exact Person-Time
The incidence rate has the person-time at risk as its denominator, so its calculation depends on how long each individual remains at risk. Equation 4.5 sets out the formula.
A person-time unit is one person observed for a defined period (for example, one person-month or one person-year). By this we mean that time and people are combined into a single quantity: one person followed for two years and two people followed for one year each both contribute 2 person-years. This is what allows people who were watched for different lengths of time to be added together in one denominator. Each person contributes time from the start of observation until they develop the disease, leave the study or reach the end of the period, whichever comes first. The reason for stopping the clock at diagnosis is that, after an individual contracts the disease of interest, they can no longer become a new case, so they are no longer at risk and no longer contribute person-time to the denominator (for first-case analyses). Worked Example 4.3 applies Equation 4.5 by adding up the exact person-time contributed by four individuals; watch how much time each person contributes before they fall ill or leave.
Worked Example 4.3: Exact Incidence Rate
Assume 4 previously healthy people were observed for exactly 1 month (30 days):
- Person 1: not sick at all → 1.00 person-months at risk
- Person 2: sick on day 10 → 0.33 person-months at risk
- Person 3: sick on day 20 → 0.67 person-months at risk
- Person 4: moved away on day 15 (lost to follow-up) → 0.50 person-months at risk
Total person-months at risk = 2.50
Total new cases = 2
I = 2 / 2.50 = 0.80 cases per person-month
The worked example shows that the two people who fell ill count in full in the numerator but contribute only the time before they fell ill to the denominator, and that the person who moved away contributes the half month during which they were observed. Calculator 4.2 repeats the calculation with other values.
3.2.3 Incidence Rate from Approximate Person-Time
When exact person-time data are unavailable, the denominator can be approximated with Equation 4.6. The starting population is reduced by half of those who fell ill and half of those who withdrew, increased by half of those who were added, and multiplied by the length of the period.
This assumes that, on average, events (illness, withdrawal, addition) occur at the midpoint of the study period. By this we mean that, when we do not know exactly when someone fell ill, left or joined, we treat them as having done so halfway through, so each of them contributes half of the period to the person-time. A person who falls ill or leaves therefore counts as half a person at risk, which is why half of them are subtracted. A person who joins also counts as half a person, but because they were not in the starting population they have to be added. If events are spread fairly evenly across the period, the early and late events balance out and the approximation is close to the exact person-time. Worked Example 4.4 applies Equation 4.6 to an occupational cohort for which exact dates are unavailable; check the sign of each half-adjustment as you follow it.
Worked Example 4.4: Approximate Incidence Rate
An occupational cohort begins a one-year period with 1,200 workers who are free of work-related back injury. During the year, 90 workers are injured, 100 leave the company, and 60 new workers join. Exact dates are unavailable, so each event is placed at the midpoint of the year.
- Approximate number at risk = 1,200 − ½(90) − ½(100) + ½(60) = 1,200 − 45 − 50 + 30 = 1,135.
- Person-time = 1,135 × 1 year = 1,135 person-years.
- I = 90 / 1,135 = 0.079 cases per person-year, or about 79 per 1,000 person-years.
The injured workers appear twice in the calculation: in full in the numerator, and as half a person each in the denominator, because they stop contributing time at risk once they are injured.
The worked example shows how a rate can be estimated from start and end counts alone, and how it is rescaled to a convenient multiplier. Calculator 4.3 carries out the approximate calculation for other values.
Whether the person-time is exact or approximate, a rate can be interpreted only when it is reported with its units, and the conventions for doing so are worth committing to memory.
Reporting a Rate in Convenient Units
A rate is often a small decimal, so it is usually multiplied by a round number of person-time units before it is reported. A rate of 0.079 cases per person-year is 0.079 × 1,000 = 79 cases per 1,000 person-years, and a rate of 18 cases in 90,000 child-years is 18 / 90,000 × 100,000 = 20 cases per 100,000 child-years. The person-time must be in one unit before the division: if some follow-up is recorded in person-months, divide it by 12 to convert it to person-years (for example, 54,000 person-months = 4,500 person-years). Whatever multiplier is chosen, the units are written out in full with the number.
The exact and the approximate methods both express new cases per unit of person-time. Because a risk cannot be computed directly from an open population, the last step is to convert a rate of this kind into a risk.
3.3 Converting a Rate to a Risk
Risk and rate are mathematically related, because both are built from the same new cases arising in the same people. The rate tells us how many cases arise per unit of time at risk, so multiplying it by the length of the interval gives an estimate of how many cases each person would accumulate over that interval. For example, a rate of 0.01 cases per person-year sustained over 5 years suggests that about 0.05, or 5%, of the group will become ill. Thus, for a closed population followed over an interval Δt, the risk is approximately the incidence rate multiplied by the length of the interval, provided that this product is small (Equation 4.7).
More precisely, assuming a constant rate, the relationship is exponential (Equation 4.8).
The exponential formula in Equation 4.8 accounts for the fact that as people become diseased, they leave the at-risk pool. By this we mean that a constant rate acts on a shrinking group: in the early years many people are still well and can fall ill, but as cases accumulate there are fewer well people left for the same rate to act on. The number of new cases each year therefore falls slowly over time, so the risk grows more slowly than the rate multiplied by the time, and it can approach 1 but never exceed it. The linear shortcut ignores this shrinking, which is harmless when only a small fraction of the group falls ill but misleading when a large fraction does. When the product I×Δt is small (less than 0.1), the simpler linear approximation in Equation 4.7 works well; when the product is larger, the approximation overstates the risk. The exponential relationship assumes that the rate stays constant over the period. Survival methods such as the Kaplan–Meier estimator apply the same idea of turning the experience of people at risk into a risk, but they do so over many short intervals and do not need the rate to be constant (Kaplan & Meier, 1958; see also Wikipedia). Worked Example 4.5 applies both equations to an incidence rate of type 2 diabetes over a short and a long period; watch the size of I × Δt in each case.
Worked Example 4.5: Converting a Rate to a Risk
Suppose the incidence rate of type 2 diabetes in a cohort of adults is 0.012 cases per person-year, and the question is what proportion of a disease-free group would develop diabetes over a given period.
- Over 5 years, I × Δt = 0.012 × 5 = 0.060. The approximation gives R ≈ 0.060, and the exact formula gives R = 1 − e−0.060 = 0.058. The two agree closely because the product is small.
- Over 40 years, I × Δt = 0.012 × 40 = 0.48. The approximation gives R ≈ 0.48, whereas the exact formula gives R = 1 − e−0.48 = 0.38. The approximation now overstates the risk by ten percentage points, because many people have already developed diabetes and left the at-risk pool.
The worked example shows that the shortcut is adequate over 5 years but overstates the risk by ten percentage points over 40 years. Before converting, put the rate and Δt in the same time unit and work out I × Δt; if the product is 0.1 or more, use the exact formula. Calculator 4.4 carries out both conversions for other values.
This section has shown how the type of population decides between a risk and a rate, described the forms in which incidence is expressed, set out how a risk and a rate are calculated, and described the relationship between them. The key takeaways below summarise these points.
Key Takeaways
- 2.2 Incidence risk and incidence rate are the two faces of incidence, and the kind of population usually tells you which one to use. A closed population has no additions and few or no losses during the study (residents of a nursing home followed for a year, for example), so you can follow the fixed group and report a risk: a proportion between 0 and 1 over a stated period. An open population has people entering and leaving throughout (such as patients who keep arriving at a cancer centre), so a head count says little about how long each person was observed, and you divide the new cases by the person-time actually accumulated at risk to get a rate. In exam questions, words such as "joined", "left", "moved away" or "followed for different lengths of time" signal an open population and a rate. Risk cannot be calculated directly from an open population; if you need one, estimate it from the rate using the risk-rate relationship in takeaway 3.3.
- 3.1 Incidence is about new cases: people who were free of the disease at the start and develop it during the period you are observing. Because it follows people from being well to becoming ill, incidence is the family of measures used to identify factors associated with becoming ill, and anyone who already has the disease at the start is left out of the population at risk. Incidence can be expressed in four ways: incident times (when each case occurred, the basis of survival analysis), an incidence count, an incidence risk (R, a proportion) and an incidence rate (I, new cases per unit of person-time). When a question describes a group that already contains some cases at the start, remove them from the population at risk and count only the cases that arise afterwards. Also check whether a study counts only first cases or every new episode (subsection 3.1), because recurrent conditions such as migraine can be counted either way.
- 3.2 An incidence risk is the number of new cases divided by the population at risk at the start, meaning the people who were free of the disease when follow-up began. If some people leave before the end (withdrawals), subtract half of them from the population at risk, on the assumption that they left halfway through on average: in Worked Example 4.2, 35 / (500 − ½ × 20) = 35 / 490 = 0.071, or 7.1% over 90 days. An incidence rate is new cases divided by person-time at risk, either added up person by person (Worked Example 4.3) or, when exact dates are missing, approximated as (start − ½ sick − ½ withdrawn + ½ added) × the length of the period (Worked Example 4.4). Common traps are forgetting the withdrawal correction and dividing a rate by a head count, so write the denominator out in full before you divide. Finish every answer with its period or its units. A risk needs its time period ("7.1% over 90 days"), and a rate needs person-time units, so convert all follow-up to one unit first (person-months ÷ 12 = person-years) and then rescale it to a convenient multiplier, as when 0.079 cases per person-year becomes 79 per 1,000 person-years.
- 3.3 Because risk and rate are linked, you can turn an incidence rate into the risk of becoming ill over a stated period with R = 1 − e−IΔt. The exact formula allows for the people who become ill leaving the at-risk pool, which is why the risk grows more slowly than the rate multiplied by the time. When I × Δt is small (below about 0.1), the shortcut R ≈ I × Δt is close enough: in Worked Example 4.5, 0.012 × 5 = 0.060, against an exact risk of 0.058. Over 40 years, however, the shortcut gives 0.48 while the exact risk is 0.38, an overstatement of ten percentage points. Before you convert, make sure the rate and Δt use the same time unit, work out I × Δt first, and use the exact formula whenever that product is 0.1 or more.
1. A factory screens 500 newly hired workers for noise-induced hearing loss and finds 20 who already have it. The other 480 are retested 2 years later, and 12 of them have developed hearing loss. What is the incidence risk, and how should it be read?
2. A cohort of 320 newly hired paramedics, none with a back injury, is followed for 12 months. During the year, 30 are injured and 40 leave the service before the year ends. Assuming that those who left did so, on average, halfway through the year, what is the 12-month incidence risk of back injury?
3. Over three years, a transit company’s bus drivers, who joined and left at different times, contributed a total of 14,400 person-months free of high blood pressure, and 36 drivers were newly diagnosed with it. What is the incidence rate per 1,000 person-years?
4. A city health unit wants to describe how often nurses have their first needle-stick injury over a five-year period. Nurses are hired and leave throughout, so some are observed for a few months and others for all five years. A colleague proposes dividing the injuries by the number of nurses on the payroll at the start. Which approach is best?
5. In a cohort of adults, the incidence rate of a chronic condition is roughly constant at 0.02 cases per person-year. A planner wants the risk for a disease-free person over 3 years and over 25 years, and proposes using R ≈ I × Δt for both. Which statement is correct?
6. A diabetes prevention programme follows 960 adults who are free of type 2 diabetes for two years. Over the two years, 72 new cases are diagnosed, 100 participants move away, and 60 new participants, all free of diabetes, join. Exact dates were not recorded, so each event is placed at the midpoint of the two years. Using the approximate person-time formula, what is the incidence rate?
✦ Pass the knowledge check with 100% to continue
Prevalence, Mortality & Other Frequency Measures
⏱ Estimated reading time: 18 minutes
Overview
Section 2 covered incidence, the measures of how quickly new cases arise. This section turns to the complementary measures that describe disease as it currently exists in a population and the harm it causes: prevalence (current cases), mortality (deaths), and burden-of-disease summaries such as DALYs that combine premature mortality with disability. Each tells you something different about a population's health. The section defines point and period prevalence and shows how prevalence depends on both incidence and the duration of disease. It then treats mortality rates as incidence rates with death as the outcome, introduces the attack rate, secondary attack rate, case fatality rate and proportional mortality, which are proportions despite their names, and ends with years of life lost, years lived with disability and DALYs, which place fatal and non-fatal health loss on one scale.
Learning Objectives
- 4.1 Define point and period prevalence, explain how prevalence differs from incidence, and calculate a point prevalence.
- 4.2 Explain how prevalence depends on incidence and disease duration, using the sink analogy and P = ID / (ID + 1), and why prevalence is a poor measure for identifying causes.
- 5.1 Explain overall and cause-specific mortality rates as incidence rates with death as the outcome.
- 5.2 Calculate and interpret attack rates, secondary attack rates, case fatality rates, and proportional mortality, and explain why each is a proportion despite its name.
- 6.1 Calculate Years of Life Lost (YLLs), Years Lived with Disability (YLDs), and Disability-Adjusted Life Years (DALYs), and explain why they add value beyond mortality and morbidity measures alone.
4.1 Point and Period Prevalence
Unlike incidence, which counts new cases, prevalence measures the number of existing cases of disease at a specific point in time (or over a period), however long ago each case began. The prevalence count is the number of individuals in a population who have an attribute or disease at a particular time. Equation 4.9 gives point prevalence as the number of existing cases divided by the population at the same point in time, so its denominator is everyone in the population at that moment, not only the people who were free of disease at some earlier start.
For example, if 75 athletes are tested for performance-enhancing drug use and 3 test positive, then P = 3/75 = 0.04 (4%). Operational decisions about case definitions and time windows matter more than this simple formula suggests (Spronk et al., 2019). Calculator 4.5 carries out the same calculation for other counts.
Period prevalence uses the same logic over an interval: it counts everyone who had the disease at any time during a stated period (for example, a calendar year) and divides by the population during that period. Like point prevalence, it is a proportion with no time units, so the moment or the period it describes is reported with it. Because it counts everyone who currently has the disease, prevalence is the measure used to plan health services, such as the staff, beds or treatments needed at a given time, whereas incidence is used to study causes and prevention.
Because prevalence counts existing cases and incidence counts new ones, the two can differ sharply even when they are measured in the same group. Suppose 400 children at a summer camp are well on day 0, and 60 develop a short diarrheal illness, lasting two to three days, over the next 14 days. The 14-day incidence risk is 60 / 400 = 0.15 (15%), because it counts every new case over the whole period. If a nurse counts the children who are ill at 9 a.m. on day 7 and finds 20 among the 400 present, the point prevalence is 20 / 400 = 0.05 (5%). The prevalence is lower because it misses the children who have already recovered and those who have not yet fallen ill. An incidence rate would divide the 60 new cases by the person-time at risk, giving the speed at which cases arise (cases per person-day), whereas the prevalence counts existing cases at one moment and has no time units.
The summer camp shows that a short illness keeps prevalence low even when new cases arise quickly. The next subsection makes this dependence on duration precise.
4.2 How Prevalence Depends on Incidence and Duration
Prevalence depends on two things: how quickly new cases arise and how long people stay ill. The reason is that the pool of existing cases is filled by new cases and emptied by recovery or death. If new cases arrive quickly, the pool grows; if each case lasts a long time, people stay in the pool longer, and it grows as well. For example, two diseases with the same incidence will have very different prevalences if one lasts a week and the other a lifetime. When the incidence rate (I) and the duration stay constant in a stable population, the inflow and outflow eventually balance, and the pool settles at a steady level. At that steady state, prevalence (P), incidence, and mean disease duration (D) are related by Equation 4.10, in which I and D must be expressed in the same time unit.
Worked Example 4.6 applies Equation 4.10 to influenza in an urban population; watch how the duration is converted to match the units of the rate.
Worked Example 4.6: Prevalence of Influenza from Incidence and Duration
If the incidence rate of influenza in an urban population is 0.3/person-year (30 cases per 100 people per year) and the mean duration of infection is 3 weeks (0.058 years), then:
P = (0.3 × 0.058) / (0.3 × 0.058 + 1) = 0.0174 / 1.0174 = 0.017 (1.7%)
So on any given day, we would expect about 1.7% of the population to have the flu.
Watch the units. The 3-week duration is converted to years (0.058) so that it lines up with the per-year incidence rate. I and D must be expressed in the same time unit, which is what makes their product I × D a pure number with no units. Feeding in 3 (weeks) next to 0.3 (per year) would inflate the answer by a factor of 52.
The worked example shows that a common, short illness produces a low prevalence, and that a units mismatch would inflate the answer many times over. Calculator 4.6 performs the calculation for other values of incidence and duration.
Interactive 4.1 represents the same relationship as water in a sink, with the incidence rate as the inflow from the tap, recovery and death as the two drains, and the prevalence as the water level. Try the disease presets first, then move one slider at a time and watch how the level settles as you open the tap or widen either drain.
🚰 Interactive 4.1: The Sink Analogy for Incidence, Duration & Prevalence
P = I × D / (I × D + 1) where D = 1 / (r + m)
The sink shows that two diseases can have very different prevalences for entirely different reasons. Diabetes has a low incidence but a very long duration: the sink fills slowly but barely drains, so prevalence is high. Pancreatic cancer has a low incidence and a very short duration because of its high mortality: the drain is wide open, so prevalence stays near zero even though the disease is deadly. It follows that anything that changes how long people stay ill, such as a treatment that prolongs survival, changes prevalence without causing or preventing a single new case.
Prevalence Is a Poor Guide to Causes
Prevalence (P) is less useful than incidence rate (I) for research into risk factors, because factors that affect either the occurrence of disease or its duration will both influence prevalence. A disease may have high prevalence simply because people survive with it for a long time, not because new cases are frequent. For example, if patients in one region receive better treatment that keeps them alive longer, more of them will be counted as existing cases at any moment, and a prevalence study comparing regions could wrongly suggest that something about that region makes people more likely to develop the disease. In this sense prevalence alone confounds occurrence and survival, mixing the two together so that they cannot be told apart, which is why it is a poor measure of disease risk.
Prevalence studies still have a part to play in measuring incidence. Serial prevalence studies, by which we mean repeated surveys of the same people at regular intervals, can be used to estimate incidence rates for diseases that are not easily detected at onset. Many infections, for example, cause no symptoms when they begin, so nobody notices the moment a person becomes a case. By testing blood samples at regular intervals, researchers can see which people were negative at one test and positive at the next, and so determine the window in which each of them became infected. Counting these conversions and dividing by the person-time between tests gives an estimate of the incidence rate.
Prevalence, incidence and duration together describe how disease occurs and persists in a population. The same measures apply when the outcome is death.
5.1 Mortality Rates
Mortality statistics are calculated in exactly the same way as P, R, and I, except that the outcome of interest is death. The term mortality rate, strictly speaking, refers to the incidence rate of mortality, but it is often used to describe the risk of mortality (Porta, 2014; Rothman, 2012), so the denominator has to be checked before the measure is named. As incidence rates, mortality rates carry person-time units, and their denominator is the person-time of the whole population at risk. Open each tab below to see how the overall and the cause-specific mortality rates are defined, and why attributing a death to one cause is harder than it looks.
Overall Mortality Rate
The overall mortality rate describes the number of individuals that die from all causes in a defined time period. It is analogous to incidence rate (I) except that the outcome is death rather than disease onset.
Cause-Specific Mortality Rate
The cause-specific mortality rate describes the number of individuals that die from (or with) a specific disease during a defined period. It is calculated the same way as I but focuses on deaths attributed to one disease.
Challenges of Cause Attribution
It is often difficult to determine the specific cause of death. For example, if a recumbent patient regurgitates, contracts aspiration pneumonia, and then dies, did they die: from the initial condition causing recumbency? Due to pneumonia? With pneumonia? The “cause” is usually deemed to be the proximate cause, the factor considered the final trigger, but this determination can be challenging.
The cause recorded is usually the proximate one, so a cause-specific rate inherits any error in that judgement. Mortality rates divide deaths by the person-time of the whole population, but several other widely used measures divide cases or deaths by much narrower groups.
5.2 Attack Rates, Case Fatality and Proportional Mortality
Several additional frequency measures appear frequently in the epidemiological literature. Most of them are technically measures of risk (proportions), a number of people divided by a number of people with no person-time in the denominator, even though they are often called “rates.” The word “rate” in these names is a long-standing habit of usage rather than a description of the calculation. For example, an attack rate of 30% means that 30 of every 100 exposed people became ill; it says nothing about how many cases arose per person-day, so it is read as a proportion and carries no time units. What separates these measures from one another is the denominator: everyone exposed for an attack rate, the contacts without the primary cases for a secondary attack rate, the people with the disease for a case fatality rate, and all deaths for proportional mortality.
5.2.1 What Each Measure Describes
The four cards below define the attack rate, the secondary attack rate, the case fatality rate and proportional morbidity or mortality. Click each card to see what the measure describes, how it is calculated and when it is appropriate, and compare the case fatality rate with the mortality rate in 5.1.
RateClick to learn more
Attack RateClick to learn more
RateClick to learn more
Morbidity/MortalityClick to learn more
Each of these definitions becomes a simple equation, and the calculations below make the different denominators concrete.
5.2.2 Calculating Each Measure
Each tab below gives the equation for one measure, a worked example from a realistic outbreak or death register, and a calculator. Open the tabs in turn and, before reading each worked example, name the group that belongs in the denominator.
Equation 4.14 gives the attack rate, the people who became ill divided by everyone exposed during the outbreak.
Worked Example 4.10 applies Equation 4.14 to an outbreak of gastroenteritis at a wedding reception; watch how comparing eaters with non-eaters of one food points to the source.
Worked Example 4.10: Attack Rate
Of the 180 guests at a wedding reception who ate the chicken salad, 54 developed gastroenteritis within 72 hours. The attack rate is 54 / 180 = 0.30, so 30% of those exposed became ill during the outbreak window. Among the 120 guests who did not eat the salad, 6 became ill, an attack rate of 6 / 120 = 0.05. The contrast between the two attack rates points to the salad as the likely source. Dividing one attack rate by the other gives the risk ratio used in outbreak investigations: 0.30 / 0.05 = 6.0, so guests who ate the salad were six times as likely to become ill as guests who did not. Each attack rate is calculated for one food at a time, with the people who ate it in the denominator.
The worked example shows that attack rates are most useful in pairs, one food at a time. Calculator 4.10 calculates the attack rate for other values.
Equation 4.15 gives the secondary attack rate, which removes the primary cases from both the numerator and the denominator.
Worked Example 4.11 applies Equation 4.15 to influenza in 40 households; watch where the 40 primary cases are subtracted.
Worked Example 4.11: Secondary Attack Rate
An influenza surveillance team follows 40 households containing 160 people in total. Each household has one primary case, and over the following week 18 other household members become ill, so there are 58 cases in all.
SAR = (58 − 40) / (160 − 40) = 18 / 120 = 0.15, so 15% of household contacts became secondary cases.
In plain words, the secondary attack rate measures how easily an infection spreads from a first (primary) case to the people around it. The primary cases are removed from both the numerator and the denominator because they were infected elsewhere and cannot become secondary cases.
The worked example shows how to isolate person-to-person spread from the infections that began it. Calculator 4.11 calculates the secondary attack rate for other values.
Equation 4.16 gives the case fatality rate, the deaths among people with the disease divided by the people with the disease.
Worked Example 4.12 applies Equation 4.16 to an outbreak of Legionnaires’ disease; note the time window that is reported with the result.
Worked Example 4.12: Case Fatality Rate
During an outbreak of Legionnaires’ disease, 250 people are confirmed as cases, and 25 of them die within 30 days of diagnosis. CFR = 25 / 250 = 0.10, so 10% of cases died within 30 days. The value describes the severity of the illness among those who had it, and the 30-day window is reported with it.
The worked example shows that the case fatality rate describes severity among the cases, not the risk of death in the population. Calculator 4.12 calculates the case fatality rate for other values.
Equation 4.17 gives proportional mortality, the deaths from one cause divided by the deaths from all causes.
Worked Example 4.13 applies Equation 4.17 to the deaths recorded in one health region; notice that the population at risk never enters the calculation.
Worked Example 4.13: Proportional Mortality
A health region records 12,000 deaths in one year, of which 3,000 are attributed to cancer. The proportional mortality for cancer is 3,000 / 12,000 = 0.25, so cancer accounts for 25% of deaths. Because the population at risk does not enter the calculation, the figure describes the share of deaths and gives no information about the risk of dying from cancer.
The worked example shows that proportional mortality describes the share of deaths and says nothing about anyone's risk of dying. That is also why and when it is used: analysts turn to proportional mortality when the population at risk is unknown or unavailable, for example when a death register records the cause of each death but there is no count of the population or its person-time. When the person-time at risk is known, the cause-specific mortality rate in 5.1 is the better measure, because proportional mortality can change simply because deaths from other causes rise or fall. Calculator 4.13 calculates proportional mortality for other values.
Across the four tabs, the measures differ only in their denominators, and choosing the wrong group, such as dividing an attack rate by the people who became ill rather than by everyone exposed, is the most common error. All of these measures, like mortality rates, count deaths or cases without regard to the age at which death occurs or the illness that does not kill, and burden-of-disease measures address both limits.
6.1 Burden of Disease: YLLs, YLDs and DALYs
Mortality, incidence, and prevalence each tell us something different, but taken individually they give an incomplete picture of a disease's impact. A condition can be deadly but rare (pancreatic cancer), or non-fatal yet extremely common and disabling (low-back pain, depression, untreated vision loss). To compare these on a single scale, the World Health Organization (WHO) and the Global Burden of Disease (GBD) Study, originated by Murray & Lopez (1997) and now updated regularly by IHME (GBD 2019 Diseases and Injuries Collaborators, 2020), express disease impact in healthy years of life lost, a common “currency” that combines premature death and time spent in less-than-perfect health (Robine, Jagger, Mathers, Crimmins, & Suzman, 2003). The burden is built from years of life lost and years lived with disability, which add up to the DALY.
6.1.1 Years of Life Lost (YLL)
YLL quantifies the burden from premature mortality. Each death is weighted by the number of additional years that person would have been expected to live had they survived. A death at age 30 contributes far more YLLs than a death at age 90, even though both count equally toward a crude mortality count. Equation 4.11 expresses YLL as a sum across ages of the number of deaths multiplied by the remaining life expectancy.
Here Nx is the number of deaths at age x, and Lx is the standard life expectancy remaining at age x, taken from a reference life table. A single reference table is used for every population, rather than each country's own life expectancy, so that a death at a given age counts as the same loss wherever it occurs; otherwise, a death in a country with low life expectancy would count as fewer years lost simply because people there tend to die younger. The reference life table of the GBD 2010 study set life expectancy at birth at about 86 years (World Health Organization, 2024); WHO and later rounds of the GBD study use their own reference life tables, so check which standard a published YLL figure uses. Worked Example 4.7 applies Equation 4.11 to road traffic deaths in one district; compare how much each age group contributes.
Worked Example 4.7: Summing YLL Across Ages
Suppose a district records 20 deaths from road traffic injuries in one year: 3 at age 20, 5 at age 40, and 12 at age 65. Suppose also that the reference life table gives a remaining life expectancy of 65 years at age 20, 45 years at age 40, and 20 years at age 65 (values rounded for illustration).
- Age 20: 3 × 65 = 195 years.
- Age 40: 5 × 45 = 225 years.
- Age 65: 12 × 20 = 240 years.
- YLL = 195 + 225 + 240 = 660 years of life lost.
The 3 deaths at age 20 contribute nearly as many years as the 12 deaths at age 65. This weighting for premature death is the information that a simple count of 20 deaths leaves out.
The worked example shows that a few early deaths can outweigh many later ones. Calculator 4.7 carries out the calculation for other numbers of deaths and life expectancies.
6.1.2 Years Lived with Disability (YLD)
YLD quantifies the burden from non-fatal health loss. Each year spent with a condition is multiplied by a disability weight (DW) between 0 (perfect health) and 1 (a health state as bad as death). By this we mean that a year lived with an illness is counted as only part of a healthy year lost, and the weight says how large that part is. For example, a year lived with a condition whose weight is 0.30 counts as 0.30 of a healthy year lost, so ten such years add up to three years of healthy life lost. The weight is needed because a year with mild hearing loss and a year with severe depression should not count the same, and without it every year of illness would be treated as equal to a year of death. Disability weights are derived from large population surveys that ask respondents to compare paired health states (Salomon et al., 2012). Equation 4.12 gives the incidence-based formula for YLD.
Here I is the number of incident cases, D is the average duration of the condition until remission or death, and DW is the disability weight. The incidence approach looks forward: it takes the people who develop the condition in a given year and counts all the years of disability they will live with it, even years that fall in the future. A prevalence-based version (used by GBD since 2010) instead looks at a single year: it multiplies the number of prevalent cases by DW directly, YLD = P × DW, so that each person living with the condition contributes one year of weighted disability for the year being measured. Worked Example 4.8 compares the incidence and prevalence approaches for a single condition; watch why they agree.
Worked Example 4.8: Incidence and Prevalence Approaches to YLD
Suppose 500 people develop a condition each year, each lives with it for 4 years on average, and its disability weight is 0.30. The incidence approach gives YLD = 500 × 4 × 0.30 = 600 years. If the population is in a steady state, about 500 × 4 = 2,000 people are living with the condition at any time, and the prevalence approach gives YLD = 2,000 × 0.30 = 600 years.
The two approaches agree here because incidence and duration have been stable long enough for prevalence to reflect them. They diverge when incidence is rising or falling.
The worked example shows that the two approaches agree only when incidence and duration are stable. Calculator 4.8 computes YLD for other values; whichever approach is used, every YLD term includes a disability weight between 0 and 1, and leaving it out is a common slip.
6.1.3 Disability-Adjusted Life Years (DALYs)
The DALY is the simple sum of years lost to early death and years lost to disability (Equation 4.13).
One DALY equals one healthy year of life lost. The total DALYs from a condition therefore measure the gap between current population health and an ideal in which everyone lives the standard life expectancy in full health (Murray et al., 2012). Worked Example 4.9 applies Equations 4.11 to 4.13 to breast cancer in a population of 100,000 women, using the same YLD inputs as Worked Example 4.8; compare the size of the two components.
Worked Example 4.9: Calculating DALYs for Breast Cancer
In a population of 100,000 women, suppose breast cancer causes 20 deaths in one year, with an average age at death of 60 (standard life expectancy at age 60 = 25 years). Suppose also that 500 women are newly diagnosed and live with breast cancer for an average of 4 years before remission or death, with a disability weight of 0.30.
YLL = 20 × 25 = 500 years
YLD = 500 × 4 × 0.30 = 600 years
DALYs = 500 + 600 = 1,100 healthy years lost in this population in one year.
Mortality alone would have shown only the 20 deaths. The DALY makes it visible that more than half of breast-cancer burden in this population is non-fatal, a fact that should shape investment in screening, treatment, and survivorship support alongside end-of-life care (Murray & Lopez, 1997).
The worked example shows that the non-fatal component can be the larger part of a disease's burden. Calculator 4.9 adds the two components for other values.
6.1.4 What DALYs Add and Where They Fall Short
Worked Example 4.9 illustrates the main reason for using DALYs beyond mortality, incidence and prevalence: they are a common currency. DALYs let you compare diseases that primarily kill with those that primarily disable on the same scale, for example ischaemic heart disease against major depression, which no single mortality or morbidity measure can do. Within that currency, premature death is weighted, because a young death contributes more YLLs than an old death, capturing the intuition that lost years matter in a way that crude mortality rates ignore. Non-fatal burden also becomes visible: conditions such as low-back pain, anxiety, hearing loss and migraine are major drivers of population health loss but are nearly invisible in mortality statistics. These properties make DALYs a tool for priority-setting and equity, since ministries of health, the WHO and global funders use them to compare cost-effectiveness across very different interventions (for example, dollars per DALY averted). Because they combine fatal and non-fatal loss, DALYs also track change over time, capturing epidemiological transitions, the shift from infectious, fatal diseases toward chronic, disabling ones, better than any single measure.
These advantages come with limits that apply whenever you read a DALY figure.
Caveats
DALYs are widely used but not value-neutral. Disability weights reflect aggregate survey responses and may not match an individual patient's lived experience; weights have also shifted across GBD revisions (Salomon et al., 2012; Vos et al., 2016). Choices about the standard life expectancy, age-weighting, and time-discounting can substantially change disease rankings (Murray & Lopez, 1997). As with any composite indicator, always check which version of the methodology produced the number you are reading, and whether comparing across years uses the same conventions.
Burden-of-disease measures complete the set of frequency measures in this section. Like the others, they are published routinely by surveillance agencies, and the additional context below describes where Canadian figures come from.
Additional Context
This part, on where Canadian frequency figures come from, is background for reading published figures rather than an exam objective. The incidence, prevalence, mortality, and case-fatality numbers in your textbooks and in the news are produced by named surveillance systems. The most important to know are the following.
- Canadian Chronic Disease Surveillance System (CCDSS): PHAC. Linked provincial/territorial physician-billing and hospital data harmonised across jurisdictions. Produces national prevalence and incidence estimates for diabetes, hypertension, mental illness, COPD, asthma, IHD, cancer, and more.
- Canadian Notifiable Disease Surveillance System (CNDSS): PHAC. Aggregated reports from provinces/territories on legally reportable communicable diseases (e.g., measles, pertussis, syphilis, HIV).
- FluWatch and RVDSS (Respiratory Virus Detection Surveillance System): PHAC, sentinel and laboratory surveillance for influenza and respiratory viruses.
- Canadian Cancer Registry: Statistics Canada / provincial cancer registries. Population-based incidence and survival data; the source for "Canadian Cancer Statistics" annual reports.
- Canadian Vital Statistics, Births and Deaths Database: Statistics Canada. Source for crude and age-standardised mortality, life expectancy, infant mortality, and cause-specific death rates.
- BC Centre for Disease Control (BCCDC) surveillance dashboards: provincial reportable-disease, STBBI, and overdose surveillance, including the overdose response indicators and weekly respiratory-virus reports.
- BC Coroners Service: primary source for unregulated drug deaths and suicide statistics in British Columbia.
- Canadian Institute for Health Information (CIHI): Discharge Abstract Database (DAD) and National Ambulatory Care Reporting System (NACRS); source for hospital-based incidence, length-of-stay, and case-fatality estimates.
Each system has its own case definition, denominator, and reporting lag. When you cite a Canadian rate, always state which surveillance system produced it, and check whether the case definition matches your research question. Worked Example 4.14 shows what is needed to interpret one such figure, the CCDSS prevalence of diagnosed diabetes; look for the three concepts from this lesson that it relies on.
Worked Example 4.14: Reading a CCDSS Diabetes Rate
The CCDSS reports that the age-standardised prevalence of diagnosed diabetes among Canadians aged 1+ was about 8.1% (2017–18 fiscal year; Public Health Agency of Canada, 2021). To interpret this number you need to know:
- Numerator (case definition): a CCDSS case is a person with one hospital discharge OR two physician claims for diabetes within a two-year window, a validated administrative algorithm with sensitivity ~86% and specificity ~97% (Hux et al., 2002).
- Denominator: all individuals registered with provincial health insurance during the fiscal year (a near-complete population frame, not a sample).
- Standardisation: rates are direct-standardised to the 2011 Canadian population, so comparisons across years and provinces aren't distorted by ageing.
Notice how all three concepts from this lesson (case definition, denominator at risk, and standardisation) are required just to read a single published number correctly.
This section has covered prevalence, mortality, the specialised proportions used to describe outbreaks and deaths, and burden-of-disease measures. The key takeaways below summarise these points, and the next section turns to the standardisation on which comparisons such as the CCDSS rate depend.
Key Takeaways
- 4.1 Prevalence describes how much disease exists in a population at a given time, counting every existing case however long ago it began, whereas incidence counts only new cases. Point prevalence is the number of existing cases at one moment divided by the population at that same moment (3 positive tests among 75 athletes gives 3 / 75 = 0.04, or 4%), and period prevalence counts everyone who had the disease at any time during a stated period. Prevalence is a proportion with no time units, so always report the moment or the period it describes. When a question gives you both a cumulative risk and a one-day count in the same group, expect them to differ: in the summer-camp example in subsection 4.1, the 14-day risk is 60 / 400 = 15% but the point prevalence on day 7 is 20 / 400 = 5%, because with a short illness many children have already recovered, or have not yet fallen ill, on the day of the count. Check your denominator too, because for point prevalence it is everyone in the population at that moment, not only the people who were free of disease at the start.
- 4.2 Prevalence depends on two things: how quickly new cases arise and how long people stay ill. In the sink analogy, the tap is the incidence rate, the two drains are recovery and death, and the water level is the prevalence, so a disease with a low incidence but a long duration, such as diabetes, can still have a high prevalence, while a rapidly fatal disease, such as pancreatic cancer, keeps the level near zero. At steady state the relationship is P = I×D / (I×D + 1), and I and D must be in the same time unit: in Worked Example 4.6, 3 weeks becomes 0.058 years to match 0.3 cases per person-year, giving a prevalence of about 1.7%. Anything that changes how long people stay ill, such as a treatment that prolongs survival or a cure that shortens the illness, changes prevalence without causing or preventing a single new case, which makes prevalence a poor measure for identifying causes. When a question asks why, name both parts (occurrence and duration) and say that the incidence rate is the better measure for research into risk factors.
- 5.1 Mortality statistics are not a new set of formulas: they are the measures you have already met, with death as the outcome instead of disease onset. The overall mortality rate is deaths from all causes divided by the person-time at risk, and the cause-specific mortality rate is deaths from one disease divided by the person-time at risk, so each is an incidence rate and carries person-time units. The term "mortality rate" is often used loosely for the risk of dying, so check the denominator before you name the measure. Keep the mortality rate separate from the case fatality rate and proportional mortality: the mortality rate divides by the person-time of the whole population at risk, the case fatality rate by the people who have the disease, and proportional mortality by the deaths alone. Remember too that the cause of a death can be hard to assign, and the cause recorded is usually the proximate one, the final trigger.
- 5.2 The attack rate, secondary attack rate, case fatality rate and proportional mortality are often called rates, but each is a proportion: a number of people divided by a number of people, with no person-time. The attack rate is the people who became ill divided by everyone exposed during the outbreak (54 / 180 = 0.30 for the chicken salad in Worked Example 4.10), and comparing the attack rates of eaters and non-eaters, one food at a time, points to the likely source. The secondary attack rate measures how easily an infection spreads from a primary case to the people around it, so remove the primary cases from both the numerator and the denominator: (58 − 40) / (160 − 40) = 0.15 in Worked Example 4.11. The case fatality rate divides deaths by the people with the disease and describes severity (25 / 250 = 10% within 30 days), whereas proportional mortality divides deaths from one cause by all deaths and tells you only the share of deaths, not anyone's risk of dying. The trap in every case is the denominator, so before you divide, name the group the measure is about (the exposed, the contacts, the cases or all deaths); a frequent error is dividing an attack rate by the people who became ill instead of by everyone exposed. Report the period with your answer.
- 6.1 Burden-of-disease measures express early death and time lived in poor health in one currency: healthy years of life lost. YLLs capture premature death by multiplying the deaths at each age by the standard remaining life expectancy at that age and adding the results (YLL = Σ Nx × Lx), so a death at 20 counts for far more than a death at 65. YLDs capture non-fatal health loss as incident cases × average duration × disability weight (YLD = I × D × DW, or prevalent cases × DW), and DALYs are simply YLL + YLD. In Worked Example 4.9, 20 deaths × 25 years = 500 YLLs and 500 cases × 4 years × 0.30 = 600 YLDs give 1,100 DALYs, which shows that more than half of the breast-cancer burden is non-fatal, something the 20 deaths alone would hide. A common slip is to leave out the disability weight, so check that every YLD term includes a weight between 0 and 1, and when you are asked what DALYs add, say that they put conditions that mainly kill and conditions that mainly disable on a single scale.
1. A new treatment keeps people with a chronic lung disease alive for much longer but does not cure them. Over the next decade the number of new cases each year stays the same, yet the prevalence of the disease doubles. A colleague concludes that something in the environment must now be causing more of the disease. What is the best response?
2. In one year, a health region whose 200,000 residents contribute about 200,000 person-years records 400 hospital admissions for confirmed influenza, and 20 of these patients die of it within 30 days. The region records 1,600 deaths from all causes that year. Which pair of values is correct?
3. In a stable population, the incidence rate of a skin infection is 0.4 cases per person-year, and an episode lasts 2 weeks on average. Using P = I×D / (I×D+1), about what proportion of the population has the infection on any given day?
4. A condition causes 10 deaths at an average age where standard life expectancy is 30 remaining years, and 200 people newly develop the condition and live with it for 5 years on average, with a disability weight of 0.20. What is the total DALY burden?
5. A ministry compares two conditions in its province. Condition X causes 50 deaths a year, mostly after age 80, and little illness before death. Condition Y causes almost no deaths, but about 20,000 people are living with the moderate disability it causes at any time. A planner proposes ranking the two conditions by their annual death counts. What would DALYs add to this comparison?
6. A residential school has 250 students, all well on 1 March and all present for the next three weeks. During those 21 days, 50 students develop a stomach illness that lasts 1 to 2 days. On the morning of 11 March, a nurse finds 10 of the 250 students ill. Which statement is correct?
7. After a community barbecue, investigators collect food histories from all 200 guests and summarise them in the table below. Which food is the most likely vehicle, and what is its risk ratio?
| Food | Ate the food: ill / total | Did not eat the food: ill / total |
|---|---|---|
| Potato salad | 32 / 100 | 8 / 100 |
| Grilled corn | 30 / 120 | 10 / 80 |
| Lemonade | 34 / 170 | 6 / 30 |
8. During an outbreak of hand, foot and mouth disease, 15 households each have one primary case, a child infected at a day-care centre. The 15 households contain 75 people in total, including the primary cases. Over the next two weeks, 21 further household members become ill. What is the secondary attack rate, and what does it measure?
9. A town of 20,000 residents has a summer outbreak of hepatitis A in which 66 people become ill, and 1 of them dies within 60 days of diagnosis. A local news report states that “hepatitis A killed 1.5% of the town this summer.” Which response is correct?
✦ Pass the knowledge check with 100% to continue
Standardisation & Confidence Intervals
⏱ Estimated reading time: 15 minutes
Overview
Sections 1 to 3 showed how to compute measures of disease frequency from raw data. This section addresses the problem that arises as soon as those measures are compared: a crude rate from one population almost never compares fairly with a crude rate from another, because populations differ in age structure and other characteristics that affect disease occurrence. Standardisation is the technique for making the comparison fair, in an indirect form that produces the standardised morbidity ratio and a direct form that re-weights each population to a common structure. Standard errors and confidence intervals then quantify how precise any estimated proportion or rate is.
Learning Objectives
- 7.1 Explain why crude rates can mislead when populations differ in age structure, and why standardisation is then necessary.
- 7.2 Distinguish indirect from direct standardisation, and calculate and interpret a standardised morbidity ratio and a directly standardised rate.
- 7.3 Calculate standard errors and approximate 95% confidence intervals for a proportion and a rate, and explain when an exact interval is preferable.
7.1 Why Crude Rates Can Mislead
When comparing disease frequency between populations, differences in host characteristics (for example, age, sex or geographic location) can confound the comparison. A population with more elderly individuals will, for example, naturally have higher rates of many diseases, simply because those diseases are more common at older ages. Standardisation adjusts for these confounders, allowing for fairer comparisons (Ahmad et al., 2001).
The reason lies in how a crude rate, the overall rate for a whole population, is built, and it is the key idea of this section.
Key Concept: Confounding by Host Factors
A population can be divided into strata based on one or more host characteristics, for example into age groups. The overall frequency of disease depends on both the size of each stratum (Hj) and the stratum-specific rates or risks (Ij or Rj). By this we mean that the crude rate is a weighted average of the stratum-specific rates, in which each stratum counts in proportion to its share of the population. A stratum with a high rate, such as the oldest age group, therefore pulls the crude rate up more strongly when it makes up a larger share of the population. As a result, if two populations have different age distributions, their crude rates may differ even if their age-specific rates are identical, because the same rates are being averaged with different weights. Standardisation removes this confounding effect by giving both populations the same set of weights.
A population whose age-specific rates are all lower can even have the higher crude rate if more of its people fall in the oldest stratum, and Worked Example 4.16 in 7.2 shows exactly this pattern. Standardisation can be carried out indirectly or directly, and the next subsection describes each method in turn.
7.2 Indirect and Direct Standardisation
The two methods swap the roles of the two populations involved. Indirect standardisation applies the rates of a standard population to the structure of the study population, whereas direct standardisation applies the study population's own rates to the structure of a standard population.
7.2.1 Indirect Standardisation
Indirect standardisation uses standard (reference) population rates applied to the study population’s structure to calculate the expected number of cases. By this we mean the number of cases the study population would have had if its people, with their own ages, had experienced the same age-specific rates as the reference population. This expectation already takes the study population's age structure into account, so an older population will have more expected cases than a younger one. The method then compares the cases actually observed with that expectation in the standardised morbidity (or mortality) ratio, SMR = observed ÷ expected. If the two numbers are equal, the population is doing exactly as well as its age structure predicts; if more cases are observed than expected, the excess cannot be explained by age, and some other factor, or chance, must account for it. The procedure has three steps. Expand each step below in order to see what it involves, then compare it with the matching equation that follows.
Obtain a set of stratum-specific rates (Isj) from a reference or standard population. For example, national age-specific disease rates.
Apply those standard rates to the study population’s time-at-risk in each stratum. The expected adjusted rate (Ie) equals the sum of Hj × Isj, where Hj is the proportion of the study population in stratum j. By this we mean that each standard rate is weighted by how much of the study population falls in that stratum, so a population with many older people gives more weight to the higher rates of the older strata. The result is the overall rate the study population would be expected to have if it experienced the standard rates. The expected number of cases is then E = T × Ie, where T is the total time at risk, which simply converts that expected rate into a number of cases.
The Standardised Morbidity (or Mortality) Ratio (SMR) is the ratio of observed cases (A) to expected cases (E): SMR = A / E.
An SMR > 1 means more cases were observed than expected; SMR < 1 means fewer. The indirect standardised rate is: Iind = Is × SMR, where Is is the overall rate in the standard population.
Equation 4.18 gives the expected cases from step 2, Equation 4.19 the SMR from step 3, and Equation 4.20 the indirectly standardised rate.
Worked Example 4.15 applies the three steps to a mining town, with the expected cases in each age stratum set out in Table 4.2; watch how the town's own person-time is combined with the provincial rates.
Worked Example 4.15: Indirect Standardisation
A mining town contributes 8,000 person-years of observation across three age strata, and 60 new cases of a respiratory disease are observed (A = 60). Provincial age-specific rates serve as the standard, and the overall provincial rate is 7.25 per 1,000 person-years (Is).
Table 4.2. Person-years, standard rates and expected cases of respiratory disease in each age stratum of the mining town.
| Age stratum | Town person-years (Tj) | Share (Hj) | Standard rate per 1,000 (Isj) | Expected cases (Tj × Isj) |
|---|---|---|---|---|
| Under 40 | 4,000 | 0.500 | 2 | 8 |
| 40 to 59 | 3,000 | 0.375 | 8 | 24 |
| 60 and over | 1,000 | 0.125 | 25 | 25 |
| Total | 8,000 | 1.000 | 57 |
- Expected rate: Ie = (0.500 × 2) + (0.375 × 8) + (0.125 × 25) = 7.125 per 1,000 person-years, so E = 8,000 × 7.125 / 1,000 = 57 expected cases.
- SMR = 60 / 57 = 1.053, so the town recorded about 5% more cases than the provincial rates predict for a population with its age structure.
- Indirectly standardised rate: Iind = 7.25 × 1.053 = 7.63 per 1,000 person-years.
The crude rate in the town is 60 / 8,000 = 7.50 per 1,000 person-years. It is slightly lower than the standardised rate because the town has a smaller share of older residents than the standard population.
The worked example shows that the town recorded slightly more cases than its age structure predicts, and that its crude rate understates its standardised rate. Calculator 4.14 carries out the three steps for other values.
An SMR describes one population at a time, and this matters when several populations are standardised against the same reference. Suppose three geographic regions with different age distributions report data on GI disease. Region 1 has a crude risk of 0.136, Region 2 has 0.125, and Region 3 has 0.088. After indirect standardisation, the SMRs are 1.12, 1.03, and 0.74 respectively. An SMR of 1.12 means Region 1 recorded about 12% more cases than the reference age-specific rates predict for a group of its age structure, while an SMR of 0.74 means Region 3 recorded about 26% fewer. Region 1’s observed risk was therefore higher than expected and Region 3’s lower, and it is worth asking what this suggests about the role of age distribution compared with true disease risk in these regions. What the SMRs cannot do is order the regions against one another.
Do Not Rank SMRs Against One Another
Each SMR uses its own region’s age structure as the weights, so SMRs from different regions are not standardised to a common population and should not be ranked against one another. Each one says only whether that region recorded more or fewer cases than the reference rates predict for its own age structure. Direct standardisation to a common standard population is the usual way to compare the regions with one another.
Indirect standardisation compares observed with expected cases using the study population's own structure. Direct standardisation reverses the roles of the study and standard populations.
7.2.2 Direct Standardisation
Direct standardisation uses the study population’s stratum-specific rates applied to a standard population distribution (Equation 4.21).
Here Tsj is the proportion of person-time in the standard population assigned to stratum j, and Ij is the observed rate in the study population for that stratum.
In plain terms, the directly standardised rate is a weighted average of the study population's own stratum-specific rates, but using the standard population's proportions as the weights. It answers a single question: what would this population's overall rate be if it had the age structure of the standard population? Because every population is re-weighted to the same structure, their standardised rates can be laid side by side and compared directly. Worked Example 4.16 applies Equation 4.21 to two regions with different age structures, the situation described in 7.1; compare the crude and the standardised rates of the two regions.
Worked Example 4.16: Direct Standardisation
Two regions report the rates of a chronic disease shown in Table 4.3. Region A contributes 30,000, 15,000, and 5,000 person-years to the three age strata, and Region B, which is older, contributes 10,000 person-years to each. The standard population places 55% of its person-time under age 40, 30% at ages 40 to 64, and 15% at age 65 and over.
Table 4.3. Standard population weights and age-specific rates of a chronic disease in Region A and Region B.
| Age stratum | Standard weight (Tsj) | Region A rate per 1,000 (Ij) | Region B rate per 1,000 (Ij) |
|---|---|---|---|
| Under 40 | 0.55 | 1.0 | 0.9 |
| 40 to 64 | 0.30 | 4.0 | 3.8 |
| 65 and over | 0.15 | 12.0 | 11.0 |
- Region A: Idir = (0.55 × 1.0) + (0.30 × 4.0) + (0.15 × 12.0) = 0.55 + 1.20 + 1.80 = 3.55 per 1,000 person-years.
- Region B: Idir = (0.55 × 0.9) + (0.30 × 3.8) + (0.15 × 11.0) = 0.495 + 1.14 + 1.65 = 3.29 per 1,000 person-years.
Region B has the higher crude rate (157 cases in 30,000 person-years, or 5.2 per 1,000, against 150 cases in 50,000 person-years, or 3.0 per 1,000, in Region A) because a third of its person-time falls in the oldest stratum. Its rate is lower than Region A’s in every age stratum, however, and once both regions are given the same age structure its standardised rate is the lower of the two.
The worked example shows that standardisation can reverse a crude comparison when the populations differ in age structure. Calculator 4.15 carries out the calculation for other values.
7.2.3 Choosing Between Indirect and Direct Standardisation
Indirect standardisation is useful when stratum-specific rates for the study population are unavailable or based on small samples, and it produces the SMR. The reason is that it borrows its rates from the large standard population and needs only the study population's person-time in each stratum and its total number of cases. For example, a small town may record only two or three cases among its oldest residents, too few to give a trustworthy age-specific rate, yet it can still be compared with provincial rates. Direct standardisation, by contrast, applies the study’s own stratum-specific rates to a standard distribution, which is why it is the method for comparing several populations with one another: every population is placed on the same structure. A drawback is that all stratum-specific rates receive equal weight regardless of precision. By this we mean that the weight each rate receives is set by the standard population, not by how many cases lie behind it, so a rate based on a handful of cases in a small stratum can still move the standardised rate considerably. Both methods are commonly used in practice. Age standardisation of cancer statistics is one well-known application (e.g., rates published by cancer registries); the WHO World Standard population (Ahmad et al., 2001) is the most widely used reference structure for direct standardisation in global health comparisons. See also Standardized mortality ratio (Wikipedia).
Standardisation makes rates comparable across populations, and any estimated proportion or rate, standardised or not, also needs a measure of its precision.
7.3 Standard Errors and Confidence Intervals
When estimating any proportion or rate, we need a measure of precision. Most estimates come from a sample or a limited period of observation, so a different sample of the same population would have given a slightly different answer purely by chance. The standard error (SE) quantifies this uncertainty around our estimate. By this we mean that it describes how much the estimate would typically vary from one sample to another: a small standard error means repeated samples would give similar answers, and a large one means they could differ considerably. For example, a prevalence based on 40 people will usually have a much larger standard error than one based on 4,000. A confidence interval built from the standard error then turns this idea into a range of values compatible with the data, which is easier to report and interpret than the standard error itself.
7.3.1 Standard Error of a Proportion
The standard error of a proportion depends on the proportion itself and on the sample size, as Equation 4.22 shows.
Here p is the estimated proportion and N is the sample size, and p enters as a decimal rather than a percentage. Worked Example 4.17 applies Equation 4.22 to a survey estimate of high blood pressure; follow each step inside the square root.
Worked Example 4.17: Standard Error of a Proportion
In a survey of 400 adults, 88 report that a health professional has told them they have high blood pressure, so p = 88 / 400 = 0.22.
SE(p) = √[0.22 × (1 − 0.22) / 400] = √(0.1716 / 400) = √0.000429 = 0.0207.
The worked example gives a standard error of about two percentage points for a sample of 400. Calculator 4.16 performs the calculation for other values.
7.3.2 Standard Error of an Incidence Rate
For an incidence rate, the standard error depends on the number of cases and the person-time at risk, as Equation 4.23 shows.
Here A is the number of cases and t is the total person-time at risk, so the standard error carries the same units as the rate. Worked Example 4.18 applies Equation 4.23 to a cohort with 2,500 person-years of follow-up; note that the square root is taken of the number of cases only.
Worked Example 4.18: Standard Error of a Rate
A cohort accumulates 2,500 person-years of follow-up, during which 40 new cases occur, so I = 40 / 2,500 = 0.016 cases per person-year (16 per 1,000 person-years).
SE(I) = √40 / 2,500 = 6.32 / 2,500 = 0.00253 cases per person-year.
The worked example shows that the precision of a rate depends on the number of cases observed. Calculator 4.17 performs the calculation for other values.
7.3.3 Confidence Intervals
An approximate confidence interval extends a multiple of the standard error on either side of the estimate, as in Equation 4.24.
Here θ is the estimate and Zα is the appropriate percentile of the standard normal distribution (e.g., 1.96 for a 95% CI) (Rothman, 2012). The multiplier sets the level of confidence: a larger multiplier, such as 2.58 for a 99% interval, produces a wider interval in exchange for greater confidence. Because the sample size sits under a square root in both standard errors, quadrupling the sample size halves the width of the interval. This means that precision improves more slowly than sample size grows; for example, going from 400 to 1,600 participants halves the interval, but halving it again would require 6,400. Worked Example 4.19 applies Equation 4.24 to the estimates from Worked Examples 4.17 and 4.18; check that both limits of the rate's interval keep the rate's units.
Worked Example 4.19: 95% Confidence Intervals
Using the two earlier estimates (Worked Example 4.17 and Worked Example 4.18) and Zα = 1.96:
- Proportion: 0.22 ± 1.96 × 0.0207 = 0.22 ± 0.041, giving a 95% CI from 0.179 to 0.261 (17.9% to 26.1%).
- Rate: 0.016 ± 1.96 × 0.00253 = 0.016 ± 0.0050, giving a 95% CI from 0.0110 to 0.0210 cases per person-year (11.0 to 21.0 per 1,000 person-years).
The worked example shows how one standard error turns each point estimate into a range. Calculator 4.18 performs the calculation for other values.
These intervals rest on a normal approximation, which does not always hold.
When Approximate Intervals Mislead
In small samples, or when the frequency of disease is very low or very high, approximate CIs can produce misleading results (e.g., negative lower limits). The reason is that the approximate interval is always symmetrical, extending the same distance above and below the estimate, whereas a proportion cannot fall below 0 or rise above 1 and a rate cannot be negative. For example, if 1 person in a sample of 50 has a disease, p = 0.02 and its standard error is about 0.020, so subtracting 1.96 standard errors gives a lower limit below zero, which is impossible. In such cases, exact CIs based on the binomial distribution (for proportions) or the Poisson distribution (for rates) are more appropriate.
This section has shown why crude rates can mislead, how indirect and direct standardisation make rates comparable across populations, and how standard errors and confidence intervals express the precision of an estimate. The key takeaways below summarise these points.
Key Takeaways
- 7.1 A crude rate, the overall rate for a whole population, depends on two things: the rate within each stratum (each age group, for example) and how much of the population falls into each stratum. Because many diseases become more common with age, an older population can have a higher crude rate even when its age-specific rates are the same as, or lower than, those of a younger population. In Worked Example 4.16, Region B has the higher crude rate (5.2 against 3.0 per 1,000 person-years) only because a third of its person-time falls in the oldest age group; its rate is lower than Region A's in every age stratum. This is confounding by host characteristics, and standardisation adjusts for it so that rates can be compared fairly. Whenever a question compares crude rates from populations with different age structures, look at the stratum-specific rates first and treat the crude comparison with caution until the rates have been standardised.
- 7.2 Indirect and direct standardisation both deal with differences in population structure, but they swap which population supplies the rates. Indirect standardisation applies a standard population's stratum-specific rates to the study population's own person-time to obtain the expected number of cases (E), then compares the observed number (A) with it as the SMR = A / E; in Worked Example 4.15, 60 / 57 = 1.05, so the mining town recorded about 5% more cases than the provincial rates predict for a population with its age structure. Direct standardisation works the other way round: it applies the study population's own stratum-specific rates to a standard population's weights, giving a weighted average that answers the question "what would this population's rate be if it had the standard age structure?" Choose indirect standardisation when the study population is small or its stratum-specific rates are unavailable or unstable, and direct standardisation when you want to compare several populations with one another. Read an SMR above 1 as more cases than expected and below 1 as fewer, and do not rank SMRs from different populations against each other, because each one is weighted by its own population's structure.
- 7.3 Every estimated proportion or rate needs a measure of its precision, and the standard error provides it: SE(p) = √[p(1 − p)/N] for a proportion, and SE(I) = √A/t for a rate, where A is the number of cases and t the person-time at risk. An approximate 95% confidence interval is the estimate ± 1.96 × SE; for 88 of 400 adults with high blood pressure, 0.22 ± 1.96 × 0.0207 gives an interval from 0.179 to 0.261. Work with decimals rather than percentages inside the square root, and give both limits of a rate's interval in the rate's units. Because the sample size sits under a square root, quadrupling it halves the width of the interval. When the sample is small or the frequency is very low or very high, the approximate interval can mislead (for example, by giving a negative lower limit), so an exact binomial interval (for a proportion) or Poisson interval (for a rate) is preferable.
1. A small farming town recorded 30 new cases of skin cancer over five years. Applying provincial age-specific rates to the town’s person-time in each age group gives 20 expected cases. Which interpretation is correct?
2. An analyst wants to report cancer incidence for (i) a remote community of 900 people, where most age groups recorded 0 to 2 cases, and (ii) the province’s five largest cities, each with hundreds of cases in every age group, and to compare the five cities with one another. Which plan is best?
3. In a survey of 900 adults, 180 report that they have been diagnosed with asthma. What is the approximate 95% confidence interval for the proportion with asthma?
4. Two districts report deaths from heart disease, summarised in the table below. A colleague notes that District B’s crude death rate is about twice District A’s and concludes that heart disease is more deadly in District B. Which response is correct?
| Age group | Standard weight (Tsj) | District A rate per 1,000 person-years | District A share of person-time | District B rate per 1,000 person-years | District B share of person-time |
|---|---|---|---|---|---|
| Under 65 | 0.80 | 2 | 0.90 | 1 | 0.50 |
| 65 and over | 0.20 | 20 | 0.10 | 15 | 0.50 |
5. Over five years, the residents of a district contribute 20,000 person-years at risk, and 50 new cases of a chronic respiratory disease are diagnosed, a rate of 2.5 per 1,000 person-years. What is the approximate 95% confidence interval for this rate?
6. The same district contributes 10,000 person-years under age 40, 6,000 at ages 40 to 64 and 4,000 at age 65 and over, and records 50 cases. The provincial rates in these age groups are 0.5, 2.5 and 5.0 per 1,000 person-years, and the overall provincial rate is 2.7. A colleague notes that the district’s crude rate (2.5 per 1,000 person-years) is below the provincial rate and concludes that the district has no excess of the disease. Which response is correct?
✦ Pass the knowledge check with 100% to continue
Final Review & Assessment
⏱ Estimated time: 20 minutes
Bringing It All Together
This lesson assembled the basic numerical vocabulary of epidemiology (Porta, 2014; Rothman, 2012). You moved from the four mathematical forms (counts, proportions, odds, and rates) through incidence (risk and rate), into prevalence and mortality, then into burden-of-disease metrics (GBD 2019 Diseases and Injuries Collaborators, 2020), and finally into standardisation (Ahmad et al., 2001) and the precision around the resulting estimates. Every comparison between groups, places or time periods depends on picking the right denominator for the question being asked.
The takeaways below pull these threads together. The decisions in this lesson are usually the difference between a defensible comparison and a misleading one: prevalence and incidence answer different questions; risks and rates require different denominators; and crude rates can hide more than they reveal whenever populations differ in age structure or other strong confounders.
Key Takeaways from this lesson
- You can tell counts, proportions, odds and rates apart by asking what is divided by what: a proportion's numerator is part of its denominator, the odds divide cases by non-cases, and only a rate has person-time in its denominator and carries units. You also know that a count means little until it is set against the population it came from. On the exam, identify the denominator before you trust the name given to a measure, because "rate" is often used loosely.
- You can choose the measure that fits a situation: a risk for a closed population followed over a defined period when the risk period is short, a rate with a person-time denominator for an open population, a long risk period or unequal follow-up, and a prevalence when the question is how many people have the disease at a point or during a period. Read each scenario for clues about who joins or leaves and how long each person is followed, and justify your choice in a sentence.
- You can calculate an incidence risk (new cases divided by the population at risk at the start, less half the withdrawals) and an incidence rate (new cases divided by exact or approximate person-time), and convert a rate into a risk with R = 1 − e−IΔt, or R ≈ IΔt when IΔt is below about 0.1. Always finish with the period for a risk or the person-time units for a rate, after putting all follow-up into one time unit.
- You can calculate a point prevalence and explain that prevalence counts existing cases whereas incidence counts new ones, and that prevalence rises with both incidence and duration, like the water level in the sink. This is why a risk and a point prevalence measured in the same group can differ, and why prevalence is a poor measure for identifying causes; when you are asked why, name both occurrence and duration.
- You can calculate and interpret attack rates, secondary attack rates, case fatality rates, mortality rates and proportional mortality, and you know that all of them except the mortality rate are proportions despite their names. Before dividing, decide who belongs in the denominator: everyone exposed for an attack rate, the contacts without the primary cases for a secondary attack rate, the people with the disease for a case fatality rate, the person-time at risk for a mortality rate, and all deaths for proportional mortality.
- You can calculate YLLs, YLDs and DALYs and explain how DALYs place fatal and non-fatal health loss on one scale of healthy years lost, so that conditions that mainly kill can be compared with conditions that mainly disable. When you calculate, weight each death by the remaining life expectancy at its age and multiply every YLD by its disability weight.
- You can explain why crude rates mislead when populations differ in age structure, calculate and interpret an SMR (observed ÷ expected) and a directly standardised rate, and attach an approximate 95% confidence interval (estimate ± 1.96 × SE) to a proportion or a rate. Remember that each SMR compares one population only with the standard rates, so use directly standardised rates to compare populations with one another, and report every estimate with its interval, since a point estimate without its precision is only half an answer.
Final Knowledge Assessment
This assessment covers all material from this lesson. You must score 100% to complete the lesson. Review the feedback for any incorrect answers and try again.
1. A clinic follows 200 adults for 6 months after a hip fracture, and 40 of them develop pneumonia. Which set of values is correctly labelled?
2. A health authority must decide how many wound-care nurses to send to its long-term care homes next month, so it needs to know how many residents have an open pressure ulcer that needs dressing. Which measure fits this purpose best?
3. A travel-health report states: “Among travellers who were free of the infection when they arrived, the 2-week incidence risk was 0.06.” Which plain-words reading is correct?
4. A fitness centre follows its members for one year for a first sports injury. At the start, 540 members are injury-free. During the year, 60 are injured, 40 cancel their membership and 20 new members join; exact dates are not recorded. Using the approximate formula, which places every event at the midpoint of the year, what is the incidence rate?
5. A school district follows 1,050 children who have never had chickenpox and are not vaccinated against it, for one school year. During the year, 90 develop chickenpox, and 100 move out of the district partway through (assume, on average, at mid-year). What is the one-year incidence risk?
6. The incidence rate of a cancer among long-term smokers is roughly constant at 4 per 1,000 person-years (0.004 per person-year). What is the risk that a cancer-free smoker develops this cancer within 10 years?
7. A student writes: “The incidence rate and the prevalence both tell us how common a disease is, so they are two names for the same thing.” Which correction is best?
8. A new antibiotic cures a common infection in 1 week instead of 4 but does nothing to prevent new infections. If the incidence stays the same, what happens to the prevalence of the infection, and why?
9. After a staff lunch, 90 people ate the sushi and 27 of them became ill within 48 hours; 40 people did not eat it, and 4 of them became ill. A colleague reports the attack rate for sushi as 87%, “because 27 of the 31 people who became ill ate the sushi.” What is the correct attack rate for sushi, and what does it mean?
10. Last year, 30% of all deaths in Region A and 20% of all deaths in Region B were from cancer. A news story concludes that residents of Region A are more likely to die of cancer. Which response is correct?
11. On a 10-day cruise, 2,000 passengers board in good health and all stay aboard. Over the cruise, 300 of them develop a norovirus illness that lasts about 2 days. At noon on day 6, 60 passengers are ill. What is the point prevalence at noon on day 6, and why does it differ from the 10-day incidence risk?
12. A public-health nurse follows the households of 40 children who caught pertussis (whooping cough) at school. These households contain 170 people in total, including the 40 primary cases, and 26 of the other household members later develop pertussis. Which statement correctly gives and interprets the secondary attack rate?
13. Over four years, 50 cases of occupational asthma occur among the staff of a bakery chain. Staff are hired and leave throughout; 1,000 different people worked there at some time, and together they contributed 2,500 person-years at risk. A manager reports the risk as 50 / 1,000 = 5% “over four years.” What is the main problem, and what should be reported instead?
14. During a study, an industrial town contributed 10,000 person-years under age 50 and 5,000 person-years at age 50 and over, and recorded 50 cases of a lung disease. The provincial (standard) rates are 1 per 1,000 person-years under age 50 and 6 per 1,000 person-years at age 50 and over. What are the expected cases and the SMR, and what does the SMR mean?
15. In one year, a district records 4 deaths from drowning at age 20, where the remaining life expectancy is 65 years, and 26 deaths from a chronic lung disease at age 85, where the remaining life expectancy is 6 years. Which statement about years of life lost (YLL) is correct?
16. On a 10-day cruise, 1,250 passengers board in good health, and 150 develop a norovirus illness that usually lasts one to two days. To find the source, a colleague proposes comparing the proportion of passengers who are ill at 8 a.m. on day 6 among those who did and did not eat at the buffet. Which approach is better, and why?
✦ Pass every section knowledge check with 100% before submitting