Parameter Estimation

Parameter Estimation: Concept Architecture

Orientation and learning roadmap

This page follows a model input from its target definition and evidence through an estimator, uncertainty assessment, and use in a health-economic model. It works through a binomial event probability, then explains why selection, dependence, and extrapolation matter as much as the numerical estimate.

Define the quantity before choosing a method

An analyst should specify whether the input is a one-year event probability, a mean resource cost, a treatment contrast, a regression coefficient, or another precisely defined quantity. The population, outcome, time horizon, treatment conditions, and scale must match the decision model or be linked through an explicitly justified transformation.

Observed data do not select the estimand automatically: a 12-month cumulative event risk is not an instantaneous hazard, and a study sample's average cost may not represent another jurisdiction's unit cost. Direct evidence, evidence synthesis, calibration, or elicitation may each contribute to an estimate when appropriate, but their methods and limitations should be recorded. Define how any missingness, censoring, clustering, and confounding are handled before interpreting the result.

From evidence to an estimate

Estimation uses a stated statistical or mathematical rule to learn unknown model quantities from data. Maximum likelihood, Bayesian inference, method of moments, regression, and evidence synthesis answer related but different questions and may require distinct assumptions about sampling and prior information.

An estimator's point value is only one product of the analysis; its uncertainty and dependence with other estimated quantities also matter. The model input may require further conversion, such as transforming an estimated hazard into a cycle probability under stated time assumptions. Avoid substituting the observed variability of individuals for the uncertainty in a population parameter.

Worked example: a one-year event probability

Suppose 24 of 80 independent patients experience a clearly defined event within one year and follow-up is complete. Under a common-probability binomial model, the likelihood for the probability $p$ is proportional to $p^{24}(1-p)^{56}$. The maximum-likelihood estimate follows from the observed event count.

$$\widehat p=\frac{24}{80}=0.30.$$

An illustrative plug-in standard error is $\sqrt{\widehat p(1-\widehat p)/80}=\sqrt{0.21/80}\approx0.0512$ under independent binomial sampling. That standard error describes uncertainty in the estimate of a one-year probability, while the binary outcomes of individual future patients remain variable even if $p$ were known exactly. A confidence interval requires a stated construction; the plug-in standard error alone should not be treated as an interval or as a causal treatment effect.

Estimate uncertainty and preserve dependence

Confidence intervals, covariance matrices, bootstrap distributions, and Bayesian posterior distributions can convey different aspects of estimation uncertainty. The chosen method should account for the sampling design, the amount of data, censoring, and any dependence among observations or among jointly estimated parameters.

For example, baseline risk and a treatment-effect coefficient estimated from one study may be correlated, while state-transition probabilities estimated from the same multinomial counts must satisfy a sum constraint. Drawing each estimate independently in probabilistic sensitivity analysis can create impossible or implausible combinations. Parameter distributions should reflect the evidence and chosen estimation model, rather than being attached to point estimates by arbitrary convention.

Bias, calibration, and external validity

Even a small standard error cannot repair an estimate derived from the wrong population, outcome definition, or measurement process. Selection bias, unmeasured confounding, incomplete follow-up, coding changes, and incompatible interventions can shift the estimate away from the decision-relevant quantity.

Calibration fits model inputs so that model predictions agree with selected aggregate targets, but several parameter combinations may reproduce the same target and remain weakly identified. External validation against evidence not used for fitting tests a different question from calibration fit. Authors should record source selection, transformations, fitted target data, uncertainty, and whether estimates were borrowed or extrapolated beyond observed conditions.

Use in health economic decisions

Event probabilities, costs, utilities, treatment effects, and survival parameters enter cost-effectiveness and budget-impact models in different places. The inputs should carry their units, timing, target population, and uncertainty into the calculation of costs and quality-adjusted life-years (QALYs).

An event probability from the worked example can inform a one-year model cycle for a comparable population, provided event definitions and competing events align. It cannot automatically be reused for a monthly cycle, a population with different baseline risk, or an intervention arm without additional modelling. Sensitivity analysis should expose which estimates materially affect the decision and which evidence gaps deserve further study.

Sources and further learning

The sources below provide the mathematical example, methodological guidance on parameter uncertainty, and decision-model context. They support transparent estimation and reporting, while a specific clinical or economic application still needs its own relevant empirical evidence.

• National Institute of Standards and Technology, Engineering Statistics Handbook, binomial distribution. This gives the maximum-likelihood estimator for a binomial probability.
• National Institute of Standards and Technology, Engineering Statistics Handbook, maximum likelihood. This explains likelihood-based estimation.
• ISPOR-SMDM Modeling Good Research Practices Task Force (2012), “Model Parameter Estimation and Uncertainty Analysis”, Value in Health. This addresses evidence, uncertainty, and decision-model inputs.
• NICE, NICE technology appraisal and highly specialised technologies guidance: the manual, economic evaluation. This sets expectations for input evidence and uncertainty in health technology evaluation.