You can make a study sound sophisticated by calling it prospective, retrospective, or ambispective. The label does not make the design defensible.
The real decision is whether your question can be answered with data that exist, data you must collect forward, or a deliberate combination of both. You also need to define who enters the cohort, when follow-up begins, which variables are reliable, how bias will be addressed, and what ethics and privacy pathway applies.
Choose the least burdensome design that can still produce interpretable evidence. Do not choose retrospective work merely because it appears fast, or prospective work merely because it appears rigorous.
First, separate time direction from intervention assignment
Retrospective and prospective describe when the relevant data arise relative to the study protocol. They do not, by themselves, tell you whether a study is interventional.
FDA’s real-world evidence framework defines a clinical trial as a study in which participants are prospectively assigned to an intervention. It defines an observational study as non-interventional. Within observational studies, FDA describes a retrospective design as identifying the population and exposure from historical data, and a prospective design as identifying the population at study start and collecting exposure and outcome data forward (FDA RWE Framework).
That distinction matters. A protocol can use historical records and still include a prospectively assigned intervention. Conversely, a prospective registry can collect data forward without assigning treatment. Do not let the time label replace an explicit statement about intervention assignment.
FDA also warns that prospective and retrospective are used variably across non-interventional research. Its March 2024 document is draft, nonbinding guidance for drug and biologic regulatory use—not a universal methods standard—but the warning is useful: define what the label means in your protocol rather than assuming everyone uses it the same way (FDA draft guidance).
Use three working definitions
For practical IIT planning, use these definitions and then describe the design in full.
| Design | Working definition | Best fit |
|---|---|---|
| Retrospective | The cohort, exposures, and outcomes are derived primarily from data generated before the study protocol begins | The needed variables and follow-up already exist and are sufficiently complete, consistent, and accessible |
| Prospective | You define the population and collect exposure, assessment, and outcome data forward from a prespecified time zero | The question requires measurements, timing, follow-up, or participant-reported outcomes not captured reliably in routine records |
| Ambispective | Under one prespecified protocol, you use a defined historical period and continue planned follow-up forward | Existing data provide a useful baseline or early outcome, but the question requires new follow-up or standardized prospective measurements |
Ambispective is a practical working label here, not a universal regulatory definition. In the protocol, describe the historical and prospective components separately, including time zero, eligibility, variables, follow-up, and analysis.
Ask seven questions before selecting the design
1. What is the exact question?
Start with the population, exposure or intervention, comparator, outcome, and time horizon. Then use the question-to-design chain in Sengi’s IIT study-design guide.
If the question depends on a symptom score that was never collected, a retrospective label cannot create it. If the outcome takes three years to occur, a fully prospective study may be scientifically clean but operationally unrealistic. The question sets the data requirement; the available data do not get to redefine the question silently.
2. Is treatment assigned by the protocol?
State yes or no. If your protocol assigns an intervention, you are making an interventional design decision even if historical data contribute to eligibility, baseline characterization, or an external comparison.
ICH E6(R3) applies to interventional trials of investigational products intended for regulatory submission, with possible application to other interventional trials according to local requirements. It emphasizes scientifically sound, operationally feasible protocols and fit-for-purpose data and processes (ICH E6(R3)). Do not apply that scope indiscriminately to every chart review.
3. What is time zero?
Time zero is the point at which eligibility, exposure classification, and follow-up align. Define it before looking at outcomes.
In retrospective work, an unclear time zero can create unequal opportunities for an outcome to be observed. In prospective work, it can create inconsistent visit windows or baseline measurements. In an ambispective design, you may need a historical index date and a distinct transition into prospective follow-up. Put both on a timeline.
4. Are the existing data fit for this question?
Do not ask whether your electronic health record contains many fields. Ask whether it contains the required exposure, outcome, covariates, dates, and follow-up with acceptable completeness and consistency.
FDA’s draft guidance identifies reliability—including accuracy, completeness, and traceability—and relevance—including key variables and a representative population—as central data-fitness considerations. It also identifies confounding, selection, follow-up errors, differential outcome measurement, and missing data as threats to non-interventional inference (FDA draft guidance).
Run a feasibility extract before committing. Check missingness by period and clinic, coding changes, duplicate records, outcome ascertainment, loss to follow-up, and whether clinically important confounders were recorded before the outcome. A large data set can still be unfit.
5. Which bias is the design buying?
Every design exchanges one problem for another.
- Retrospective: faster access to outcomes, but limited control over what was measured, when it was measured, and why patients received different care.
- Prospective: standardized measurements and planned follow-up, but recruitment, attrition, behavior change, and operational burden can distort the sample or execution.
- Ambispective: a shorter path to additional follow-up, but historical and prospective measurements may not be comparable and only some historical patients may be reachable.
Do not write “bias will be minimized” as a generic promise. Name the likely bias, its direction if known, the design control, the analysis control, and the residual limitation.
6. Can your team execute the design?
Prospective data are not automatically better if visits are missed, instruments are inconsistently administered, or follow-up stops when the coordinator becomes overloaded.
Map recruitment, consent where applicable, assessments, follow-up, data entry, monitoring, analysis, and closeout to named people and realistic time. Use the sample-size guide to connect precision with feasibility rather than treating the target number as detached from the available population.
An ambispective design requires both capabilities: reliable historical extraction and disciplined prospective follow-up. If your team can do only one well, the mixed design adds failure points rather than flexibility.
7. What ethics and privacy pathway applies?
Retrospective does not mean ethics-free, consent-free, or privacy-free.
In the Canadian TCPS 2 framework, research involving humans generally requires REB review before it begins. Epidemiological chart-review research generally involves private health information. For secondary use of identifiable information without consent, Article 5.5A sets conditions that the REB must be satisfied are met; secondary use of non-identifiable information still requires REB review under Article 5.5B. Research relying exclusively on anonymous information may be exempt under Article 2.4 when linkage, recording, or dissemination does not generate identifiable information (TCPS 2 (2022)).
Those are Canadian policy examples, not universal rules. Your institution, jurisdiction, intervention, funding, and data custodian may impose different or additional requirements. Ask the responsible REB or ethics office; do not decide exemption yourself.
Prospective participant contact commonly adds recruitment, consent, burden, safety, and withdrawal considerations. Use Sengi’s informed-consent guide as a design aid, then apply the governing requirements for your study.
When each design is usually the better choice
Choose retrospective when the outcome has already occurred, the required variables were captured consistently, the cohort and time zero can be reconstructed, and the remaining bias is acceptable for the intended claim.
Choose prospective when you need standardized assessments, new biospecimens, participant-reported outcomes, controlled visit timing, active safety follow-up, or variables that routine records do not contain reliably.
Choose ambispective when the historical component can answer a defined part of the question and prospective follow-up adds a necessary outcome—not when you are trying to rescue incomplete records after analysis starts.
The strongest objection to this three-part framework is correct: the labels can oversimplify. The answer is not to abandon them, but to subordinate them to the protocol. Report the cohort construction, setting, relevant dates, variables, data sources, missing-data approach, bias controls, limitations, and generalizability. The STROBE checklist provides a useful reporting structure for cohort, case-control, and cross-sectional observational studies; it does not repair a weak design.
Lock the decision before analysis
Write the design rationale into your protocol before the final extract or prospective enrollment begins. Include:
- the question and intended inference;
- intervention assignment, if any;
- cohort-entry rule and time zero;
- historical and prospective periods;
- data provenance and fitness checks;
- confounders and bias controls;
- missing-data and sensitivity analyses;
- ethics, privacy, consent, and data-access pathway;
- feasibility limits and stop conditions.
For retrospective analyses, prespecification protects you from repeatedly changing eligibility, exposure windows, or outcomes until a favorable result appears. FDA’s RWE framework specifically identifies up-front transparency and reproducibility as concerns in retrospective observational research (FDA RWE Framework).
Your design is not the label at the top of the protocol. It is the set of decisions that determines who is studied, what is measured, when it is measured, what comparison is credible, and which conclusion the data can support.
Take the next step
Put your question, data timeline, time zero, and intervention assignment on one page. Then test whether the least burdensome design can still provide the variables, follow-up, bias controls, and participant protections the question requires.
