What Actually Moves "Would You Recommend"
A key-driver analysis of nearly every reporting US hospital, asking which parts of the patient experience drive whether someone would recommend the hospital. Built from 2.3 million completed CMS patient surveys.
Survey results pulled from the CMS Provider Data Catalog, rolled up in Python, and rendered live with Apache ECharts. Importance is measured with relative weights, which stay stable even though the experience dimensions are heavily correlated.
file:// path, serve the folder over HTTP (the page fetches a JSON roll-up from the same origin).
Importance against performance
Each experience dimension placed by how much it drives recommendation against how well hospitals currently score on it. The dashed lines mark the averages. A dimension that is both important and low-scoring (upper left) is where attention pays off most; a high score on a weak driver is not worth chasing.
Key-driver importance-performance map
Vertical: importance (share of explained recommendation variance). Horizontal: current top-box score.
Communication leads, and the environment matters less despite low scores
On the left, importance as each dimension's share of the recommendation variance the model explains. Communication dimensions lead. On the right, the current top-box scores. Quietness scores lowest, near 56%, yet it is one of the weakest drivers, so leading with it would be a mistake.
Drivers ranked by importance
Relative weight, the share of explained variance each dimension accounts for.
Current performance by dimension
Top-box score. The lowest-scoring dimension is highlighted.
The geography of recommendation
Where patients are most and least likely to recommend their hospital, by state. The spread is wide, and it does not track population or hospital count, so it points to differences in care and operations rather than size.
Share who would definitely recommend, by state
Survey-weighted across each state's reporting hospitals. Darker is higher.
Methodology & data
Source: CMS Provider Data Catalog, Patient survey (HCAHPS) - Hospital (dataset dgck-syfz), released 28 April 2026, retrieved 29 June 2026. The analysis keeps the 3,956 hospitals that report every dimension used, covering 2.3 million completed surveys.
Definitions: performance is the survey-weighted top-box score (for recommendation, the share who would definitely recommend). Importance is the Johnson relative weight, which splits the model R-squared (0.57) into non-negative shares per driver. Relative weights are used instead of raw regression coefficients because the experience dimensions are strongly correlated, which makes plain coefficients unstable and can even flip their sign.
Pipeline: the script resolves the current CMS release from the stable dataset id, downloads the survey file, pivots it to one row per hospital, computes survey-weighted means and relative weights in Python and statsmodels, and writes a compact JSON. The page renders that roll-up client-side with Apache ECharts. Re-running the pipeline refreshes the dashboard from the latest CMS data, so it updates as CMS publishes new releases.
References: AHRQ CAHPS Hospital Survey documentation; Doyle, Lennox, and Bell, BMJ Open (2013), on patient experience and outcomes; Reichheld, Harvard Business Review (December 2003), on the recommend question.
PythonpandasstatsmodelsRelative weights Apache EChartsChoroplethSurvey weighting