Retrospective Study: Design, Bias, and STROBE Reporting

A retrospective study is an observational study that uses data that already exist when the research question is asked: medical records, registries, insurance claims, school records, or databases collected for another purpose. The outcomes have already happened, and the researcher looks back to see how they relate to earlier exposures or characteristics.

Retrospective studies are fast, cheap, and often the only practical way to study rare outcomes or long latency periods. Their weakness is that you are stuck with data nobody designed for your question, which makes bias the central problem to manage and to report.

Retrospective vs prospective

The difference is timing relative to the start of the study:

The labels are not perfectly standardized. A cohort assembled from 2015–2019 records and followed forward in the same records to 2024 is still "retrospective" in common usage because all events occurred before the study began. The STROBE explanation paper advises against relying on the labels alone and asks authors to state exactly when and how the data were collected. Do that in your Methods and the reader will not have to guess.

Two main retrospective designs

Retrospective cohort study

You define a cohort from past records, classify people by exposure at a defined baseline, and then use the records to see who later developed the outcome.

• Example: all adults who started drug A or drug B at one hospital between 2019 and 2022, followed in the records for 30-day readmission.
• Measures: risk, risk difference, relative risk, or, with follow-up time and censoring, a hazard ratio from Cox regression.
• Can study several outcomes from one exposure.

Case-control study

You start from the outcome: select people who have it (cases) and comparable people who do not (controls), then look back at their past exposures.

• Example: patients with a rare drug reaction and matched patients without it, compared on prior medication use.
• Measure: the odds ratio, because you fixed the number of cases and controls and cannot estimate risk directly.
• Efficient for rare outcomes; can study several exposures for one outcome.

Our case-control study guide walks through matching and a full odds ratio calculation.

Case-control studies are usually retrospective with respect to exposure, but not every retrospective study is a case-control study. A frequent reviewer complaint is a paper that calls a retrospective cohort a "case-control study" because it compares two groups. The test: did you sample by outcome (case-control) or by exposure / everyone in a source population (cohort)?

Other retrospective designs include cross-sectional analyses of existing data and chart-review case series without a comparison group.

Strengths

• Speed and cost: data already exist, so a study can be done in months.
• Rare outcomes and long latency: you do not have to wait decades for events to accumulate.
• Large samples: registries and claims databases can include hundreds of thousands of people.
• Real-world populations: routine care data include patients usually excluded from trials (older, multimorbid, pregnant).
• Hypothesis generation: a good retrospective study can justify a prospective study or trial.

Limitations and the biases to address

Selection bias

Who ended up in the records, and who you included, may relate to both exposure and outcome. Examples: a tertiary referral hospital sees sicker patients; excluding patients with incomplete charts may exclude those who transferred out because they deteriorated. Use explicit eligibility criteria and a flow diagram showing how many records were screened, excluded (with reasons), and analyzed.

Information (measurement) bias

Variables were recorded for care, not research. Diagnoses can be coded inconsistently, exposures can be undocumented, and coding practices change over time. In case-control studies, recall bias is a risk when exposure comes from interviews: cases tend to search their memories harder than controls. State how each variable was defined and, if possible, validate key variables against a sample of full records.

Confounding

Without randomization, exposed and unexposed groups differ in ways that affect the outcome. A special case in clinical data is confounding by indication: sicker patients are more likely to receive a treatment, which makes the treatment look harmful. Address confounding with multivariable regression, stratification, matching, or propensity score methods, and draw a causal diagram (DAG) to decide which variables to adjust for.

Missing data

Missingness is rarely random in routine records. Report the amount missing for each variable, and prefer multiple imputation over complete-case analysis when missingness is substantial and plausibly related to observed variables.

Time-related biases

• Immortal time bias: classifying people as "exposed" based on something that happened after baseline. Define exposure at time zero, or model it as time-varying.
• Survivorship bias: including only people who survived long enough to be recorded.

Causal inference

A retrospective study can show an association. Claims about causation need careful design, adjustment, sensitivity analyses, and cautious language ("was associated with," not "caused").

A worked example: retrospective cohort

A team reviews records of 1,240 adults discharged after heart failure admission between 2019 and 2022. 480 were discharged on drug A and 760 on drug B. The outcome is 30-day readmission.

• Risk with drug A: 62 / 480 = 12.9%
• Risk with drug B: 76 / 760 = 10.0%
• Risk difference: 2.9 percentage points
• Crude relative risk: 0.129 / 0.100 = 1.29, 95% CI 0.94 to 1.77
• Chi-square test of independence: χ²(1, N = 1,240) = 2.53, p = .112 (you can reproduce it with the chi-square calculator)

Interpretation: the crude data are compatible with anything from a 6% lower to a 77% higher readmission risk on drug A. Before reading anything into the point estimate, the team must ask why patients got drug A. If clinicians prescribed it to sicker patients, the crude RR is confounded by indication. The next step is an adjusted model (for example, logistic or log-binomial regression with age, ejection fraction, kidney function, and prior admissions) and a sensitivity analysis.

How to report a retrospective study: STROBE

STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) is the standard reporting guideline for cohort, case-control, and cross-sectional studies. It is a 22-item checklist, with versions tailored to each design. Many journals require a completed checklist at submission. For studies using routinely collected health data (electronic records, claims, registries), the RECORD extension adds items on code lists, data linkage, and data cleaning.

Key STROBE items for a retrospective study:

1. Title and abstract: name the design ("a retrospective cohort study").
2. Setting: locations, dates of recruitment, exposure, follow-up, and data collection.
3. Participants: eligibility criteria and how participants were selected; for case-control studies, how cases and controls were chosen and matched.
4. Variables: define all outcomes, exposures, confounders, and effect modifiers, including codes used.
5. Data sources: where each variable came from and how it was measured.
6. Bias: describe efforts to address potential sources of bias.
7. Study size: explain how it was reached.
8. Statistical methods: including confounding control, missing data, and sensitivity analyses.
9. Participants flow: numbers at each stage, ideally a flow diagram.
10. Main results: unadjusted and adjusted estimates with 95% CIs, and which confounders were adjusted for.
11. Limitations: sources of bias or imprecision, with direction and likely magnitude.

Reporting sentences you can adapt

Methods:

Only write the ethics sentence if it matches what your IRB actually decided.

Results (APA 7):

Limitations:

For a fuller picture of variable coding, tables, and survival outputs in clinical datasets, see the clinical data analysis guide.

Common mistakes

1. Calling a retrospective cohort a case-control study (or the reverse).
2. No flow diagram, so reviewers cannot see how the sample was built.
3. Reporting only adjusted estimates, or only crude ones. Report both.
4. Choosing covariates by p value screening rather than subject knowledge.
5. Silent complete-case analysis with large amounts of missing data.
6. Causal language in the abstract and conclusion.
7. No sample size justification. Even with fixed records, state how many events you had and whether that supports the planned model.

FAQ

Is a retrospective study qualitative or quantitative?

Usually quantitative, since it analyzes existing records statistically. Retrospective qualitative work (for example, analysis of archived documents) exists but is a different tradition.

Where does a retrospective study sit in the evidence hierarchy?

Below randomized trials and generally below well-conducted prospective cohort studies, mainly because of measurement and confounding problems. A carefully designed and reported retrospective study can still be strong evidence, especially when trials are impossible.

Do retrospective studies need ethics approval?

In most countries, yes, or at least a formal exemption decision. Ethics committees often waive individual consent for de-identified record review, but that is their decision, not the researcher's.

Can a retrospective study show causation?

It can support causal conclusions only with strong design and assumptions: clear temporal order, adjustment for confounders guided by a causal model, and sensitivity analyses. Most papers should report associations.

Is a retrospective study the same as a chart review?

A chart review is a data collection method. It can feed a retrospective cohort, a case-control study, or a descriptive case series.

From records to a draft paper

Once your extracted dataset is clean and de-identified, Data2Paper's Generate Paper can produce baseline tables, crude and adjusted estimates, and a draft Methods and Results structured around the analysis. You can generate and preview the first half free; the full draft is a one-time $19.90. It is an AI-generated draft, so check every number, add your STROBE checklist, and verify citations yourself.