Retrospective Cohort Study: Example and Bias

Q: What is the meaning of retrospective cohort study?

It is a cohort comparison built from data that were already recorded when the study began, so the investigators look back at exposure and outcomes rather than waiting for them.

Q: How is it different from a chart review?

A cohort study defines exposed and unexposed groups and compares outcomes over follow-up; a chart review may simply extract and describe information from records.

Q: Is a retrospective cohort study a good design?

It is efficient and suited to rare exposures and outcomes with long latency, but its quality is limited by the records available. Use a bias tool and report to STROBE.

A retrospective cohort study follows a group of people from a past starting point to a past outcome, using records that already exist. This guide defines the design, shows a worked example with the arithmetic, explains why the reporting guideline STROBE avoids the word “retrospective”, and sets out the biases and the level-of-evidence question that matter when you read or write one.

What is a retrospective cohort study? (definition and meaning)

A retrospective cohort study identifies a defined group, classifies members by an exposure recorded at the time, and compares how often an outcome later occurred in exposed and unexposed members, with both exposure and outcome already documented when the investigators start. The investigators do not wait for events; they reconstruct follow-up from sources such as medical records, registries, administrative claims or employment files. It is a cohort study in every analytic respect, and differs from a prospective one only in when the data were collected relative to when the study was conceived.

Cohort studies assess multiple exposures and multiple outcomes, and suit rare exposures or questions where randomization is impractical or unethical. Euser and colleagues (2009) describe the main trade-off: prospective cohorts offer higher accuracy of measurement, retrospective cohorts offer higher efficiency.

Define retrospective cohort study in one line

Exposure groups are assembled from existing records, and outcomes are counted over follow-up time that has already happened.

Retrospective cohort study example

Illustrative composite, with invented numbers for teaching. It is not a real study. A hospital analyst wants to know whether patients discharged on a particular medication class were readmitted more often within one year. Using the electronic record, she lists 1,000 patients discharged in a past calendar year: 400 were prescribed the medication (exposed) and 600 were not (unexposed). Everyone’s readmission status is already in the record. Another example of a retrospective cohort study in the same style: a registry of workers classified by past job title, followed through archived health files to see who later developed a disease.

	Outcome: yes	Outcome: no	Total
Exposed	a = 60	b = 340	400
Unexposed	c = 36	d = 564	600

The arithmetic

Risk in exposed = a / (a + b) = 60 / 400 = 0.15 (15%).
Risk in unexposed = c / (c + d) = 36 / 600 = 0.06 (6%).
Risk difference = 0.15 – 0.06 = 0.09.
Relative risk (RR) = 0.15 / 0.06 = 2.5.
Standard error of ln(RR) = √(1/60 – 1/400 + 1/36 – 1/600) = √0.04028 = 0.2007.
95% CI = exp(ln 2.5 ± 1.96 × 0.2007) = exp(0.9163 ± 0.3934) = 1.69 to 3.70.

Read this as: in these invented records, exposed patients had 2.5 times the one-year risk of the outcome, with a 95% confidence interval from 1.69 to 3.70. Because a cohort starts from exposure, it yields risks and a relative risk. A case-control study, which starts from outcome, estimates an odds ratio; on the same table the odds ratio would be (60/340) / (36/564) = 2.76, close to but not equal to the RR because the outcome is not rare. When follow-up lengths differ, analysts use incidence rates per person-time instead of simple proportions.

Reading guidance

Result	Meaning
RR > 1, interval excludes 1	Higher risk in the exposed group; consistent with association, not proof of cause.
RR = 1, or interval includes 1	No clear difference in risk at this sample size.
RR < 1, interval excludes 1	Lower risk in the exposed group.

How to tell if a study is retrospective or prospective

Ask when the data were collected relative to when the research question was formed. If exposure and outcome were already recorded before the investigators posed the question, the cohort is retrospective; if participants are enrolled and followed forward to outcomes that have not happened yet, it is prospective.

Why STROBE avoids the words

The STROBE Explanation and Elaboration paper (Vandenbroucke et al., 2007) recommends that authors “refrain from simply calling a study ‘prospective’ or ‘retrospective’ because these terms are ill defined.” It lists three usages in the literature: (a) all cohort studies are prospective and “retrospective” is reserved for case-control studies; (b) cohort studies are prospective or retrospective according to the timing of data collection relative to when the idea for the study was developed; and (c) case-control studies are prospective or retrospective by whether exposure data already existed when cases were selected. The practical rule is to state how and when the data were collected instead of relying on the label. See also the dictionary entry on the retrospective study and the comparison prospective vs retrospective study.

Retrospective cohort vs case-control: how to remember the difference

A cohort study starts from exposure and looks at outcomes; a case-control study starts from outcome and looks back at exposure. A cohort gives risks and RR; a case-control gives odds and OR. The three-way split between cohort, case-control and cross-sectional designs is the one STROBE uses for its checklists. The detailed comparison is in case-control study vs cohort study, and the case-control design is covered in the case-control study guide.

Can you have a prospective case-control study?

Yes, under usage (c) above: a case-control study can be called prospective when exposure data are collected only after cases are selected, and retrospective when the exposure data already existed. Because of this ambiguity, describe the data collection explicitly.

Assumptions and common mistakes

Misclassified or missing exposure. Records were made for care or administration, not research, so exposure and outcome definitions may be incomplete. Report how missing data were handled.
Selection from records. Who is in the database, and who dropped out of it, can differ from the population you wish to describe.
Confounding by indication. Euser and colleagues flag confounding by indication and selection bias as key concerns in cohort studies: people receive a treatment because of their condition, so the groups differ before exposure. Adjustment and design choices reduce but may not remove it.
Immortal time. If exposure status is assigned using information that could only be known after follow-up began, the exposed group cannot have had the outcome during that interval. Define the start of follow-up for both groups identically. (Editorial interpretation of standard cohort methodology.)
Recall. Record-based exposure avoids most recall bias, but not bias from how records were created.
Confusing it with a chart review. A retrospective chart review may only describe cases; a cohort study compares defined exposed and unexposed groups over follow-up.

What level of evidence is a retrospective cohort study?

There is no single universal level. Evidence hierarchies place observational cohort studies below well-conducted randomized trials for questions about treatment benefit, but where a given study sits depends on the framework used, the question (treatment, prognosis, harm) and the study’s risk of bias, which can move its credibility up or down. This is an editorial interpretation rather than a ranking from one authority. To judge an individual study, appraise it with a tool designed for the design, such as the Newcastle-Ottawa Scale or ROBINS-I.

How to report it (STROBE and an APA-style sentence)

STROBE provides a “Checklist: cohort studies” on the STROBE statement website (strobe-statement.org/checklists), based on the STROBE Statement (von Elm et al., 2007). A reporting sentence in APA style might read: “In this retrospective cohort, risk of one-year readmission was higher among exposed than unexposed patients (15% vs 6%; RR = 2.50, 95% CI [1.69, 3.70]).” Add a sentence naming the data source and the dates when records were created.

How to calculate it in Excel, R or Python

Excel (a = B2, b = C2, c = B3, d = C3): relative risk =(B2/(B2+C2))/(B3/(B3+C3)); lower limit =EXP(LN(RR)-1.96*SQRT(1/B2-1/(B2+C2)+1/B3-1/(B3+C3))), and use + for the upper limit.

a<-60; b<-340; c<-36; d<-564
rr <- (a/(a+b))/(c/(c+d))
se <- sqrt(1/a - 1/(a+b) + 1/c - 1/(c+d))
c(rr, exp(log(rr) + c(-1,1)*1.96*se))

Python:

import math
a,b,c,d = 60,340,36,564
rr = (a/(a+b))/(c/(c+d))
se = math.sqrt(1/a - 1/(a+b) + 1/c - 1/(c+d))
print(rr, math.exp(math.log(rr)-1.96*se), math.exp(math.log(rr)+1.96*se))

Frequently asked questions

What is the meaning of retrospective cohort study?