Identifiability of Placebo and Disease-Progression Parameters

Why Identifiability Matters

In a placebo-controlled progression study, an observed change over time may reflect both a response associated with placebo and the underlying course of disease. If both components change during the same study period, a model may fit the observed trajectory well while providing little information about how much change belongs to each component. The issue is not merely whether the model converges: it is whether the data can distinguish the parameters the model is intended to estimate.

Here we examine a baseline-centered mean trajectory with a placebo response that approaches a plateau and a linear progression component. The placebo time constant is treated as known for the main analysis. This isolates the question of whether placebo amplitude and progression slope can be separated; estimating the onset rate as well generally makes the problem more demanding.

A Placebo-Plus-Progression Model

Let \(t\) denote time since baseline, in weeks. Let \(B\) be the expected baseline outcome, \(P\) the amplitude of the placebo response, \(\beta\) the linear progression slope, and \(k\) a fixed placebo onset-rate constant. The model is

The units of \(P\) and \(B\) are outcome units, while \(\beta\) has units of outcome units per week and \(k\) has units of inverse weeks. The sign of \(\beta\) depends on the outcome scale: for an outcome where larger values mean worse disease, worsening progression often has \(\beta>0\). This interpretation does not change the identifiability argument.

We assume the functional forms are adequate over the study period, \(k\) is fixed rather than estimated from these data, observation times are known, and the mean observations have finite error variance. For the structural argument, we consider the expected means at the scheduled times, not random deviations around them. If \(B\) is estimated rather than known, it is also a parameter that must be distinguished from the time-varying components.

Why the Model Can Be Identifiable Yet Difficult to Estimate

Subtracting the baseline mean gives the expected change

The two contributions are separable in principle when their time functions differ enough across the sampled times. At early times, however, the placebo function is nearly linear. A Taylor expansion about \(t=0\) gives

When \(kt\) is small, the leading term dominates, so \(\Delta\mu(t)\) is approximately \((Pk+\beta)t\). The data then mainly inform the combined slope \(Pk+\beta\), rather than its two components. Later observations can reveal the curvature of the placebo response, but only if they occur at times when that curvature is appreciable and the observations are sufficiently precise.

A short justification of structural identifiability makes the distinction clear. Suppose \(B\) is known and there are two post-baseline times \(0<t_1<t_2\). The coefficients of \(P\) and \(\beta\) form the matrix with rows \((f(t_1),t_1)\) and \((f(t_2),t_2)\). Its determinant is

For \(t>0\), the ratio \(f(t)/t\) decreases strictly when \(k>0\). To see this, its derivative is

because \(e^{kt}>1+kt\) for \(kt>0\). Thus \(D>0\) when \(t_1<t_2\), and the two columns are linearly independent: the two parameters have a unique solution from exact means. This does not guarantee useful precision. If the times are close together or both are early, the determinant can be small and noise can make the separate estimates unstable.

Using Sensitivity Correlations to Diagnose Practical Identifiability

For nonlinear models, local sensitivity describes how the predicted observations change as parameters change. In this model, the sensitivities to \(P\) and \(\beta\) at time \(t_i\) are \(f(t_i)\) and \(t_i\). If these sensitivity columns are nearly proportional across the observed times, the data respond similarly to changes in the two parameters. The model then has difficulty distinguishing them.

A weighted sensitivity correlation can summarize this overlap. If \(s_{P,i}\) and \(s_{\beta,i}\) are the two sensitivities and \(w_i\) are nonnegative observation weights, define

This is a measure of alignment between sensitivity columns, not a probability and not automatically the correlation between fitted parameter estimates. Values of \(|r_J|\) near 1 indicate strong local overlap; values nearer 0 indicate more distinct sensitivity patterns. The parameter covariance from a fitted model must be checked separately.

Sampling Schedules and Alternatives

Sampling times should be chosen to reveal differences in the time functions, not merely to increase the number of observations. Consider an early-only schedule at weeks 1, 2, and 3. With \(k=0.20\) per week, the placebo sensitivities are approximately 0.1813, 0.3297, and 0.4512, while the progression sensitivities are 1, 2, and 3. These sequences are close to proportional: their uncentered sensitivity correlation is

A hypothetical schedule at weeks 0, 2, 8, and 20 has placebo sensitivities approximately 0, 0.3297, 0.7981, and 0.9817. For the positive-time observations, the corresponding sensitivity correlation is about 0.943. The later observations show the placebo component approaching its plateau while the linear progression component continues to increase, reducing local overlap. They may therefore help distinguish the contributions, although their value also depends on measurement precision, participant burden, and whether the assumed trajectory remains credible over that longer period.

Alternatives should be considered in light of the scientific question. More informative timing is preferable when feasible. External information may support fixing \(k\), but the resulting estimates of \(P\) and \(\beta\) are conditional on that fixed value. Simplifying the model, such as replacing one component with a constant shift, can improve estimability but changes what the components mean. A prior distribution or constraint can stabilize estimates, but it does not make the data themselves more informative; conclusions then depend partly on the added assumptions.

Failure Modes and Limitations

Inspect estimated parameter correlations, standard errors, confidence intervals, sensitivity patterns, and whether estimates change sharply when a time point is removed or the initial values are changed. A fitted trajectory may look smooth while its component parameters remain weakly identified. Also check whether the estimated parameters sit on bounds: a boundary estimate can reflect weak information or an overly restrictive model rather than clear evidence that the parameter has that value.

• Confusing a good fit with identifiable components. Many combinations of \(P\) and \(\beta\) can fit a single endpoint equally well.
• Treating nonzero determinant as evidence of adequate precision. A very small determinant still permits a unique exact solution, but noisy estimates can be highly unstable.
• Ignoring uncertainty in \(k\). The analysis above fixes \(k\). Estimating it jointly adds another source of similarity among time courses, particularly when observations are all early.
• Assuming later sampling always solves the problem. Later observations help only if the model remains appropriate and measurements contain useful information at those times.

The structural result depends on the specified functional forms and fixed \(k\). If the true placebo response or natural history differs from the model, apparent separation of \(P\) and \(\beta\) may be misleading. Identifiability analysis cannot establish that the model is scientifically correct; it assesses whether the parameters can be distinguished within that model and design.