Why Timing and Confirmation Belong in the Titration Rule
After a dose change, the clinical response may not immediately reflect the new dose. A response can take days or weeks to approach its eventual level, and a single measurement can differ substantially from the patient's underlying response because of day-to-day variation, measurement error, or both. If a titration rule reacts to an early or unusually high or low reading, it can change the dose for the wrong reason.
A practical rule therefore needs more than a response target. It should specify when to assess, which measurements to summarize, and how much evidence is enough to confirm a dose adjustment. These choices should be prespecified and clinically interpretable. They are part of the decision rule, not merely details of data collection.
Representing Delay and Averaging Over a Window
One useful working model describes a response that approaches a new plateau after a dose change at time zero:
Here, \(\mu_0\) is the expected response just before the change, \(\mu_{\infty}\) is the expected long-run response under the new dose, and \(\tau\) is a time constant in the same units as \(t\). At time \(\tau\), the remaining difference from the new plateau is \(e^{-1}\), or about 36.8%, of its initial size. This is an illustrative response model, not a universal description of clinical time courses.
If the decision uses the average expected response over a continuous window from \(a\) to \(b\), the relevant quantity is not simply the response at the window's endpoint. Integrating the model and dividing by the window length gives:
The formula follows because the integral of \(e^{-t/\tau}\) from \(a\) to \(b\) is \(\tau(e^{-a/\tau}-e^{-b/\tau})\). In practice, a clinical rule may instead use the arithmetic mean of scheduled measurements in the window. That mean estimates the window response under suitable sampling and measurement assumptions; it is not automatically identical to the continuous-time average.
A confirmation rule must also account for noise. If measurements in a window are \(Y_i=\mu_i+\epsilon_i\), where the errors have mean zero and variance \(\sigma^2\), then averaging \(n\) independent measurements with a common underlying mean reduces the variance of their average to \(\sigma^2/n\). Positive pairwise correlations generally make the reduction smaller than under independence, while negative correlations can make it larger; in general, the effect depends on the correlations. For equal variance and correlations \(\rho_k\) between measurements \(k\) time steps apart:
Thus, more readings do not necessarily provide proportionally more information. Closely spaced readings can share the same underlying fluctuation, and a window that spans a changing response can average together clinically different states.
Worked Example: Averaging a Delayed Response
Suppose, hypothetically, that a clinical response score is expected to decline after a dose change according to \(R(t)=3+6e^{-t/10}\) points, with \(t\) in days. Lower scores are favorable, and the response target is an average of 3.6 points or lower. A planned assessment window contains three readings on days 20, 25, and 30. Find the average expected response in the window.
The model gives \(R(20)=3+6e^{-2}\approx3.8120\) points, \(R(25)=3+6e^{-2.5}\approx3.4925\) points, and \(R(30)=3+6e^{-3}\approx3.2987\) points. The arithmetic mean, calculated from the unrounded model values, is:
The window mean is below the 3.6-point threshold, even though the day-20 expected response is above it. Under a rule based on this window mean, the criterion is met for this window. It would not follow that the response was already adequate on day 20, nor that a persistence requirement has been met: those conclusions require the rule's other specified conditions and, if required, a second qualifying window.
Set the Rule Before Applying It
A well-defined rule makes the timing and evidence requirements explicit. The exact values depend on the clinical response, the expected delay, measurement properties, and consequences of acting too early or too late. In particular, the assessment window should be chosen to represent a clinically meaningful response state, not simply to maximize the number of readings.
State whether a higher or lower value is favorable, the target or threshold, and the units. Identify whether the rule concerns an individual measurement, a window summary, or another predefined summary.
Use prior clinical or model-based knowledge to identify when the response is expected to be sufficiently informative. Do not treat a measurement taken soon after a dose change as though it necessarily reflects the new dose's long-run effect.
Give the start and end of the window, the required measurement schedule, and the summary calculation. State how missing or invalid measurements affect whether a window can be evaluated.
Specify whether one qualifying window is sufficient or whether the criterion must persist across multiple windows. Ensure that the rule does not silently count overlapping windows as independent confirmation.
Make the earliest decision time clear, including the time needed to complete the measurement window. Keep the response-based timing rule distinct from any safety-specific action rules.
Worked Example: A Two-Window Rule for a Noisy Measure
Consider a hypothetical score for which lower values are favorable, with a target window mean of 10 points or lower. Four readings are collected in each weekly window. For illustration, suppose the measurement errors are independent, approximately normal, and have a standard deviation of 1.5 points, and assume the underlying response is roughly stable within each window. The operational rule requires two successive weekly means at or below 10 points.
In the first week, the readings are 9.2, 10.6, 10.1, and 11.0 points. Their mean is \((9.2+10.6+10.1+11.0)/4=40.9/4=10.225\) points, so this window does not qualify. In the next two weeks, the readings are 10.4, 9.8, 10.7, 10.3 and 9.9, 10.2, 10.0, 9.7 points. The means are \(41.2/4=10.300\) points and \(39.8/4=9.950\) points, respectively. Only the last window qualifies, so the two-window confirmation requirement is not met.
The standard error of a four-reading mean under the stated independence assumption is \(SE=1.5/\sqrt{4}=0.750\) points. This describes the measurement uncertainty around a stable underlying mean; it does not change the operational threshold unless the rule explicitly incorporates uncertainty. For example, if the true mean were 9.5 points, the chance that one window mean would nevertheless exceed 10 is \(1-\Phi((10-9.5)/0.75)=1-\Phi(0.667)\approx0.252\), where \(\Phi\) is the standard normal cumulative distribution function. A confirmation rule can reduce the chance of acting on one unusually favorable window, but two-window probabilities cannot be calculated by multiplying single-window probabilities unless the windows are independent. Persistent within-person variation can make successive windows correlated.
Choosing Adjustment Timing
The earliest sensible adjustment time depends on both the response dynamics and the assessment process. In the exponential model, the fraction of the eventual change achieved at time \(t\) is \(1-e^{-t/\tau}\). For example, at \(3\tau\) the response has completed \(1-e^{-3}\approx0.9502\), or about 95.0%, of its modeled change. That does not mean every response should be assessed at \(3\tau\): the acceptable degree of settling depends on how much residual change could alter the decision.
A further delay arises because a window average includes measurements from before its endpoint. Suppose, hypothetically, that the response changes from a baseline of 0 toward a new level of 1 with \(\tau=8\) days. If measurements are summarized over days 21 through 28, the average fraction of the eventual change is:
The response at day 28 alone has reached \(1-e^{-28/8}=1-e^{-3.5}\approx0.9698\), or 97.0%, of the modeled change. The window average is lower, about 95.2%, because it also includes earlier days. If a decision requires a window average near the new plateau, using only the endpoint's expected response would overstate how settled that window is. In a real application, uncertainty in the estimated time course and variation between individuals also matter.
Worked Example: Scheduling a Window After a Dose Change
Assume, hypothetically, that a response follows the preceding model with \(\tau=8\) days, and the decision rule is intended to use a seven-day average that reflects at least 95% of the eventual change. Compare windows from days 14–21 and 21–28.
For days 14–21, the average fraction is:
This window reflects about 88.4% of the eventual change, below the 95% design criterion. For days 21–28, the average is:
The later window meets the stated criterion. Under these assumptions, a rule requiring that seven-day average should not make its response-based decision before day 28. This is a planning calculation, not proof that a patient's response has stabilized by day 28; the rule should be checked against observed data and plausible variation in response delay.
Diagnostics, Alternatives, and Limitations
Before using a rule, check whether its assumptions are plausible. Plot measurements against time since each dose change and inspect whether the response is still trending during the proposed window. Compare observed within-window variability with the variability used to plan the rule. Check for missing measurements, irregular timing, floor or ceiling effects, and changes in measurement conditions. If measurements are serially correlated, the nominal precision from treating them as independent will be too optimistic.
A fixed-time single reading is simple and may be appropriate when response delay and measurement variability are small relative to the decision threshold. It is vulnerable to transient or atypical readings. A window mean can reduce random noise, but may obscure a continuing trend or average across an important transition. A persistence requirement can reduce decisions driven by one anomalous window, but delays action and does not provide the expected protection when successive windows are strongly correlated. Model-based prediction can explicitly represent delay and uncertainty, while a transparent threshold-and-window rule may be easier to implement and communicate. The introductory material on Bayesian individualization during titration provides a complementary framework for using individual-level information; it does not remove the need to define assessment timing and measurements.
The exponential model is a simplification: clinical response may have multiple delays, nonlinear dose dependence, or a time course that changes between individuals. A rule designed from one typical time constant may therefore assess some patients too early and others later than necessary. Sparse data make it difficult to distinguish slow response from noisy measurement. If the response remains systematically changing across successive windows, a fixed-window interpretation is questionable; revise the model or rule rather than labeling the trend as random noise. The exact response and safety decisions are distinct. Safety holds and restart criteria require their own explicit rules.
Common Mistake
Wrong reasoning: “The latest measurement is below target, so the dose is working and we should adjust now.” This ignores whether the response had time to develop, how variable a single reading is, and whether the protocol requires a window summary or persistence.
Correction: Apply the prespecified rule to the measurements collected at the prespecified times. Confirm that the delay, window completeness, summary, and persistence conditions are satisfied before making a response-based decision. An early or noisy observation can be informative, but it is not automatically confirmation.
Check Your Understanding
Answer each question using the definitions and calculations in this tutorial.
- For an exponential response with time constant \(\tau\), what fraction of the eventual change has occurred at \(t=2\tau\)?
- Why can four closely spaced readings provide less information than four independent readings?
- A response is still trending across a proposed averaging window. Why might its mean be a poor summary for a stable-response decision?
- For a seven-day average under the hypothetical model with \(\tau=8\) days, which window met the 95% criterion in the worked example?
Solutions: (1) The fraction is \(1-e^{-2}\approx0.8647\), or 86.5%. (2) If all pairwise correlations are positive, the variance of the mean is greater than \(\sigma^2/n\), the independent-measurement result; negative correlations can make it smaller, and with mixed correlations the direction depends on their values. (3) The mean combines different response levels over time and may conceal that the response has not yet settled. (4) The days 21–28 window, whose average represented about 95.2% of the eventual change; days 14–21 represented about 88.4%.