6.2 STROBE & RECORD for Observational and Routinely Collected Data
Key Takeaways
STROBE is a 22-item reporting checklist for cohort, case-control, and cross-sectional studies; it improves transparent reporting but is not a study-design prescription or quality score.
RECORD extends STROBE for studies using routinely collected health data and calls for transparent reporting of code lists or algorithms, validation where performed, linkage, selection, access, and data-cleaning methods.
Propensity scores, regression, weighting, and target-trial approaches can address measured confounding but do not prove that residual or unmeasured confounding has been eliminated.
Immortal time bias arises when exposure classification gives one group outcome-free time by design; prevention requires alignment of eligibility, treatment assignment, and time zero or a suitable time-varying analysis.
A defensible RWE report distinguishes reporting guidance from regulatory fitness-for-purpose expectations and describes limitations rather than claiming that checklist compliance validates the evidence.
STROBE & RECORD for Observational and Routinely Collected Data
Observational evidence can describe treatment patterns, long-term outcomes, uncommon harms, resource use, and effectiveness in populations that may differ from randomized-trial participants. Its value depends on a clear account of where the data came from, how the cohort and variables were constructed, which biases were anticipated, and what uncertainty remains. A polished effect estimate does not compensate for an opaque design.
STROBE: a reporting framework
The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement contains 22 reporting items for cohort, case-control, and cross-sectional studies. It asks authors to identify the design, explain the setting and participant selection, define variables, describe bias and study size, report statistical methods, account for participants, present descriptive and outcome data, distinguish unadjusted from adjusted estimates, and discuss limitations and generalizability.
STROBE is a reporting guideline. It does not prescribe one correct design, require a particular statistical package, certify study quality, or guarantee causal validity. Authors should use the checklist to make their actual decisions visible. A publication professional can map each item to the draft and supporting evidence, but should not insert methods that the investigators did not perform.
Important reporting practices include:
- state the design in the title or abstract when appropriate;
- give eligibility criteria, sources, selection methods, dates, and follow-up;
- define outcomes, exposures, predictors, confounders, and effect modifiers;
- explain efforts to address bias and missing data;
- report participant numbers through the analysis and consider a flow diagram;
- report both unadjusted and adjusted estimates when relevant, identifying variables and rationale;
- distinguish prespecified, sensitivity, subgroup, and exploratory analyses;
- discuss direction and likely magnitude of important limitations.
RECORD: the extension for routinely collected data
The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement extends STROBE for data such as health records, administrative claims, registries, and linked databases. These sources were often created for care, billing, or administration rather than the publication question, so readers need to see how records became analyzable variables and a study cohort.
RECORD calls for reporting the codes or algorithms used to classify exposures, outcomes, confounders, and effect modifiers. When complete lists cannot be shared, the report should explain why. Authors should also say whether algorithms were developed or validated by the investigators and, if so, report the validation methods and results. This is more accurate than saying every RECORD report must publish every code unabridged or produce new sensitivity, specificity, and predictive-value estimates. Validation information should be reported when validation was performed or an external validation is relied on.
For linked data, describe the databases, linkage level and methods, and assessment of linkage quality. If linkage is performed by a third party and details are restricted, report the available information and the restriction. Describe data access, cleaning, exclusions, and cohort derivation sufficiently for readers to understand selection and potential error. Privacy and governance statements must reflect the actual jurisdiction, data controller, ethics determination, and agreements; RECORD itself does not create a legal basis for data processing.
Confounding is addressed, not erased
Confounding by indication occurs when clinical factors that influence treatment also influence the outcome. Regression adjustment, matching, stratification, propensity-score weighting, instrumental-variable analysis, negative controls, and target-trial emulation may be useful, depending on the estimand and assumptions. These methods can reduce imbalance in measured variables. They do not automatically control unmeasured confounding, measurement error, model misspecification, lack of positivity, or selection into the database.
A transparent report should identify the causal question, time zero, treatment strategies, follow-up, outcome, estimand, covariates, and censoring rules. It should explain why adjustment variables were chosen rather than selecting them solely by statistical significance. Diagnostics such as balance after weighting can show whether measured covariates were balanced; they cannot prove exchangeability.
Immortal time and other time-related biases
Immortal time is a span during which, under the study definition, a person must remain alive or event-free to enter the exposed group. For example, classifying people as treated only after they fill two prescriptions gives future treated patients guaranteed survival until the second fill if earlier time is counted as exposed. Solutions may include aligning eligibility, assignment, and follow-up at one time zero; treating exposure as time varying; using a landmark design with its tradeoffs; or emulating a target trial.
The manuscript should also address prevalent-user bias, informative censoring, time-varying confounding, outcome latency, depletion of susceptibles, and differential follow-up where relevant. The correct response is design-specific; deleting early deaths after seeing the data is not a general cure.
Missingness, misclassification, and sensitivity analysis
Explain the amount and pattern of missing data, the assumptions behind the selected method, and how missingness was handled. Complete-case analysis can be biased and inefficient. Multiple imputation is not automatically valid; its variables and assumptions should be reported. For claims or EHR data, absence of a code may mean absence of disease, lack of capture, different coding, or care outside the network. Discuss the likely direction of exposure and outcome misclassification.
Sensitivity analyses should test plausible threats to the main inference: alternate definitions, exposure grace periods, negative controls, quantitative bias analysis, missing-data assumptions, or unmeasured-confounding scenarios. Label analyses honestly and avoid presenting robustness to several model specifications as proof that all bias is absent.
Regulatory and publication use
When RWE is intended for a regulatory purpose, current agency guidance and engagement govern data relevance, reliability, study design, and documentation. STROBE or RECORD compliance can improve the report but does not by itself establish regulatory fitness for purpose. Similarly, journal acceptance does not convert an association into a causal effect.
Important
On exam scenarios, favor transparent cohort construction, code and linkage reporting, appropriate bias control, and candid limitations. Reject answers that say a statistical method “corrects” all confounding or that a reporting checklist validates the underlying study.
A claims-database study uses diagnosis and procedure algorithms to define exposure and outcome. Which action best reflects the RECORD extension?
State only that standard codes were used because code details are proprietary.
Generate new diagnostic-accuracy estimates for every variable even when no validation study was performed.
Treat use of a validated database as a substitute for describing cohort construction.
Report the codes or algorithms and their availability, describe any validation that was performed or cited, and explain restrictions when details cannot be shared.
High-risk patients are preferentially prescribed a new anticoagulant, and the same risk factors predict stroke. What is the best interpretation of propensity-score adjustment?
It can address imbalance in measured confounders, but residual and unmeasured confounding may remain and should be discussed.
It proves treatment groups are exchangeable once the propensity model converges.
It removes the need to report crude estimates and covariate definitions.
It converts the observational study into a randomized trial.
A treated cohort is defined by survival long enough to receive two cycles, while early deaths are assigned to the untreated cohort. What is the central problem and a defensible response?
Recall bias; ask clinicians to reconstruct treatment choices from memory.
Immortal time bias; align eligibility, treatment assignment, and time zero or use an appropriate time-varying strategy.
Publication bias; report only statistically significant outcomes.
Random error; enlarge the typeface of the Kaplan–Meier plot.
Sections you finish are checked off in the contents.