Parameter Distribution

Parameter Distribution: Concept Architecture

Orientation and learning roadmap

This page starts with the uncertainty represented by a distribution over a model input, then shows how evidence, bounds, and dependence guide its specification. It works through a treatment-event probability, follows draws into a health-economic model, and distinguishes parameter uncertainty from patient heterogeneity and simulation noise.

Specify the uncertain input

An economic model may require a probability of an event, a mean cost, a treatment effect, or a baseline hazard that is estimated imperfectly from evidence. The analyst must identify the input's estimand, scale, population, time period, and source data before selecting a distribution to represent uncertainty about its value.

A distribution for an uncertain population event probability is different from the Bernoulli distribution of an individual patient's event conditional on that probability. Likewise, uncertainty about a mean cost is not identical to the distribution of individual patient costs. A probabilistic sensitivity analysis (PSA) generally draws uncertain input values and propagates them through the model to examine uncertainty in outputs and decisions.

Choose support and evidence together

The possible values of the input constrain plausible distribution families, but support alone does not determine the correct shape or spread. Estimation data, sampling model, prior assumptions where applicable, joint covariance, and clinical knowledge should justify the actual parameters used.

When the evidence consists of individual-level data or model regression coefficients, sampling from a fitted joint model or bootstrap may be more faithful than assigning separate textbook distributions to each point estimate. A structural choice, such as whether a health state exists in the model, is a distinct uncertainty unless explicitly represented by a defensible model-averaging scheme. Reporting should specify distribution family, numerical parameters, evidence source, and any truncation or transformation.

Worked example: uncertainty in an event probability

Suppose 30 events occur among 100 independent patients observed for a defined interval under a binomial sampling model. Under an explicitly chosen uniform $\operatorname{Beta}(1,1)$ prior for the common event probability $p$, the posterior distribution is $\operatorname{Beta}(31,71)$; the prior is an assumption rather than a fact supplied by the data.

$$p\mid\text{data}\sim\operatorname{Beta}(1+30,1+100-30) =\operatorname{Beta}(31,71), \qquad E(p\mid\text{data})=\frac{31}{102}\approx0.304.$$

A PSA draw may select a plausible value of $p$ from this posterior and pass it into a one-year transition calculation that uses the same event definition. It should not draw a new $p$ for every individual in the same simulated population unless the model explicitly represents that additional between-person variation. An alternative prior or different effective sample size changes the uncertainty distribution, and the simple binomial model may be inadequate when outcomes are clustered or heterogeneous.

Propagate dependence and transformations

Parameters estimated from the same trial, regression, or clinical process can be correlated. If treatment and comparator event probabilities share a baseline risk, drawing them independently can discard a meaningful relationship even when each marginal distribution looks reasonable.

One defensible approach draws a vector of coefficients from its estimated joint distribution and then transforms it into probabilities, hazards, or costs inside each PSA iteration. Probabilities for mutually exclusive transitions must sum to at most one or follow a joint construction that guarantees valid totals. A hazard ratio sampled on its log scale should be exponentiated before application, and uncertainty in the baseline survival curve may need to be drawn with the relative effect rather than held fixed without justification.

Check the model and communicate results

Plotting input draws can reveal impossible values, implausible tails, unintended point masses, or broken correlations. Compare implied intervals and moments with source evidence, and check that transformed outputs preserve the intended clinical and mathematical constraints.

Parameter distributions describe uncertainty conditional on the chosen evidence and statistical assumptions; they do not eliminate structural uncertainty, missing-data bias, or external-validity concerns. Monte Carlo error is the additional numerical fluctuation from using finitely many draws, which can be reduced by more simulations without narrowing the underlying parameter distributions. A cost-effectiveness acceptability curve summarises decision uncertainty after propagation, not the distribution of any single input.

Health economic interpretation

In a model of costs and quality-adjusted life-years (QALYs), the joint input draws generate a distribution of incremental costs, incremental effects, and net benefit across plausible parameter sets. An expected net benefit can be estimated by averaging over draws, while the spread helps characterise decision uncertainty under the model's assumptions.

Choosing distributions arbitrarily can make a PSA look rigorous while distorting the evidence base. The aim is to encode what is known about each input and its dependence with other inputs, then show how that uncertainty affects the decision. Distributional choices with material influence should be tested in sensitivity analyses and explained to readers.

Sources and further learning

The linked methodological and institutional sources support fitting uncertainty distributions to evidence and preserving dependence between inputs. They also distinguish the statistical representation of uncertainty from the simulation machinery used to propagate it.

• Briggs (2005), “Probabilistic Analysis of Cost-Effectiveness Models: Statistical Representation of Parameter Uncertainty”, Value in Health. This examines evidence-based selection of distributions for model parameters.
• ISPOR-SMDM Modeling Good Research Practices Task Force (2012), “Model Parameter Estimation and Uncertainty Analysis”, Value in Health. This gives recommendations for probabilistic modelling and disclosure.
• NICE, NICE technology appraisal and highly specialised technologies guidance: the manual, economic evaluation. This requires justified distributions and consideration of correlation in PSA.