Delayed Effects and Hysteresis in Exposure-Response

Why Exposure and Response May Not Move Together

A concentration-effect relationship can appear different depending on whether concentration is rising or falling. For example, the same plasma concentration may correspond to a smaller response on the rising part of a dosing profile than on the falling part. A plot of response against concentration then forms a loop rather than a single curve. This pattern is called hysteresis and indicates that current concentration alone may not summarize the timing of the response.

A loop is a description of the observed time pattern, not a mechanism. Delayed equilibration between plasma and the site of action, slow change in a response variable, and effects of earlier exposure can all produce temporal mismatch. Disease progression, time-dependent background effects, measurement timing, or model misspecification can also create an apparent loop. The modeling goal is therefore to represent plausible dynamics and assess whether the available observations support them, rather than to infer a mechanism from the loop alone.

The three model structures below address different versions of the timing problem. An effect-compartment model represents delayed equilibration between measured plasma concentration and a latent effect-site concentration. A turnover model represents a response that changes gradually because it is produced and removed over time. A distributed-lag model represents the response as depending on a weighted history of exposure. These are alternative modeling assumptions, not interchangeable labels for every loop.

Effect-Compartment Models

Let \(C(t)\) be plasma concentration and \(C_e(t)\) be an unobserved effect-site concentration. The basic effect-compartment model assumes that \(C_e(t)\) moves toward plasma concentration at a rate proportional to their difference:

Here \(k_{e0}\) has units of inverse time. The model assumes that this equilibration rate is constant over the modeled period and that the effect-site concentration is adequately linked to the measured plasma concentration through this first-order process. A concentration-effect function is then applied to \(C_e(t)\), for example a saturable function for a stimulatory response:

In this expression, \(E_0\) is the response at zero effect-site concentration, \(E_{\max}\) is the maximum increase above \(E_0\), and \(EC_{50}\) is the effect-site concentration producing half of that increase. A decreasing response can be represented by a negative \(E_{\max}\), provided the resulting response range is scientifically meaningful.

Turnover Models

A response can lag even when the concentration at the site of action tracks plasma concentration promptly. For example, a pharmacologic stimulus may alter the rate at which a biomarker is produced, while removal of that biomarker continues over time. A simple indirect-response or turnover model describes the response \(R(t)\) with an input rate and a first-order loss rate. One stimulatory form is:

The model assumes a saturable concentration effect on production, first-order removal, and parameters that remain constant during the modeled period. With no drug effect, the baseline is \(R_0=k_{\mathrm{in}}/k_{\mathrm{out}}\). At constant concentration \(C\), the steady response is \(R_{\mathrm{ss}}=R_0[1+S_{\max}C/(SC_{50}+C)]\). The response approaches that steady value at a rate governed by \(k_{\mathrm{out}}\); therefore a slow turnover rate can create a substantial delay.

Distributed-Lag Models

A distributed-lag model directly summarizes exposure over a chosen past window. In continuous time, one possible form is \(Y(t)=\beta_0+\beta_1\int_0^L w(u)C(t-u)\,du+\epsilon(t)\), where \(L\) is the maximum lag, \(w(u)\) is a prespecified or estimated weighting function, and \(\epsilon(t)\) is residual variation. The model assumes that the selected history and weighting function adequately represent the exposure contributions relevant to the outcome. It need not claim a literal effect compartment or a particular production mechanism.

With measurements at discrete time points, a finite distributed lag can be written as \(Y_t=\beta_0+\sum_{j=0}^{m}\beta_j C_{t-j}+\epsilon_t\). The coefficients describe the contributions assigned to current and earlier concentrations, conditional on the other included lag terms. Because nearby concentrations are often highly correlated, estimating many unrestricted lag coefficients can be unstable. A shorter prespecified window, a smooth weighting function, or scientifically justified constraints can reduce the number of effective parameters.

Choosing and Assessing a Dynamic Model

The model should correspond to the scientific source of delay that is plausible for the response and to the temporal resolution of the data. An effect-compartment model is often compact when a single equilibration timescale between plasma and an effect site is a useful approximation. A turnover model is appropriate when the response has a production-loss interpretation or a clearly gradual return toward baseline. A distributed lag is flexible when exposure history matters but a specific kinetic mechanism is not established. Each model should include a suitable baseline and any necessary time or background terms; otherwise, temporal trends unrelated to drug exposure can be attributed to the exposure dynamics.

A useful assessment begins by checking whether the observed timing is consistent with the fitted dynamics. Plot observed responses and model predictions against time alongside the concentration profile, and examine whether residuals retain a systematic time pattern or show different behavior on rising and falling exposure. Compare predictions at similar concentrations reached at different times: a model intended to represent hysteresis should capture the response difference without leaving a consistent directional pattern in the residuals. Such checks diagnose temporal mismatch; they do not by themselves identify the biological mechanism.

Parameter precision and sensitivity also matter. If \(k_{e0}\) is very large, the effect compartment nearly tracks plasma concentration, and the model may be difficult to distinguish from an immediate concentration-effect model. If the observation schedule is sparse relative to the delay, many values of \(k_{e0}\) may imply similar predictions. In turnover models, baseline, production, and removal parameters can be weakly identified without observations that capture baseline and response recovery. In distributed-lag models, highly correlated concentrations at adjacent times can yield uncertain individual lag coefficients even when the combined prediction is stable. Report uncertainty and avoid mechanistic claims stronger than the data support.

Common Pitfall and Limitations

Wrong reasoning: “The concentration-effect plot is counterclockwise, so the drug must have a delayed effect-site equilibration.” A counterclockwise loop can be compatible with a delayed response, but it is not unique evidence for an effect compartment. Turnover, a history-dependent response, or unmodeled time trends can also produce such a pattern. Correction: describe the loop as evidence that current concentration alone may be inadequate, compare plausible dynamic structures, and examine whether the sampling schedule and data can distinguish their predictions.

All three approaches require adequate information about concentration and response over time. Inaccurate exposure estimates, sparse or misaligned sample times, unmodeled changes in disease or background therapy, and influential observations can obscure or imitate delay. More flexible models are not automatically better: extra parameters can be weakly identified, and an apparent improvement in fit may not yield reliable predictions outside the observed dosing and follow-up pattern. These models describe exposure-response timing; they do not, by themselves, establish a causal effect or resolve confounding.