outcomes-attribution - SKILL.md Agent Skill

name: outcomes-attribution description: Attribute health outcomes to specific clinical interventions, programs, or policy changes using causal inference methods. Use when determining which interventions drove observed improvements, allocating credit across concurrent programs, or defending outcome claims to payers and regulators.

metadata: display_name: "Outcomes Attribution" short_description: "Attribute health outcomes to clinical interventions" default_prompt: "Analyze my outcomes attribution and recommend clear next actions" version: "1.0.0" tags: - healthcare

icon_path: "assets/icon.png"

Outcomes Attribution

Overview

This skill applies causal inference and attribution methodologies to determine whether observed health outcome changes can be credibly linked to specific interventions. It employs difference-in-differences, instrumental variables, interrupted time series, and multi-touch attribution models adapted from epidemiology and health services research to produce defensible attribution conclusions.

When to Use

Determining which program or intervention caused an observed outcome improvement
Allocating credit when multiple concurrent interventions could explain results
Building evidence for value-based care contract performance narratives
Defending outcome claims in RADV, MSSP, or quality program audits
Deciding resource allocation among competing programs based on attributed impact

Required Inputs

Input	Description	Format
Outcome metrics	Time-series outcome data (admissions, costs, quality scores)	Panel data
Intervention timeline	Start/end dates, scope, intensity of each intervention	Event log
Patient exposure	Which patients were exposed to which interventions and when	Enrollment/assignment
Covariates	Demographics, comorbidity, SES, provider characteristics	Patient-level data
External factors	Policy changes, seasonal patterns, market events	Context log
Comparison data	Outcomes for unexposed or differently-exposed populations	Claims/outcomes

Methodology

Step 1 — Map the Intervention Landscape

Create a comprehensive timeline of all interventions active during the evaluation period:

Program launches, modifications, and terminations with exact dates
Policy or formulary changes affecting the population
Provider network additions or losses
External events (public health emergencies, competitor market entry)
Gradual rollouts vs. abrupt implementations (matters for method selection)

Classify each intervention by: target population, mechanism of action (clinical, behavioral, financial, structural), and expected outcome pathway.

Step 2 — Select Attribution Method

Choose the appropriate causal inference approach based on data structure and intervention characteristics:

Method	When to Use	Requirements
Difference-in-differences (DID)	Clear treatment/control groups, pre/post data	Parallel trends assumption
Interrupted time series (ITS)	Single population, abrupt intervention start	≥ 8 pre-intervention time points
Propensity score matching (PSM)	Observational data, selection bias concerns	Rich covariate data
Instrumental variables (IV)	Unmeasured confounding suspected	Valid instrument available
Regression discontinuity (RD)	Assignment based on a threshold variable	Sharp/fuzzy assignment cutoff
Multi-touch attribution	Multiple concurrent interventions	Patient-level exposure data

Step 3 — Test Causal Assumptions

Validate the assumptions underlying the chosen method:

DID: Test parallel trends in pre-intervention period; plot pre-trends visually and test with placebo DID
ITS: Test for autocorrelation (Durbin-Watson), seasonality, and non-stationarity
PSM: Verify covariate balance (SMD < 0.1), common support, and no unmeasured confounders
IV: Test instrument relevance (F-statistic > 10) and exclusion restriction
Document assumption violations and their potential impact on conclusions

Step 4 — Estimate Treatment Effects

Calculate the causal effect of each intervention:

Average Treatment Effect on the Treated (ATT): Impact on patients who actually received the intervention
Intent-to-Treat (ITT): Impact on all assigned patients regardless of engagement
Local Average Treatment Effect (LATE): Impact on patients whose behavior was changed by the instrument
Report point estimates with 95% confidence intervals and p-values
Calculate effect sizes (Cohen's d, risk ratios, NNT) for clinical interpretability

Step 5 — Decompose Multi-Intervention Attribution

When multiple programs operate simultaneously, allocate outcome changes:

Sequential decomposition: Attribute based on temporal ordering of intervention starts
Shapley value attribution: Game-theoretic fair allocation across all intervention combinations
Mediation analysis: Trace outcome through intermediate pathways specific to each intervention
Exposure-weighted attribution: Weight by patient-level dose/engagement across programs
Sum attributed effects and compare to total observed change; report unexplained residual

Step 6 — Conduct Sensitivity and Robustness Checks

Stress-test attribution conclusions:

Vary measurement windows (6, 12, 18-month post-intervention)
Test alternative comparison groups or matching specifications
Run placebo tests (apply method to periods/populations with no intervention)
Conduct E-value analysis for unmeasured confounding (how strong would a confounder need to be to nullify results?)
Report Rosenbaum bounds for PSM sensitivity

Step 7 — Produce Attribution Report

Structure conclusions with appropriate certainty language:

Strong attribution (multiple methods converge, assumptions validated): "The evidence strongly supports that Program X caused a Y% reduction in Z"
Moderate attribution (single method, assumptions reasonable): "Program X is likely associated with a Y% reduction in Z, though residual confounding cannot be fully excluded"
Weak attribution (assumption violations, limited data): "Program X may have contributed to the observed Y% reduction in Z, but definitive causal claims are not supported by available evidence"

Output Specification

Attribution Report:
├── Intervention Timeline (visual map of all active interventions)
├── Method Selection Rationale (chosen approach with justification)
├── Assumption Validation (test results for causal assumptions)
├── Treatment Effect Estimates (ATT, ITT with CI and p-values)
├── Multi-Intervention Decomposition (attributed share per program)
├── Sensitivity Analysis Results (robustness across specifications)
├── Attribution Confidence Assessment (strong/moderate/weak per claim)
├── Unexplained Residual Analysis (portion not attributed to known interventions)
└── Recommendations (resource allocation implications)

Analysis Framework

Attribution Confidence Scoring

Factor	Weak (1)	Moderate (2)	Strong (3)
Study design	Pre-post only	Quasi-experimental	Randomized or natural experiment
Comparison group	None	Historical or unmatched	Well-matched concurrent
Assumption validity	Violations detected	Assumptions plausible	Assumptions verified
Effect consistency	Single specification	Consistent across some	Robust across all
Biological plausibility	Unclear mechanism	Plausible pathway	Established causal pathway

Total score: 5-8 = weak, 9-12 = moderate, 13-15 = strong attribution confidence.

Common Attribution Pitfalls

Regression to the mean: High-cost patients selected into programs naturally regress, mimicking program effect
Selection bias: Patients who enroll/engage differ systematically from those who don't
Secular trends: Background improvements (declining smoking, better medications) mistaken for program impact
Contamination: Control group partially exposed to intervention components
Hawthorne effect: Observation itself changes behavior

Examples

Example 1 — Care Management Program Attribution A health system runs a CHF care management program alongside a new cardiology telehealth initiative. CHF admissions dropped 25%. ITS analysis on care management (launched 6 months earlier) attributes 18% reduction (95% CI: 12-24%). DID analysis adding telehealth as a second intervention attributes an additional 5% (95% CI: 1-9%). Residual 2% consistent with background trend. Shapley decomposition: 72% care management, 20% telehealth, 8% secular.

Example 2 — VBC Contract Performance Defense An ACO must demonstrate that its 8% total cost reduction is attributable to its interventions (not market changes) for shared savings. Apply DID against non-ACO-attributed patients in the same market. Estimate ATT of 6.2% (p=0.003). Remaining 1.8% explained by regional utilization trend. E-value analysis shows an unmeasured confounder would need RR > 2.8 to nullify, supporting robust attribution.

Guidelines

Always prefer quasi-experimental methods over simple pre-post comparisons
Report ITT as primary analysis; per-protocol as secondary (avoids healthy-participant bias)
Never claim causation from correlation — clearly state the method and its limitations
When multiple programs operate concurrently, attribution must sum to ≤ 100% of observed change
Document all analytical choices (method, specification, comparison group) to enable replication

Validation Checklist

Causal method is appropriate for the data structure and intervention type
Key assumptions are explicitly tested and reported
Confidence intervals are reported for all effect estimates
Sensitivity analyses cover at least 3 alternative specifications
Attribution claims use appropriate certainty language
Unexplained residual is reported and interpreted
Report is structured for peer review or audit scrutiny

HIPAA Compliance

This skill processes Protected Health Information (PHI) for health care operations including program evaluation and quality improvement. All analyses must comply with HIPAA Privacy and Security Rules. Apply minimum necessary data access standards. Patient-level exposure and outcome data must be stored in access-controlled environments. Report aggregate attribution results only; apply minimum cell-size suppression (n ≥ 11) for sub-group analyses. De-identify data per 45 CFR §164.514 before sharing with external evaluators or auditors.