SA
Savvy Analytic Solutions
Carole Bonner, PhD
Portfolio — Sample Work

Applied Analytics

Healthcare quality & payer performance — CMS readmissions, quality measurement, and causal inference, built end-to-end in Stata.

Two case studies working the CMS readmissions space from opposite directions.

2 Case Studies

Overview

Read either case in full

The CMS Hospital Readmissions Reduction Program reduces all Medicare inpatient payments by up to 3% for hospitals with excess 30-day readmissions across six conditions. That makes readmission measurement one of the few places where a quality metric maps directly onto a payment mechanism — and where getting the analysis right has real money attached.

Case 01 asks whether a modifiable intervention actually causes lower readmissions, or only appears to because healthier, better-insured patients are the ones receiving it. Case 02 asks whether a hospital's excess-readmission flag reflects something specific to that hospital at all, once volume and measurable quality performance are accounted for.

Case 01 — Causal InferenceDoctoral quant-methods framing

Does follow-up care actually prevent readmissions — or does it just look that way?

Five-hospital system, FY2023, 2,500 discharges. A naïve comparison says 7-day post-discharge follow-up cuts readmission by 9.2 points. After correcting for who actually receives follow-up, the real effect is meaningfully smaller — and still worth acting on.

Headline
−6.5 ppDoubly-robust effect
2.61E-value robustness
Analytic sequence
01
Naïve logistic regression — establish the biased baseline
02
Propensity model — positivity and covariate balance checks
03
Doubly-robust IPW + teffects ipwra
04
Subgroup effects and E-value sensitivity analysis
Potential impact — $72K–$96K/yr avoided in the highest-risk segmentView full analysis
Case 02 — Multilevel ModelingReal public CMS data

Is a hospital's excess-readmission flag a real signal, or just volume?

CMS HRRP FY2026 merged with the Unplanned Hospital Visits file — 2,477 hospitals, 8,037 hospital-condition records. Once discharge volume and eight quality measures enter the model, most of the apparent hospital effect disappears. Then the model is applied to one real hospital.

Headline
92%Variance reduction
2,477Hospitals modeled
Analytic sequence
01
Reconcile and merge two CMS files on a common CCN
02
Condition-specific logistic models across six conditions
03
Pooled melogit — hospital random intercept, quality covariates
04
Case profile via empirical Bayes BLUPs — one named driver
Potential impact — one actionable gap named, not a blunt benchmark flagView full analysis
Carole Bonner — both analyses built end-to-end in Stata 19. Case 01 uses a simulated instructional dataset constructed from published literature parameters; Case 02 uses real, public CMS data from data.cms.gov. Full methods, tables, and limitations are documented in each linked write-up.