Retrospective Cohort Study for a
Hospital Discharge Program

RESEARCH QUESTION

What the Team Set Out to Learn

This retrospective cohort study example follows a hospital medicine team that wants to know whether booking a follow-up visit before discharge is associated with unplanned readmission within 30 days. The data already exist in the electronic health record, and the team needs an analysis a reviewer can reconstruct from the extraction rules to the final estimate.

The data were collected for care, not for research, and patients were not assigned to a follow-up visit at random. The question is therefore framed as an association, and the plan says in advance how far an observational comparison of this kind can go, because patients booked for follow-up may differ from others in ways the record does not capture.

STUDY DESIGN AND DATA

Routine Records With Rules Set Before the Outcome

A retrospective cohort of about 6,000 patients discharged from medical wards, drawn from routinely collected record data. The extract holds demographics, diagnoses, comorbidity, prior admissions, length of stay, discharge destination, selected laboratory values, whether a follow-up visit was booked at discharge, and readmission within 30 days. Many patients have several discharges, so the unit of analysis, the time zero and the exposure definition are written down before any outcome is examined.

METHOD AND ASSUMPTIONS

Why Weighted Odds Ratios, and What Must Hold

The outcome is binary, so the model is logistic, and the comparison is adjusted for baseline differences through propensity-score weights built from covariates measured before discharge. Exposure is defined as a visit scheduled at discharge, not as a visit attended, because attending requires surviving and staying out of hospital, which would add immortal time and flatter the comparison. Balance is judged from standardized differences after weighting and from the overlap of the score distributions, not from a significance test.

Weighting adjusts only for measured covariates, so the plan reports an E-value, which states how strong an unmeasured confounder would have to be to explain the association. Some laboratory values are missing through clinical practice, not by chance, so multiple imputation under a stated missing-at-random assumption is compared with a complete-case analysis. Death before readmission removes patients from risk, so a competing-event analysis is added as a sensitivity check, and the covariate list is limited by the number of readmissions.

OUTPUT AND INTERPRETATION

How the Results Are Read

The output has a patient flow table showing how the extraction rules shaped the cohort, a covariate balance plot before and after weighting, and the adjusted odds ratio with its 95% confidence interval beside the unadjusted estimate. A weighted risk difference is reported with it, because an odds ratio for a common outcome can overstate how large the difference in risk is.

Sensitivity analyses show how far the conclusion depends on the choices made: trimmed weights, the imputation compared with complete cases, and the competing-event analysis. The write-up uses association wording throughout and states what a record-based comparison can and cannot support about a discharge program.

REPORTING

What the Researcher Received

A written report structured to the STROBE checklist and its RECORD extension for routinely collected health data, with statistical wording for the methods, matched to the output, for the team to check and incorporate, in the target journal's format. The research question, substantive interpretation and conclusions remain with the research team.