Population health surveys: sampling, weights and design effect
- Concept
What a country knows about smoking, contraception or hypertension often comes from surveys of a few thousand homes. Reading them means knowing the sampling.
What it is
Population health surveys interview, and sometimes examine, a probability sample of households to estimate the health of a whole population. The sample is drawn from a sampling frame, usually in stages, strata and clusters; each respondent carries a weight equal to the inverse of their chance of selection, adjusted for non-response. Clustering widens confidence intervals (design effect), and answers that rely on self-report can be biased. Colombia's examples include the demographic and health survey (ENDS) and DANE's quality-of-life survey (ECV); in the United States, national surveys such as NHANES add physical examinations and laboratory tests to interviews.
Key facts
- Multistage cluster sampling: areas first, then households within areas, then a person within each household.
- Stratification (by region, urban or rural) improves precision; clustering lowers it.
- Colombia's ENDS rounds ran in 1990, 1995, 2000, 2005, 2010 and 2015, about every 5 years, on fertility, contraception, and maternal and child health.
- DANE's health-related surveys include the national quality-of-life survey (ECV) and the national survey of psychoactive substance use (ENCSPA).
- NHANES invites about 5 000 people a year and combines interviews with examinations in mobile centres and laboratory tests.
Worked example
- Made-up department: 750 000 urban and 50 000 rural adults. Survey: 1 500 urban (1 in 500) and 500 rural (1 in 100) in 100 clusters.
- Weights = 1 / probability of selection: 500 for each urban and 100 for each rural respondent.
- Self-reported hypertension: 300 urban (20 %) and 150 rural (30 %). Unweighted: 450 / 2 000 = 22.5 %.
- Weighted: (300 × 500 + 150 × 100) / 800 000 = 165 000 / 800 000 ≈ 20.6 %; rural people had been oversampled.
- Clustering gives a design effect of 2.0: the 2 000 interviews carry the information of 1 000 simple random ones.
- SE = √(0.206 × 0.794 / 1 000) ≈ 0.013; 95 % CI ≈ 20.6 % ± 2.5 points (18.1 % to 23.1 %).
- Blood pressure measured in a subsample gives 28 %: self-report counts only people who know their diagnosis.
- Reading: report the weighted 20.6 % with the wider interval, and treat self-reported prevalence as a floor.
Survey and the other sources of population health data
| Source | Who is counted | What it gives best | Main weakness |
|---|---|---|---|
| Population health survey | A probability sample of households | Prevalence of conditions and behaviours, including people outside health services | Self-report bias, non-response, sampling error |
| Census | Everyone, in principle | Population size and structure: the denominators | Few health questions; every several years |
| Vital statistics | Every birth and death registered | Mortality and fertility rates | Under-registration; cause of death coding |
| Notifiable disease surveillance | Cases reported by services | Timely trends and outbreaks | Misses people who never consult |
Markers
- Sampling frame
- Weights
- Deff
- ENDS
- ECV
- NHANES
In practice
- When citing a survey prevalence, check that it is weighted and that its interval accounts for the design.
- Compare rounds of the same survey for trends only if the questions and the sampling did not change.
- Local planning uses surveys together with registries and surveillance, because each misses different people.
Common traps
- A large sample guarantees a correct estimate: size reduces random error but not bias from a poor frame, non-response or self-report
- Unweighted percentages are fine for a national figure: groups sampled at different rates must be weighted back to their true share
- A cluster sample of 2 000 is as precise as a simple random sample of 2 000: clustering usually inflates the variance (design effect above 1)
- Self-reported hypertension or diabetes equals true prevalence: it misses undiagnosed people, so measured surveys find more
- A survey showing that people who exercise have less depression proves exercise prevents it: a cross-sectional survey cannot show which came first
Test yourself
Think of the answer, then tap the question to see it.
A national survey oversamples rural households. Why must the analysis use weights?
Rural respondents are a larger share of the sample than of the population; without weights the estimate leans towards rural values. Weights restore each group's true share.
A survey of 3 000 people has a design effect of 1.5. What is its effective sample size, and what does that mean?
3 000 / 1.5 = 2 000: its estimates are as precise as those of a simple random sample of 2 000, so intervals are wider than a naive calculation suggests.
Why does a survey that measures blood pressure usually find more hypertension than one that asks about it?
Many people with hypertension have never been diagnosed, so they answer no; measurement detects them.
Which people does a household survey usually miss?
Those outside the frame: people living in institutions (prisons, care homes, barracks), homeless people and areas that are unsafe or hard to reach.
Connections
Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.
Sources: ENDS y Estudios a Profundidad - Profamilia (Profamilia, 2026); DANE - Estadísticas por tema: Salud (Departamento Administrativo Nacional de Estadística (DANE), Colombia, 2026); DANE - Encuesta nacional de calidad de vida (ECV) (Departamento Administrativo Nacional de Estadística (DANE), Colombia, 2026); About NHANES | National Health and Nutrition Examination Survey | CDC (Centers for Disease Control and Prevention, 2026).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.
