Testing Pharmacist-Led Discharge Medication Reconciliation Against 30-Day Heart Failure Readmission: A Quasi-Experimental Plan
Student Name
American College of Education
RES5303: Research Methods and Applied Statistics in Healthcare
Module 4 Assignment
Instructor Name
September 8, 2025
The Practice Problem and the Research Question
Bayview Regional Medical Center is a 268-bed nonprofit hospital inside a four-hospital system serving a metropolitan area of roughly 700,000 people. The hospital and the patients here are composites written for teaching; no real facility, employee, or patient appears. In the 12 months ending June 30, 2024, Bayview recorded 1,043 index discharges with a principal diagnosis of heart failure and 234 all-cause inpatient readmissions within 30 days, a rate of 22.4%. The other three hospitals in the system averaged 19.1% over the same period. A pharmacy review of 60 randomly pulled heart failure charts found at least one unreconciled medication difference at discharge in 41 charts, 22 of them involving a diuretic or a beta blocker dose.
The question is written as one sentence: among adults discharged alive with a principal diagnosis of heart failure from Bayview Regional, does pharmacist-led discharge medication reconciliation paired with a 72-hour telephone review, compared with usual nurse-delivered discharge teaching, change the proportion of index discharges followed by an all-cause inpatient readmission within 30 days? The denominator is index discharges rather than patients, counted once each, because a patient with three admissions in a year would otherwise land three times in the numerator and once in the denominator. Observation stays are recorded separately rather than folded into the outcome, since converting short readmissions into observation status lowers the headline rate without changing what happens to patients.
It is worth naming what the question does not ask. This is a comparative question about a service, not a mechanism question about why patients stop taking medication; adherence is measured as a possible mediator but is not the endpoint. The plan also does not ask whether the pharmacist position pays for itself, which would need cost data, a defined payer perspective, and a horizon longer than 30 days. Payment penalties tied to excess readmission supply the reason the hospital is interested, but a study built to detect a difference in a clinical outcome cannot be reported at the end as a return on investment.
Design and Sampling Plan
The design is quasi-experimental: a nonequivalent comparison group with a pretest and posttest period, analyzed as a difference in differences. Bayview Regional receives the pharmacist-led service; Ridgeway Memorial, a 240-bed sister hospital with similar heart failure volume, continues usual discharge teaching. A randomized trial was considered first and rejected for a stated reason rather than for convenience. One pharmacist embedded with the cardiology service delivers it, and those patients share discharge educators, a standing order set, and a floor, so randomizing patients inside one hospital would leak the intervention into the control arm and push the estimate toward no difference.
Enrollment is consecutive rather than selected. Every discharge meeting the criteria enters: age 18 or older, principal diagnosis of heart failure, discharged alive to home or to home with home health. Transfers to another acute hospital, discharges to hospice, discharges against medical advice, and stays longer than 30 days are excluded, each exclusion counted so the flow diagram reconciles. A two-proportion power calculation set at a two-sided alpha of .05 and 80% power, testing a fall from 22% to 15%, calls for 482 index discharges per arm. Each hospital produces roughly 520 qualifying discharges in 12 months, so one enrollment year clears that figure with little room to spare, and the plan says in advance that a true reduction of 3 percentage points would very likely be missed.
Covariates come from the electronic record and the system data warehouse through a report written before enrollment opens: age, sex, insurance class, ejection fraction where recorded, Charlson comorbidity index, length of stay, inpatient admissions in the prior 180 days, and a neighborhood-level social risk score. Fidelity is measured, not assumed. Two indicators are tracked monthly against a target of 85%: the share of enrolled discharges carrying a documented pharmacist reconciliation note and the share with a completed 72-hour call. Because the primary analysis compares everyone discharged at one hospital with everyone discharged at the other, weak fidelity would shrink the observed effect rather than break the comparison, and the fidelity numbers are what make a small effect interpretable.
The Analysis and the Assumptions Behind It
The unadjusted comparison comes first, because a reader deserves raw numbers before a model touches them. Readmission by hospital in the post period is tested with a Pearson chi-square on a two by two table, reported as a risk difference in percentage points with a 95% confidence interval beside it, never as a p value alone. Chi-square rests on independent observations and expected cell counts of at least five; restricting each patient to a first qualifying discharge protects the first, and expected counts near 100 satisfy the second. A significant p value says a difference this large is unlikely under chance; only the interval says how large it might be.
The primary adjusted analysis is a multivariable logistic regression carrying a hospital by period interaction term, and that interaction is the estimate of interest, reported as an adjusted odds ratio with a 95% confidence interval. Four assumptions are checked rather than asserted. Linearity in the logit for continuous covariates is examined with restricted cubic splines, and age is entered in categories if the curve bends. Multicollinearity is screened with variance inflation factors, with 5 as the action threshold. Complete separation is inspected in the model output before any result is read. Events per variable governs how many covariates are allowed: roughly 220 readmissions in the smaller arm would tolerate more, but eight are fixed in advance, since a model tuned after seeing outcomes stops being a test.
Difference in differences carries one more assumption that no software reports: absent the intervention, the two hospitals would have moved in parallel. It cannot be proven, only supported, so eight quarters of baseline readmission rates are plotted and the two slopes compared before the post period opens. Missing data is settled in advance for the same reason. Ejection fraction is absent in about 14% of records, and dropping those patients would bias the estimate if the gaps cluster among the sickest, so multiple imputation by chained equations with 20 imputed sets is primary and complete-case analysis is a sensitivity check. One outcome is tested; 7-day readmission and observation stays are labeled exploratory and carry intervals only.
Threats to Validity and What They Would Cost
The largest internal threat is selection, and difference in differences answers it only partly. Ridgeway Memorial serves a larger share of dual-eligible patients, about 31% against 22%, and its baseline rate sits higher for reasons that have nothing to do with pharmacists. Subtracting each hospital from its own baseline removes a fixed gap; it does not remove a differing trend, and if one hospital's payer mix is shifting faster, the estimate absorbs that shift. History is the second: the system's transitional care call center is due to expand during the enrollment year, so its go-live date at each hospital is recorded and a sensitivity model drops those months. Regression to the mean is the third: 22.4% may be a bad year rather than a level.
Two threats sit outside the model entirely. The outcome counts readmissions inside this system only, so a patient readmitted to a competitor 12 miles away is recorded as not readmitted. That undercounts both arms, and unevenly if referral patterns differ between the two service areas; the state health information exchange feed can partly close the gap, and its match rate belongs in the results rather than an appendix. The second is the label itself, since medication reconciliation covers anything from a five-minute list check to a 40-minute counseling visit, which is why fidelity is measured. The 30-day window is a payment rule rather than a clinical boundary, so a null result at 30 days is not evidence that the service does nothing.
External validity is narrow by construction: one system, one diagnosis, one staffing ratio, an urban and suburban catchment. A result here would inform a hospital with a comparable pharmacist presence and would say little about a 25-bed critical access hospital with no clinical pharmacist on site. Stating that boundary is not modesty for its own sake. It fixes what the finding will be allowed to claim, a defensible estimate of association whose causal reading holds only if the parallel trend survives inspection. The plan that promises less at the start is the one whose results survive the first hard question at a quality meeting.
References
Agency for Healthcare Research and Quality. (2016). Designing and delivering whole-person transitional care: The hospital guide to reducing Medicaid readmissions (AHRQ Publication No. 16-0047-EF). https://www.ahrq.gov/patient-safety/settings/hospital/resource/guide/index.html
Faul, F., Erdfelder, E., Lang, A.-G., & Buchner, A. (2007). G*Power 3: A flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior Research Methods, 39(2), 175-191.
Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression (3rd ed.). Wiley.
Peduzzi, P., Concato, J., Kemper, E., Holford, T. R., & Feinstein, A. R. (1996). A simulation study of the number of events per variable in logistic regression analysis. Journal of Clinical Epidemiology, 49(12), 1373-1379.
Polit, D. F., & Beck, C. T. (2021). Nursing research: Generating and assessing evidence for nursing practice (11th ed.). Wolters Kluwer.
Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and quasi-experimental designs for generalized causal inference. Houghton Mifflin.
How this RES 5303 Module 4 example is structured
In many sections this RES5303 Module 4 assignment in the Research Methods and Applied Statistics in Healthcare course asks for a written research plan on one practice problem, with the design justified, the analysis named, and threats to validity addressed; your course instructions and rubric decide the exact form. The example is ordered the way a reviewer reads a plan. It opens with the problem and the question, so the reader knows what is being estimated, and in what denominator, before any method appears. The second section defends the design choice against the alternatives and states the sample size the question needs. The third names the analysis and attaches a check and a remedy to each assumption. The last section treats validity threats as consequences with a mitigation and a residual, which is where a plan either earns trust or loses it.
RES5303 Module 4 questions, answered
What does RES5303 Module 4 usually ask for?
American College of Education does not publish deliverable names module by module, so treat this as the common shape rather than a fixed name. In many sections a Module 4 assignment in a graduate research methods and applied statistics course asks for a written plan on one practice problem, with the design justified, the analysis named, and validity threats addressed. Your course instructions and rubric decide the exact form.
Do I need real hospital data to write a research plan for this module?
No. Build a composite setting with realistic numbers and say once, in plain language, that it is a composite. Public rates from federal agencies make the scenario believable, but what gets scored is whether your design, your sample size, and your analysis fit the question you wrote. Never lift figures from a real employer and never describe an identifiable patient.
Should I name the statistical test or just describe the analysis?
Name it, then state what it assumes and how you would check each assumption. Writing chi-square with a risk difference and a confidence interval, or logistic regression with linearity in the logit, multicollinearity, and events per variable checked, shows a reader that you could run the plan. A described analysis with no named test reads as a summary of a methods chapter.
Write yours, or have the desk draft it
This paper is an original model document written by our desk, not a submitted student paper and not an official American College of Education document. Read it for the moves, then write your own to the instructions in your classroom. If you want one built to your exact prompt and rubric, the first custom sample is free and arrives in 24 to 48 hours.