Stable Natural-History Models for No-Drug Response

What “Stable” Means in a No-Drug Model

A no-drug response is the outcome expected in the absence of active drug, given the background conditions represented in the model. A stable natural-history model is the simplest case: the expected response does not systematically change over the observation period. Individual measurements can still vary. The key distinction is between a changing expected response and fluctuations around an expected response that remains fixed.

For example, repeated symptom scores may move up and down even if the underlying average symptom level is unchanged. Treating every movement as disease progression mistakes variability for a trend. Conversely, averaging across time can hide a real trend if the expected response is changing. Earlier material on baseline, placebo, and natural-history response distinguishes these components; here the focus is the fixed-mean natural-history case and what repeated observations can tell us about it.

In the second model, the conditional mean for subject \(i\), given that subject’s stable deviation, is

It does not depend on \(t_{ij}\). The subject may have a persistently higher or lower response than the population average, but the model does not assign that difference a systematic time trend. The observed response can nevertheless vary at each measurement occasion through \(\epsilon_{ij}\).

Assumptions and Interpretation

The fixed-mean model makes a claim about the expected response, not about identical observations. A common version of the repeated-measures model assumes \(E(b_i)=0\), \(E(\epsilon_{ij})=0\), and that the subject effects and measurement deviations are independent. It also commonly assumes that the deviations have a common variance, written \(\operatorname{Var}(b_i)=\sigma_b^2\) and \(\operatorname{Var}(\epsilon_{ij})=\sigma_w^2\). The term “within-subject” variance refers to \(\sigma_w^2\); \(\sigma_b^2\) describes stable between-subject variation.

For the simplest calculations, assume the \(\epsilon_{ij}\) are independent across occasions within a subject, and that different subjects are independent. These assumptions are stronger than the constant-mean assumption itself. If successive deviations are correlated, the expected response may still be stable, but standard errors and variance estimates based on independence can be wrong. Normal distributions for \(b_i\) and \(\epsilon_{ij}\) are sometimes assumed for likelihood-based estimation or exact small-sample inference; the mean-and-variance interpretation does not require normality.

A no-drug model also depends on what “no drug” means in the study. If participants receive placebo, the observed control response may include a placebo component. The stable natural-history term is not automatically the entire observed control-arm mean. The interpretation must match the model components and the conditions under which observations were collected.

Why Repeated Measurements Have Two Sources of Variation

The random-intercept model separates a subject’s stable level from occasion-specific fluctuation. For two distinct occasions \(j\) and \(k\) for the same subject, under the independence assumptions above,

The covariance is positive because both measurements contain the same \(b_i\). Thus repeated observations from one subject are not independent when subjects differ in their stable levels. The correlation between two occasions is \(\sigma_b^2/(\sigma_b^2+\sigma_w^2)\) under this model. This correlation is not evidence of progression: it can arise simply because some subjects tend to remain above the population mean and others below it.

For \(N\) subjects with \(m\) measurements each, the grand mean averages both sources of variation. Under the stated independence assumptions, its variance is

More measurements per subject reduce the contribution of within-subject fluctuation to the grand mean, but they do not provide as much independent information about the population mean as the same number of measurements from new subjects. Each additional subject contributes a new draw of the stable subject effect; another visit to an existing subject does not.

Estimating Stable Differences from a Small Repeated-Measures Set

When repeated measurements are available for several subjects, compare variability within subjects with variability between their averages. The following balanced data set illustrates the distinction. The values are hypothetical symptom scores in points.

Checking Whether Stability Is Plausible

A stable model should be checked against the ordering of observations in time, not just against their overall spread. Plot individual observations or subject means against time. Under the model, residuals after accounting for stable subject levels should fluctuate around zero without a consistent temporal pattern. Runs of increasing residuals, a shared upward or downward shift across subjects, or a change concentrated at later visits can indicate that the fixed-mean assumption is inadequate.

Other diagnostics matter too. Check whether residual spread is similar across subjects and visits, whether subjects have markedly different patterns, and whether measurements are missing in a way related to worsening or improvement. Assay drift, seasonal changes, changes in assessment procedures, or concurrent care can create time patterns that a fixed natural-history mean does not explain. Serial correlation can also make successive deviations move together even when there is no systematic change in their expected value.

Alternatives, Limitations, and a Common Pitfall

The simplest alternative to a random-intercept stable model is a single fixed mean with independent residuals. It may be adequate when between-subject differences are negligible or when the analysis concerns one subject. A random-intercept model is more appropriate when subjects have stable differences and repeated observations. A time-dependent model is needed when the expected response changes systematically with time; a time-varying placebo component may also be relevant when the observed response includes placebo response. These alternatives answer different questions: adding a stable subject effect does not represent progression, while adding a time trend changes the expected-response assumption.

A fixed-mean model is a useful reference, not a guarantee that untreated response is truly constant. It can miss slow changes over a short observation window, abrupt changes between visits, or distinct subgroups with different time patterns. Sparse visits may make a temporal pattern hard to detect. Likewise, an apparently flat group average can conceal subjects changing in opposite directions. Diagnostics and study design are therefore part of interpreting the model, not optional decoration.

• Pitfall: treating every fluctuation as progression. A sequence of different measurements does not itself establish that the expected response changed. Under a stable model, observations vary around a fixed expected value.
• Pitfall: treating repeated observations as independent subjects. Measurements on the same subject share the stable subject effect. Ignoring this dependence can understate uncertainty.
• Pitfall: calling every no-active-drug observation “natural history.” A control response may also reflect placebo or other study components. Interpret the model term according to what participants actually received and what components are included.