How the Star Rating Predictor Works

The interactive Star Rating predictor lets you adjust four quality inputs and watch a predicted Medicare Star Rating update in real time. This post explains what is happening under the hood.

The right model for ordered outcomes

Star Ratings are ordinal: 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 tens of millions of dollars. A jump from 4 to 5 does not carry the same financial weight.

Standard linear regression treats the outcome as continuous and assumes equal spacing. Multinomial logistic regression ignores the ordering entirely. Ordinal logistic regression — specifically the proportional odds model introduced by McCullagh (1980) — respects both the ordering and the non-equal spacing.

The model estimates cumulative probabilities:

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

where $\sigma(x) = \frac{1}{1 + e^{-x}}$ is the logistic function, $\alpha_k$ are threshold parameters for each cumulative split, and $\beta$ is a shared coefficient vector that captures each input's effect across all thresholds.

This is the proportional odds assumption: the effect of each predictor is constant regardless of which threshold you are crossing. It means one set of coefficients tells the full story.

From cumulative probabilities to a prediction

The cumulative model gives $P(Y \geq k)$ for each threshold. Individual star-level probabilities come from differencing:

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

$$P(Y = k) = P(Y \geq k) - P(Y \geq k+1) \quad \text{for } k = 2, 3, 4$$

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

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

Calibration to CMS weights

The demo uses four inputs that map to real CMS Star Rating domains:

Input	CMS Domain	CMS Weight	Coefficient ($\beta$)
HEDIS Composite Rate	Part C — Process & Outcome	1–3×	+0.08 per pp
CAHPS Member Satisfaction	Part C — Patient Experience	2×	+1.50 per unit
Medication Adherence	Part D — Drug Measures	3×	+0.06 per pp
Readmission Rate	Part C — Intermediate Outcome	3×	−0.08 per pp

The coefficients are calibrated to satisfy two constraints simultaneously:

1. Weight proportionality: Each input's total contribution across its realistic slider range is proportional to its share of CMS weighted points (out of 81 total for MA-PD contracts).
2. Distribution calibration: At median slider positions, the predicted distribution matches the 2025 MA-PD star distribution — average 3.92 stars, approximately 42% of contracts at 4 stars or above, roughly 2% at 5 stars.

The four intercepts ($\alpha_2 = -10.1$, $\alpha_3 = -11.76$, $\alpha_4 = -15.1$, $\alpha_5 = -18.17$) position the thresholds so that a plan at the 50th percentile on all inputs lands around 3.2 stars.

The "What Would Move the Needle?" calculation

For each input, the model simulates a feasible one-year improvement:

Input	Simulated Step	Rationale
HEDIS Composite	+5 pp	Achievable with targeted gap closure
CAHPS Satisfaction	+0.3 points	Typical gain from member experience initiatives
Medication Adherence	+5 pp	Consistent with MTM program impact
Readmission Rate	−2 pp	Achievable with care transition programs

It computes $E[Y \mid x_j + \Delta] - E[Y \mid x_j]$ for each input and reports the two highest-impact levers. Because the logistic function is nonlinear, the highest-impact lever changes depending on where the plan currently sits — it is largest near the ordinal cut-points and smallest in the tails.

CMS Reward Factor

Plans maintaining 4 or more stars for three consecutive years receive up to +0.4 stars added to the summary rating before final rounding. The demo toggle simulates this additive bonus.

What the model does not include

1. No real training data. Coefficients are synthetic, calibrated to CMS weight shares and the published star distribution. A production model would be trained on contract-level measure scores.
2. Simplified inputs. CMS uses 42 measures across 9 domains. The four inputs here represent roughly 40% of total weighted points; the remaining 60% is absorbed into the intercepts.
3. No case-mix adjustment. CMS applies a Categorical Adjustment Index (CAI) for plans serving dual-eligible, disabled, or low-income populations. D-SNP plans should treat these predictions as pre-adjustment estimates.
4. No improvement measures. CMS awards credit for year-over-year improvement, which requires longitudinal data not modeled here.
5. Adherence treated as a single composite. The three PDC measures (diabetes, RAS, statins) have different baseline distributions and intervention profiles.

Despite these simplifications, the model accurately represents the correct statistical framework, the correct relative importance of each domain, and the nonlinear sensitivity structure inherent in ordinal logistic regression.

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.