Constant Placebo Shifts and Visit-Specific Placebo Effects

Why the Placebo Term Needs a Time Structure

Repeated pharmacodynamic (PD) measurements can change after baseline even when a participant receives no active drug. A model may represent that background response using a placebo component in addition to baseline and natural disease progression. The key choice here is how that placebo contribution varies across scheduled visits: is it one common post-baseline shift, or may it differ from visit to visit?

These are alternative assumptions about the mean response, not two names for the same model. A constant shift predicts the same placebo contribution at every post-baseline visit. A visit-specific model allows each scheduled visit to have its own contribution, without requiring a smooth or monotone pattern. Neither choice alone establishes why a response occurred.

The sign of a shift depends on the outcome scale. If higher scores indicate worse symptoms, a negative \(\delta\) is a reduction in the score. If higher values indicate improvement, the same sign has a different clinical interpretation. Always state the scale and direction rather than calling a shift beneficial by its sign alone.

For clarity, define \(r_{ij}\) as the observed outcome minus the model-predicted contribution from all components other than placebo. In a placebo arm, that prediction may include baseline and the chosen natural-history term; in an active arm, it may also include the specified drug effect. The remaining \(r_{ij}\) are the values used here to illustrate estimation of the placebo contribution. A model fitted to actual repeated measurements must also account for its residual-error structure.

Two Models and Their Assumptions

Under ordinary equal-weight least squares, the constant-shift model chooses one value of \(\delta\) to represent the post-baseline residuals. Its key structural assumption is that the mean placebo contribution is unchanged across the scheduled post-baseline visits. It does not assert that every participant has the same observed response: individual residuals can vary around that common mean.

The visit-specific model assigns a separate \(\delta_j\) to each post-baseline visit. It assumes only that the mean contribution is common across participants at a given visit, subject to the rest of the model. It does not impose a particular trajectory between visits, and it does not define an effect at an unscheduled time without an additional interpolation or time-course assumption.

The baseline convention matters for both forms. Since \(P_0=0\), baseline residuals do not estimate a post-baseline placebo shift. The fitting objective and all sums used to estimate these shifts therefore range over \(j=1,\ldots,J\), not over baseline and post-baseline observations together.

Least-Squares Estimates and Their Relationship

Suppose the other modeled components have been specified, the remaining residuals are \(r_{ij}\), and ordinary least squares gives each post-baseline observation equal weight. Let \(n_j\) be the number of observed residuals at visit \(j\), and let \(\bar r_j\) be their mean. The constant model minimizes the sum of squared residuals over post-baseline observations:

Differentiating this objective with respect to \(\delta\) and setting the derivative to zero gives the pooled mean of the post-baseline residuals:

In the visit-specific model, each \(\delta_j\) minimizes the squared residuals at its own visit. Consequently, its least-squares estimate is the mean residual for that visit:

The reduction has a useful exact decomposition. Let \(N=\sum_{j=1}^{J}n_j\) and let \(\bar r=\widehat{\delta}\) be the pooled post-baseline mean. For any visit, write \(r_{ij}-\bar r=(r_{ij}-\bar r_j)+(\bar r_j-\bar r)\). Squaring and summing gives a within-visit term, a between-visit term, and a cross-product. The cross-product is zero because \(\sum_i(r_{ij}-\bar r_j)=0\). Thus:

The first term is the visit-specific model’s residual sum of squares. The second is the extra sum of squares incurred by requiring one common shift. It is nonnegative, so the visit-specific fit cannot worsen the least-squares objective. This is a mathematical comparison of fitted sums of squares, not a model-selection rule that accounts for uncertainty or complexity.

Diagnostics, Alternatives, and Limitations

A constant shift is most defensible when the post-baseline visit means, after accounting for the other modeled components, are reasonably consistent with a common value. Plotting residuals against scheduled visit is a direct diagnostic: a clear pattern of means rising or falling across visits is evidence against a flat mean structure. Also inspect within-visit residual spread and unusual observations. A visit-specific model can capture differing means, but it does not correct a misspecified natural-history component, systematic measurement problems, or an inappropriate residual covariance model.

A practical comparison is to report the visit-specific mean residuals alongside the pooled constant estimate and the corresponding residual sums of squares. The visit-specific model will always fit at least as well by unpenalized least squares, so a smaller sum of squares alone should not decide between them. Model selection or inference requires a specified error distribution and covariance structure, and must account for the additional parameters. When observations are repeated within people, treating all measurements as independent can understate uncertainty even when the mean structure is appropriate.

The principal alternatives have distinct tradeoffs. A constant shift is parsimonious and gives a simple prediction at every post-baseline visit, but it can conceal time variation. Visit-specific effects make fewer assumptions about the means at the observed visits, but use more parameters and offer no principled prediction between or beyond those visits. A structured time-course model can impose a smooth pattern, but its form and assumptions must be justified separately; it is not implied by the visit-specific estimates.

Finally, placebo and natural-history components can be difficult to distinguish from data if both are allowed to explain similar changes. These terms are model components, not labels that the observed measurements can assign with certainty. Their interpretation depends on the study design and on the other components held in the model. A fitted placebo shift should therefore be reported as a model-based mean contribution on the outcome scale, with its reference time and assumptions stated.

Common Mistake

• Wrong reasoning: “Baseline is part of the repeated-measures data, so it should be included when averaging residuals to estimate the placebo shift.”
• Correction: The model defines \(P_0=0\), while \(\delta\) or \(\delta_j\) describes post-baseline placebo contributions. Estimate these parameters using only observations at \(j=1,\ldots,J\); including a baseline residual changes the objective and generally biases the calculation toward zero.
• Another common error: “The visit-specific model has a lower residual sum of squares, so it must be the better scientific model.” Its lower sum of squares follows from adding parameters. Assess the time pattern, uncertainty, covariance assumptions, and scientific purpose before choosing a form.