Tolerance and Adaptation in Safety Exposure-Response

Why safety effects may change during treatment

An adverse effect associated with exposure may be less pronounced after sustained treatment than it was early in treatment. One explanation is pharmacodynamic tolerance: at the same current exposure, the response becomes smaller as a consequence of prior exposure. But a declining event rate in the observed data does not establish tolerance. People at higher risk may have stopped treatment or left follow-up, changing who remains under observation. Alternatively, the event may simply be detected or reported less often later.

The key comparison is therefore not just whether event rates decline with calendar time. It is whether the exposure-related response changes with exposure history, after accounting for relevant time trends, exposure, and changes in the observed population and event ascertainment. This tutorial develops one way to express that change and clarifies what the model can—and cannot—show.

An exposure-history attenuation model

Let \(x_i(t)\) be the current exposure for individual \(i\) at time \(t\), scaled by a reference exposure so that it is dimensionless. Let \(H_i(t)\) represent prior exposure history, also on a dimensionless scale. One possible history measure gives more weight to recent exposure than to distant exposure:

Here \(\tau_H>0\) is a history time scale, measured in units of time. With sustained exposure \(x_i(t)=1\), \(H_i(t)\) increases toward 1; after exposure stops, it declines. This is a model choice, not a universal biological law. A cumulative history measure or another weighting function could be more appropriate for a particular safety question.

For an adverse-event rate, a simple conditional model is:

In this expression, \(\lambda_i(t)\) is the event rate per unit time while individual \(i\) is under observation and at risk; \(\lambda_0(t)\) is a baseline rate that may vary over time; \(Z_i(t)\) is a vector of measured covariates; \(u_i\) is an individual-specific log-frailty term; \(\beta\) is the exposure effect at \(H_i(t)=0\); and \(\kappa\) is the attenuation parameter. In this parameterization, \(\beta\) is a log rate ratio per unit of normalized current exposure when history is zero. The model assumes that the exposure contribution to the log rate is \(\beta x_i(t)\exp[-\kappa H_i(t)]\).

For a given current exposure \(x\) and history \(H\), the rate ratio relative to zero current exposure, with the other model terms held fixed, is

Thus, when \(\beta x>0\), a positive \(\kappa\) moves this rate ratio toward 1 as history increases. When \(\beta x<0\), the rate ratio also moves toward 1, but from below. The model describes attenuation of an exposure association; it does not require that exposure increase the event rate.

Justifying the attenuation interpretation

The interpretation follows directly from the history-dependent exposure term. For fixed \(x\), define \(g(H)=\beta x\exp(-\kappa H)\), the exposure contribution to the log rate. Its derivative with respect to history is:

If \(\kappa>0\), then \(\exp(-\kappa H)>0\). Consequently, \(g(H)\) decreases when \(\beta x>0\), and increases when \(\beta x<0\). In either case, its absolute magnitude is \(|g(H)|=|\beta x|\exp(-\kappa H)\), which decreases as \(H\) increases. This distinction matters: saying that the contribution itself always decreases would be incorrect when \(\beta x<0\).

Several assumptions are needed before this parameterization can be interpreted as evidence consistent with tolerance. The chosen \(H_i(t)\) must capture the relevant exposure memory; current exposure and history must be measured sufficiently well; and the log-rate form must be a reasonable description over the studied exposure range. The analysis must also address time trends and relevant differences between individuals. For example, an apparent history effect could arise because the baseline event rate changes over time, or because individuals with greater underlying risk are no longer represented in later risk sets.

Most importantly, exposure history and calendar time often increase together. If all participants have the same exposure schedule, the data may not distinguish an effect of history from a general time trend. Estimating \(\kappa\) therefore requires informative variation in exposure histories, such as differences in timing, interruption, or intensity, and adequate overlap in current exposure across those histories. A fitted positive \(\kappa\) is not, by itself, proof of a biological mechanism.

Separate within-person adaptation from changing risk sets

A decline in a population-average rate can occur even if no individual’s response changes at a given current exposure. Consider a hypothetical study with two risk groups. Individuals in the high-risk group have a constant rate of 6 events per 100 person-days, while those in the low-risk group have a constant rate of 2. Suppose half of the observed person-time is initially from each group. The overall rate is

events per 100 person-days. Later, suppose high-risk individuals are less likely to remain under observation, so only 20% of observed person-time is from that group. The later overall rate becomes

events per 100 person-days. The population-average rate fell by 30%, since \((4-2.8)/4=0.30\), although both groups’ rates stayed constant. This is survivor selection (or risk-set selection), not within-person tolerance. The mechanism could involve discontinuation, withdrawal, an event that changes subsequent observation, or another process that changes group composition.

Separate tolerance from changing ascertainment

Observed events may not equal all events that occurred. Suppose the true event rate is \(\lambda(t)\), and each event is detected and recorded with probability \(q(t)\). If missed events are not replaced by false positive reports, the observed rate is \(q(t)\lambda(t)\). A change in \(q(t)\) can therefore look like a change in the true rate. Changes in visit schedules, laboratory testing, reporting practices, or symptom questioning can all affect ascertainment.

Estimating attenuation and checking its interpretation

The model’s history effect can be understood by comparing exposure effects across histories, holding current exposure and other model terms fixed. For example, the log rate ratio for current exposure \(x\), relative to zero current exposure, is \(\beta x\exp(-\kappa H)\). If \(x\neq0\), \(\beta\neq0\), and comparable estimates are available at histories \(H_1\) and \(H_2\), the ratio of the two log rate ratios is \(\exp[-\kappa(H_2-H_1)]\). Taking logarithms gives a way to express the implied attenuation:

In application, the log rate ratios are estimated quantities, with uncertainty; this algebra is an interpretation aid, not a substitute for fitting the model and evaluating uncertainty.

Useful diagnostics focus on whether the model’s explanation is plausible rather than on whether \(\kappa\) has the expected sign alone. Examine overlap in current exposure at different histories; if high history occurs only at late calendar times, history and time may be difficult to separate. Compare observed and predicted event patterns across time, current exposure, and history. Check whether conclusions are sensitive to reasonable alternatives for the history definition and time trend. Investigate whether discontinuation or loss to follow-up differs by measured risk, and whether event surveillance changed during the study. Sparse data at some combinations of exposure and history can also make the estimated attenuation unstable.

Alternatives, limitations, and common mistakes

The exposure-history model is not the only way to represent a changing exposure association. A model with \(\kappa=0\) assumes a constant exposure effect across histories. A time-varying exposure coefficient can represent a change over calendar time without attributing it specifically to prior exposure. A model with a separate history term can represent cumulative toxicity or delayed effects, rather than attenuation of the current exposure effect. These alternatives answer different questions; choosing among them requires a safety rationale and adequate data, not just a better numerical fit.

Even a well-fitting attenuation model has limitations. The form \(\exp(-\kappa H)\) imposes a particular pattern: the magnitude of the exposure contribution approaches zero as history grows. It may not capture an initial adaptation followed by a plateau, rebound after treatment interruption, or different adaptation rates across individuals. A single average \(\kappa\) can obscure heterogeneity. In addition, an estimated \(\kappa\) is an association parameter conditional on the model and observed data. Interpreting it as a pharmacodynamic mechanism requires assumptions and evidence beyond the fitted curve.

• Mistake: treating every declining event rate as tolerance. A changing risk-set composition or reduced event detection can produce the same pattern. Examine those processes before attributing the decline to adaptation.
• Mistake: saying a positive \(\kappa\) always makes the exposure contribution numerically smaller. It reduces the contribution’s magnitude. When \(\beta x<0\), the contribution increases toward zero.
• Mistake: treating exposure history as interchangeable with calendar time. History and time can be strongly associated, but they represent different explanations. Without variation that separates them, the data may not identify which explanation is supported.
• Mistake: interpreting a fitted history effect as proof of mechanism. Model fit does not rule out selection, ascertainment changes, or unmeasured time-varying factors.