Module 2 · Types of Studies
Two studies. Same question. Opposite designs.
Both of these set out to ask whether smoking causes lung cancer. Read them:
Study A
Researchers took 30,000 healthy adults, recorded who smoked and who didn't, and followed them for fifteen years to see who developed lung cancer.
Study B
Researchers took 400 lung cancer patients and 400 cancer-free patients, and compared how much each group had smoked in the past.
Same question, similar wording. Are these the same kind of study?
Read the fingerprint
Here's the good news: you will not have to memorise a list of study names. Every study design is just the result of a few design choices — and each choice leaves a fingerprint in the way the study is described.
Learn to spot the fingerprints, and you can name any study in the world from a single sentence of its methods — and, more importantly, know what it's allowed to claim.
Every study answers three questions, and its answers decide its type:
- Did the researcher assign the exposure — or just watch?
- Is there a comparison group at all?
- Where does the study start — from the exposure, the outcome, or neither?
Three questions. Walk them in order and the study names itself. Let's take them one at a time.
Fork 1: assign or observe?
Question 1: did the researcher assign the exposure, or just watch?
This is the deepest divide of all — and you already know why it matters. When the researcher assigns who gets the treatment, they can do it by chance, making the groups identical and shutting out confounding. When they only watch who happened to get it, they can't.
Experimental — the researcher assigns the exposure (gives the drug, withholds it).
Observational — the researcher only watches what people already do or receive.
The fingerprint is in the verbs. "Patients were assigned / allocated / randomised to…" → experimental. "Patients who took / chose / happened to receive…" → observational. The researcher's hand, or just their eye.
Inside experimental: the coin toss
Step into the experimental branch. One more question splits it:
Randomised — assignment was by chance → Randomised controlled trial (RCT). The coin toss is what makes the groups identical, so nothing hidden can confound the result. This is the rung you met at the top of the trial world last lesson.
Not randomised — the researcher assigned the treatment, but chose who got what → non-randomised controlled trial. Without the coin toss, the groups can differ from the start — confounding sneaks back in. An assigned study, but with its main protection switched off.
The fingerprint: the single word randomised (or "allocated by chance"). Its presence is the difference between the strongest design and a weaker cousin.
Read each method. Tap the design — and watch for the fingerprint.
Fork 2: is there a comparison group?
Now the observational branch — where most real-world HTA evidence lives. Question 2 splits it:
Analytical — there's a comparison group (exposed vs unexposed, or diseased vs not). You can actually compare, so you can hunt for a link.
Descriptive — no comparison group. Just a description of some patients. You can describe, but you can't compare — so you can't establish that anything caused anything.
The fingerprint: is anyone being compared to anyone? "Compared with…", "versus…", "relative to controls" → analytical. "We describe a series of patients who…", with no control group → descriptive.
The analytical branch is where the three observational designs from your hook live. They differ on Question 3 — and this is the one everyone gets wrong.
The heart: where does the study start?
Question 3: where does the study start — and which way does it travel?
This single question separates the three analytical designs. Watch the direction of the arrow.
exposure
(cases vs controls)
(same moment)
Cohort — starts from exposure. Sort people by what they're exposed to, then follow them forwards in time to see who develops the outcome. Arrow points forward.
Case-control — starts from the outcome. Take people who already have the disease (cases) and people who don't (controls), then look backwards to compare their past exposure. Arrow points backward.
Cross-sectional — starts from neither. Measure exposure and outcome at the same moment — a snapshot. No arrow; one slice of time.
Cohort — enters by EXPOSURE
Case-control — enters by OUTCOME
Here's the cleanest way to tell cohort from case-control — and it's not about time, it's about the door people walk in through:
- Cohort: people enter sorted by their exposure (smokers / non-smokers).
- Case-control: people enter sorted by their disease (cancer / no cancer).
Ask one question — what defines who's in the study, their exposure or their outcome? — and cohort and case-control come apart instantly.
Three observational designs. For each, find the door — and the direction.
The time trap: prospective vs retrospective
Time for the mistake almost everyone makes. People assume that if a study uses old records — looking into the past — it must be a case-control study. It isn't necessarily.
"Using employment records from 1985 to 2005, researchers identified workers exposed to a chemical, then traced each worker forward through the records to see who later developed disease." What design is this?
No comparison group: descriptive
Back to the descriptive branch we set aside — the bottom of the ladder.
Sometimes there's no comparison at all. A doctor notices something striking and writes it up:
- Case report — one remarkable patient.
- Case series — a handful of similar patients.
The fingerprint is an absence: no control group, nobody to compare against. "We describe five patients who developed an unusual reaction after…" — and that's it.
These aren't worthless — a sharp case series has raised the first alarm on many a drug side-effect. But with no comparison, they can only describe, never show cause. They're where evidence starts, not where it settles.
Choosing the design — it's not a ranking
One more idea, and it rescues you from a trap from last lesson: the "higher" design isn't always the right one. The design has to fit the question.
- Studying a rare disease? A cohort is hopeless — you'd follow a million people to find a handful of cases. A case-control study starts from those rare cases, so it's the practical winner.
- Tracking a new exposure's effects over time? A cohort is built for exactly that.
- Just need to know how common something is right now? A cross-sectional snapshot answers it directly.
The best design is the one that fits the question — not the one highest on a ladder. Remember: a rung is a starting presumption, not a verdict.
Assemble: the full map
Put all three questions together and the whole family tree appears — the one you just built, branch by branch:
And the fingerprints, in one place to keep:
| Design | The tell | What it can show |
|---|---|---|
| RCT | "randomly allocated" | Strongest evidence of cause |
| Non-randomised trial | assigned, but not by chance | Weaker — confounding possible |
| Cohort | enters by exposure, runs forward | Risk over time; can suggest cause |
| Case-control | enters by outcome, looks back | Links for rare diseases; not direct risk |
| Cross-sectional | everything measured at once | What travels together — never what came first |
| Case series / report | no comparison group | A first alarm — never proof |
Why this matters for HTA
Here's where it lands on your desk. When a manufacturer's evidence isn't a clean RCT — and increasingly it isn't, especially the "real-world evidence" you'll meet in M11 — it's one of these observational designs. The moment you can name the design, you know exactly which enemy to hunt:
- A cohort comparing drug-takers to non-takers? The takers chose differently — look for confounding.
- A case-control asking patients to recall past use? People with the disease remember harder — look for recall bias.
- A cross-sectional snapshot claiming one thing caused another? It literally can't — there's no time order in a snapshot.
Naming the design isn't trivia. It tells you, before you read a single result, what the study is allowed to claim — and where it's most likely trying to claim too much.
Types of studies, in one breath
- You don't memorise study names — you read fingerprints, by asking three questions.
- Did the researcher assign, or watch? → experimental vs observational.
- Is there a comparison group? → analytical vs descriptive.
- Where does it start? → cohort (from exposure, forward), case-control (from outcome, backward), cross-sectional (snapshot).
- The big trap: direction of reasoning (cohort vs case-control) is not the same as when data were collected (prospective vs retrospective).
- The best design fits the question — it isn't simply the highest rung.
Name the design, and you already know what it can prove — and what it's quietly pretending to. That's half of critically reading any study.
You've now met the designs. The next pieces of M2 sharpen the tools they rely on: how randomising and blinding actually shut out bias, and how to tell chance, bias and confounding apart for real.