Decision Model

A decision model is a mathematical framework that structures a clinical, policy, or resource allocation decision by mapping out the possible pathways, their associated probabilities, costs, and outcomes. Decision models synthesize evidence from multiple sources, including clinical trials, observational studies, and expert opinion, to estimate the expected consequences of different choices under conditions of uncertainty. They are widely used in health technology assessment, health economics, and clinical guideline development to compare alternative strategies when direct head to head evidence is unavailable.

Healthcare decisions are complex, involving trade offs between costs and benefits, competing risks, time horizons that extend beyond the duration of clinical trials, and uncertainty about key parameters. Direct experimental comparison of every possible treatment strategy is not feasible. Decision models fill this evidence gap by integrating available data into a coherent framework that projects long term outcomes, estimates cost effectiveness, and identifies the conditions under which one strategy is preferred over another. Health technology assessment bodies such as NICE in England, Canada's Drug Agency (formerly CADTH), and ICER in the United States rely on decision models to inform coverage and reimbursement decisions that affect millions of patients.

Building a decision model begins with defining the decision problem: the population, the interventions being compared, the relevant outcomes, and the time horizon. The modeler then selects an appropriate model structure. Common structures include decision trees (for decisions with a short time horizon and a limited number of health states), Markov models (for chronic conditions where patients transition between health states over time), microsimulation models (which track individual patients through a series of events), and discrete event simulations (which model events occurring at irregular intervals).

The model is populated with input parameters drawn from the best available evidence. Transition probabilities (the likelihood of moving from one health state to another) come from clinical trials, cohort studies, or registry data. Costs are obtained from published cost studies, claims databases, or institutional accounting systems. Health state utilities (quality of life weights) are derived from patient surveys using instruments such as the EQ 5D. When evidence is lacking, expert elicitation may provide estimates.

Sensitivity analysis is integral to decision modeling. One way sensitivity analysis varies individual parameters across plausible ranges to determine which inputs most influence the model's conclusions. Probabilistic sensitivity analysis (PSA) simultaneously varies all uncertain parameters according to their probability distributions across thousands of simulations, generating a distribution of outcomes that reflects parameter uncertainty. Results are often presented as cost effectiveness acceptability curves, which show the probability that each strategy is cost effective across a range of willingness to pay thresholds.

Model validation involves verifying the mathematical accuracy of the model (internal validation), comparing model predictions against independent observed data (external validation), and assessing whether the model produces results consistent with other published models of the same decision (cross validation). Transparent reporting follows guidelines such as the CHEERS checklist (Consolidated Health Economic Evaluation Reporting Standards).

Decision models are appropriate when comparing strategies that cannot be directly compared in a single trial, when the decision requires integrating evidence from multiple sources, when the time horizon of interest extends beyond available trial data, or when cost effectiveness considerations are relevant. They are standard in health technology assessments and are increasingly required by reimbursement authorities. Decision models are not appropriate when a single high quality trial directly answers the decision question, or when the model would require so many assumptions that the results are more speculative than informative.

A common error is building a model that is more complex than the available evidence can support, creating an illusion of precision. Another mistake is failing to conduct adequate sensitivity analysis, presenting point estimates without acknowledging the uncertainty surrounding key parameters. Structural assumptions (such as the choice between a Markov model and a microsimulation) can significantly influence results, and researchers sometimes make these choices for convenience rather than for conceptual fit with the disease process. Using outdated or geographically inappropriate input data (such as costs from one country applied to another) without adjustment undermines the model's relevance. Finally, models that lack transparency in their reporting, omitting details about structure, assumptions, and data sources, cannot be evaluated or replicated by others.

In a hypothetical example, a health technology assessment team builds a Markov model to compare three treatment strategies for newly diagnosed type 2 diabetes: metformin alone, metformin plus an SGLT2 inhibitor, and metformin plus a GLP 1 receptor agonist. The model tracks patients through health states (stable diabetes, microvascular complications, cardiovascular event, post cardiovascular event, death) over a 40 year horizon using one year cycles. Input parameters are drawn from 12 clinical trials, two national claims databases, and published utility studies. Probabilistic sensitivity analysis over 10,000 iterations shows that the SGLT2 inhibitor combination has a 68% probability of being cost effective at a willingness to pay threshold of $100,000 per quality adjusted life year.