Delayed Drug Effects in Time-to-Event Models

Why a Drug Effect Can Lag Behind Plasma Concentration

A plasma concentration can rise and fall before the pharmacologic effect reaches its maximum. If an event hazard is linked directly to plasma concentration, the model predicts that the hazard follows the same time course as that concentration. A delayed-effect model instead lets a hypothetical effect-compartment concentration respond gradually to plasma concentration, and links the hazard to this changing effect-site quantity.

The effect compartment is a mathematical representation of equilibration delay; it need not correspond to a separately measured physical compartment. Its rate constant controls how quickly the modelled effect-site concentration responds. This distinction matters when predicting event timing: a delayed hazard may continue to rise even while plasma concentration is already falling.

Definition: In a first-order effect-compartment model, plasma concentration $C_p(t)$ drives an effect-compartment concentration $C_e(t)$ according to $$ \frac{dC_e(t)}{dt}=k_{e0}\bigl(C_p(t)-C_e(t)\bigr), $$ where $k_{e0}>0$ is the equilibration rate constant, with units of inverse time. A log-linear delayed hazard model is $$ h(t\mid C_e)=h_0(t)\exp\bigl(\beta C_e(t)\bigr), $$ where $h_0(t)$ is the hazard in the absence of drug effect and $\beta$ has units inverse to concentration. The hazard is the instantaneous event rate among individuals still event-free at time $t$.

The model assumes that the effect compartment responds linearly and with a constant rate to the difference between plasma and effect-compartment concentrations. It also assumes that the specified hazard link is appropriate and that the baseline hazard represents the non-drug time course. With a log-linear link, the hazard ratio relative to the same baseline at time $t$ is $\exp(\beta C_e(t))$. If $\beta>0$, higher $C_e(t)$ raises the hazard; if $\beta<0$, it lowers it.

Deriving the Delayed Concentration Profile

The differential equation can be solved for any sufficiently regular input profile. Multiplying both sides by the integrating factor $e^{k_{e0}t}$, integrating from 0 to $t$, and dividing by that factor gives

C_e(t)=C_e(0)e^{-k_{e0}t} +k_{e0}\int_0^t e^{-k_{e0}(t-s)}C_p(s)\,ds.

This expression shows that the current effect-compartment concentration is a weighted contribution from prior plasma concentrations, with more recent concentrations generally receiving greater weight. The weights decay at rate $k_{e0}$. Thus the effect compartment smooths and delays changes in the plasma input; it is not, in general, just the plasma curve shifted by a fixed number of hours.

For a single dose with a one-compartment, first-order decline, let $C_p(t)=C_0e^{-k_{el}t}$ and suppose $C_e(0)=0$, where $k_{el}$ is the plasma elimination rate constant. Substitution into the integral gives

C_e(t)=\frac{k_{e0}C_0}{k_{e0}-k_{el}} \left(e^{-k_{el}t}-e^{-k_{e0}t}\right), \qquad k_{e0}\ne k_{el}.

The sign of the coefficient and the order of the exponentials must be considered together. For either ordering of $k_{e0}$ and $k_{el}$, the resulting concentration is nonnegative. When the two rates are equal, the limiting solution is $C_e(t)=k_{e0}C_0t e^{-k_{e0}t}$; the formula with a zero denominator should not be used.

The concentration reaches its positive-time maximum when its derivative is zero. Differentiating the solution shows that the peak time satisfies $k_{e0}e^{-k_{e0}t}=k_{el}e^{-k_{el}t}$, so

t_{\text{peak}}= \frac{\ln(k_{el}/k_{e0})}{k_{el}-k_{e0}}.

This time is positive for either ordering of the two distinct positive rate constants. If $k_{e0}$ is small relative to $k_{el}$, equilibration is slow and the effect can peak well after plasma concentration begins to decline. Under a monotone log-linear hazard link, the drug-related hazard ratio peaks at the same time as $C_e(t)$, provided $\beta\ne0$; for $\beta<0$, the hazard is lowest there rather than highest.

Worked Example: Finding a Delayed Peak and Hazard Ratio

Suppose a hypothetical single dose produces $C_p(t)=4.00e^{-0.500t}$ mg/L, with time in hours. Let $k_{e0}=0.200$ h$^{-1}$ and use the log-linear hazard effect with $\beta=0.400$ L/mg. Assume $C_e(0)=0$.

The predicted peak time is $$ t_{\text{peak}}= \frac{\ln(0.500/0.200)}{0.500-0.200} =\frac{\ln(2.5)}{0.300} =3.054\text{ h}. $$

At this time, $e^{-0.500(3.054)}\approx0.2172$ and $e^{-0.200(3.054)}\approx0.5429$. Substitution into the effect-compartment solution gives $$ C_e(3.054)= \frac{0.200(4.00)}{0.200-0.500}(0.2172-0.5429) \approx0.8685\text{ mg/L}. $$ The coefficient is negative and the parenthesized difference is negative, so their product is positive.

The hazard ratio relative to the baseline hazard at the peak is $$ \exp\bigl(0.400\text{ L/mg}\times0.8685\text{ mg/L}\bigr) =\exp(0.3474)\approx1.415. $$ Thus, under this hypothetical model, the instantaneous hazard at the effect-compartment peak is about 1.415 times the baseline hazard at that time. This is a relative hazard, not an event probability.

How the Rate Constants Shape the Delay

The plasma elimination rate $k_{el}$ and equilibration rate $k_{e0}$ have different roles. The former determines how quickly plasma concentration declines; the latter determines how quickly the effect compartment tracks its input. Changing either can change both the effect-compartment peak time and its magnitude. The delayed model therefore does not imply that the effect is always smaller than the direct concentration effect, or that its peak always occurs at one fixed delay.

Worked Example: A Faster Effect Compartment

Consider a second hypothetical profile with $C_p(t)=4.00e^{-0.200t}$ mg/L, $k_{el}=0.200$ h$^{-1}$, and $k_{e0}=0.800$ h$^{-1}$. Use $\beta=0.400$ L/mg and $C_e(0)=0$.

Here, the effect compartment equilibrates faster than plasma concentration declines. The peak time is $$ t_{\text{peak}}= \frac{\ln(0.200/0.800)}{0.200-0.800} =\frac{\ln(0.25)}{-0.600} =2.310\text{ h}. $$

At this time, $e^{-0.200(2.310)}\approx0.6300$ and $e^{-0.800(2.310)}\approx0.1575$. Therefore, $$ C_e(2.310)= \frac{0.800(4.00)}{0.800-0.200}(0.6300-0.1575) \approx2.520\text{ mg/L}. $$ Using the unrounded peak time gives approximately $2.5198$ mg/L.

The peak hazard ratio is consequently $$ \exp\bigl(0.400\times2.5198\bigr)\approx2.740. $$ The result differs from the first example because the plasma profile and equilibration rate differ. It illustrates why the delay and peak effect must be calculated from the specified rates rather than inferred from a generic notion of “hysteresis.”

From a Hazard Profile to an Event-Time Prediction

A hazard ratio describes an instantaneous rate among those still event-free; it does not by itself give the probability of an event by a specified time. For a hazard $h(t)$, cumulative hazard through time $T$ is $\int_0^T h(t)\,dt$, survival is the exponential of the negative cumulative hazard, and event probability through $T$ is one minus survival. These quantities depend on the whole hazard trajectory, not just its peak.

Worked Example: Comparing Delayed and Direct Concentration Effects

For a hypothetical comparison, let $C_p(t)=4.00e^{-0.500t}$ mg/L, $k_{el}=0.500$ h$^{-1}$, $k_{e0}=0.200$ h$^{-1}$, $\beta=0.400$ L/mg, and constant baseline hazard $h_0=0.0500$ h$^{-1}$. Compare a hazard linked to $C_e(t)$ with one linked directly to $C_p(t)$. Use Simpson’s rule with values at 0, 1, and 2 hours as an approximate integration over the two-hour interval.

The plasma concentrations at those times are $4.000$, $4e^{-0.5}=2.426$, and $4e^{-1}=1.472$ mg/L. The effect-compartment solution gives $C_e(0)=0$, $C_e(1)=0.5659$, and $C_e(2)=0.8065$ mg/L. The corresponding delayed-model hazard multipliers are $$ \exp(0.400C_e)\approx(1.000,\;1.254,\;1.381). $$ Thus the delayed hazard is still increasing at 2 hours in these sampled values, even though plasma concentration has fallen from 4.000 to 1.472 mg/L.

For the delayed model, Simpson’s rule estimates the integral of the hazard multiplier as $$ \frac{2}{6}\left(1.000+4(1.254)+1.381\right) \approx2.466\text{ h}. $$ The estimated cumulative hazard is $0.0500\times2.466=0.1233$. Hence the estimated event probability by 2 hours is $$ 1-e^{-0.1233}\approx0.1160. $$

Under the direct-effect model, the multipliers at 0, 1, and 2 hours are $\exp(0.400C_p)\approx(4.953,\;2.639,\;1.801)$. The same Simpson approximation gives $$ \frac{2}{6}\left(4.953+4(2.639)+1.801\right) \approx5.770\text{ h}. $$ The resulting cumulative hazard is approximately $0.2885$, and the estimated event probability by 2 hours is $1-e^{-0.2885}\approx0.2506$. In this hypothetical example, the delayed model predicts lower event risk over the interval because it has not yet reached the direct model’s high early hazard. These are model-based approximations, not observed study results; a three-point numerical integration can be crude, especially when the hazard changes rapidly.

Assumptions, Diagnostics, and Alternatives

The effect-compartment model is most defensible when a first-order equilibration process is a reasonable representation of the delay and when the concentration input is adequately described. In an analysis, the concentration profile should be checked against the pharmacokinetic model, and predicted effect-site profiles should be examined over the observed dosing and follow-up times. Implausible negative values, peaks occurring at unexpected times, or sensitivity to small changes in the rate constants can indicate a coding error, a poorly supported model, or a mismatch between the assumed and actual concentration dynamics.

Event-time diagnostics should also compare predicted event patterns with observed event-time data while accounting for censoring. Systematic early or late prediction errors can suggest that the delay structure, baseline hazard, or concentration-effect link is inadequate. A well-fitting concentration profile alone does not establish that the hazard model is appropriate. Likewise, an apparent delayed pattern in observed data does not uniquely identify an effect compartment; other mechanisms can produce a lag.

The main alternative is to link hazard directly to $C_p(t)$. That model is simpler and is appropriate when the event hazard responds effectively instantaneously to plasma concentration. The effect-compartment alternative is useful when the effect is expected to equilibrate gradually. More flexible lag structures may be considered when a single first-order process is inadequate, but additional flexibility requires data capable of distinguishing those models. A first-order effect compartment also has limitations: it imposes a smooth, single-timescale delay and does not, by itself, represent tolerance, disease progression, a delayed onset with a fixed duration, or a cumulative-dose mechanism.

Common pitfall: Do not treat $C_e(t)$ as a fixed time-shifted copy of $C_p(t)$, or reverse the exponentials in the single-dose solution without also checking the coefficient’s sign. Use the differential equation or its integral solution, verify that predicted concentrations are nonnegative, and distinguish an instantaneous hazard ratio from an event probability accumulated over time.

Check Your Understanding

Try each problem, then compare your reasoning with the solution included.

For $C_p(t)=C_0e^{-k_{el}t}$, what equation determines the peak time of $C_e(t)$?
Solution: Set the derivative of the effect-compartment solution to zero. This gives $k_{e0}e^{-k_{e0}t}=k_{el}e^{-k_{el}t}$, and therefore $t_{\text{peak}}=\ln(k_{el}/k_{e0})/(k_{el}-k_{e0})$ when the rates are distinct.
With $k_{el}=0.400$ h$^{-1}$ and $k_{e0}=0.100$ h$^{-1}$, is the peak time positive, and what is it?
Solution: $t_{\text{peak}}=\ln(0.400/0.100)/(0.400-0.100)=\ln(4)/0.300=4.621$ h, which is positive.
If $\beta=0.300$ L/mg and $C_e(t)=1.20$ mg/L, what is the hazard ratio relative to baseline under the log-linear model?
Solution: The exponent is $0.300\times1.20=0.360$, so the hazard ratio is $e^{0.360}\approx1.433$. This is an instantaneous relative hazard, not a 43.3% event probability.
Why can the delayed hazard still be rising while plasma concentration is falling?
Solution: The effect compartment responds to the gap $C_p(t)-C_e(t)$. While $C_p(t)>C_e(t)$, its derivative $k_{e0}(C_p-C_e)$ is positive, so $C_e(t)$, and a positively linked hazard, can rise even as $C_p(t)$ declines.