Reporting Performance and Stability in TRIPOD+AI & Riley Framework Clinical Prediction Models: A Stata-Centered Code+Framework

This article integrates the TRIPOD+AI reporting standards with the latest Riley/Collins/Ensor stability framework. It shifts the focus from just "average" performance to "individual" reliability using the pm-suite in Stata.

---

Introduction

In modern clinical prediction, showing that a model is "accurate on average" is no longer enough. Under TRIPOD+AI, you must report both Performance (how well the model works for the population) and Stability (how much an individual’s risk estimate changes if the training data were slightly different).

---

1. The Core Distinction

• Performance (Average): "On average, how close are we to the truth?" (AUC, Calibration, Net Benefit).

• Stability (Reliability): "If I re-ran this study with a different sample, would this specific patient get the same risk score?"

---

2. The 8 Required Outputs

To be fully compliant with the Riley/Collins logic, your results section should include the following three pillars:

Pillar A: Performance (Population-Average)

1. ROC Curve / C-statistic: Can the model separate cases from non-cases?

2. Calibration Plot: Does the predicted risk match observed risk across the spectrum?

3. Decision Curve Analysis (DCA): Does the model provide higher Net Benefit than "treat all" or "treat none" at clinical thresholds?

Pillar B: Stability (Individual & Decision Reliability)

4. Prediction Instability Plot: Visualizes the "wiggle" of individual risks across bootstrap re-fits.

5. Average MAPE (Stability Index): The mean absolute difference between original and bootstrap risks. Target: $< 0.02$ (context-dependent).

6. 95% Uncertainty Interval (UI): The range (2.5th to 97.5th percentile) of risk for a single patient across re-fits.

7. Classification Instability Plot: Shows "threshold flipping"—how often a patient moves from "low risk" to "high risk" across model developments.

Pillar C: Stability (Population-Level)

8. Calibration Instability Plot: A "spaghetti plot" of calibration curves from bootstrap re-fits to show if the model's reliability is volatile.

---

3. The Stata Toolchain: pm-suite

The authoritative tools for this workflow are maintained by Joie Ensor and the Riley/Collins team.

Installation

Stata

* Performance & Utilities
ssc install pmcalplot, replace
net install dca, from("https://raw.github.com/ddsjoberg/dca.stata/master/") replace

* The Stability Suite (Riley/Ensor)
net from https://joieensor.github.io/pm-suite/
net install pmstabilityplots, replace
net install pmstabilityss, replace  // For sample size planning

Mapping Requirements to Commands

Requirement	Stata Command	Key Output
Calibration	pmcalplot	Observed vs. Predicted
Clinical Utility	dca	Net Benefit
Individual Stability	pmstabilityplots	Prediction Instability Plot & MAPE
Decision Stability	pmstabilityplots	Classification Instability
Calib. Stability	pmstabilityplots	Spaghetti Calibration Curves

---

4. Implementation Workflow

Step 1: Fit and Assess Performance

Stata

logistic outcome x1 x2 x3
predict p_app, pr

* Standard Performance
lroc
pmcalplot p_app outcome, count
dca outcome p_app, xstop(0.5)

Step 2: Assess Stability (The Riley Method)

Using pmstabilityplots automates the bootstrap re-development process. It re-estimates the model parameters multiple times to see how much the individual predictions change.

Stata

* Stability Assessment (e.g., 200 bootstrap reps)
* 'threshold' defines the point for Classification Instability
pmstabilityplots outcome x1 x2 x3, reps(200) threshold(0.2)

This command generates the three critical figures:

1. Prediction Instability Plot: Highlighting the MAPE and 95% UIs.
   

2. Classification Instability: Visualizing how many patients cross the 20% risk threshold.
   

3. Calibration Instability: Showing the variation in the calibration intercept and slope.

---

5. Reporting Template (Methods Section)

"Model performance was evaluated via discrimination (C-statistic), calibration (calibration plots), and clinical utility (Decision Curve Analysis). To ensure the reliability of individual-level predictions, we performed a stability analysis according to the Riley/Collins framework. We quantified prediction instability using Mean Absolute Prediction Error (MAPE) and individual 95% uncertainty intervals (UI). Decision stability was assessed via classification instability plots at a clinical threshold of [X%]. All stability analyses were performed in Stata using the pmstabilityplots package (pm-suite), involving [200] bootstrap re-development cycles."

---

6. R Crosswalk: pminternal

If collaborating with R users, the pminternal package provides the exact same framework:

Item	R Function (pminternal)
Prediction Instability	prediction_stability()
Stability Index (MAPE)	mape_stability()
Decision Stability	dcurve_stability()
Calibration Stability	calibration_stability()

Would you like me to generate a mock Results table showing how to present the MAPE and 95% UI for different risk strata?