Forecast Horizon, Extrapolation, and Interval Widening

Why Forecast Horizon Matters

A forecast made shortly after an individual's last concentration measurement is not generally as uncertain as one made much later. Uncertainty can accumulate as uncertain individual parameters affect the predicted trajectory; future residual or process variation can add further spread. But “farther ahead” does not guarantee a wider interval on the measurement scale. A declining concentration can approach zero, so its absolute interval width may shrink even while uncertainty about its rate of decline becomes more consequential.

The key is to specify what is being forecast, what information is conditioned upon, and which model assumptions carry the forecast beyond the observed data. The uncertainty in an individual forecast is conditional on the individual's observations and on the specified future dosing and covariate history. A long-horizon prediction is therefore not simply a short-horizon prediction with a larger error bar.

Definition: Let \(t_{\mathrm{last}}\) be the time of an individual's last observation and \(t^*\) a future or target time. The forecast horizon is \(h=t^*-t_{\mathrm{last}}\). A target within the time range of observations is an interpolation target; a target outside that range is an extrapolation target. A forecast distribution describes uncertainty about the specified future quantity, conditional on the available individual data, the model, and the assumed future regimen and covariates.

Interpolation and extrapolation are about the target's location relative to observed times, not whether it occurs after the last observation. A target after the last observation but before an earlier observation's time range ends is unusual for a single individual's chronological profile; more generally, interpolation means predicting within supported time or covariate ranges, while extrapolation means going beyond them. A target after the last observation is often extrapolation in time, but can still be supported by a well-established mechanistic model and known future inputs. Conversely, a forecast at a familiar time can extrapolate in dose, covariates, or another important dimension.

How Uncertainty Changes With Horizon

For an individual \(i\), a future observation \(Y_i(t^*)\) has a predictive distribution that averages over uncertainty in that individual's parameters and, where appropriate, future residual or process variation. In schematic form, if \(\theta_i\) denotes the individual's parameters and \(D_i\) the observed data, then

p\{Y_i(t^*)\mid D_i\} = \int p\{Y_i(t^*)\mid \theta_i\}\, p(\theta_i\mid D_i)\,d\theta_i .

This is not the same as evaluating the model once at a single point estimate of \(\theta_i\). Uncertainty in \(\theta_i\) matters because different plausible parameter values can imply increasingly different future trajectories. The future observation may also vary around each trajectory because of residual error or process variability. The introductory material on posterior predictive distributions establishes this conditioning perspective; here the focus is how its consequences depend on horizon and extrapolation.

A short derivation makes the horizon effect explicit in a simple case. Suppose log concentration after the last observation is \(Z(h)=a-kh+\epsilon\), where \(a\) is the log concentration at the last observation, \(k\) is an uncertain elimination rate, and \(\epsilon\) is independent future residual error with variance \(\sigma^2\). Conditional on the observed data, let the variances of \(a\) and \(k\) be \(V_a\) and \(V_k\), and their covariance be \(C_{ak}\). Then

\begin{aligned} \operatorname{Var}\{Z(h)\mid D_i\} &=\operatorname{Var}(a-kh+\epsilon\mid D_i)\\ &=V_a+h^2V_k-2hC_{ak}+\sigma^2. \end{aligned}

The \(h^2V_k\) term shows why uncertainty in the rate can have an increasing effect on log concentration as the horizon grows. The covariance term can either reduce or increase the variance, depending on its sign. Even in this simple model, the variance need not increase at every horizon. Nonlinear models add further complications: the predicted quantity may be curved in the parameters, model discrepancies may accumulate, and the scale on which uncertainty is summarized matters.

Worked Example: A Log-Linear Forecast at Two Horizons

Hypothetical setup. An individual's last concentration is 10 mg/L. Conditional on the observed data, the elimination rate is \(k\sim N(0.10,0.02^2)\) per hour. For a future observation, suppose log concentration follows \(\log C(h)=\log(10)-kh+\epsilon\), where \(\epsilon\sim N(0,0.10^2)\) independently of \(k\). This assumes a constant elimination rate, no intervening dose or covariate change, and a stable log-scale residual variance.

Because a linear combination of independent normal variables is normal, the forecast log concentration has mean \(\log(10)-0.10h\) and standard deviation \(\sqrt{(0.02h)^2+(0.10)^2}\). At \(h=4\) hours, the standard deviation is \(\sqrt{0.0064+0.0100}=0.1281\). At \(h=12\) hours, it is \(\sqrt{0.0576+0.0100}=0.2600\). The increasing rate uncertainty contribution is visible: its standard deviation contribution is \(0.02h\).

At 12 hours, the log-scale forecast mean is \(\log(10)-1.20=1.1026\). A central 90% normal interval uses \(1.645\) standard deviations on each side:

1.1026\pm1.645(0.2600) = (0.6749,\ 1.5303).

Exponentiating the endpoints gives an approximate 90% prediction interval of \((1.964,\ 4.619)\) mg/L, or \((1.96,\ 4.62)\) mg/L to three significant figures. The endpoints are asymmetric around the forecast on the concentration scale because exponentiation is nonlinear. This is an interval for a future concentration under the stated model and conditioning information, not an interval for the elimination rate.

Extrapolation and the Assumptions Behind a Forecast

An extrapolated prediction relies on the model remaining useful where the data provide less direct support. For an NLME forecast, that reliance can involve several distinct assumptions:

Structural form. The chosen structural model continues to describe the individual's concentration-time relationship. A one-compartment decline, for example, may miss distribution phases or time-varying clearance.
Parameter behavior. The individual's parameters are treated as stable over the forecast unless the model explicitly allows changes. Disease progression, organ-function changes, or adherence changes can violate this assumption.
Future inputs. The dose history, dosing times, and relevant covariates are known or correctly specified. A forecast under a planned regimen is conditional on that plan being followed.
Error behavior. The residual or process-variation model is assumed to remain appropriate at the target time. A constant residual variance may not be credible across a wide concentration range.
Individual information. The conditional distribution of individual parameters is adequate. Sparse or weakly informative data can leave different parameter combinations plausible, producing different distant trajectories.

A model can produce a numerically precise interval while these assumptions are doubtful. The interval represents the uncertainty included in the specified model; it does not automatically include uncertainty about an incorrect structural form, an unrecorded dose, or a new physiological process.

Worked Example: Interpolation Versus Extrapolation in a Decline Model

Hypothetical setup. Suppose concentrations of 8 mg/L at hour 2 and 4 mg/L at hour 10 are described by a one-compartment exponential decline with no intervening dose: \(C(t)=C(2)\exp\{-k(t-2)\}\). Assume the two values are treated as exact for this illustration, so the fitted rate is

\hat{k}=\frac{\log(8/4)}{10-2} =\frac{\log 2}{8} =0.08664\ \text{hour}^{-1}.

At hour 8, which lies between the two measured times, the model gives \(C(8)=8\exp\{-0.08664(6)\}=8(2^{-0.75})\approx4.757\) mg/L. This is interpolation in time. At hour 18, beyond the measured range, it gives \(C(18)=8\exp\{-0.08664(16)\}=8(2^{-2})=2.000\) mg/L. This is extrapolation.

The arithmetic does not make the hour-18 forecast as well supported as the hour-8 value. The extrapolation assumes that the same exponential decline continues for another eight hours, that no dose is given, and that the rate remains constant. The example also has no parameter or residual uncertainty because the two values were treated as exact; a real forecast would need to account for uncertainty in the fitted individual parameters and future observations. A perfect fit to two points does not verify the model's long-term behavior.

Interval Width, Scale, and Diagnostics

There is no universal rule that an individual prediction interval must widen monotonically with \(h\). The answer depends on the target scale, the model, and the source of uncertainty. For a declining concentration, uncertainty in a rate may increase relative uncertainty while absolute concentrations and absolute interval widths shrink. Conversely, growth, accumulation, or stochastic processes can make absolute uncertainty grow. On a log scale, increasing spread corresponds to multiplicative rather than additive differences on the concentration scale.

Worked Example: A Nonlinear Decline Can Have Shrinking Absolute Uncertainty

Hypothetical setup. Suppose the forecast is \(Y(h)=A\exp(-kh)+\epsilon\), with \(A\mid D_i\) approximately normal with mean 8 mg/L and standard deviation 1 mg/L, \(k\mid D_i\) approximately normal with mean \(0.15\) per hour and standard deviation \(0.02\) per hour, and independent additive future error \(\epsilon\sim N(0,0.30^2)\) mg/L. Assume \(A\) and \(k\) are conditionally independent. The normal approximations are plausible here because the parameter standard deviations are small relative to their means, but the calculation below is a first-order approximation, not an exact distribution.

For \(f(A,k)=A\exp(-kh)\), the gradient is \((\exp(-kh),-Ah\exp(-kh))\). The first-order variance approximation adds the error variance to the parameter-induced variance:

\operatorname{Var}\{Y(h)\mid D_i\} \approx e^{-2kh}\{V_A+A^2h^2V_k\}+\sigma_\epsilon^2.

At \(h=8\) hours, the forecast mean is \(8e^{-1.2}=2.410\) mg/L. Substituting \(V_A=1\), \(V_k=0.02^2\), and \(\sigma_\epsilon^2=0.09\) gives

\operatorname{Var}\{Y(8)\mid D_i\} \approx e^{-2.4}\{1+8^2(8^2)(0.02^2)\}+0.09 =e^{-2.4}(2.6384)+0.09 \approx0.3294.

Thus the approximate standard deviation is \(0.5739\) mg/L, and an approximate central 90% interval is \(2.410\pm1.645(0.5739)=(1.466,\ 3.354)\) mg/L. At \(h=0\), the mean is 8 mg/L and the standard deviation is \(\sqrt{1+0.09}=1.044\) mg/L. In this example the absolute standard deviation decreases between these horizons because the declining trajectory scales down parameter uncertainty. Yet the relative contribution from \(A\) and \(k\) grows: its approximate coefficient of variation is \(1/8=12.5\%\) at \(h=0\), compared with \(\sqrt{1/64+8^2(0.02^2)}=20.3\%\) at \(h=8\). Absolute width and relative uncertainty answer different questions.

Useful diagnostics therefore examine more than one width at one time. Plot forecast medians and interval endpoints across the intended horizons, preferably on both the measurement and a scientifically meaningful transformed scale. Inspect simulated individual trajectories: implausible negative concentrations, abrupt changes, or unrealistic persistence can reveal an unsuitable error or structural model. Check whether the forecast depends strongly on a small number of poorly identified parameters or on extrapolated covariate values. Compare model-based trajectories with observed profiles over the supported range, while recognizing that good fit there cannot prove distant behavior.

If an interval narrows over time, do not label it automatically as a calculation error; establish whether the mean trajectory approaches a boundary such as zero and whether relative uncertainty is increasing. If it widens sharply, determine whether that reflects parameter sensitivity, process noise, a dose change, or a numerical problem. These checks explain the behavior of the forecast; they do not establish that its long-horizon coverage is adequate.

Alternatives, Limitations, and a Common Pitfall

A plug-in forecast substitutes point estimates for individual parameters and follows one trajectory. It can be useful as a concise central prediction but omits parameter uncertainty unless that uncertainty is added separately. A posterior predictive forecast averages over plausible individual parameters and future variation, giving a distribution of possible future observations. For a long extrapolation, a useful complement is a scenario comparison: calculate forecasts under scientifically plausible alternative structural forms, parameter assumptions, or future regimens. Such comparisons expose sensitivity to assumptions, but do not turn an uncertain assumption into a known fact.

The posterior predictive approach is only as credible as its conditioning data, model, and future-input assumptions. Long-horizon predictions may be especially sensitive to weakly identified rates, correlated parameters, model misspecification, and changes in an individual's physiology or adherence. Numerical integration or simulation can approximate a complicated forecast distribution, but computational accuracy alone cannot repair unsupported extrapolation.

Common pitfall: “The interval is wider at a longer horizon, so it must be more realistic”; or, conversely, “the interval is narrower, so uncertainty has disappeared.” Neither conclusion follows from width alone. Check the scale, the modeled sources of uncertainty, the trajectory's behavior, and the assumptions required beyond the data. A narrow interval can omit important model uncertainty, while a shrinking absolute interval can coexist with growing relative uncertainty.

The practical question is not simply whether an interval widens. It is whether the forecast distribution adequately reflects uncertainty relevant to the decision, and whether its assumptions remain defensible over the full horizon. That requires explicit horizon-by-horizon interpretation and caution when the prediction leaves the range directly supported by an individual's observations.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If uncertainty in an elimination rate increases the log-scale variance in proportion to \(h^2\), does that guarantee an increasing absolute interval width for concentration?
A target time is between two measured times, but its dose is twice the doses observed in the profile. Is it interpolation or extrapolation in every relevant sense? Explain.
Why can a plug-in forecast understate uncertainty in a distant individual prediction?
Name two assumptions that should be checked before relying on a long-horizon concentration forecast.

Solutions. (1) No. Exponentiation and a declining mean can make absolute concentration intervals shrink even as log-scale or relative uncertainty grows. (2) It is interpolation in time but extrapolation in dose, so the model's behavior at the higher dose is not directly supported by those observations. (3) A plug-in forecast fixes uncertain individual parameters at point estimates and so does not propagate their uncertainty into the future distribution. (4) Examples include constant structural form and stable individual parameters; correct future doses and covariates; and an appropriate residual or process-variation model. State which assumptions matter for the particular forecast.