Delayed Placebo Onset and Persistent Response

Why Model a Delayed, Persistent Placebo Response?

A placebo-associated change need not appear fully at the first follow-up, nor must it fade soon afterward. A response may develop gradually over several visits and then remain near a plateau. A model that forces an immediate change or a rapid return toward baseline can misrepresent that pattern, even if it describes one or two observed visits reasonably well.

The earlier material on placebo-response components separates the placebo contribution from baseline and natural-history change. Here the focus is the shape of the placebo contribution over time. The model below describes a smooth, gradual onset and a persistent long-term effect; it does not by itself establish that an observed time trend is caused by placebo rather than by another component.

The normalization by \(1-L(0)\) is important: it makes the placebo contribution exactly zero at \(t=0\), even though the sigmoid \(L(t)\) itself is generally above zero there. The term “delayed” describes a gradual transition centered around \(t_d\), not a literal period with exactly no response followed by a sudden start. For a positive \(t_d/s\), the initial change can be small, while the main rise occurs later.

The sign of \(E_P\) follows the outcome scale. For a symptom score where decreases indicate improvement, a negative \(E_P\) represents a persistent reduction in symptoms. For an outcome where improvement means an increase, a positive value may be appropriate. The units of \(P(t)\) and \(E_P\) are the units of the outcome; \(t_d\) and \(s\) have units of time.

What the Parameters Imply

The amplitude \(E_P\) sets the eventual size and direction of the modeled placebo change. The time \(t_d\) is the inflection point of \(L(t)\), where the underlying sigmoid changes from accelerating to decelerating. The scale \(s\) controls the width of the transition: a smaller \(s\) produces a steeper rise around \(t_d\), while a larger \(s\) spreads the rise over a longer period.

Because the function is normalized at time zero, \(t_d\) is not exactly the time when half the total normalized response has occurred. That time can be derived. If \(g(t_{1/2})=1/2\), then $$ \frac{L(t_{1/2})-L(0)}{1-L(0)}=\frac{1}{2} \quad\Longrightarrow\quad L(t_{1/2})=\frac{1+L(0)}{2}. $$ The inverse of \(L(t)\) is \(t=t_d+s\log\left(\frac{L(t)}{1-L(t)}\right)\), so $$ t_{1/2} =t_d+s\log\left(\frac{q}{1-q}\right), \qquad q=\frac{1+L(0)}{2}. $$ This distinction is usually small when \(L(0)\) is close to zero, but it matters when reporting a parameter precisely.

The function is monotone and bounded for the stipulated parameter values. Differentiating gives $$ g'(t)=\frac{L(t)\{1-L(t)\}}{s\{1-L(0)\}}>0. $$ Since \(L(t)\) approaches 1 as \(t\) increases, \(g(t)\) approaches 1 and \(P(t)\) approaches \(E_P\). These properties encode a gradual onset and a persistent plateau. They do not imply that every patient has the same response or that the observed outcome itself must plateau; other response components and residual variation may also contribute.

Comparison with Immediate-Onset and Waning Patterns

The sigmoid can be compared with two simpler time courses. An immediate-onset exponential approach to a plateau is \(P_I(t)=E_P\{1-\exp(-k_{\mathrm{on}}t)\}\), where \(k_{\mathrm{on}}>0\) is a rate with units of inverse time. Its response begins changing from \(t=0\) and approaches \(E_P\). A rapidly waning pattern can be represented illustratively by \(P_W(t)=E_P\{1-\exp(-k_{\mathrm{on}}t)\}\exp(-k_{\mathrm{w}}t)\), where \(k_{\mathrm{w}}>0\) is a waning rate. This curve rises initially and then diminishes toward zero. These equations are comparison shapes; choosing among them requires data over time, not just a preferred verbal description.

Interpreting the Half-Response Time

A parameter should be interpreted according to its mathematical definition rather than its name. For example, \(t_d\) marks the sigmoid’s inflection, not exactly the time of half the total normalized response. The half-response time follows from the normalized function itself.

Assumptions, Diagnostics, and Limitations

The model is most interpretable when time is measured consistently from the start of the placebo period, the outcome scale and direction are stated, and the response can reasonably be treated as an additive component. It assumes a smooth, monotone transition toward a stable plateau, with one common time-course shape and amplitude for the observations being modeled. Those are substantive assumptions, not automatic consequences of using a sigmoid. If the measured outcome also changes through baseline-related variation or natural history, a placebo curve alone does not explain the full trajectory.

Inspect predictions alongside observed outcomes at each visit, preferably with uncertainty intervals where available. A systematic early mismatch may indicate that the model’s onset is too fast or too slow. Persistent late residuals that trend in one direction may signal a poor plateau assumption; a clear downturn after an initial rise is inconsistent with a monotone persistent sigmoid. Sparse visits near the transition can make \(t_d\) and \(s\) difficult to distinguish: a later, sharper rise and an earlier, broader rise may predict similar values at the observed times. A fitted amplitude that changes substantially when a few visits are removed is another warning that the available follow-up does not adequately define the plateau.

Compare candidate shapes on the same outcome scale and at the same observation times. The immediate-onset exponential is more economical if change begins promptly and approaches a plateau. The sigmoid is useful when the pattern suggests a gradual transition centered later in follow-up. A waning function is more appropriate when the response rises and then recedes. Model comparison should consider residual patterns, uncertainty in parameter estimates, and whether the implied time course is scientifically plausible; a better numerical fit alone does not prove a mechanism.

A smooth sigmoid does not represent a literal delay with exactly zero effect before a specified time. If the observations support a sharp threshold or an abrupt change, this gradual function may be a poor description. Conversely, sparse early observations may not establish whether onset was gradual or immediate. The plateau \(E_P\) may also be weakly determined if follow-up ends before the response is close to its asymptote. Finally, a population-average curve does not describe between-person differences in response. Such heterogeneity requires a separate modeling decision and should not be inferred from the sigmoid’s smooth shape alone.