Why Parameter Precision and Prediction Precision Differ
A QSP model can fit observed responses well even when several parameter values remain uncertain. This is not necessarily a contradiction. Different parameter combinations can compensate for one another, producing nearly the same output in the observed setting. The parameters may therefore be poorly estimated individually while a particular model prediction is well determined.
The distinction matters when interpreting a calibrated model. A narrow confidence interval for a predicted concentration at a specified time does not imply that every underlying parameter is precisely estimated. Conversely, a precise estimate of one parameter does not guarantee that predictions far from the observed conditions are precise.
Suppose a procedure yields a joint \(95\%\) confidence region \(C\) for \(\theta\). For a scalar prediction \(g(\theta)\), the region can be mapped through the model:
The image \(\mathcal{G}\) is a confidence set for the model output; its infimum and supremum define an enclosing interval, whose endpoints may be unattained or infinite. If the joint region contains the true parameter vector in at least \(95\%\) of repeated applications under the stated assumptions, its mapped set contains the corresponding true output at least as often. This set may be conservative, and its width depends on the particular output being examined.
A Compensatory Exponential Model
Consider the hypothetical concentration model \(C(t)=A\exp(-kt)\), where \(A\) is the concentration at time zero in mg/L, \(k\) is the elimination rate constant in h\(^{-1}\), and \(t\) is time in hours. The model is assumed to describe the mean concentration adequately over the times considered. Its logarithm is \(\log C(t)=\log A-kt\). At a fixed time, an increase in \(A\) can be offset by an increase in \(k\), leaving the predicted concentration nearly unchanged.
Worked Example: Different Parameters, Nearly the Same Prediction
Suppose two plausible fitted parameter pairs from a hypothetical analysis are \((A,k)=(10.5\ \text{mg/L},0.4\ \text{h}^{-1})\) and \((9.5\ \text{mg/L},0.3\ \text{h}^{-1})\). Compare their predicted concentrations at 1 hour and 4 hours.
At 1 hour, the first pair gives:
The second pair gives \(C_2(1)=9.5\exp(-0.3)=7.04\ \text{mg/L}\), to three significant figures. The unrounded values are approximately 7.03836 and 7.03777 mg/L. Thus the two parameter pairs differ, but their 1-hour predictions are nearly identical.
At 4 hours, however, \(C_1(4)=10.5\exp(-1.6)\approx2.120\ \text{mg/L}\), whereas \(C_2(4)=9.5\exp(-1.2)\approx2.861\ \text{mg/L}\). The compensatory relationship that preserves the prediction at 1 hour does not preserve it at 4 hours. A prediction can be estimable at one time and poorly estimable at another.
This example uses two parameter pairs for illustration; two points alone do not establish a confidence region or its coverage. In a real analysis, the relevant object is the full joint region from the specified fitting and uncertainty procedure. Marginal intervals for \(A\) and \(k\) do not retain the dependence between them. Combining their separate endpoints as if every pairing were plausible can therefore give misleading predictions.
Projecting a Joint Region Through the Model
The following hypothetical region shows how a prediction can remain constrained while its parameters vary. Assume that a specified procedure has produced the joint \(95\%\) region \(C=\{(A,k):0.2\leq k\leq0.8,\ |\log A-k-\log(7.03836)|\leq0.02\}\), with \(A\) measured in mg/L and the numerical value of \(k\) in h\(^{-1}\). The region is deliberately shaped to express compensation: as \(k\) changes, \(A\) changes so that \(\log A-k\) remains within a narrow range. The coverage statement is conditional on the procedure and the model assumptions; the numerical region is hypothetical.
Worked Example: A Narrow Prediction From a Broad Parameter Region
For \(t=1\) hour, the model output is \(C(1)=A\exp(-k)\). Taking logarithms gives \(\log C(1)=\log A-k\), which is constrained by the region to lie within 0.02 of \(\log(7.03836)\). Therefore:
so the projected range is approximately 6.899 to 7.181 mg/L. Individual parameter values vary substantially: for example, near \(k=0.2\), the central relation gives \(A\approx7.03836\exp(0.2)=8.597\) mg/L, whereas near \(k=0.8\), it gives \(A\approx15.66\) mg/L. The 1-hour prediction remains relatively constrained because the relevant combination \(\log A-k\) is constrained.
At 4 hours, \(\log C(4)=(\log A-k)-3k\). The smallest output over the stated region occurs at the smallest allowed \(\log A-k\) and largest \(k\); the largest occurs at the largest allowed \(\log A-k\) and smallest \(k\). Thus:
which is approximately 0.626 to 3.941 mg/L. The much wider range illustrates that prediction precision is specific to the output and conditions, not a general property of the fitted model.
The justification is direct: if the true \(\theta\) lies in \(C\), then its output \(g(\theta)\) must lie in the image \(\mathcal{G}\). Consequently, a confidence region for all parameters can support a confidence set for a prediction without requiring each parameter to have a narrow interval. The reverse implication does not hold: a narrow image for one output does not confine parameters that the output barely depends on.
Mean-Response Uncertainty and Future Observations
A predicted mean response and a future observation are different quantities. The model output \(g(\theta)\) describes the mean under the model. A future observed value may also include observation noise. A confidence interval for the mean response accounts for uncertainty about that mean; a prediction interval for one future observation must account for both parameter uncertainty and observation variability.
For a simple approximate calculation, suppose the predicted mean is approximately normal with standard error \(SE_{\text{mean}}\), and independent future observation error is normal with standard deviation \(\sigma\). Under these assumptions, the future observation has approximate variance \(SE_{\text{mean}}^2+\sigma^2\). This addition is justified by independence and the variance-addition rule; it is not a universal formula for every nonlinear model or error structure.
Worked Example: Adding Future Observation Noise
At a specified time in a hypothetical QSP model, suppose the estimated mean concentration is 7.038 mg/L, its standard error from parameter uncertainty is 0.350 mg/L, and future observation errors are independent normal errors with standard deviation 0.600 mg/L. These assumptions permit a normal approximation. The approximate standard deviation for a future observation is:
Using 1.96 as the normal \(95\%\) multiplier, an approximate prediction interval is:
By comparison, an approximate interval for the mean response using only its stated standard error is \(7.038\pm1.96(0.350)\), or approximately [6.352, 7.724] mg/L. The prediction interval is wider because it includes future observation noise. Neither interval is exact unless its normal and variance assumptions are suitable.
Prediction Uncertainty Can Grow Outside the Data Range
A small uncertainty range at measured times does not guarantee precise extrapolation. The model may be constrained by the observed data over a limited time window while alternative parameter values imply sharply different behavior later. Nonlinear transformations can also make the output range asymmetric, even when uncertainty in a parameter is roughly symmetric. A point prediction alone hides both effects.
Worked Example: Uncertainty in a Late-Time Concentration
For a separate hypothetical calculation, take the point estimates \(A=9.5\) mg/L and \(k=1.0\) h\(^{-1}\). At 4 hours, the plug-in mean prediction is:
The point estimate is not a measure of its own uncertainty. To illustrate the effect of parameter uncertainty, suppose a hypothetical joint confidence region has minimum and maximum late-time predictions of 0.1105 and 0.2732 mg/L, respectively. The resulting mapped range is [0.1105, 0.2732] mg/L; it is not a future-observation prediction interval because no observation-error component has been added. The endpoints would need to be obtained by optimizing the model output over the actual joint region, not by assuming that parameter endpoints can be paired arbitrarily.
The central estimate 0.1740 mg/L is within that range, but reporting only the central estimate would conceal meaningful uncertainty. If the prediction is used to assess a clinically relevant concentration threshold, the range may change the decision even though the fitted curve appears well behaved at earlier times.
Diagnostics, Alternatives, and Limitations
Several checks help distinguish genuine prediction precision from an artifact of reporting or model assumptions. Examine the joint parameter confidence region rather than only separate parameter intervals. Map the region through each output that matters, including relevant times, doses, or scenarios. Check whether the prediction range becomes wide near or beyond the observed conditions. Compare the predicted mean range with a future-observation interval when the decision concerns an individual future measurement.
Residual patterns and systematic disagreement between observations and model predictions can indicate that the assumed model or error distribution is inadequate. A narrow prediction range conditional on a misspecified model is not reliable evidence of real-world predictive precision. Likewise, numerical optimization over the parameter region must locate the true extrema closely enough; a sparse grid can miss a maximum or minimum, especially for a nonlinear output.
The main alternatives answer different questions. A marginal confidence interval for one parameter addresses that parameter alone and can hide compensatory dependence. A joint parameter region preserves that dependence and can be projected to a model output. A pointwise confidence interval for the mean concerns the mean at one specified condition; a prediction interval concerns a future observation and includes observation variability. A plug-in prediction evaluates the model only at the fitted parameter estimate and, by itself, describes neither kind of uncertainty interval.
These distinctions are conditional on the chosen model, observation design, error assumptions, and confidence-region procedure. A projected confidence set covers the model-defined output when the parameter region has its stated coverage and the model is adequate. It does not include uncertainty about an incorrect structural model, unrepresented biological variability, or a change in experimental conditions unless those uncertainties are explicitly incorporated.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- Two parameter pairs give nearly the same concentration at 1 hour but different concentrations at 4 hours. What does this show about prediction estimability?
- Why can separate confidence intervals for two parameters give a misleading range for a model prediction?
- A confidence interval for a mean response has a width of 2 mg/L. Should a future observation interval generally be expected to be narrower, equal, or wider? State the assumptions behind your answer.
- For a joint parameter region \(C\), explain why mapping its values through \(g(\theta)\) yields a confidence set for the corresponding model output.
Solutions. (1) Prediction estimability depends on the particular output and time: compensation can preserve one prediction without preserving another. (2) The marginal intervals discard parameter dependence; endpoint combinations may be implausible or absent from the joint region. (3) It is generally wider when independent observation noise with positive variance is added to parameter uncertainty; this conclusion uses the stated variance and error assumptions. (4) Whenever the true parameter lies in \(C\), its model output must lie in the mapped set \(\{g(\theta):\theta\in C\}\), so the mapped set inherits at least the region's coverage under the model assumptions.