Modeling Change From Baseline Without Losing the Baseline Signal

Why the Choice of Response Scale Matters

A treatment response can be expressed as the value observed at follow-up or as the difference between that value and baseline. These descriptions sound like alternative ways to record the same outcome, and algebraically they are: if the same baseline value is subtracted from every follow-up observation, the change is fully determined by the two measurements. But the model fitted to the response determines what patterns are represented and what assumptions are made. Subtracting baseline can remove a useful signal, impose a particular baseline relationship, or create dependence when baseline is measured with error.

The distinction matters in PK/PD analysis because a response may vary with the subject’s starting state as well as with concentration or treatment. A change score can make a treatment effect easier to communicate, but it does not automatically adjust for baseline differences or make baseline irrelevant. The key question is whether the absolute-response model and the change model encode the same baseline relationship.

Algebraic Relationship and Its Assumptions

Suppose the follow-up response is modeled as a linear function of baseline \(B_i\) and a treatment indicator \(T_i\):

Here \(B_i\) is the baseline value used in the model, \(T_i=1\) for treatment and \(T_i=0\) for the comparison condition, and \(\varepsilon_{it}\) is a mean-zero follow-up error. If the change score subtracts this same \(B_i\), then \(D_{it}=Y_{it}-B_i\), so substitution gives:

Thus the models are exact algebraic re-expressions when they use the same baseline quantity, the same observations, and compatible error assumptions. In particular, an absolute model with a baseline coefficient \(\beta_B\) corresponds to a change model with baseline coefficient \(\beta_B-1\). A change model that omits baseline is not generally equivalent: it implicitly sets \(\beta_B-1=0\), or \(\beta_B=1\), in this simple linear formulation.

This equivalence is conditional on the chosen model form. If the absolute model includes nonlinear baseline effects, treatment-by-baseline interactions, or other covariates, the change model must include the corresponding transformed terms to preserve the same predictions. For example, if the absolute model contains \(\beta_{BT}B_iT_i\), that interaction remains in the change model. Subtracting baseline does not erase the need to represent such effects.

Why Subtracting a Noisy Baseline Can Induce Dependence

The algebra above treats the baseline value as the same quantity in both models. In practice, an observed baseline may differ from the subject’s underlying baseline state. Write the observed baseline as \(Y_{i0}=B_i+u_{i0}\), where \(B_i\) is the underlying state and \(u_{i0}\) is baseline measurement error. Suppose the follow-up observation is \(Y_{it}=B_i+\Delta_i+u_{it}\), where \(\Delta_i\) is the subject’s true change and \(u_{it}\) is follow-up measurement error. The observed change is then:

The baseline error appears with a minus sign in the change. If the errors at baseline and follow-up are independent of each other and of the true baseline and true change, then:

Even if true change is independent of true baseline, so that \(\operatorname{Cov}(B_i,\Delta_i)=0\), the covariance is negative by the baseline-error variance. A subject whose baseline measurement error happens to be high will tend to have a smaller measured change, because that same high error was subtracted. This is a mathematical source of baseline-change dependence; it is not, by itself, proof that subjects with high true baseline values respond less.

When Change Scores Obscure Baseline Effects

A change score can be useful when the scientific quantity of interest is explicitly the amount of within-subject change. But a model that relates change only to treatment may conceal systematic differences in change associated with baseline. It may also combine a treatment contrast with baseline imbalance if the treatment groups start at different levels.

A baseline-adjusted follow-up model instead estimates follow-up response conditional on baseline. Its treatment coefficient describes the modeled difference in follow-up response between treatment conditions at the same baseline value, assuming the specified model is adequate. A change model can yield the same treatment contrast if it includes the appropriately transformed baseline terms. The problem is not using change as the outcome; it is imposing an unsupported restriction by omitting relevant baseline dependence.

Choosing and Checking a Formulation

The main alternatives are an absolute follow-up model with baseline as a predictor, a change model that retains the necessary baseline terms, and a change-only model. When models are specified to be algebraically equivalent, they should give the same fitted values and treatment contrasts for the same data and fitting method, subject to the same error structure. The change-only model is simpler, but that simplicity comes from a restriction, not from a guaranteed property of change scores.

Before interpreting a fitted result, check the following:

• Define the baseline quantity. Clarify whether the model subtracts one observed baseline value, an average of repeated baseline measurements, or another estimate. The algebra and error implications depend on what is subtracted.
• Check baseline patterns. Plot change against baseline and inspect whether the mean change varies with baseline. A trend may reflect true baseline-dependent response, measurement error, or both; the plot alone cannot distinguish them.
• Compare implied predictions. For representative baseline values, calculate predicted absolute outcomes and changes. Confirm that a proposed change model preserves the intended baseline relationship.
• Inspect residual behavior. Examine residuals against fitted values, baseline, time, and treatment. Patterns may signal a misspecified mean relationship or nonconstant residual variance. A good-looking residual plot does not remove measurement error in baseline.
• Consider baseline imbalance. In a randomized study, chance imbalance can occur. A crude change contrast may still reflect that imbalance when change depends on baseline; report the modeled contrast and the baseline assumptions clearly.

A change score is limited by the information in the measurements from which it is constructed. It does not by itself separate biological response from natural time trends, placebo response, or measurement error. Those are distinct components that require an explicit model. Nor does a negative baseline-change association alone prove a biological mechanism: the subtraction itself can induce such an association.