Cross-Sectional, Prospective Cohort and Retrospective Case-Control Studies: A Quick Reference
Three observational designs, and one structural difference that explains almost everything else about them. It is the direction of inquiry: whether exposure and outcome are measured at once, whether exposure is followed forward to outcome, or whether the outcome comes first and exposure is traced back. What each design can estimate, and which bias threatens it, both follow from that.
- Which measure of association belongs to which design, and why
- Why a cross-sectional study cannot measure incidence
- Which bias threatens each design, and where it enters
- “A cohort study establishes cause and effect.” A prospective cohort establishes that exposure preceded outcome, which is a necessary condition for causation and not a sufficient one. Confounding remains, and no amount of follow-up removes it.
- “Prevalence shows how many new cases there are.” Prevalence counts existing cases at one moment. Incidence counts new cases arising over a period. They are a stock and a flow, and a snapshot can only see the stock.
What this article covers
This is a reference article rather than a tutorial. It sets the three designs side by side, states what each can and cannot support, and sets out the comparison as a table.
All three are observational, which means the investigator records what happened rather than assigning who is exposed. That single constraint is what separates them from experimental designs, and it is why confounding is a live concern throughout.
The designs where the investigator does assign the exposure are covered in experimental research design, and the designs where assignment is partial or outside the investigator’s control in quasi-experimental design. Measuring association without manipulating anything is covered in correlational research design.
This article stays with the three observational designs named in its title.
The three designs, told apart by direction
Most confusion between these designs dissolves under a single question. At the moment of data collection, what did the investigator already know about each participant?
Cross-sectional: both at once
Exposure and outcome are measured at the same moment, in a sample drawn without regard to either. The result is a snapshot of who has what, right now.
Because nothing is measured twice, the design cannot show which came first. A sample in which smokers have more cough shows that the two travel together, not that one preceded the other.
Prospective cohort: exposure first, then wait
Participants are grouped by exposure before the outcome has occurred, and then followed forward. The investigator knows who was exposed and does not yet know who will develop the outcome.
That ordering is the design’s whole advantage. It establishes temporality directly, because the exposure was recorded before the outcome existed.
Retrospective case-control: outcome first, then look back
Participants are selected because they have the outcome, or because they do not, and exposure is then ascertained for both groups. The direction of inquiry runs backwards from effect to possible cause.
This is efficient for rare outcomes, because cases can be assembled directly rather than awaited. It also creates the design’s two characteristic problems, which are how controls were chosen and how exposure is remembered.
Sampling on exposure leaves the outcome free to vary, so its frequency can be counted in each group. Sampling on outcome fixes the number of cases in advance, so the frequency of the outcome cannot be counted at all.
That is the structural reason the designs yield different measures of association, and it is covered in full below.
Figures: comparing observational study designs
Figure 1 shows the direction of inquiry in cross-sectional, cohort and case-control studies. Figure 2 explains why cohort studies can report risk ratios while case-control studies report odds ratios. Authors can use them to choose the correct design and effect measure, and to describe both accurately.
The same two variables, exposure and outcome, approached three different ways. Everything else about these designs follows from this.
What to look for. The arrows point in opposite directions, and that is the whole difference. The cohort investigator knows who was exposed and waits to see who develops the outcome. The case-control investigator knows who has the outcome and works back to exposure. The cross-sectional investigator knows both at once and therefore knows neither first.
Schematic. Retrospective cohort studies also exist, in which the exposure and outcome have both already occurred but the grouping is still by exposure; the direction of inquiry stays forward.
The same two-by-two table, filled two different ways. Whichever margin is fixed by design is the one that cannot afterwards be estimated.
What to look for. The arithmetic is identical on both sides; the meaning is not. In the cohort table the row totals arise from the sampling, so dividing within a row yields a genuine risk. In the case-control table the number of cases and controls was decided in advance, so the same division yields a number reflecting the recruitment ratio rather than anything about the population.
Schematic. The odds ratio is not merely a fallback: under density sampling it estimates the incidence rate ratio directly, which is covered in the text below.
The quick reference table
The three designs are set out below on the dimensions that distinguish them. Each row is discussed at greater length in the sections that follow.
| Cross-sectional | Prospective cohort | Retrospective case-control | |
|---|---|---|---|
| Participants selected by | Neither exposure nor outcome | Exposure status | Outcome status |
| Direction of inquiry | Neither; both measured at once | Forward, exposure to outcome | Backward, outcome to exposure |
| Also known as | “Also called prevalence studies”1 | “The gold standard of observational research”, in one assessment1 | — |
| Measures of frequency available | Prevalence | Incidence; “the only observational study that can calculate incidence”1 | Neither; the number of cases is fixed by design |
| Measure of association | Prevalence ratio or odds | Risk ratio, rate ratio, hazard ratio | “The measure of risk that is calculated in case-control studies is the odds ratio”1 |
| Temporality | Not generally demonstrable17 | Demonstrated directly1 | Depends on the quality of exposure ascertainment |
| Characteristic bias | Selection bias in sampling; information bias7 | Loss to follow-up, which is selective3 | Recall bias; control selection13 |
| Suits a rare outcome | No | “Not good for rare diseases”1 | Yes; “the most efficient study design for rare diseases”1 |
| Number of exposures examinable | Many | Few, fixed at recruitment | Many, from one set of cases |
| Number of outcomes examinable | Many | Many, from one cohort | One, fixed by case definition |
| Time and cost | “Relatively inexpensive” and “timely”1 | “Expensive” and “time intensive”1 | “Relatively fast execution compared to cohort studies”1 |
| Reporting checklist | STROBE covers all three. It “addresses the three main types of observational studies: cohort, case-control and cross-sectional studies”3 | ||
The pairing of design to scarcity is only half documented. That a case-control study suits a rare outcome is stated directly in the methodological literature.1 The symmetrical claim, that a cohort suits a rare exposure, could not be verified in any source consulted for this article.
It follows from the structure of the designs rather than from a citation: an exposed group can be recruited deliberately, whereas controls drawn from a general population will seldom carry a rare exposure. It is presented here as a structural inference, not as a finding.
What each design can measure
Prevalence and incidence are not the same quantity
Prevalence is a stock and incidence is a flow. The authoritative definitions make the distinction turn on the numerator: “prevalence includes new and pre-existing cases whereas incidence includes new cases only”, and “the key difference is in their numerators”.6
Incidence also carries a denominator that prevalence does not. The incidence proportion is “the proportion of an initially disease-free population that develops disease”, while an incidence rate “incorporates time directly into the denominator”.6 Prevalence, by contrast, is “the proportion of persons in a population who have a particular disease or attribute” at a moment.6
A single round of measurement cannot observe a flow. Cross-sectional studies are “also called prevalence studies”, because what they capture is “the proportion of participants with disease at a given point in time” rather than “the accumulation of new cases over time”.1 A review of the design lists among its weaknesses “the inability to assess incidence”.7
The positive statement is narrower than it is often rendered. Cohort studies are described as “the only observational study that can calculate incidence”.1
Temporality, and how firmly it can be claimed
A prospective cohort records exposure before the outcome exists, and so it has “the ability to demonstrate temporality”.1 For a cross-sectional study the sample “is assessed for both exposure and outcome at a single point in time”, with the consequence that “temporality cannot be demonstrated”.1
Sources differ on how absolute that last statement should be. The reporting guideline’s own account is softer, noting that in cross-sectional work “the time order of exposure and outcome may sometimes be clear”.3 A review of the design likewise says it is “difficult to determine whether the exposures preceded or followed the outcomes” and that establishing causal relationships is “relatively difficult”, rather than impossible.7
The defensible position sits between them. Temporality is not generally available from a cross-sectional design, and where a particular study can argue for it, the argument has to be made explicitly rather than assumed from the data structure.
A cohort design buys the time order, and the time order strengthens a causal argument without completing it. The same source that calls these designs “stronger measures for demonstrating causation” also states the limit directly: “causal conclusions cannot be proven from an observational study”.1
What remains is confounding, which the reporting guideline glosses as “confusion of effects” and describes as “crucial” to take into account in observational work.3 Adjustment reduces it and does not abolish it.
Why case-control studies report an odds ratio
The reason is structural rather than conventional. In a case-control study “selection into the study is intentionally conditional on the occurrence/non-occurrence of an outcome event”, with the result that “denominators from which rates could be calculated are not measured”.5
Figure 2 above shows the arithmetic consequence. The number of cases is a recruitment decision, so dividing cases by the total in a row describes the recruitment ratio rather than any population quantity. The odds ratio is what survives that constraint, and it is accordingly described as “the measure of risk that is calculated in case-control studies”.1
The first is the familiar one, that a case-control study simply cannot estimate relative risk. The second is the correction that has become almost as common, that choosing controls from those still at risk makes the odds ratio an estimate of the incidence rate ratio.
The more careful treatment makes the second conditional rather than automatic. Consistency for a rate ratio requires that “the case-control sampling is time-matched” and holds “if the HR is constant over the study period”; without that condition the sample odds ratios “are consistent for a weighted mean of HRs”.5
The same authors caution against the other common shortcut, noting that “the same caution applies when using population ORs as approximations to risk ratios”.5 What a given odds ratio estimates therefore depends on how controls were sampled, which is a reason to report that sampling scheme rather than to leave it implicit.
The biases, each in its own place
Bias here has a technical meaning: “a systematic deviation of a study’s result from a true value”.3 The three designs are threatened by different kinds, and the usual error is to attribute a bias to the wrong design.
Selection bias arises twice, by two different routes
The general statement is that “selection bias affects the internal validity of a study”.3 How it enters depends on how participants were obtained.
In a case-control study it enters through recruitment, where the danger is that “the probability of including cases or controls is associated with exposure”.3 The remedy is a principled choice of comparison group: “in general, controls should reflect the population from which the cases arose”.3
In a cohort study it enters later, through attrition. “Loss to follow-up occurs when participants withdraw from a study” before the end of observation,3 and it becomes a bias rather than a nuisance when “loss to follow-up occurs selectively in exposed individuals”.3 Describing a cohort as minimising selection bias therefore inverts the position: attrition is the design’s characteristic selection problem.
In a cross-sectional study it enters through sampling, where “selection bias from non-random sampling may result in flawed measure of prevalence”.1
Recall bias belongs to the retrospective design
Information bias is the broader category, occurring “when systematic differences in the completeness or the accuracy of data” arise between groups.3 Recall bias is the form it takes when participants differ in how readily they report exposures, “or if they are more or less willing to report them”.3
The mechanism is specific to looking backward from a known outcome. Case-control studies are “at increased risk for bias, particularly recall bias”, and the reason given is the participants’ “known case status”: a person who has the disease is “more likely to remember an exposure prior to becoming a case”.1
No amount of statistical adjustment addresses this, because the measurement itself differs between the groups being compared. The defences available are design choices such as using records made before the outcome occurred, and honest discussion of what remains.
Confounding sits underneath all three
Confounders are “variables that either conceal or falsely demonstrate associations between the exposure and outcome”.1 A standard illustration is “tobacco use confounding the relationship between coffee intake and cardiovascular disease”.1
Design can reduce the problem without removing it, since “matching on one or more potential confounders allows for minimization of those factors”.1 Matching carries an analytical obligation in return: in a matched case-control study, “a crude analysis of the odds ratio, ignoring the matching, usually leads to an estimation that is biased towards unity”.3
What the reporting checklist asks for
STROBE was written for exactly these designs, being addressed to “the 3 main study designs of analytical epidemiology: cohort, case-control, and cross-sectional studies”.2 It covers all three in one checklist, with branches where the designs differ. For bias it asks authors to “describe any efforts to address potential sources of bias”,4 and that item names no design, so the burden of saying which bias applies falls on the author.
The design-specific obligations sit in the participant items. A case-control study must “give the eligibility criteria, and the sources and methods of case ascertainment and control selection”, together with “the rationale for the choice of cases and controls”.4 A cohort study must give its eligibility criteria and selection methods and “describe methods of follow-up”.4 Matched studies in either design must report the matching criteria.4
On results, the checklist asks for unadjusted and adjusted estimates “and their precision (e.g., 95% confidence intervals)”, and for authors to “make clear which confounders were adjusted for and why they were included”.4 It does not prescribe which measure of association each design should report, which is why the preceding section matters.
Choosing between them
The choice is usually settled by two facts about the research question, and only afterwards by time and budget.
The first fact is what has to be estimated. A question about how common a condition is in a population is answered by a cross-sectional study, and by nothing else on this page. A question about how often a condition arises requires follow-up, which means a cohort.
The second fact is which of the two variables is rare. The two designs fail at opposite ends of that question, and the asymmetry is the most useful selection rule available.
| Situation | Suitable design | Reason |
|---|---|---|
| The outcome is rare | Case-control | Cases are recruited directly rather than awaited, so no cohort has to be large enough to generate them |
| The exposure is rare | Cohort | An exposed group can be assembled deliberately, whereas few controls in a case-control study would carry the exposure |
| Several outcomes are of interest | Cohort | One exposed group can be followed for any number of outcomes |
| Several exposures are of interest | Case-control | One set of cases can be questioned about any number of exposures |
| Only current burden is required | Cross-sectional | A single round of measurement answers the question |
| Temporal order is disputed | Cohort | Exposure is recorded before the outcome exists |
Cost and duration follow from those choices rather than driving them. A cross-sectional study is the quickest and cheapest of the three because it measures once. A prospective cohort is the slowest and most expensive because it waits. A case-control study sits between them, since the outcome has already occurred when recruitment begins.
Budget is a legitimate constraint and it is not a methodological argument. A case-control study selected because a cohort was unaffordable still has to answer for control selection and for how exposure was ascertained.
The practical consequence is that the reasons for the choice belong in the methods section, stated as design decisions rather than as apologies.
Common mistakes
| Mistake | Design | Why it is wrong | The accurate statement |
|---|---|---|---|
| Reporting incidence from a snapshot | Cross-sectional | Prevalence counts existing cases at a moment; incidence counts new cases over a period | A cross-sectional study estimates prevalence and cannot estimate incidence |
| Claiming causation from follow-up | Cohort | Temporal order is necessary for causation but not sufficient, and confounding persists | The design establishes temporality and association |
| Reporting a risk ratio from case-control data | Case-control | Sampling on outcome fixes the number of cases, so no risk can be computed | The odds ratio is the measure the design yields |
| Stating flatly that case-control cannot estimate relative risk | Case-control | An overcorrection; the relationship depends on how controls were sampled | The interpretation follows from the sampling scheme |
| Attributing minimal selection bias to cohorts | Cohort | Loss to follow-up is itself a form of selection bias | Attrition is the characteristic selection problem of a cohort |
| Omitting recall bias from a retrospective design | Case-control | It is the signature threat where exposure is reported after the outcome is known | Recall bias requires explicit discussion |
| Describing controls without saying how they were chosen | Case-control | The source population for controls determines what the odds ratio means | Control selection belongs in the methods, in detail |
| Citing whole books without page numbers | Any | An attribution that cannot be located cannot be checked | Specific claims require locatable sources |
One structural fact separates these three designs: what the investigator already knows at the moment of sampling. Selecting on neither variable gives a snapshot, which yields prevalence and not incidence. Selecting on exposure and waiting gives temporality and a risk ratio. Selecting on outcome fixes the number of cases, which is why no risk can be computed and the odds ratio is what remains. Each design also carries its own characteristic bias, and attributing the wrong one is the commonest reporting error of the three.
How to apply this to your work
- State the quantity to be estimated — prevalence, incidence, or an association — before considering any design.
- Identify which variable is rare, the exposure or the outcome, since that governs feasibility.
- Check whether temporal order is in dispute; if it is, only a design that records exposure first will settle it.
- Name the measure of association the chosen design can support, and check it against the analysis plan.
- For a case-control study, specify the control source and the sampling scheme, because both govern interpretation.
- For a cohort study, plan for attrition and record who was lost and when.
- Name the characteristic bias of the chosen design in the protocol, not only in the discussion.
- Work through the STROBE checklist for the relevant design branch before submission.
Common questions
What is the main difference between cohort and case-control studies?
The direction of inquiry. A cohort study groups participants by exposure and follows them forward to see who develops the outcome. A case-control study selects participants because they have or do not have the outcome, and then establishes what their exposure was. Everything else, including which measure of association is available, follows from that difference.
Can a cross-sectional study measure incidence?
No. Prevalence counts existing cases at a moment and incidence counts new cases arising over a period, so “the key difference is in their numerators”.6 A single round of measurement cannot observe a flow, and a review of the design lists “the inability to assess incidence” among its weaknesses.7 Cohort studies are described as “the only observational study that can calculate incidence”.1
Does a cohort study prove causation?
No. It establishes that exposure preceded outcome, which is a necessary condition for causation and not a sufficient one. Temporality makes a causal argument stronger, but the position is stated plainly in the methodological literature: “causal conclusions cannot be proven from an observational study”.1 Confounding remains, and adjustment reduces it without abolishing it.
Why do case-control studies report odds ratios instead of relative risk?
Because the sampling fixes the number of cases. Selection is “intentionally conditional on the occurrence/non-occurrence of an outcome event”, so “denominators from which rates could be calculated are not measured”.5 Without those denominators no risk can be computed, and the odds ratio is the measure that survives the constraint.
Is it true that a case-control study can never estimate relative risk?
No, and the opposite shortcut is equally unsafe. Whether an odds ratio estimates a rate ratio depends on the sampling: consistency requires that “the case-control sampling is time-matched” and holds “if the HR is constant over the study period”, failing which the odds ratios “are consistent for a weighted mean of HRs”.5 The practical consequence is that the control-sampling scheme belongs in the methods.
Which bias threatens which design?
Recall bias threatens retrospective case-control studies, because participants have “known case status” and a person with the disease is “more likely to remember an exposure prior to becoming a case”.1 Loss to follow-up threatens cohorts, and becomes a bias when it “occurs selectively in exposed individuals”.3 Non-random sampling threatens cross-sectional estimates of prevalence.1
Do cohort studies minimise selection bias?
No, and the claim inverts the position. A cohort’s characteristic selection problem arrives through attrition: “loss to follow-up occurs when participants withdraw from a study” before observation ends,3 and it distorts the result when the loss is related to exposure.3 Reporting how many were lost, and when, is a checklist requirement rather than a courtesy.
Which reporting checklist applies to these designs?
STROBE, which “addresses the three main types of observational studies: cohort, case-control and cross-sectional studies”.3 It is one checklist with design-specific branches rather than three separate ones: items on participants and matching split by design, while the items on bias, estimates and confounders apply to all three.4
Declared interest: Pubrica provides research and medical writing support commercially. Nothing on this page should be read as suggesting that a design can be made to answer a question it is not structured to answer. Where the quantity required is incidence and the available data are cross-sectional, the honest course is to change the question or collect follow-up.
References and sources
- Thiese MS. Observational and interventional study design types; an overview. Biochemia Medica. 2014;24(2):199–210. read 21 October 2026 from the publisher-hosted copy at HRČAK.View the full text
- von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP, for the STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Medicine. 2007;4(10):e296. doi:10.1371/journal.pmed.0040296 read 21 October 2026. The checklist itself is served as an image on this page and was read from the EQUATOR copy listed below.View the statement
- Vandenbroucke JP, von Elm E, Altman DG, Gøtzsche PC, Mulrow CD, Pocock SJ, et al. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration. PLoS Medicine. 2007;4(10):e297. doi:10.1371/journal.pmed.0040297 read 21 October 2026. The page is long enough that the commentary on items 14 to 18 could not be retrieved; nothing from those items is used here.View the explanation and elaboration
- STROBE Statement. Checklist of items that should be included in reports of observational studies. EQUATOR Network. read 21 October 2026.View the checklist
- Kerr S, Greenland S, Jeffrey K, Millington T, Bedston S, Ritchie L, et al. Understanding and reporting odds ratios as rate-ratio estimates in case-control studies. Journal of Global Health. 2023;13:04101. read 21 October 2026 from the University of Strathclyde repository.View the full text
- Centers for Disease Control and Prevention. Principles of Epidemiology in Public Health Practice, 3rd edition. Lesson 3, Section 2: Morbidity frequency measures. read 21 October 2026 from the CDC archive.View the lesson
- Wang X, Cheng Z. Cross-sectional studies: strengths, weaknesses, and recommendations. CHEST. 2020;158(1 Suppl):S65–S71. read 21 October 2026 from a university-hosted copy of the published article; the publisher’s own site was not reachable.View the full text
Further reading. Parker RA, Berman NG. Planning Clinical Research. Cambridge: Cambridge University Press; 2016. doi:10.1017/CBO9781139024716. And Barría RM, editor. Cohort Studies in Health Sciences. London: IntechOpen; 2018 — an edited volume, so individual chapters carry their own authorship. Both bibliographic records were checked on 21 October 2026. Neither was consulted at page level, so no claim on this page rests on either, and whole-book citations without page numbers cannot be verified in principle.
How to cite this article
Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.
This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
Work through the question first, then the search.
THE QUESTION [ ] Problem written as a plain sentence before any structure applied [ ] Population narrowed only where the narrowing is clinically meaningful [ ] Intervention described as it would actually be delivered [ ] Comparator named explicitly, including what "usual care" means [ ] Outcome measure and time point both specified [ ] Question checked against an intervention frame before committing to PICO [ ] Variant chosen correctly: time is PICOT, study design is PICOS THE SEARCH [ ] Three concepts chosen: population, intervention, study design [ ] Comparator left out, unless it is placebo reported consistently [ ] Outcome left out, unless defined and reported consistently in abstracts [ ] Synonyms gathered for each concept and combined with OR [ ] Concepts combined with AND, and no further concepts added [ ] Controlled vocabulary and free-text terms both included [ ] MeSH terms exploded where appropriate [ ] Vocabulary half rebuilt for each database rather than reused [ ] Result count treated as a screening problem, not a reason to add terms [ ] Strategy revised iteratively and the final version recorded THE WRITE-UP [ ] PICO reported as the structure used to define the question [ ] No claim made that the framework improved retrieval
We help you frame an answerable question, build and document a search strategy that will stand up in a methods section, and choose the right framework when PICO does not fit.