Point Estimate

Point Estimate: Concept Architecture

Orientation and learning roadmap

This page traces one reported number back to its target quantity, data, and estimating rule. It then compares two effect scales in a health example and explains why precision, bias, and model propagation matter alongside a point estimate.

Name the quantity represented by the number

A report might give one value for an event probability, mean cost, hazard ratio, treatment effect, or another unknown target. To interpret the number, identify its population, outcome, time interval, scale, comparator, and method of estimation.

An estimator is the rule applied to possible datasets, such as the sample mean for a population mean; the estimate is the numerical result for the dataset actually observed. The target parameter may be fixed in a frequentist model or assigned a posterior distribution in a Bayesian analysis, but one summary value never describes all uncertainty by itself. A rounded number can also hide important units or a transformation, such as whether a treatment coefficient is on the log-hazard or hazard-ratio scale.

Worked example: two valid summaries of the same counts

Suppose 18 of 100 patients assigned to treatment and 25 of 100 assigned to a comparator experience a specified adverse event within one year, with complete ascertainment. The sample risks are $\widehat p_T=18/100=0.18$ and $\widehat p_C=25/100=0.25$.

The two numbers summarise the same observed groups on different scales and cannot be compared as though $-0.07$ and $0.72$ measured the same quantity. Randomisation, if actually used and preserved, supports a causal assignment comparison under relevant follow-up assumptions; counts alone do not establish that design. Neither point estimate establishes the true effect, clinical importance, or cost effectiveness by itself.

Attach precision to the estimate

Repeated sampling would generally give different point estimates even if the same study design were followed. A standard error measures a model-based approximation to this sampling variability, while an interval estimate reports a range under an explicitly stated method and assumptions.

For independent binomial groups in the example, a simple plug-in standard error for the risk difference is

$$\widehat{\operatorname{SE}}(\widehat p_T-\widehat p_C) =\sqrt{\frac{0.18(0.82)}{100}+\frac{0.25(0.75)}{100}} \approx0.0579.$$

An approximate normal 95% interval would be $-0.07\pm1.96(0.0579)$, roughly $-0.183$ to $0.043$ in this teaching example. The interval is a consequence of the assumed independent sampling and approximation, not a guarantee about every repeated trial or a probability statement about the fixed effect. A design with clustering, weighting, missing outcomes, or covariate adjustment needs uncertainty calculations aligned to that design.

Sources of error beyond the standard error

Precision does not imply accuracy: a large biased study can report a narrow interval around a wrong target. Misclassification, selective follow-up, confounding in nonrandomised comparisons, outcome timing, and transport to a new population all require separate appraisal.

Choice of estimator also matters for small samples, rare events, skewed costs, and ratios with unstable denominators. For Bayesian work, posterior mean, median, and mode are different possible point summaries, and a credible interval conveys posterior uncertainty under the model and prior. Reporting should identify which summary was chosen rather than using the phrase the estimate as though it were unique.

Point estimates in health economic models

A base-case cost-effectiveness model often uses a set of central input estimates to calculate expected costs and quality-adjusted life-years (QALYs). That deterministic calculation is useful for tracing mechanisms and checking arithmetic but does not reflect uncertainty in every input or correlations between inputs.

For a nonlinear model, running it at each input's mean need not equal the mean output over a joint parameter distribution: $f(E[\theta])$ can differ from $E[f(\theta)]$. Probabilistic sensitivity analysis propagates the joint uncertainty and can estimate expected net benefit and decision probabilities under the model. Point estimates remain important as transparent summaries, while interval estimates and scenario analyses show what the single reported value leaves unresolved.

Reporting and common misunderstandings

State the effect measure, units, confidence or credible interval method, data source, population, and time horizon next to a numerical result. When comparing alternatives, use the same direction of subtraction and a common scale; a negative risk difference is favourable only when the event is harmful and the group order is stated.

A $p$ value is not a point estimate of effect size, and a confidence interval is not a set of equally likely parameter values. An estimate of association is not automatically causal, and two identical point estimates can carry very different uncertainties because sample size, design, or data quality differs. Rounding should preserve material decision differences without suggesting spurious precision.

Sources and further learning

The statistical sources below distinguish point and interval estimation in applied health data, while the health economic guidance addresses the limits of relying only on central model inputs. Each application still requires source-specific appraisal of bias and relevance.

• National Institute of Standards and Technology, Engineering Statistics Handbook, quantitative techniques. This explains sample-based point estimates and intervals.
• US Centers for Disease Control and Prevention, Field Epidemiology Manual, “Analyzing and Interpreting Data”. This describes effect measures and intervals in epidemiological analysis.
• ISPOR-SMDM Modeling Good Research Practices Task Force (2012), “Model Parameter Estimation and Uncertainty Analysis”, Value in Health. This explains why decision models must examine uncertainty beyond point estimates.