Worked Analysis of a Placebo-Controlled Progression Dataset

From Placebo-Arm Records to a Progression Model

Longitudinal placebo-arm data contain more than a natural-history trajectory. A measured outcome may start at a nonzero baseline, change over time in response to trial participation and other placebo-associated influences, and also progress because of the underlying disease. A useful analysis makes these components explicit rather than attributing every observed change to disease progression.

The example here uses a continuous clinical score measured at baseline and at several follow-up times in a hypothetical placebo-controlled trial. The analysis below fits the placebo arm. It estimates the mean baseline, a saturating placebo time course, and linear progression. It also accounts for repeated observations from the same participant and assesses both parameter uncertainty and predictive performance. It does not estimate a drug effect from placebo-arm records.

To define the likelihood used in the worked analysis, assume \(b_i\sim N(0,\tau^2)\), \(\epsilon_{ij}\sim N(0,\sigma^2)\), and that all participant random intercepts and measurement errors are mutually independent. The follow-up times are treated as fixed, and the same \(\tau^2\) and \(\sigma^2\) apply across participants and occasions. These assumptions give the model a full probability distribution, not merely a mean and variance. If the score has a different clinical direction, the signs and interpretations of \(P\) and \(g\) must be adjusted accordingly.

At \(t=0\), the two time-dependent terms are zero, so the expected score is \(B\). This makes the baseline interpretation explicit. The term \(P(1-e^{-kt})\) is zero at baseline and approaches \(P\) as time increases; \(k\) has units of inverse years. The progression term \(gt\) changes linearly with time, and \(g\) has units of score points per year. The components are additive, so \(P\) and \(gt\) describe distinct contributions to the modeled mean, not necessarily distinct biological causes.

Why Repeated Measurements Require a Mixed-Effects Likelihood

Repeated outcomes from one person are not independent under this model: they share \(b_i\). For participant \(i\), collect the observations into a vector \(Y_i\), and let \(\mu_i\) be the vector of model means without \(b_i\). Integrating over the normally distributed random intercept gives a multivariate normal distribution:

Here \(\mathbf{1}\) is a vector of ones and \(I\) is an identity matrix of the appropriate size. Thus two different measurements from the same participant have covariance \(\tau^2\), while each observation has variance \(\tau^2+\sigma^2\). Different participants are independent. The marginal log-likelihood, summed over participants, is

where \(m_i\) is the number of observations for participant \(i\). Maximum likelihood fitting chooses \(B,P,k,g,\tau^2,\sigma^2\) to maximize this expression, subject to sensible constraints such as \(k>0\), \(\tau^2\geq0\), and \(\sigma^2>0\). Because the mean depends nonlinearly on \(k\), numerical optimization is generally needed.

The likelihood derivation also explains why fitting a straight line to every record as if observations were independent is inappropriate: it assigns too much independent information to clusters of measurements from the same participant. Fitting only the visit means can also lose information about within-participant variability and the covariance structure. A valid analysis checks that the model has converged, that variance estimates are plausible, and that the fitted parameters are identifiable from the observation times and data.

Worked Analysis of Hypothetical Trial Data

Consider a hypothetical study with 24 participants assigned to placebo and measured at baseline and years 2, 4, and 6. For illustration, the table shows visit means and the corresponding fitted population means from a Gaussian mixed-effects maximum-likelihood fit. The fitting uses the participant-level records, not only the visit means. The visit summaries are provided to make the predictions and residuals transparent.

The fitted values follow directly from the estimates. At year 2, for example, \(e^{-0.476}\approx0.6213\), so the fitted mean is \(50.1+7.80(1-0.6213)+0.720\approx53.774\) points. At year 4, \(e^{-0.952}\approx0.3859683\), giving \(50.1+7.80(1-0.3859683)+1.440\approx56.329\) points. At year 6, using the unrounded exponential in the calculation gives approximately 58.18965, which rounds to 58.190; the residual is \(57.9-58.18965\approx-0.2897\) points, or \(-0.290\) to three decimals. Rounding intermediate values can change the last displayed digit, so retain precision until the final result.

Predictive Checks and Failure Modes

A fitted mean can be plausible while the model fails to describe individual trajectories. A first check is to plot observed-minus-fitted residuals against time and against fitted values. A systematic curve over time may indicate that the assumed progression shape or placebo time course is inadequate. A widening residual spread suggests nonconstant measurement variability. Residuals within participants can also reveal patterns that a random intercept alone does not capture, such as serial correlation or participant-specific slopes.

For the Gaussian mixed-effects model, the marginal variance of one observation is \(\tau^2+\sigma^2\). A marginal standardized residual divides the observation's deviation from its fitted population mean by \(\sqrt{\tau^2+\sigma^2}\). This is a screening diagnostic, not an independent standard-normal test: residuals from the same participant are correlated, and parameters were estimated from the data. More complete predictive checks simulate replicate datasets from the fitted model, retaining the visit schedule and participant structure, then compare features such as visit means, within-person changes, and the spread of individual trajectories with those observed.

An important limitation is that a random intercept allows participants to differ in their baseline level but assumes they share the same placebo and progression parameters. If participants show genuinely different rates of progression, residual patterns may persist even when the population mean curve looks adequate. Adding random slopes or further complexity may help, but requires enough information to estimate the additional variation reliably. Likewise, if baseline is measured with substantial error, treating it as a perfectly known covariate or intercept can distort interpretation; the measurement model must reflect the study design.

The model is also not the only reasonable approach. A linear-only trajectory omits the saturating placebo term; it may be a useful simpler comparison if the observed time window cannot distinguish the components. A nonlinear disease trajectory instead of linear \(gt\) may be appropriate when the data support curvature, but should not be added solely to improve a few residuals. Generalized estimating equations offer a population-averaged approach with a working covariance structure and robust standard errors, while the Gaussian mixed-effects likelihood explicitly models participant-level heterogeneity and supports individual predictions under its assumptions. These approaches answer related but not identical questions. Model comparison should consider fit, diagnostics, parameter identifiability, and the scientific meaning of the terms, not just a single fit statistic.

Common Mistake: Treating the Components as Automatically Identifiable

A common error is to assume that a good numerical fit proves the baseline, placebo magnitude, placebo rate, and progression rate are each separately determined by the data. The model's components can trade off: over a short follow-up window, a slowly saturating placebo term may resemble a linear trend, making \(P\), \(k\), and \(g\) difficult to distinguish. Baseline is informed by baseline observations, while separation of the time-dependent terms depends on follow-up duration, visit timing, and signal relative to noise.

The correction is to inspect uncertainty and parameter dependence, examine whether different starting values lead to similar solutions, and assess predictions over the observed time range. Wide intervals, unstable estimates, or strongly correlated estimates indicate weak identification even if the optimizer reports convergence. If the purpose is to predict within the observed time window, predictions may be more stable than individual parameters; extrapolating beyond that window is especially uncertain.