Decision Analytic Modeling in Health Economics
- Introduced by
- Pauker & Kassirer (medical decision analysis, Massachusetts General Hospital)
- Year
- 1975
- Type
- Method
- Family
- health economic evaluation framework
What does it do?
Decision analytic modeling is a systematic framework for comparing health interventions by integrating evidence on probabilities, outcomes, costs, and patient preferences into a quantitative model.
Developed by Pauker and Kassirer in 1975, decision analysis structures clinical uncertainty and economic trade-offs, enabling transparent comparison of treatment options and identification of optimal strategies. Used in health technology assessment, clinical practice guideline development, and resource allocation decisions.
The intuition behind the method
A clinician faces a decision: should a 65-year-old with newly diagnosed hypertension initiate antihypertensive therapy or observe (watchful waiting)? Treatment reduces cardiovascular risk but has side effects and costs. No single trial directly answers this question for this patient because outcomes depend on individual risk profile, medication efficacy in their subgroup, and preferences. Decision analysis quantifies the trade-off: What is the probability this patient benefits? At what cost? What is quality of life on vs off treatment? By integrating all available evidence (RCTs, epidemiology, cost data) into one model, decision analysis calculates expected value of each option. Expected value = sum of (probability of outcome × value of outcome). A treatment with 80% probability of preventing stroke (worth 2 QALYs) and 20% probability of causing side effects (worth −0.5 QALYs) has expected value = 0.8×2 + 0.2×(−0.5) = 1.5 QALYs. If alternative (watchful waiting) has expected value 1.0 QALYs, treatment is preferred by expected value criterion. Decision analysis forces explicit reasoning and permits systematic comparison.
History
Pauker & Kassirer (1975, 1976) pioneered decision analysis in medicine, applying expected value theory (von Neumann & Morgenstern) to clinical decision-making. Early applications: when to treat asymptomatic conditions, when to perform diagnostic tests. Weinstein & Fineberg (1980) published influential textbook. By 1990s, decision analysis was established in clinical epidemiology and health services research. Briggs, Claxton, and Sculpher (2000s) extended decision analysis to economic evaluation (integrating costs, QALYs, CEA/CBA). Decision trees gave way to Markov models for chronic disease (1990s). Probabilistic sensitivity analysis (Monte Carlo simulation) became standard by 2000s, enabling uncertainty quantification. Today, decision modeling is routine in HTA submissions and guideline development; tools (TreeAge, R packages) have democratized access beyond statisticians.
When to use it
When comparing health interventions and decision-makers need systematic evidence integration. Decision analysis is standard for: (1) clinical practice guideline development (comparing treatment strategies). (2) health technology appraisal (reimbursement decisions). (3) resource allocation (prioritizing interventions). (4) research prioritization (identifying where more data would most improve decision confidence). (5) patient counseling (shared decision-making tools). Used when: multiple options exist, outcomes uncertain, preferences and costs matter, and transparent framework is valued. Not needed for: straightforward clinical comparisons with clear winner (e.g., two drugs with identical efficacy, one much cheaper—intuitive decision); rare outcomes where modeling adds little value.
Limitations and considerations
- Parameter uncertainty: model is only as good as input probabilities and utilities; if data are sparse or biased, model output unreliable.
- Assumes rational decision-making: expected value approach assumes decision-makers are risk-neutral; actual preferences often non-linear (loss aversion, hope bias).
- Simplifies reality: model necessarily abstracts disease complexity; important clinical factors may be omitted or averaged, losing patient-specific nuance.
- Requires substantial data: building rigorous model demands data on probabilities (trials, registries), utilities (preference surveys), and costs (accounting records); data often incomplete.
- GIGO (Garbage In, Garbage Out): if input assumptions are wrong or controversial, output is misleading despite mathematical rigor.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Integrates evidence transparently: probabilities, outcomes, costs, utilities explicitly stated and linked, enabling stakeholder critique and updating as evidence evolves.
- Handles uncertainty systematically: sensitivity and probabilistic analysis quantify impact of uncertainty on decisions, guiding where more research is most valuable.
- Facilitates communication: decision tree visualizes problem structure; stakeholders (clinicians, patients, payers) understand trade-offs and value judgments.
- Flexible: accommodates diverse evidence types (RCTs, observational, expert opinion) and can be updated as new data emerge.
Applications
- Clinical guideline development: cardiology guideline for acute coronary syndrome treatment: decision model compares PCI (percutaneous coronary intervention) vs medical management vs emergency bypass. Model integrates RCT efficacy, procedural complications, recovery time, QoL; outputs recommend PCI for specific patient profiles (e.g., age <70, symptom duration <12 hours).
- HTA submission: pharmaceutical company submits new diabetes drug; decision model compares drug vs standard therapy over 30-year horizon. Markov model tracks disease states (A1c control, complications). Output: ICER $25,000/QALY. NICE recommends reimbursement, informing pricing negotiations.
- Screening program: public health agency uses decision model to evaluate cervical cancer screening strategies (Pap smear annually vs every 3 years vs HPV-based screening). Model compares costs, life-years saved, QALYs, harms (over-treatment). Output: HPV-based every 5 years most efficient, guides national policy.
- Shared decision-making: patient with early prostate cancer (low-risk, asymptomatic): decision model compares active surveillance vs radical prostatectomy. Tree incorporates prostate cancer natural history, surgery risk, functional outcomes (continence, erectile function), quality of life. Patient can see expected value of each option tailored to their preferences and risk tolerance.
- Research priority-setting: model of infectious disease control compares vaccination, treatment, and surveillance strategies. Probabilistic sensitivity analysis identifies parameters with highest uncertainty. Value of information (VOI) analysis calculates: 'How much would perfect data on X reduce decision uncertainty?' Guides research funding to studies with highest value.
How it works
How it works
Step 1: Define the decision problem. What are the decision alternatives? (e.g., Screen for colorectal cancer (colonoscopy) vs FOBT (fecal occult blood test) vs no screening). Who is the patient population? (e.g., age 50–75, average risk). What is the time horizon? (10-year outcome? Lifetime?). Step 2: Map the structure. Identify all chance events (disease progresses vs regresses, test is positive vs negative, adverse events occur vs not). Identify outcomes of interest (alive/dead, symptom-free vs symptomatic, cost in dollars, QALYs). Step 3: Populate with probabilities. Where do probabilities come from? RCTs (primary evidence), cohort studies, national registries, meta-analyses. Example: If RCT shows colonoscopy detects colorectal cancer in 90% of cases and prevents death in 95% of detected cases, use P(detection) = 0.90, P(prevented death | detected) = 0.95. For disease progression (not in any trial), use published epidemiology. Step 4: Assign outcomes. Utilities: quality-of-life weights (0–1 scale) for each health state (normal health = 1, cancer-diagnosed = 0.7, post-treatment = 0.8). Costs: screening cost ($300), cancer treatment cost ($50,000), ongoing monitoring ($500/year). Step 5: Build the model. Use decision tree (short horizon, discrete events) or Markov (long horizon, recurring states). Software: TreeAge Pro (commercial), Excel (basic), R/Python (flexible). Step 6: Validate. Run baseline scenario (current standard care) and compare outputs to published literature (does model match epidemiology?).
Formula
Define decision problem: intervention options (e.g., screen vs not screen, treat A vs treat B vs watchful waiting), patient population, time horizon. Identify chance events (test results, disease progression, treatment response, adverse events) and outcomes (health states, mortality, quality of life, costs). Obtain probabilities from trials, literature, clinical expertise. Assign utilities and costs to outcomes.
Decision Tree Example: Colorectal cancer screening in 50-year-old. Root: Decision node (Square): [Screen vs No Screen]. Branch 1 (Screen): Colonoscopy performs. Chance node (Circle): [Cancer found (10% prevalence) vs Not found (90%)]. If cancer found: treatment, outcome = QALY = 0.85 (cancer case with treatment, long-term survival). Cost = $300 (colonoscopy) + $50,000 (treatment) = $50,300. If not found: routine care. Outcome = QALY = 1.0 (cancer-free). Cost = $300. Expected value (Screen) = 0.10 × (0.85 QALYs, $50,300) + 0.90 × (1.0 QALYs, $300) = [0.10×0.85 + 0.90×1.0] = 0.985 QALYs, [0.10×50,300 + 0.90×300] = $5,300 + $270 = $5,570 cost. Branch 2 (No Screen): Routine care, outcome = 1.0 QALY (assuming no undiagnosed cancer; conservative). Cost = $0. Expected value (No Screen) = 1.0 QALYs, $0 cost. ICER (Screen vs No Screen) = ($5,570 − $0) / (0.985 − 1.0) = negative denominator (screening gives fewer QALYs?). This suggests screening reduces QALYs due to misdiagnosis/over-treatment risk. Sensitivity analysis: if P(cancer_prevalence) = 20% instead of 10%, re-calculate. ICER becomes positive, screening now beneficial. Threshold analysis: at what cancer prevalence does screening break-even (equal QALY)? Solve: 0.985 becomes ≥ 1.0, find P(cancer) threshold. Probabilistic sensitivity analysis: assign distributions to all parameters (cancer prevalence, detection rate, treatment efficacy), run 1,000 simulations, report acceptability curve.
Formula
Build decision tree (short time horizon) or Markov model (long-term): nodes represent decisions (square) and chance events (circles). Branches assign probabilities. Calculate expected value: recursively multiply probabilities along each pathway, sum to root node. Expected value per arm = sum of (probability × outcome value) across all pathways. Compare expected values across decision options.
Optimal decision: Choose intervention with highest expected net benefit (QALY − cost/WTP). If ICER < willingness-to-pay (WTP) threshold (e.g., £20,000/QALY), intervention is cost-effective and recommended. Sensitivity analysis findings guide robustness. One-way sensitivity (vary one parameter): tornado chart shows which parameters most influential. If decision changes when a parameter varies ±20%, that parameter is critical to assumptions. Two-way sensitivity: co-vary two parameters (e.g., sensitivity and specificity of test) to map decision landscape. Probabilistic sensitivity: distribution of plausible ICERs reveals uncertainty range. Acceptability curve answers: at what WTP threshold is intervention most likely cost-effective? If curve shows 85% probability cost-effective at £20,000/QALY, decision is robust; if 40%, high uncertainty warrants caution or more research. Subgroup analysis: run model separately for patient subgroups (age, disease severity, comorbidity). Decisions may differ: e.g., screening recommended for high-risk patients but not low-risk. Threshold analysis: identify break-even values. Example: 'Intervention is cost-effective if treatment efficacy is ≥40%.' This guides what evidence is needed to make a confident decision.
Formula
Optimal decision: choose option with highest expected value (QALYs, cost-effectiveness ratio). Conduct sensitivity analysis: vary probabilities, costs, utilities; identify which parameters most influence decision. Threshold analysis: find break-even values (at what probability does decision switch?). Report acceptability curve: probability each decision is optimal across range of willingness-to-pay thresholds.
Common pitfalls
- Over-modeling: including too many states/branches, inflating complexity without improving decision insight. Start simple; add detail if sensitivity analysis shows it matters.
- Anchoring to single estimates: using point-estimate probabilities without uncertainty ranges. Always conduct sensitivity analysis; decisions should be robust to reasonable parameter variations.
- Ignoring heterogeneity: applying average model output to all patients, ignoring subgroup differences. Subgroup analysis essential; some patients benefit more/less.
- Failing to validate: not comparing model outputs to published epidemiological data. Always calibrate baseline scenario to external data to catch model errors.
- Misinterpreting expected value: treating expected value as guarantee. Expected value is average outcome; actual patient may experience best or worst outcome. Communicate appropriately.
Frequently asked questions
What is the difference between a decision tree and a Markov model?
Decision tree: short time horizon (e.g., immediate post-treatment period), single pathway per decision arm, each branch terminal endpoint. Used for acute decisions (test or not, treat or not) with outcomes occurring within trial timeframe. Markov model: long time horizon (chronic disease over years/decades), patients cycle through health states repeatedly, transitions defined probabilistically. Used when disease progression over time is relevant. Often combined: decision tree determines initial choice (e.g., screen yes/no); then Markov model projects long-term consequences for chosen arm.
How are utilities assigned if no direct data available?
Methods: (1) Use published utility weights (EQ-5D, SF-6D, HUI tariffs) for similar health states. (2) Conduct focused utility elicitation (time trade-off, visual analog scale, discrete choice) with patient/general population sample. (3) Expert judgment: clinicians estimate, acknowledging uncertainty. (4) Mapping: if trial collected QoL data on disease-specific scales, use published mapping algorithms to convert to utilities. Data gaps should trigger sensitivity analysis: vary utilities ±20% and see if decision changes.
Why is sensitivity analysis so important in decision models?
Model output (ICER, recommendation) depends on input parameters (probabilities, utilities, costs). Parameters are estimated with uncertainty. Sensitivity analysis tests robustness: if decision is robust to plausible parameter ranges, confidence is high; if decision flips with small changes, sensitivity is high and decision is fragile. One-way sensitivity (vary one param), two-way, and probabilistic (vary all simultaneously) show which parameters drive results. This guides where more data collection is most valuable.
What is 'threshold analysis' and when is it useful?
Threshold analysis identifies break-even values: 'At what parameter value does the decision switch?' Example: 'If treatment efficacy is ≥60%, intervention is cost-effective at £20,000/QALY threshold; below 60%, not cost-effective.' Thresholds guide interpretation: if estimated efficacy is 70% (above threshold), recommendation is robust; if 58% (near threshold), more data on efficacy needed. Thresholds also inform pricing: company can calculate, 'At what price is our drug cost-effective for payer?' Cost thresholds drive negotiation.
How does expected value differ from expected utility?
Expected value: expected cost/QALY outcome (mathematical average). Expected utility: value adjusted for individual risk preferences. Example: decision with 50% chance of QALY outcome 10 and 50% chance of 0 (expected value = 5 QALYs). Risk-neutral person values this equally to guaranteed 5 QALYs. Risk-averse person may prefer certain 3 QALY over risky gamble; their expected utility < expected value. Most decision models use expected value (risk-neutral assumption); shared decision-making models may adjust for patient risk preference.
Related methods
Sources
Pauker, S. G., & Kassirer, J. P. (1975). Therapeutic Decision Making: A Cost-Benefit Analysis. New England Journal of Medicine, 293(5), 229-234.
Open DOI referenceCopied.Briggs, A. H., Claxton, K., & Sculpher, M. J. (2006). Decision Modelling for Health Economic Evaluation. Oxford: Oxford University Press.
Open publication referenceCopied.Drummond, M. F., Sculpher, M. J., Claxton, K., Stoddart, G. L., & Torrance, G. W. (2015). Methods for the Economic Evaluation of Health Care Programmes (4th ed.). Oxford: Oxford University Press.
Open publication referenceCopied.