Baseline and Natural-History Models for Precision Planning

Why Baseline and Natural History Matter for Precision

A planned treatment contrast depends not only on the expected treatment effect, but also on how variable the endpoint is in the population under study. Baseline measurements and placebo or natural-history data can help characterize that variability and the expected course without active treatment. Used carefully, they can improve planning; used uncritically, they can make a design look more precise than the evidence supports.

The central distinction is between the expected trajectory and variability around that trajectory. A placebo group may show a mean decline over time, for example, but that mean decline alone does not tell us how dispersed individual endpoint values are. Conversely, a baseline-endpoint relationship can help predict endpoint variability, but does not establish what the untreated mean trajectory will be in a new trial.

For planning, the most useful baseline relationship is often a linear regression of the endpoint on baseline. A simple model is

Here \(i\) indexes participants, \(\mu_B\) is the target-population mean at baseline, \(\beta\) is the baseline-outcome slope, and \(\sigma_e\) is the residual standard deviation. Centering baseline makes \(\alpha\) the expected follow-up endpoint at the population mean baseline. The model assumes a linear conditional mean, a common slope, constant residual variance, and independent participants. Normal residuals are additionally needed for exact small-sample normal-theory calculations, but not for the variance identity below.

From Baseline Association to Endpoint Variability

If the linear model is appropriate, the residual variance can be estimated from the endpoint variance and the baseline-endpoint correlation. Let \(\sigma_T\) and \(\sigma_B\) denote the population standard deviations of \(Y_T\) and \(Y_B\), and let \(\rho\) be their correlation. The linear regression slope is \(\beta=\rho\sigma_T/\sigma_B\). The variance of the linear prediction is \(\beta^2\sigma_B^2=\rho^2\sigma_T^2\). Because total endpoint variance is the sum of explained and residual variance under the model,

This calculation uses the population linear projection. In practice, sample estimates of the endpoint standard deviation and correlation provide planning estimates, not known population values. The relation is useful when the baseline and endpoint are measured in a comparable population and at the intended assessment time.

The calculation also shows how the baseline association can inform a planned treatment contrast. If both randomized groups have \(n\) participants, share a common baseline slope and residual variance, and the contrast is estimated with baseline adjustment, a large-sample planning approximation is

For an unadjusted comparison of endpoint means under equal group sizes and a common endpoint variance, the corresponding standard error is approximately \(\sqrt{2\sigma_T^2/n}\). These expressions isolate the variability contribution to precision; they do not determine the treatment effect, and they are not a substitute for a complete analysis plan. The baseline-informed approximation relies on the stated regression assumptions and on estimating the treatment contrast consistently with that model.

Using Placebo and Natural-History Trajectories

Placebo or natural-history data provide evidence about the expected outcome course at specified times. A simple starting point is a mean trajectory, such as a linear model for the expected endpoint over time. The trajectory can support selection of a plausible control mean at the planned endpoint and can reveal whether a constant rate of change is a reasonable approximation over the observed interval.

A fitted mean curve does not by itself estimate endpoint variability. Variability must be assessed from participant-level deviations around the relevant model or from an appropriate endpoint standard deviation. In repeated measurements, those deviations may be correlated within a participant; treating measurements as independent would overstate the information. For the present planning task, the key is to use the control trajectory for its mean-course information and a suitable participant-level estimate for endpoint dispersion.

Quantifying the Contribution to Planned Precision

The difference between total and residual endpoint variability can be consequential for planning. The next example compares standard errors under an unadjusted endpoint calculation and a baseline-informed calculation. It illustrates the variability contribution only; the assumed treatment contrast and all other design choices remain separate inputs.

Model Fit, Diagnostics, and Limits

A planning model should be challenged before its estimates are used. For the baseline-outcome model, inspect a plot of follow-up endpoint against baseline and residuals against fitted values and baseline. Curvature suggests that a straight-line conditional mean may be inadequate; a fan-shaped residual pattern suggests nonconstant variance. Large influential observations can materially change the estimated slope, correlation, or residual standard deviation. A strong association in one subgroup but a different association in another may invalidate a common-slope assumption.

For a natural-history trajectory, inspect the observed means and their uncertainty across time, along with participant-level residual patterns where available. Systematic curvature warns against assuming a linear trend; an apparent change in trend may also reflect changes in who remains observed. A small residual sum of squares for a few aggregate means is not enough to validate an individual-level model. Extrapolating past the observed follow-up period requires additional scientific justification.

Historical and natural-history evidence also requires transportability: the population, endpoint definition, assessment schedule, background care, and data collection should be sufficiently comparable to the planned trial. Differences in these features can shift the expected control course or alter endpoint dispersion. A placebo trajectory is an estimate for the population and trial conditions in which it was observed; it does not guarantee the same course in a future trial. Uncertainty in the estimated standard deviation, correlation, slope, and trajectory should therefore be considered in planning, rather than treating estimates from a finite dataset as exact.

The main alternatives answer different planning questions. An unadjusted endpoint standard deviation is straightforward and avoids relying on a baseline model, but does not use the predictive information in baseline. A baseline-informed residual variance can give a more relevant precision estimate when its relationship is stable and well fitted, but is sensitive to model misspecification and transportability. A placebo trajectory gives direct evidence about the expected control course, while a natural-history cohort may be more vulnerable to differences in care, selection, and measurement. These sources can inform one another, but a mean trajectory cannot substitute for a variability estimate, and a baseline correlation cannot substitute for evidence about the control mean over time.

In practice, report the source and assessment time for each planning input, the model used, its assumptions, and the diagnostics that support it. When evidence is limited, compare plausible values rather than presenting a single estimated precision as certain. The introductory material on model-based endpoint planning motivates matching the endpoint to the scientific question; here, the additional task is to make the variability and control-course assumptions explicit enough to scrutinize.