Delayed Peaks and Lag-Time Misspecification

Why Peak Timing Can Reveal a Misspecified Input

A concentration profile can have a plausible terminal decline and still rise later than the fitted model predicts. If the model assumes absorption begins immediately, it may predict an early peak even when observed concentrations remain low initially and reach their maximum later. That timing discrepancy can indicate that the input begins after a delay, but it does not establish that explanation by itself: slow absorption, sparse sampling, and other sources of model misspecification can also affect the apparent peak.

The question is therefore not simply whether a lagged model fits better. It is whether the data contain enough information to distinguish a genuine delay before systemic input from a different absorption pattern or from uncertainty about when the peak occurred. The residual-pattern principles in the earlier material on residual patterns help localize the discrepancy; here the focus is specifically on delayed peaks and lag-time interpretation.

Zero-Lag and Lagged First-Order Absorption

For a single oral dose in a one-compartment model with first-order absorption and elimination, let \(D\) be dose, \(F\) bioavailability, \(V\) apparent volume of distribution, \(k_a\) the absorption rate constant, and \(k_e\) the elimination rate constant. With absorption beginning at time zero, the model-predicted concentration for \(t\geq 0\) is

This equation assumes a dose enters an absorption compartment at time zero, absorption and elimination follow first-order rates, and the parameters remain constant over the profile. It also assumes the one-compartment structure is adequate for the concentration process being described. A population model may allow parameters to vary between individuals, but individual variation does not change the defining zero-lag assumption: input starts immediately.

A fixed absorption lag \(t_{\mathrm{lag}}\geq 0\) changes the input onset. Under the same assumptions, concentration is zero before the lag and follows the same concentration-time function, shifted in time, afterward:

The zero before onset is a statement about the model's systemic input and resulting drug concentration, not necessarily about what an assay reports. In real data, baseline concentrations, assay limits, residual error, and other sources of variation affect observed values. A model that predicts a lag should therefore be checked against the measurement process and the sampling design, not just against a curve drawn through the observations.

Deriving the Predicted Peak Time

For the zero-lag model, the peak occurs where the concentration stops increasing and starts decreasing. Differentiating the concentration equation with respect to \(t\) gives a derivative proportional to \(-k_e e^{-k_e t}+k_a e^{-k_a t}\). Setting it equal to zero and rearranging yields the peak time.

This is the predicted peak time for the stated model and parameter values, not a universal peak time for an individual. It depends on both absorption and elimination. In the lagged model, the same post-onset curve is shifted, so its predicted peak is \(t_{\mathrm{lag}}+t_{\max}\). This shift is exact under the model assumptions; it does not mean that a lag is the only way to obtain a later peak. Changing the absorption rate, for example, also changes the predicted peak time, though it changes the curve’s shape as well.

What Evidence Supports a Lag?

The most informative pattern is often the early portion of the profile. A zero-lag model predicts rising concentrations from the start of input. Repeated early observations that are unexpectedly low, followed by a rise and a later peak, can produce a time-ordered residual pattern: observed-minus-predicted residuals tend to be negative early and then move toward zero or become positive near the rise or peak. The exact pattern depends on the residual definition and error model. It should be assessed across individuals or occasions where possible, rather than inferred from one noisy profile.

A lag is more plausible when a common onset delay accounts for the early mismatch and the later part of the profile is reasonably described by the same post-onset absorption and elimination shape. It is less persuasive when the proposed lag merely improves one peak while producing systematic errors in the rising phase, peak height, or decline. As emphasized in the earlier material on separating structural misfit from random-effect misfit, a population-level timing pattern should not automatically be attributed to one structural parameter if individual variability or data limitations could explain it.

• Check the sampling schedule. If no samples were taken before or around the predicted peak, the observed maximum may only bracket the true peak. Do not treat the time of the largest sampled concentration as the exact peak time.
• Check parameter dependence. A late peak under one fitted \(k_a\) and \(k_e\) does not prove a lag. The estimated peak time also depends on those parameters and their uncertainty.
• Check the early data and assay. Values below quantification limits or close to assay noise may not distinguish a true low concentration from an unobserved early rise.
• Check the whole profile. Review residuals over time, not only the peak. A lag should improve the coherent timing pattern without leaving a new systematic pattern elsewhere.

Comparing a Lag With a More Flexible Input

A lagged first-order model makes a specific claim: input is absent until \(t_{\mathrm{lag}}\), then follows the assumed first-order process. A more flexible input process can describe gradual, distributed, or otherwise non-instantaneous onset. These approaches are not interchangeable. A lag shifts the entire post-onset profile; a flexible input can alter the rising-phase shape as well as the apparent peak timing.

Use the simpler lagged model when a distinct delay is biologically and operationally plausible, the early data help locate onset, and the shifted post-onset shape remains adequate. Consider a more flexible input when the mismatch is not well described by a simple time shift—for example, if the rise remains systematically too sharp or too gradual after the estimated onset. The specific alternatives have different assumptions and are treated separately in later material; here, the key point is to avoid interpreting every delayed peak as evidence for one particular input mechanism.

Identifiability, Diagnostics, and Common Pitfalls

Lag time is difficult to estimate when there are few samples before the peak. If the first concentration is measured only after input has begun, many combinations of lag and absorption rate may generate similar observations. A longer lag paired with a faster post-onset rise can sometimes resemble a shorter lag paired with slower absorption. This is a practical identifiability problem: the data may not contain enough information to estimate the two features separately with useful precision.

Look for warning signs such as a lag estimate that changes substantially when early samples are removed or added, a wide interval, strong parameter correlation, or estimates pressed against a boundary. Also examine whether the estimated lag is plausible given dose administration, formulation, and sample timing. A numerically improved fit is not enough if the parameter is unstable or the residual pattern remains structured.

A common mistake is to say, “The observed peak is later than the zero-lag prediction, so the difference between the two times is the lag.” That reasoning treats a sampled maximum as the true peak, assumes \(k_a\) and \(k_e\) are known without uncertainty, and ignores alternative input shapes. The correction is to treat the discrepancy as a diagnostic clue: evaluate early residuals, account for sampling and parameter uncertainty, fit the lagged model, and assess whether it improves the full profile with an interpretable and identifiable lag.