Short answer: random sampling is how you pick who is in a study; random assignment is how you decide what happens to them once they’re in it. Random sampling buys external validity (the right to generalise your findings to a wider population). Random assignment buys internal validity (the right to claim your treatment, not some other variable, caused the outcome you measured). A study can have either, both, or neither — and most published research has only one of the two, which is the single most common source of over-claiming in results sections.
The distinction in one table
| Random sampling | Random assignment | |
|---|---|---|
| Question it answers | Who ends up in the study? | What condition does each participant get? |
| Mechanism | Every member of a defined population has a known, non-zero (often equal) chance of being selected from a sampling frame | Every enrolled participant has a known, non-zero (often equal) chance of being allocated to each condition (treatment, control, arm A/B/C) |
| Validity it supports | External validity — generalisability of results to the population the sample was drawn from | Internal validity — confidence that the manipulated variable, not a confound, produced the observed effect |
| What it protects against | Selection bias in who gets studied | Selection bias in who gets which treatment (confounding) |
| Where it happens | At recruitment, before anyone is enrolled | After enrolment, when conditions are allocated |
| Typical study type | Surveys, prevalence studies, large observational cohorts | Randomised controlled trials (RCTs), lab and field experiments |
| Does it require the other? | No — a randomly sampled population can then be assigned to conditions non-randomly | No — most RCTs recruit a convenience sample and randomise within it |
Random sampling, defined
Random sampling is a method for drawing a subset of individuals from a defined population such that every member of that population has a known probability of being selected. The purpose is representativeness: a properly drawn random sample lets you estimate population parameters (a mean, a proportion, a prevalence) with a quantifiable margin of error, and lets you generalise findings from the sample back to the population it was drawn from. See the CASRAI guide on simple random sampling for the mechanics of drawing one, and random sampling vs convenience sampling for how a non-random sample changes what you can claim. Stratified sampling is a variant that improves precision by sampling randomly within predefined subgroups rather than the population as a whole.
Random sampling says nothing about what happens to people once they’re in the study. A randomly sampled group of patients could then be split into treatment and control by clinic location, by physician preference, or by whoever showed up first — none of which is random assignment, and all of which reintroduce the confounding that random sampling was never designed to control.
Random assignment, defined
Random assignment (also called random allocation) is a method for placing already-enrolled participants into experimental conditions such that each has a known probability of ending up in each condition, independent of their own characteristics. The purpose is causal inference: because assignment to condition is unrelated to any pre-existing trait of the participant, any measured difference between groups at the end of the study can be attributed to the manipulated variable rather than to the kind of person who happened to end up in each group. This is the logical foundation of the randomised controlled trial and, more generally, of experimental (as opposed to observational) research design — see the CASRAI experimental design entry and the RCT definition. The formal case for randomisation as the safeguard against confounding in experiments traces to Ronald Fisher’s The Design of Experiments (1935), and it remains the reason regulators and journals treat the RCT as the strongest single-study design for causal claims — see also RCT vs observational study.
Random assignment says nothing about who was eligible to be assigned in the first place. A trial that randomises 200 volunteers recruited from one hospital’s outpatient list has excellent internal validity for that group of 200, and no guaranteed claim to represent patients at other hospitals, in other countries, or with different demographics — because the volunteers were not randomly sampled from that wider population.
The 2×2: what you get with each combination
Crossing the two randomisations produces four distinct study profiles. Almost all methodological confusion in this area comes from treating “randomised” as a single property, when a study report needs to specify which of the two — or both — it means.
1. Random sampling + random assignment (both)
What you get: the strongest possible design — internal validity and external validity together. The causal effect you find in the sample is both trustworthy (random assignment) and generalisable to the source population (random sampling).
Worked example: a national health authority draws a random sample of 5,000 adults from the national population register, then randomly assigns each participant to receive a reminder text message or no reminder before a scheduled screening appointment, measuring attendance. Because the sample was randomly drawn from the national register, the attendance effect generalises to the national population; because assignment to the reminder condition was random, the effect can be attributed to the reminder itself.
Reality check: this combination is uncommon and expensive. Randomly sampling an entire target population and then getting every one of them to consent to experimental manipulation is logistically difficult, which is why most experiments settle for design #2 below.
2. Random assignment only, no random sampling
What you get: strong internal validity, weak/undetermined external validity. This is the profile of the great majority of published RCTs, lab experiments, and A/B tests.
Worked example: a university lab recruits 120 undergraduate volunteers (a convenience sample — not randomly drawn from any defined population) and randomly assigns them to a memory-training task or a control task. The difference in recall scores between groups can be causally attributed to the training, because assignment was random. Whether that effect holds for adults generally, older adults, or non-students is not established by this design — that is an external-validity question the study cannot answer on its own.
Why this is so common: random assignment is entirely within the researcher’s control once participants have agreed to take part. Random sampling from a full target population usually is not — there is rarely a complete sampling frame of “everyone who could someday take this drug” or “everyone who might use this product” to sample from.
3. Random sampling only, no random assignment
What you get: strong external validity, no basis for causal claims. This is the profile of most surveys, prevalence studies, and observational cohort studies.
Worked example: a national statistics agency draws a random sample of households from a postal address register and measures both smartphone ownership and self-reported wellbeing. Because the sample was randomly drawn, the association observed generalises to the national population with a known margin of error. But because no one was randomly assigned to own or not own a smartphone — ownership is a self-selected, pre-existing characteristic — the study cannot establish that smartphone ownership causes any wellbeing difference. Confounds (income, age, employment) were never controlled by randomisation.
Why researchers accept this trade-off: many real-world variables (income, diagnosis, occupation, prior exposure) cannot ethically or practically be randomly assigned. Random sampling is often the only randomisation available, which is exactly why observational designs report associations, not causal effects — see internal vs external validity for how reviewers and readers should weigh that trade-off.
4. Neither random sampling nor random assignment
What you get: the weakest combination for both generalisability and causal inference, though such studies can still be useful for hypothesis generation, feasibility, or description.
Worked example: a clinician reviews outcomes for the last 40 patients who happened to receive a new surgical technique at their own hospital, compared informally to their memory of outcomes under the old technique. The patients were not randomly sampled (they are whoever presented at that hospital in that period) and were not randomly assigned to technique (the surgeon chose based on case suitability). Neither the causal claim (“the new technique caused better outcomes”) nor the generalisation (“this holds for patients elsewhere”) is supported by the design alone — though the observation may justify a properly randomised follow-up study.
Why the two get conflated
Three habits of language drive the confusion:
- “Randomised” is used as a one-word badge of rigor. A press release or abstract that says a study was “randomised” almost always means random assignment (an RCT), not random sampling — but readers without methods training often assume it implies both, including generalisability the study never claimed.
- “Random sample” is sometimes used loosely for “convenience sample I didn’t hand-pick.” Recruiting whoever responds to a flyer or an online panel invitation is not random sampling in the statistical sense (there is no defined sampling frame with known selection probabilities) — see random sampling vs convenience sampling for the exact line between the two.
- Reviewers ask “was it randomised?” without specifying which randomisation they mean. The precise, reviewer-proof phrasing is “participants were randomly <sampled from / assigned within> <population/pool>” — stating both halves removes the ambiguity.
How to state your own design precisely
When writing a methods section, name both randomisations explicitly and separately, even when one of them didn’t happen:
- “Participants were recruited via convenience sampling from three outpatient clinics and randomly assigned to intervention or control using a computer-generated allocation sequence.” — correctly claims internal validity only.
- “Households were randomly sampled from the national postal address file; no experimental manipulation was applied.” — correctly claims external validity only, for an observational design.
- “A random sample of 5,000 registered patients was drawn from the regional health register and randomly assigned to receive the reminder intervention or usual care.” — correctly claims both.
The same discipline applies when reading someone else’s study: check both properties before accepting either a causal claim or a generalisation claim at face value. A well-designed sample size and power calculation tells you the study was adequately sized to detect its effect — it says nothing about which of the two randomisations, if either, the study actually used.
Frequently asked questions
Is an RCT automatically a random sample?
No. “Randomised controlled trial” refers to random assignment to trial arms. The trial’s participants are typically a convenience or eligibility-screened sample of volunteers who consented to enrol, not a random sample of the target patient population. This is why RCT results are described as internally strong but require separate justification (or replication in other settings) to generalise.
Can you randomly assign without randomly sampling?
Yes — this is in fact the normal case. Almost every lab experiment and most clinical trials randomly assign within a convenience sample of people who agreed to participate; the sample itself was never drawn at random from a larger population.
Can you randomly sample without randomly assigning?
Yes — this is the standard design for surveys, censuses, and observational cohort studies, where there is no manipulated variable to assign at all, only characteristics that already exist in the population.
Which one matters more for my study?
That depends on your research question. If the goal is to estimate a causal effect, prioritise random assignment. If the goal is to estimate a population parameter (a prevalence, an average, a proportion) accurately, prioritise random sampling. Studies that need to do both need both randomisations, which is why design #1 above is rare and valuable when it’s achievable.
Does random assignment eliminate the need for a large sample?
No. Random assignment controls for confounding, not for statistical power. A study can be perfectly randomised and still be underpowered to detect its effect — that is a separate question addressed by a sample size and power calculation.
