Reporting Model Uncertainty and Precision Without False Certainty

Why Uncertainty Needs More Than a Point Estimate

A parameter estimate is a summary of what the fitted model says, but it does not show how precisely the data determine that value. A clearance estimate of 5.2 L/h, for example, means something different if its standard error is 0.2 L/h than if its standard error is 2.0 L/h. Reporting uncertainty helps readers judge the strength and limits of a quantitative conclusion; it does not convert an estimated model into certainty about the true system.

For pharmacometric reports, the uncertainty summary should be selected to match the parameter and estimation method, then interpreted on a scale that readers can understand. A standard error is useful for describing local estimation precision. A confidence interval often gives a more direct view of the range of parameter values compatible with the data and method. Neither summary, by itself, captures every source of uncertainty, such as whether the selected structural model is appropriate.

Wald Intervals and Their Conditions

A common uncertainty summary for a parameter estimate \(\hat{\theta}\) is the approximate Wald interval. If the estimator is approximately normally distributed around the true parameter \(\theta\), and its standard error \(SE(\hat{\theta})\) is estimated reliably, an approximate two-sided 95% interval is:

For a vector of estimated parameters, an approximate covariance matrix describes the estimated variances and covariances. The standard error for parameter \(\theta_j\) is the square root of the \(j\)th diagonal element of that matrix. The Wald interval uses that standard error for the parameter in question; it does not show how strongly parameters are correlated with one another.

The normal approximation is most credible when the estimator is near an interior optimum, the likelihood is reasonably well behaved around that optimum, the parameter is identifiable from the data, and the estimated covariance matrix is reliable. It also depends on the assumed model and estimation method being appropriate for the analysis. These conditions are not guaranteed merely because software reports a standard error.

Worked Examples on Different Scales

A symmetric interval on one scale need not remain symmetric after transformation. This matters when a model estimates a coefficient on a log scale but the report expresses the result as a ratio or multiplicative effect. If \(\beta\) is a log-ratio coefficient, then the ratio is \(\exp(\beta)\). Because the exponential function is increasing, exponentiating both endpoints preserves their order and yields an interval on the ratio scale.

Weak Precision and Clinically Meaningful Interpretation

An interval can be mathematically valid yet too wide to support a useful decision. Precision is not determined by whether an interval includes a familiar reference value. It depends on how much the data constrain the parameter relative to the range of values that would matter scientifically or clinically. A narrow interval can still be uninformative if it lies in an irrelevant range; a broad interval may show that the analysis has not resolved an important question.

For a coefficient reported as a ratio, values below 1 represent a decrease and values above 1 an increase relative to the specified reference contrast. The interval should therefore be read across its full range, not summarized only as “includes 1” or “does not include 1.” The same principle applies to parameters on an additive scale: compare the interval with plausible and consequential values in the parameter’s units.

Why Transformation Produces Asymmetry

The interval transformation above follows directly from monotonicity. Let \((L,U)\) be an interval for a log-scale parameter \(\beta\), where \(L<U\). Since \(\exp(x)\) is strictly increasing, every value satisfying \(L<\beta<U\) also satisfies \(\exp(L)<\exp(\beta)<\exp(U)\). Thus the transformed interval is \((\exp(L),\exp(U))\). The endpoints remain correctly ordered, but their distances from the transformed estimate are generally unequal.

A different approximation is the delta method. For a smooth function \(g\), a first-order Taylor expansion near an estimate \(\hat{\theta}\) gives \(g(\theta)\) approximately equal to \(g(\hat{\theta})+g'(\hat{\theta})(\theta-\hat{\theta})\). This implies the approximate standard error:

For \(g(\theta)=\exp(\theta)\), this gives \(SE\{\exp(\hat{\theta})\}\approx\exp(\hat{\theta})SE(\hat{\theta})\). A symmetric Wald interval built from that transformed-scale standard error is not generally the same as exponentiating the original-scale interval. When a report needs an interval for a ratio from a log-scale estimate, transforming the endpoints is a straightforward way to preserve the original interval procedure’s coverage properties under a monotone transformation. State which approach was used.

Choosing and Reporting an Uncertainty Summary

The standard error and confidence interval answer related but different reporting needs. An SE is compact and useful for showing local precision or supporting calculations on the estimation scale. An interval more directly displays the range and direction of uncertainty, especially after a transformation. Relative standard error, often expressed as \(100\times SE(\hat{\theta})/|\hat{\theta}|\), can help compare relative precision across positive estimates, but it is unstable or unhelpful when the estimate is near zero and does not replace an interval.

A Wald interval is not the only possible interval. A likelihood-based interval can use the shape of the likelihood away from the estimate and may be more informative when that shape is asymmetric. Its validity still depends on the likelihood and method used, and the report should name the method rather than presenting every interval as interchangeable. For an interval that is highly asymmetric, constrained by a boundary, or inconsistent with the apparent likelihood shape, a symmetric Wald interval may be a poor summary.

For reproducible reporting, identify the parameter and scale, give the estimate and its units, state the uncertainty measure and method, and explain the interpretation in context. If an estimate is reported after exponentiation, say what contrast or multiplicative quantity the ratio represents. Keep adequate digits to avoid rounding obscuring the interval’s meaning, while avoiding false numerical precision.

Diagnostics, Failure Modes, and Limits

Before presenting an uncertainty summary, check whether the estimation output supports it. A covariance matrix that cannot be calculated, is non-positive definite, or has extremely large entries can indicate instability. Strong parameter correlations, flat likelihood directions, or parameters that change substantially with small analysis changes can signal weak identifiability. In such cases, a neat-looking SE should not be treated as evidence of precision.

Boundary parameters deserve particular caution. Variance components and other nonnegative quantities cannot take negative values, but a symmetric Wald interval can extend below zero. Do not silently truncate the lower endpoint to zero and report it as an ordinary Wald interval: that changes the procedure and its coverage properties. Instead, report the limitation and use an uncertainty method appropriate to the parameter and prespecified analysis, if available. If no reliable interval is available, say so rather than implying precision that the analysis did not establish.

All these summaries are conditional on the chosen model and analysis. A confidence interval for a parameter does not quantify uncertainty over alternative structural models, data-handling choices, or assumptions not represented in the interval calculation. Nor is parameter uncertainty the same as residual variability or the distribution of individual predictions. Keep those quantities distinct in the narrative.

• Common mistake: “The 95% confidence interval means there is a 95% chance that the true parameter is in this interval.” Correction: The stated coverage describes the long-run performance of the interval procedure under its assumptions; the parameter is treated as fixed.
• Common mistake: “The interval includes the null value, so the parameter is zero.” Correction: Inclusion means the interval does not exclude that value under this procedure; it does not show that the value is true or that other values are ruled out.
• Common mistake: Reporting only a relative standard error for an estimate near zero. Correction: The ratio can become unstable; provide an uncertainty summary on a suitable scale and explain its limits.
• Common mistake: Treating an interval for a population parameter as a range containing most patients. Correction: A parameter interval describes uncertainty in an estimate, not patient-to-patient variability or a prediction interval.