Building a Browser-Based Star Rating Predictor: Methodology & Evidence

TL;DR for Executives

• What it does: The interactive demo predicts a Medicare Star Rating (1–5) from four quality inputs using a statistical model that runs entirely in your browser.
• Why it matters: Plans at ≥4 stars qualify for Quality Bonus Payments (QBPs) — a 5% premium increase worth $50M+ annually for large contracts. The 4-star threshold is the single most consequential cliff in managed care economics.
• What it shows you: Which quality lever — HEDIS, CAHPS, medication adherence, or readmission — delivers the most star improvement at your plan’s current performance level, so you can prioritize investment.

The interactive demo on this site lets visitors adjust four healthcare quality inputs and watch a predicted Medicare Star Rating update in real time. This post documents the statistical approach, explains why ordinal logistic regression is the right tool for the job, and cites the CMS methodology and published research that inform the model’s design.

What the Demo Does

Four sliders control inputs that map to real CMS Star Rating domains:

The output includes:

• A predicted overall Star Rating on a 1–5 scale with full probability distribution
• Quality Bonus Payment context — whether the predicted rating would qualify for QBPs, how far from the threshold, and the financial implications
• Plain-language impact guidance — which quality lever offers the largest star gain from a feasible improvement at the current performance level
• An ordinal threshold visualization showing where the linear predictor falls relative to cut-points
• National percentile markers (25th, 50th, 75th) on each slider for benchmarking

An optional CMS Reward Factor toggle adds up to +0.4 stars, simulating the bonus CMS grants to plans maintaining ≥4 stars for 3+ consecutive years.

The Financial Stakes: Quality Bonus Payments

The 4-star threshold is the most consequential cliff in Medicare Advantage economics. Plans at ≥4 stars qualify for Quality Bonus Payments — a 5% increase in their CMS benchmark payment. For a plan with 200,000 members, this translates to roughly $50–80 million annually in additional revenue.

The stakes are asymmetric:

• Below 3.5★: No QBP eligibility. At ≤2.5★ for 3 consecutive years, CMS may terminate the contract entirely.
• 3.5–3.99★: Plans round to 3.5★ and receive no bonus, despite being close. This is the “dead zone.”
• ≥4.0★: Full QBP eligibility. The reward factor can push plans that sustain 4★+ even higher.

This cliff function is why quality improvement ROI is heavily nonlinear — a 0.1-star improvement from 3.9 to 4.0 can be worth tens of millions, while the same improvement from 4.3 to 4.4 has essentially zero marginal financial impact.

The demo surfaces this context directly below the predicted rating, so users can immediately see where their plan stands relative to the QBP cliff.

Why Ordinal Logistic Regression?

Star Ratings are an ordinal outcome: 1 < 2 < 3 < 4 < 5, but the distances between levels are not equal. A plan jumping from 3 to 4 stars triggers quality bonus payments worth millions; 4 to 5 does not carry the same marginal financial impact.

Standard linear regression treats the outcome as continuous and assumes equal spacing. Multinomial logistic regression ignores the ordering entirely. Ordinal (proportional odds) logistic regression respects both the ordering and the non-equal spacing by modeling cumulative probabilities:

$$P(Y \geq k \mid X) = \sigma(\alpha_k + \beta^T X)$$

where $\sigma$ is the logistic function $\sigma(x) = \frac{1}{1 + e^{-x}}$, $\alpha_k$ are threshold parameters (intercepts) for each cumulative split, and $\beta$ is a shared coefficient vector.

This is the proportional odds model introduced by McCullagh (1980), the standard approach for ordinal outcomes in biomedical research.

Foundational Reference

McCullagh, P. (1980). Regression Models for Ordinal Data. Journal of the Royal Statistical Society, Series B, 42(2), 109–142. Wiley Online Library · Author PDF

McCullagh showed that the proportional odds model provides interpretable odds ratios that summarize the effect of each predictor across all threshold cutpoints, making it ideal for ordered rating scales.

CMS Star Ratings Methodology

CMS publishes annual Technical Notes that detail exactly how Star Ratings are calculated. The key points relevant to this demo:

Measure Structure

The 2025 Star Ratings include up to 42 measures across 9 domains for MA-PD contracts. Each measure receives a weight from 1 to 5. CMS assigns individual measure stars using either clustering or relative distribution methods, then computes weighted summary ratings.

CMS (2024). Medicare 2025 Part C & D Star Ratings Technical Notes (Updated 10/03/2024). PDF

CMS (2024). 2025 Medicare Advantage and Part D Star Ratings Fact Sheet. CMS Newsroom

Medication Adherence: Triple-Weighted and Highly Sensitive

Three medication adherence measures — for diabetes medications, RAS antagonists (hypertension), and statins (cholesterol) — each carry a weight of 3 in the overall Star Rating. Together they account for 9 of 81 total weighted stars (11.1%) in the 2026 ratings cycle.

Key nuance: CMS measures the percentage of members achieving PDC ≥80% (Proportion of Days Covered at or above 80%), not the plan’s mean PDC. This is a binary threshold applied at the member level, then aggregated. The three sub-measures have meaningfully different baseline performance profiles — statin adherence typically runs 5–8 percentage points higher than diabetes medication adherence across plans — so interventions should be targeted by measure, not treated as a monolith.

As little as 3 percentage points in PDC can separate a 3-star from a 5-star rating on the RAS antagonist adherence measure, making this the single most actionable lever for many plans.

A 2025 decade-long analysis confirmed that plans performing well on adherence measures have substantially higher rates of achieving ≥4-star overall ratings and qualifying for quality bonus payments:

Akinbosoye, O. E., et al. (2025). The influence of medication adherence on Medicare Star Ratings: A decade-long analysis of health plan performance. Journal of Managed Care & Specialty Pharmacy, 31(5), 512. JMCP · PMC

Plan All-Cause Readmissions Now Triple-Weighted

For 2025, CMS increased the weight of the Plan All-Cause Readmissions (PCR) measure from 1 to 3, making it one of the most heavily weighted Part C measures. PCR assesses the percentage of hospital stays followed by an unplanned readmission within 30 days.

This change contributed to the average overall MA-PD Star Rating dropping to 3.92 (down 0.15 stars) — the lowest in four years — and the number of 5-star contracts falling from 38 to just 7.

Operational note for D-SNP plans: Readmission is driven heavily by social determinants of health — housing instability, food insecurity, lack of transportation — making it substantially harder to move than clinical quality or adherence measures for plans serving dual-eligible populations. The model’s coefficients treat a percentage-point improvement in readmission equivalently to one in HEDIS, but the operational cost and difficulty of achieving that improvement can differ by an order of magnitude depending on population acuity.

CMS (2024). 2025 Medicare Advantage and Part D Star Ratings Fact Sheet. CMS Newsroom

AdhereHealth (2024). Conquering the Plan All-Cause Readmissions (PCR) Measure. Article

CAHPS Patient Experience Measures

CAHPS survey scores evaluate member satisfaction across dimensions like provider communication, access to care, and customer service. These measures have been weighted between 2× and 4× in recent years. For 2026, CMS reduced the weight from 4 back to 2, but CAHPS remains a significant contributor to overall ratings.

Measurement noise caveat: CAHPS scores are derived from a sample survey with response rates that vary widely across plans (typically 20–45%). Low response rates amplify sampling noise, meaning that a plan’s true satisfaction level may differ meaningfully from its reported score. The model treats CAHPS as a precise input, but users should interpret the output with this measurement uncertainty in mind — particularly for plans with small enrollment or low response rates.

MA-PDP CAHPS (2025). Scoring and Star Ratings. ma-pdpcahps.org

Health Management Associates (2024). HMA Analysis of Medicare Advantage Star Rating Challenges. HMA Blog

Reward Factor

CMS applies a Reward Factor of up to +0.4 stars to plans that have maintained ≥4-star overall ratings for three or more consecutive years. This bonus is additive — applied to the weighted summary rating before final rounding — and can be the difference between a 3.5-star and a 4-star plan qualifying for quality bonus payments.

The demo includes a toggle to simulate this factor, reflecting its material impact on final ratings.

Improvement Measures

CMS also evaluates year-over-year improvement in certain measures. Plans showing statistically significant improvement may receive credit even if their absolute level remains below the top cut-point. The current demo does not model improvement measures, as they require longitudinal data (prior-year performance), but they represent a meaningful path to higher ratings for plans in the 3.0–3.5★ range that are on an upward trajectory.

Published Precedent: Ordinal Logistic Regression for Star Ratings

Multiple published studies have applied ordinal logistic regression to Medicare star ratings or closely related healthcare quality outcomes.

Kurian et al. (2021) — Hospital Star Ratings

Kurian, N., Maid, J., Mitra, S., Rhyne, L., Korvink, M., & Gunn, L. H. (2021). Predicting Hospital Overall Quality Star Ratings in the USA. Healthcare, 9(4), 486. PMC · MDPI · PubMed

• Applied ordinal logistic regression with stepwise variable selection to 4,519 U.S. hospitals
• Found that 20 performance measures (from 57 total) contained all the relevant information for star rating prediction after accounting for correlation
• Used multiple imputation to handle missing data, enabling inference even when not all measures are available
• Demonstrated that the proportional odds model provides interpretable odds ratios for the relationship between quality measures and star rating levels

Hohmann et al. (2018) — Part D Star Ratings

Hohmann, N., Hansen, R., Garza, K. B., Harris, I., Kiptanui, Z., & Qian, J. (2018). Association between Higher Generic Drug Use and Medicare Part D Star Ratings: An Observational Analysis. Value in Health, 21(10), 1186–1191. PubMed · ScienceDirect

This is the most direct methodological precedent: Hohmann et al. explicitly used ordinal logistic regression to model Medicare Part D summary and domain star ratings (1–5) as a function of generic drug dispensing rates, controlling for contract type and enrollment. They found higher generic dispensing was associated with higher summary star ratings (adjusted OR 1.08, 95% CI 1.04–1.12).

Xie et al. (2023) — Adherence Predicts Clinical Outcomes

Xie, Z., St. Clair, P., Goldman, D. P., & Joyce, G. F. (2023). Is There a Relationship Between Part D Medication Adherence and Part C Intermediate Outcomes Star Ratings Measures? Journal of Managed Care & Specialty Pharmacy, 29(8), 918–925. PMC

Analyzed 366 contracts and found medication adherence measures explained 27–29% of variance in related clinical outcomes ($R^2$ = 0.27–0.29). Each unit increase in adherence increased odds of top-quartile performance by factors of 4.13–4.69, providing strong empirical support for adherence as a key predictor.

Lingsma et al. (2021) — Why Ordinal Beats Binary

Lingsma, H. F., Bottle, A., Middleton, S., Kievit, J., Steyerberg, E. W., & Marang-van de Mheen, P. J. (2021). Ordinal Outcome Analysis Improves the Detection of Between-Hospital Differences in Outcome. BMC Medical Research Methodology, 21(1), 4. PMC

Demonstrated that ordinal outcome analysis substantially improves statistical power for detecting quality differences compared to dichotomizing outcomes (e.g., “4+ stars vs. below”). This provides methodological justification for treating star ratings as ordinal rather than collapsing them into binary categories.

Together, these studies validate using a small number of composite quality inputs in an ordinal logistic framework to predict a 1–5 star outcome.

How the Demo Model Works

Architecture

The demo implements a cumulative logit model (proportional odds) entirely in client-side JavaScript. The core equation:

$$P(Y \geq k \mid X) = \sigma(\alpha_k + \beta_1 \cdot \text{HEDIS} + \beta_2 \cdot \text{CAHPS} + \beta_3 \cdot \text{MedAdherence} + \beta_4 \cdot \text{Readmission})$$

Four threshold parameters ($\alpha_2, \alpha_3, \alpha_4, \alpha_5$) define the cumulative splits. The shared coefficient vector $\beta$ captures each input’s effect across all thresholds (the proportional odds assumption).

Converting Cumulative Probabilities to Category Probabilities

The cumulative model yields $P(Y \geq k)$ for each threshold. Individual category probabilities are obtained by differencing:

$$P(Y = 1) = 1 - P(Y \geq 2)$$

$$P(Y = 2) = P(Y \geq 2) - P(Y \geq 3)$$

$$P(Y = 3) = P(Y \geq 3) - P(Y \geq 4)$$

$$P(Y = 4) = P(Y \geq 4) - P(Y \geq 5)$$

$$P(Y = 5) = P(Y \geq 5)$$

The expected value $E[Y] = \sum_{k=1}^{5} k \cdot P(Y = k)$ gives the predicted star rating displayed in the UI.

Coefficient Calibration

The synthetic coefficients are calibrated to satisfy two constraints simultaneously:

1. CMS weight proportionality: Each input’s total contribution across its realistic slider range is proportional to its share of CMS weighted points.
2. Distribution calibration: At median slider positions, the predicted distribution matches the 2025 MA-PD star distribution (avg 3.92, ~42% at ≥4★, ~2% at 5★).

CMS Weight Share Analysis

From the 2025 Technical Notes, the 81 total weighted points for MA-PD contracts break down as:

The demo’s four inputs represent 40% of total CMS weighted points. The “other” 60% is held constant (absorbed into the intercepts), a standard approach when building a reduced-form predictor from the most policy-relevant measures.

Resulting Coefficients

These proportions track the CMS weight shares when accounting for scale differences: CAHPS has fewer scale units but high weight per measure, while readmission is a single measure despite its triple weight.

Intercept Calibration

The four intercepts ($\alpha_2 = -10.1$, $\alpha_3 = -11.76$, $\alpha_4 = -15.1$, $\alpha_5 = -18.17$) are set so that at median input values:

$$P(Y \geq 2) \approx 0.99, \quad P(Y \geq 3) \approx 0.95, \quad P(Y \geq 4) \approx 0.40, \quad P(Y \geq 5) \approx 0.03$$

This produces an expected rating of ~3.2★ at default slider positions (representing a middling plan), consistent with the real-world distribution where most plans cluster between 3.0 and 4.5 stars.

Reward Factor

When the CMS Reward Factor toggle is enabled, +0.4 is added to $E[Y]$, capped at 5.0. This models the CMS mechanism where sustained high performance receives an additive bonus before final rounding.

Impact Guidance

The demo computes scenario-based impact guidance for each input by simulating a feasible one-year improvement:

For each input, the model computes $E[Y \mid x_j + \Delta] - E[Y \mid x_j]$ and reports the two highest-impact levers in plain language (e.g., “Improving medication adherence from 70% → 75% would shift the predicted rating by +0.3★”).

Because the logistic function is nonlinear, impact varies across the input space — it is largest near the ordinal cut-points and smallest in the tails. This means the highest-impact lever changes depending on where your plan currently sits.

Threshold Visualization

The ordinal threshold bar shows the four cut-points ($-\alpha_2, -\alpha_3, -\alpha_4, -\alpha_5$) on a linear predictor scale, with a moving indicator for the current $z$ value. This makes visible the core mechanism of the proportional odds model: the linear predictor $z$ crossing successive thresholds shifts probability mass from lower to higher star categories.

Limitations and Disclaimers

This demo uses synthetic weights that are not trained on actual CMS contract-level data. The model is illustrative:

1. No real training data. Actual Star Rating prediction would require contract-level measure scores linked to final ratings, which CMS publishes but which would need proper train/test splitting and cross-validation.

2. Simplified inputs. Real Star Ratings incorporate 42 measures across 9 domains. Our 4 inputs represent the most influential categories but absorb the remaining ~60% of weighted points into the intercepts. This is a reduced-form model, not a full specification.

3. Proportional odds assumption. We assume the effect of each predictor is constant across all threshold cutpoints. In practice, this assumption should be tested with a Brant test or graphical method (Harrell, 2001). For star ratings, the assumption is plausible for HEDIS and adherence (monotonic quality measures) but may be violated for readmission rate, where plans with extreme values may face nonlinear threshold effects.

4. No case-mix adjustment (CAI). This is the most important limitation for D-SNP and other plans serving high-acuity populations. CMS applies a Categorical Adjustment Index (CAI) that adjusts measure-level stars for the proportion of enrollees who are dual-eligible (Medicare + Medicaid), disabled, or low-income subsidy recipients. CAI adjustments can shift individual measure stars by up to a full level. Plans serving predominantly dual-eligible populations should interpret the demo’s predictions as pre-adjustment estimates — actual final ratings may differ substantially. A future version could include a D-SNP population toggle to approximate CAI effects.

5. No interaction terms. The model assumes additive effects. In reality, HEDIS and medication adherence are partially overlapping constructs (adherence measures feed into HEDIS composites for some plans), and the joint effect of poor readmission + poor CAHPS may be super-additive. A production model would test for interactions.

6. Input ranges are bounded. Slider ranges reflect the 5th–95th percentile of observed plan performance, not the full theoretical range. This improves face validity but means the model cannot represent extreme outlier plans.

7. No improvement measures. CMS awards credit for statistically significant year-over-year improvement, which can boost ratings for plans on an upward trajectory even if absolute levels remain modest. This requires longitudinal data and is not modeled here.

8. Adherence is treated as a single composite. The three PDC measures (diabetes, RAS, statins) have different baseline distributions and different intervention profiles. Statin adherence typically runs 5–8 points higher than diabetes adherence. The demo’s single slider is a weighted average, but operational improvement strategies should target each measure independently.

Despite these simplifications, the demo accurately represents:

• The correct statistical framework (ordinal logistic regression) for a 1–5 ordered outcome
• The correct direction and CMS-proportional relative importance of each input domain
• The probabilistic nature of quality prediction (showing full distributions, not just point estimates)
• The nonlinear sensitivity structure inherent in logistic models (impact varies by position on the curve)
• The material impact of CMS’s reward factor and QBP threshold on plan economics

References

1. McCullagh, P. (1980). Regression Models for Ordinal Data. JRSS Series B, 42(2), 109–142.
2. CMS (2024). Medicare 2025 Part C & D Star Ratings Technical Notes.
3. CMS (2024). 2025 Medicare Advantage and Part D Star Ratings Fact Sheet.
4. Kurian, N. et al. (2021). Predicting Hospital Overall Quality Star Ratings in the USA. Healthcare, 9(4), 486.
5. Hohmann, N. et al. (2018). Association between Higher Generic Drug Use and Medicare Part D Star Ratings. Value in Health, 21(10), 1186–1191.
6. Xie, Z. et al. (2023). Is There a Relationship Between Part D Medication Adherence and Part C Intermediate Outcomes Star Ratings Measures? JMCP, 29(8), 918–925.
7. Lingsma, H. F. et al. (2021). Ordinal Outcome Analysis Improves the Detection of Between-Hospital Differences in Outcome. BMC Medical Research Methodology, 21(1), 4.
8. Akinbosoye, O. E. et al. (2025). The influence of medication adherence on Medicare Star Ratings. JMCP, 31(5), 512.
9. Harrell, F. E. (2001). Regression Modeling Strategies. Springer.
10. SAS Institute (2005). Using the Proportional Odds Model for Health-Related Research. SUGI 30 Proceedings, Paper 205-30.