Sampling in Clinical Research
Start from the one fact that makes statistics necessary at all: you can almost never measure everyone you care about. There is no register of every pregnant woman in South Africa with her haemoglobin, her blood pressure, and her delivery outcome attached. So research takes a small, reachable group, measures it carefully, and then reasons backwards — from the few you saw to the many you did not. That reasoning is called inference, and sampling is simply the rule by which you chose the few. Everything in this chapter is one idea worked out in detail: the sample is allowed to stand in for the population only to the extent that you can defend how it was selected.
That single idea has a sharp consequence. A study can use the correct statistical test, produce a beautifully narrow confidence interval, and still be completely wrong, because the test only describes the sample it was given. If the sample was the wrong group, the maths is precise about the wrong thing. In O&G this matters at every level of evidence — a clinic survey of contraception uptake, a cohort of women with preterm prelabour rupture of membranes (PPROM), a case-control study of stillbirth, a diagnostic accuracy study of a pre-eclampsia biomarker, a national maternal mortality audit. In each, the result you can trust is set as much by who got into the sample as by what was measured in them.
The Primary FCOG task is therefore to learn the language and the logic, in that order. The language: target population, source population, sampling frame, eligibility, sample, responders, analysed denominator, probability sampling, non-probability sampling, sampling error, selection bias, non-response, clustering, generalisability. The logic: how each of these can widen or narrow the gap between your sample and the group you actually want to talk about. The exam usually tests this indirectly — it gives you results from a tertiary unit, a small audit, or a household survey and asks whether they apply to a district hospital population. The rest of this chapter builds the tools to answer that, one layer at a time, starting from who you are studying and ending with how many of them you need.
Population, Frame, Sample and Denominator
Because inference runs from the sample back to the population, the first job is to name both ends of that chain precisely. Sampling starts before recruitment: you must state exactly who you want to make claims about, and exactly where your participants can actually be found. Those two are rarely identical, and the gap between them is where most sampling problems live.
| Term | Meaning | O&G example | Why it matters |
|---|---|---|---|
| Target population | The group you want to generalise to | Pregnant women in South African public-sector care | Defines external validity |
| Source population | The accessible population from which the sample is drawn | Women booking at selected district clinics | May differ from target population |
| Sampling frame | The list or system used to identify eligible people | Clinic booking register, theatre logbook, birth register | Incomplete frames miss people systematically |
| Eligible population | Those meeting inclusion/exclusion criteria | Singleton pregnancies booking before 20 weeks | Criteria shape applicability |
| Sample | Those selected or recruited | 800 women selected from registers | Selection method determines sampling error |
| Responders/retained | Those who provide usable data | 710 completed questionnaire and delivery follow-up | Non-response can bias results |
| Analysed denominator | Those included in final analysis | 690 with haemoglobin and outcome data | Exclusions after recruitment can distort estimates |
External validity — whether the result travels to the people you care about — depends on the distance between target population, source population, sample, and analysed denominator. Each step in that chain can quietly shrink who the result really describes. A tertiary fetal medicine sample may be excellent for describing severe placental disease referred to a specialist service, but it should not be used to estimate the prevalence of abnormal umbilical artery Doppler in all pregnancies, because the source population was already filtered by referral before anyone was sampled.
With the two ends of the chain named, the next question is how you reach into the source population to pick your sample. There are two fundamentally different families of method, and the difference decides whether honest inference is even possible.
Probability Sampling
In probability sampling, every eligible unit has a known, non-zero probability of being selected, and selection is left to chance rather than to convenience or judgement. This is the family that makes formal inference legitimate, because if you know each person's chance of being chosen, you can estimate how much your result might have wobbled by luck of the draw — the sampling error. Without a known selection probability there is no defensible way to put a confidence interval on a population estimate.
| Method | How it works | Strength | Main trap |
|---|---|---|---|
| Simple random sampling | Select individuals randomly from a complete list | Conceptually clean; easy inference | Needs a complete sampling frame |
| Systematic sampling | Select every kth eligible person after a random start | Practical in registers or clinics | Hidden periodicity can bias selection |
| Stratified sampling | Divide into strata, then sample within each | Ensures representation of important groups | Requires correct weighting if strata oversampled |
| Cluster sampling | Randomly sample groups, then individuals within groups | Efficient for clinics, hospitals, districts | People within clusters are correlated |
| Multistage sampling | Sample in stages, e.g. provinces -> districts -> clinics -> women | Feasible for large surveys | Needs weights and complex analysis |
Stats SA household surveys illustrate the principle: households are scientifically selected, cannot simply be replaced by easier households, and represent the population only because the sampling design and weights are respected. In clinical research, the same discipline applies. If selected women are replaced whenever they are absent, the final sample may over-represent women who are easier to contact, more stable, or less controlled by partners.
Stratification
Stratification is useful when important subgroups might otherwise be too small. An antenatal anaemia prevalence survey may stratify by district, urban/rural setting, or level of care. A contraception survey may stratify by age group because adolescent uptake is clinically important and may be under-represented in a simple sample.
Oversampling a stratum is allowed, but analysis must account for it. If adolescents are deliberately sampled at twice their population proportion, the raw sample prevalence cannot be quoted as district prevalence without weighting.
Cluster sampling and design effect
Cluster sampling is common in health-services research because it is easier to sample clinics than scattered individual patients. The price is loss of statistical efficiency, and the reason is worth understanding rather than memorising. Most statistical tests assume that observations are independent — that knowing one woman's outcome tells you nothing about the next woman's. Clustering breaks that assumption. Women attending the same clinic share staff, protocols, geography, socioeconomic context, referral pathways, and record systems, so their outcomes resemble each other more than they resemble women picked at random across the whole province. Each extra woman from an already-sampled clinic therefore adds less new information than a genuinely independent woman would. The strength of this within-cluster resemblance is captured by the intracluster correlation coefficient (ICC): at ICC = 0 the cluster members behave as independent individuals, while a positive ICC means they partly duplicate one another. Even a small ICC can sharply inflate the required sample size when clusters are large, because the effective sample size — the number of truly independent units — is much smaller than the headcount.