Response-Guided Titration With Delayed Measurements

Why the Measurement Time Matters

A titration rule can be mathematically clear and still make the wrong decision if it treats a newly received result as a measurement of the current dose. A laboratory value, symptom score, or other clinical response may be collected days before it is reported. In addition, the response itself may take time to reflect a dose change. These are distinct delays: reporting delay is the time from collection to availability, while response delay is the time required for the measured outcome to reflect a changed regimen.

The central design task is therefore not just to specify how a response changes the dose. It is to specify which regimen history the response represents, when the response is collected, when it becomes available, and whether it is still appropriate to use for a dose decision. This builds on state-update rules and bounded titration: the update must operate on the correct state and must respect the allowed dose range.

Representing Delayed Measurements in the Algorithm

Let \(D_k\) denote the dose chosen at decision occasion \(k\), and let \(Y_j\) denote a response collected at occasion \(j\). Let \(a(j)\) be the occasion at which the result of \(Y_j\) becomes available. In general, \(a(j)>j\). The result must retain its collection label \(j\), even when it arrives at occasion \(a(j)\). The dose in effect when the result arrives is not automatically the dose that generated the response.

A useful record can be written as \((j,a(j),Y_j,v_j,H_j)\), where \(v_j\) identifies the regimen version associated with the observation and \(H_j\) records the relevant dose history. If response depends on cumulative exposure, dose timing, or more than one previous dose, \(H_j\) must describe that history; a single dose label is not enough. A single-version label is appropriate only when the clinical and pharmacologic assumptions make the response attributable to that version—for example, the dose was held long enough for the selected response metric to represent that regimen.

The rule also needs an explicit policy for pending or stale results. A conservative policy is to allow a simple response-threshold update only when the result is available, its source regimen is still the current regimen, and the response has had the required time to reflect that regimen. Stale results can still inform a model or clinical review, but they should not be substituted into a direct current-dose rule without accounting for the intervening regimen history.

Here \(g\) is the chosen update rule, and \(\operatorname{clip}\) restricts a proposed dose to the interval from \(D_{\min}\) to \(D_{\max}\). Eligibility is part of the algorithm, not an informal judgement added after the calculation. It can require a matching dose version, a minimum time on that regimen, and no unresolved change that makes the response history ambiguous. In a clinical application, safety rules and monitoring plans also need to be specified; the equation alone is not a treatment recommendation.

A Model-Based Justification

Consider a simple response model for a setting where the response is measured after the regimen has been held long enough to represent a single dose version:

Here \(\alpha\) is an intercept, \(\beta\) is the response change per unit of dose, and \(\varepsilon_j\) is unexplained variation. To interpret the fitted mean response for a particular dose history, the needed assumption is not merely that the errors average to zero over all observations. The relevant condition is conditional mean zero given the information that determined the dose and describes the response history.

The implication follows by taking conditional expectations of both sides of the response model. The dose and intercept are known conditional on the stated information, so \(E(Y_j\mid H_j,D_{v_j})=\alpha+\beta D_{v_j}+E(\varepsilon_j\mid H_j,D_{v_j})\), which reduces to the displayed mean under the conditional assumption. An unconditional statement \(E(\varepsilon_j)=0\) is insufficient: errors could have mean zero overall yet have systematically different means for different dose histories. In adaptive titration, this distinction matters because past responses may determine later doses.

This linear model is a simplifying example, not a universal response model. If the response depends on prior doses, time since dose change, or other covariates, those predictors belong in the model or in \(H_j\). The dose label then identifies a regimen history rather than claiming that one isolated dose caused the entire observed response. Even with a correctly specified model, delayed results do not automatically establish a causal dose effect; interpretation also depends on the design, covariates, and assumptions about unmeasured influences.

Building a Delayed-Response Rule

Worked Examples

Diagnostics, Alternatives, and Limitations

The most useful diagnostic is an audit trail that makes the timeline reconstructable: dose-version start and stop times, collection and report times, response value and units, relevant prior doses, decision time, and the result that triggered each update. A table or software log should make it possible to answer, for every change, “Which observation was used, and under which regimen history was it collected?” Missing or contradictory version labels are a design failure, not a minor clerical inconvenience.

Other warning signs include repeated updates from the same result, a dose change made before the expected response delay has elapsed, results arriving out of order, and a long queue of pending observations. Check that a result is used once unless the algorithm explicitly permits repeated use, and define what happens when a result is missing, invalid, or late. If responses reflect overlapping regimens, a single source-dose label is inadequate; retain the full history or wait for an assessment under a sufficiently stable regimen.

There are three broad approaches. A wait-for-result rule holds the regimen until an eligible response arrives; it is straightforward to interpret but can delay adjustment. A history-aware model uses the dose history and response timing to interpret delayed outcomes; it may use information more efficiently but relies on a suitable model and assumptions. An unqualified latest-result rule updates from whatever value most recently arrived; it is simple but can confuse old responses with current effects and should not be used without a justification that makes the result current-relevant.

No single delay policy is appropriate for every response or drug. The result may reflect a gradual transition rather than one dose version, and measurement error, adherence, changes in clinical status, or missing assessments can complicate attribution. A model can also be wrong or poorly identified when dose histories and response delays are highly correlated. The algorithm should therefore state its assumptions and define review or fallback actions for cases in which its provenance checks fail. Selecting sparse sampling schedules and concentration targets are separate design questions; here the emphasis is on correct timing and attribution for response-guided updates.

Common Mistake