Prospective and Retrospective Cohorts

Overview

Prospective and retrospective cohorts share the same longitudinal logic: define a cohort, establish exposure and eligibility, and compare subsequent outcomes. The labels describe when data and events occur relative to study initiation, not whether the design is a cohort.[1,2]

In a prospective cohort, participants are enrolled before the outcomes of interest occur and are followed using a planned protocol. In a retrospective or historical cohort, cohort entry, exposures and some or all outcomes have already occurred when investigators assemble the data. An ambidirectional cohort combines historical follow-up with new prospective observation.

The choice affects measurement control, speed, cost, available confounders, missingness, outcome verification and governance. Neither direction is inherently superior; validity depends on whether the available information supports the intended inference.

Why this matters

A database with millions of records can produce narrow confidence intervals but still lack valid exposure timing, baseline eligibility or important confounders. Conversely, a carefully designed prospective cohort may fail if recruitment is selective or follow-up is incomplete.

Common misconceptions include:

• “prospective” means unbiased;
• “retrospective” means case-control;
• routinely collected codes are equivalent to research definitions;
• absence of a code means absence of disease;
• the first observed record is disease onset;
• adding future follow-up corrects historical measurement problems; and
• large sample size compensates for systematic error.

Key concepts

Prospective cohort

Eligibility, exposure measurement and follow-up procedures are designed before most outcomes occur. Researchers can standardise instruments, collect patient-reported factors, measure confounders and monitor retention. Disadvantages include time, cost, behavioural changes from participation and delayed results.

Retrospective cohort

The cohort is reconstructed from records generated for clinical, administrative, occupational or registry purposes. Data may provide long follow-up quickly and avoid participant recall for recorded exposures. Limitations include changing definitions, missing variables, undocumented care outside the system and inability to alter historical procedures.

Ambidirectional cohort

Historical data establish baseline and earlier outcomes, while participants or records are followed prospectively from a defined transition date. This can extend follow-up or add new measurements. The historical and prospective phases may have different ascertainment and must not be treated as one uniform process without assessment.

When to use each implementation

DECISION GUIDE

Choosing cohort direction

A prospective implementation is attractive when:

• critical exposures or confounders are not routinely recorded;
• standardised biological or patient-reported measurements are required;
• outcomes are frequent enough for feasible follow-up;
• active retention and adjudication are important; and
• resources support long-term operations.

A retrospective implementation is attractive when:

• reliable historical exposure and outcome dates exist;
• latency is long or outcomes have already accumulated;
• the source population and data coverage can be reconstructed;
• coding changes and missingness can be characterised; and
• linkage and governance are feasible.

An ambidirectional implementation is attractive when historical data provide an efficient foundation but future outcomes or measurements remain important.

Research question and methodological fit

For the postpartum diabetes question, historical linked data can estimate recorded diagnoses over five years and compare testing-completion groups. The target population is women delivering in participating hospitals whose maternity, laboratory and outpatient data are observable.

If diet, breastfeeding, weight change and family history are essential confounders but absent historically, a retrospective cohort may not support the intended causal interpretation. It may still answer a descriptive service-prognosis question. The question should be narrowed rather than pretending unavailable variables were controlled.

Design, setting and study population

DESIGN PROFILE

Prospective cohort implementation

Core structure: Eligible participants are enrolled and measured under a protocol before outcomes accrue.

Best suited for: New measurements, standardised follow-up, uncommon exposures and detailed confounder assessment.

Important limitations: Time, cost, attrition, recruitment effects and delayed event accrual.

DESIGN PROFILE

Retrospective cohort implementation

Core structure: Cohort entry, exposures and outcomes are reconstructed from existing dated sources.

Best suited for: Long latency, large populations, service evaluation and efficient use of registries or electronic records.

Important limitations: Data were not collected for the research question; coverage, provenance and changing definitions can threaten validity.

The same target-trial or cohort specification should be written for both options: eligibility, exposure strategies or groups, time zero, follow-up, outcome, causal or prognostic contrast and analysis.[2]

Evaluating retrospective data fitness

Population coverage

Determine which people and services enter the database. A hospital electronic record may miss private laboratories, primary care and migration. Coverage can change over time after mergers or interface upgrades.

Data provenance

For every key field, identify who recorded it, why, when, in which system and whether values were overwritten. A diagnosis code may support billing, problem lists or confirmed disease, with different meanings.

Temporal resolution

Dates must be precise enough to order eligibility, exposure and outcome. Month-only dates may be inadequate when classification windows are short. Imported historical diagnoses can appear on the migration date rather than true onset.

Measurement stability

Laboratory methods, units, reference ranges, code systems and clinical guidelines may change. Harmonisation should preserve original values and document transformations.

Missingness and absence

“No record” can mean no event, care elsewhere, incomplete capture or a test not performed. Missingness indicators and coverage periods should be defined explicitly.

PROTOCOL CHECKLIST

Historical cohort data audit

• Map source systems and coverage by site and calendar period.
• Define a minimum look-back for baseline disease and exposure.
• Validate key codes, dates and laboratory units.
• Identify changes in coding, instruments and care pathways.
• Quantify missingness and discontinuities before analysis.
• Define continuous enrolment or observable-data requirements where appropriate.
• Prespecify linkage, deduplication and conflicting-source rules.
• Retain provenance and raw values for audit.

Sampling and recruitment

Prospective cohorts recruit participants, so approach and consent can create selection. Retrospective cohorts often include all eligible records, but database inclusion, linkage success and complete-case restrictions are still selection mechanisms.

A hybrid consent model may use a waiver for historical records and consent for new questionnaires or specimens. Investigators should distinguish the record-based source cohort from the recontacted subset; results requiring new consent may apply to a more selected population.

Variables, constructs and measurements

Comparing available measurements

illustrative

Table 1. Prospective and retrospective data availability

Domain	Prospective cohort	Historical records	Design implication
Testing completion	Standardised tracking	Laboratory interfaces vary	Harmonise source coverage
Body mass index	Measured at planned visits	Often missing after delivery	Residual confounding likely
Diabetes outcome	Scheduled testing and adjudication	Diagnosis and laboratory codes	Detection may depend on care use
Healthcare contact	Captured by study and linkage	Only within participating system	Define observable coverage
Participant barriers	Repeated questionnaire	Not routinely available	Historical causal question may be narrower

Note: The table is illustrative.

Procedures and data collection

Prospective procedures

Use standard operating procedures, calibrated measurements, staff certification, visit windows, retention plans, event adjudication and real-time data-quality monitoring.

Retrospective procedures

Freeze source extracts, retain query code, version data dictionaries, validate linkage, perform duplicate review, create unit and code crosswalks, and compare extracted counts with source-system reports. A clinician or domain specialist should review implausible temporal patterns.

Validation studies

A validation subsample can compare coded exposure or outcomes against source records or an independent reference. Report positive predictive value, sensitivity or agreement where appropriate, but recognise that a selected validation sample may not represent all sites and periods.

Bias, validity and quality assurance

Prospective designs can have Hawthorne effects, differential retention and protocol drift. Retrospective designs can have left truncation, incomplete look-back, misclassified onset, missing confounders and changes in coding. Both can suffer confounding and differential surveillance.

APPRAISAL CHECKLIST

Appraising cohort direction and data fitness

• Does the chosen direction fit the latency and available event numbers?
• Can the source population and observable period be defined?
• Are time zero and exposure known without future information?
• Are key variables measured for the research purpose or validated adequately?
• Have site and calendar changes been mapped?
• Is missingness distinguished from absence?
• Are historical and prospective phases analysed with their measurement differences in mind?

Ethical and governance considerations

Interface with statistical analysis

Both implementations can use risk, rate or survival methods when their data support the same estimand. Retrospective cohorts may require calendar-period adjustment, site effects, delayed entry and sensitivity analyses for look-back and outcome definitions. Prospective cohorts may need repeated-measures or time-varying exposure models.

Data size should not drive an excessively complex model before measurement validity is established. Report the number of eligible records at each data-processing step, person-time coverage and the proportion with valid exposure, outcome and confounder data.

Emulating a target trial with routine dataMMed extension

A target-trial protocol clarifies eligibility, treatment or exposure strategies, assignment procedures, time zero, follow-up, outcome, causal contrast and analysis before records are manipulated. It reduces design choices that use future information, although it cannot create unmeasured confounders or complete data coverage.[2,3]

Choosing the most defensible design rather than the fastest design

Investigators can compare options in a design matrix covering population coverage, exposure validity, confounder availability, outcome completeness, latency, cost and decision timeline. A retrospective cohort may be the best primary design when dates and coverage are strong; a small prospective validation cohort can then quantify measurement limitations. Alternatively, a prospective cohort may be necessary when the central construct does not exist in routine data. The final choice should be justified against the research question, not merely by the largest obtainable sample.

The protocol should also state the date on which historical data are considered complete. Later data refreshes can change diagnoses, linkage and follow-up. Versioned extracts, frozen analysis datasets and a documented refresh policy are necessary so that results can be reproduced and differences between analyses can be explained.

Worked example

WORKED EXAMPLE

Choosing an ambidirectional postpartum cohort

The team audits records from 2016 onward. Laboratory coverage is reliable from 2018, but outpatient diagnoses before 2019 are incomplete at two sites. They therefore define cohort entry for deliveries from 2019–2021, require documented system coverage during the six-month exposure window and preserve site-specific data provenance.

Historical follow-up continues through 2026. A subset consenting to recontact completes new questionnaires and measurements from 2027 onward. The investigators treat this as an ambidirectional design: record-based findings apply to the full eligible historical cohort, whereas patient-reported analyses apply to the consenting subset.

A validation sample compares algorithm-defined diabetes with clinician-reviewed records. Sensitivity analyses vary the look-back and outcome definition. The study does not claim that absent historical breastfeeding data were controlled.

Suggested protocol or reporting wording

Common mistakes

Knowledge check

What makes a historical-record study a cohort rather than a case-control study?

Answer and reasoning

Participants are defined as an at-risk cohort based on eligibility and exposure, and outcomes are then ascertained over time. Sampling is not performed according to outcome status.

Why is a larger retrospective cohort not automatically more valid than a smaller prospective cohort?

Answer and reasoning

Systematic misclassification, incomplete coverage and unmeasured confounding do not disappear with sample size. Larger data can make a biased estimate more precise.

What is an ambidirectional cohort?

Answer and reasoning

It combines a historical cohort phase with prospective follow-up or new measurement from a defined transition point. Measurement differences between phases must be documented.

Why is a minimum look-back period used?

Answer and reasoning

It provides a defined opportunity to identify pre-existing disease, prior exposure or confounders. It does not guarantee complete history and may select people with stable data coverage.

Key take-home messages

SUMMARY

Key take-home messages

• Prospective and retrospective cohorts share the same longitudinal design logic.
• Direction describes timing relative to study initiation, not inherent quality.
• Historical data require explicit assessment of coverage, provenance, timing and measurement stability.
• Prospective data permit standardisation but remain vulnerable to recruitment and attrition.
• Ambidirectional cohorts should preserve distinctions between historical and new measurements.
• The research question should be narrowed when essential variables are unavailable rather than overstating adjustment.

STROBE reporting should describe data sources, measurement comparability, follow-up and missing-data handling for either direction.[4,5]

References

1. Setia MS. Methodology series module 1: cohort studies. Indian J Dermatol. 2016;61(1):21–25. https://doi.org/10.4103/0019-5154.174011
2. Hernán MA, Robins JM. Using big data to emulate a target trial when a randomised trial is not available. Am J Epidemiol. 2016;183(8):758–764. https://doi.org/10.1093/aje/kwv254
3. Hernán MA, Sauer BC, Hernández-Díaz S, Platt R, Shrier I. Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses. J Clin Epidemiol. 2016;79:70–75. https://doi.org/10.1016/j.jclinepi.2016.04.014
4. von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology statement. PLoS Med. 2007;4(10):e296. https://doi.org/10.1371/journal.pmed.0040296
5. Vandenbroucke JP, von Elm E, Altman DG, Gøtzsche PC, Mulrow CD, Pocock SJ, et al. STROBE: explanation and elaboration. PLoS Med. 2007;4(10):e297. https://doi.org/10.1371/journal.pmed.0040297