Baseline Drift Confounded With Drug-Induced Change

Why Baseline Drift Matters

An indirect response may change gradually after dosing because the drug alters production or loss and the response turns over over time. But a gradual change is not necessarily drug-induced: the untreated response can also rise or fall during observation. If a model assumes a stationary baseline when the baseline is drifting, the model may attribute some or all of that trend to the drug. The resulting drug-effect and turnover estimates can then be biased, even if the fitted curve looks plausible.

The key diagnostic question is not simply whether the response changes after dosing. It is whether the change differs from what would have happened over the same period without drug. Pre-dose measurements and a suitable drug-free control can help answer that question, provided the trend is shared and the measurements are comparable.

A Turnover Model With Baseline Drift

Let \(Y(t)\) be the response, measured in response units, and let \(k\) be a positive first-order turnover rate in inverse time. Suppose that, without drug, the equilibrium response changes linearly from \(B_0\) at time zero, with slope \(g\). Thus the time-varying equilibrium is \(B_0+gt\). Let \(q(t)\) denote the drug-induced change in the response input, in response units per unit time. For example, inhibition of production makes \(q(t)\) negative. A model that includes both drift and drug effect is

This representation assumes first-order turnover with constant \(k\), an untreated equilibrium that changes linearly over the modeled time interval, and an additive drug effect on the input. It also assumes that the initial response is at the initial untreated equilibrium. These are modeling assumptions, not conclusions guaranteed by observing a gradual response. The input \(q(t)\) can vary with concentration; the equation does not require it to be constant.

Write \(x(t)=Y(t)-B_0\) for the change from the initial response. Subtracting \(B_0\) from the differential equation gives

Both the drug input and drift affect the response through the same turnover process. The drift contributes the forcing term \(kgt\); it is not generally equal to a straight line \(gt\) in the observed response, because the response takes time to follow its changing equilibrium.

Deriving the Drift Contribution

Multiplying the equation for \(x(t)\) by the integrating factor \(e^{kt}\) and integrating from zero to \(t\) yields

The first integral is the drug-driven contribution. To evaluate the second, substitute \(u=t-s\), so \(s=t-u\):

Therefore the drift contribution is \(g[t-(1-e^{-kt})/k]\). If the drug input is constant, \(q(t)=a\), its contribution is \(a(1-e^{-kt})/k\). Together, these give

This closed form requires constant \(a\) and \(g\), fixed \(k\), and the initial condition \(x(0)=0\). With a time-varying drug input, use the integral expression instead. For short times, expansion of the exponential gives

Thus the drift contribution initially grows approximately as \(gkt^2/2\), whereas a constant drug input initially contributes approximately \(at\). These different early-time shapes can help distinguish them if measurements are sufficiently informative. Over a limited time range, however, sparse observations and flexible drug-effect parameters can make their contributions difficult to separate.

How Omitting Drift Changes an Apparent Drug Effect

Consider fitting a stationary-baseline model to data generated by both a constant drug input and drift. To isolate the issue, suppose \(k\) is known and held fixed. The stationary model predicts \(x(t)=a_{\text{fit}}(1-e^{-kt})/k\). At any one time point, it can match the observed change by setting

This is an apparent input estimate at that time, not necessarily one constant estimate that fits all observations. It equals the true input \(a\) only if the drift contribution is zero or otherwise has no effect at the observation time. When \(k\) is also estimated, the stationary fit may additionally change its estimate of turnover to accommodate the unmodeled curvature. The direction and size of those changes depend on the sampling times, the sign of drift relative to the drug effect, and the exposure profile.

Using Controls to Separate Drift From Drug Effect

A drug-free control observed on the same time scale can help estimate the untreated trend. If the control and treated groups share the same drift, start from comparable baselines, and have the same turnover behavior, subtracting the control change from the treated change cancels the shared drift in this additive model. The assumptions matter: a control cannot correct a trend that differs systematically between groups.

Without a control, repeated pre-dose observations can reveal whether the response was already trending before exposure. Such observations do not prove that the same trend would continue after dosing, but they provide evidence against assuming a stationary baseline without checking. Change from baseline alone does not remove drift: it subtracts the initial value \(B_0\), while a subsequent time trend remains in \(x(t)\).

Model Checks, Alternatives, and Limitations

A practical analysis should ask whether the baseline assumption is supported by the design and observations, not just whether the treatment response can be fitted. Useful checks include examining pre-dose response trajectories, plotting drug-free control trajectories over the treatment period, and checking whether a fitted drug effect is required to explain changes that also occur without drug. Compare a stationary-baseline model (\(g=0\)) with a model that estimates or specifies drift, and check whether the estimated drug and turnover parameters are sensitive to that choice. A good fit by itself does not establish that the baseline is stationary or that the drug caused the modeled change.

The main alternatives answer different questions. A stationary-baseline model is simpler and can be appropriate when controls or prior measurements support stability. A model with an explicit time trend represents a systematic baseline change but introduces at least one additional parameter, which may be difficult to estimate when sampling is sparse. A drug-free control or randomized comparison provides design-based evidence about the untreated trajectory, while relying only on the treated trajectory requires stronger modeling assumptions. None of these approaches automatically resolves confounding if the trend and drug effect have very similar time courses.

This model is deliberately limited. A linear trend may be unsuitable over a long interval, and changes in turnover rate, nonlinear baseline behavior, or a time trend in measurement may require a different representation. The additive formulation also may not suit responses that must remain positive or whose drug effect is proportional to baseline. When the model is weakly identified, a plausible numerical estimate of \(g\), \(k\), or drug effect does not guarantee that the data distinguish their contributions. Earlier material on convolution solutions and turnover-rate identifiability provides useful context for understanding why different inputs can produce similar response trajectories.

Common Mistake

Wrong reasoning: “The response fell after dosing, so the drug must have caused the fall; subtracting the baseline value has already adjusted for baseline.” Subtracting \(B_0\) only sets the initial change to zero. It does not remove a subsequent decline in the untreated response.

Correction: A post-dose change can contain both drug and drift contributions. Evaluate evidence for the untreated trajectory using pre-dose or control data, and assess whether the drug-effect estimate changes when drift is represented. If the design and observations do not distinguish the two, report that limitation rather than treating the fitted drug effect as uniquely determined.