Using causal inference methods on observational data to estimate the impact of missed post-discharge follow-up on 30-day hospital readmissions — without randomizing patients to worse care.
Lakeview's 30-day all-cause readmission rate put the five-hospital network at risk of $2.8M/year in HRRP Medicare penalties. The CMO suspected gaps in post-discharge follow-up were a major driver, but the proposed transition-of-care program would cost $1.6M annually.
The question: is the causal impact of missed follow-up large enough to justify the investment?
An A/B test was not ethical — you cannot randomize patients to receive no follow-up. The team needed observational causal inference methods.
IPW reweights the sample so patients who received follow-up "look like" those who missed it. Combined with regression, the estimate is doubly robust — consistent if either model is correct.
Within comparable acuity groups, follow-up was clearly protective. The naïve overall comparison masked this because high-acuity patients dominated the "missed" group.
| Component | Estimate |
|---|---|
| Discharges missing follow-up/year | ~22,800 |
| Excess readmissions attributed to gap | ~7,000/year |
| Cost per readmission | $12,400 |
| Total attributable cost | $86.8M/year |
| Reach assumption (conservative) | 70% |
| Effectiveness assumption | 80% |
| Projected cost avoidance | $48.6M/year |
Program cost: $1.6M/year. Even if the true effect were half the estimate, the program would pay for itself many times over. Break-even requires only a 1.0 pp readmission reduction.
Phase 1 (Months 1–3): Pilot targeting Medicaid/uninsured patients with HF and COPD — the highest-return segment. Follow-up contact rate rose from 39% → 78%. Early readmission data: 28.1% → 19.3%.
Phase 2 (Months 4–12): System-wide rollout across all five hospitals, 7-day TOC nurse coverage. Annualized program cost: $1.6M.
Subgroup findings: Medicaid patients showed a larger effect (−34.1 pp vs. −25.6 pp for commercial). HRRP-targeted conditions (HF, COPD) showed effects of −36.2 pp and −33.8 pp — concentrating the financial return where CMS penalties are highest.
Follow-up studies planned: Interrupted time series (pre-post impact), dose-response analysis (does a second contact add value?), and formal cost-effectiveness with QALYs for payer negotiations.
Sensitivity analysis showed an unmeasured confounder would need to be 8× stronger than acuity + insurance combined to reduce the estimate below break-even. Three independent methods (OLS, IPW, doubly-robust) converged on a 26–31 pp effect. The CFO approved the investment in a single meeting.