Baseline Imbalance and Treatment-Effect Interpretation

Why Baseline Imbalance Changes the Comparison

Randomization makes treatment assignment independent of baseline characteristics in expectation across repeated randomizations. It does not make the treatment and control groups identical in every realized trial. A chance difference in baseline response, disease severity, or another prognostic measure can therefore affect a simple comparison of outcomes. The key question is not whether the baseline means happen to match, but how the analysis uses baseline information and what treatment contrast it estimates.

Consider a continuous outcome measured before assignment and again at follow-up, with lower values representing improvement. A difference in follow-up means compares the groups at the endpoint. A difference in mean changes compares how much each group changed from its own baseline. A baseline-adjusted regression compares follow-up outcomes while accounting for baseline. These estimates need not be numerically equal in a finite sample.

A change score \(Y_i-B_i\) is meaningful only when subtraction is meaningful—for example, when baseline and follow-up measure the same quantity on the same scale. Regression adjustment is more general: baseline and follow-up variables may use different units. Their definitions and units must be clear, and the chosen model must adequately describe their relationship.

Three Contrasts and Their Relationship

Write \(\bar{B}_1,\bar{B}_0\) for the treatment- and control-group baseline means, and \(\bar{Y}_1,\bar{Y}_0\) for their follow-up means. Define \(\Delta_B=\bar{B}_1-\bar{B}_0\) and \(\Delta_Y=\bar{Y}_1-\bar{Y}_0\). The unadjusted follow-up contrast is \(\Delta_Y\). When baseline and follow-up are measured on the same scale, the contrast in mean changes is

For a common-slope baseline-adjusted model, use

Here \(B^*\) is a reference baseline value, \(\beta_B\) is the expected follow-up difference associated with a one-unit baseline difference at fixed treatment assignment, and \(\tau\) is the model-based treatment-versus-control contrast at the reference baseline. The model assumes a linear baseline relationship over the range studied, a common baseline slope in both groups, and an additive treatment contrast that does not vary with baseline. The residual \(\epsilon_i\) has conditional mean zero given treatment and baseline. Standard model-based uncertainty calculations additionally require an appropriate residual variance model; for the usual pooled-variance standard error, residual variance is assumed constant and participant errors are independent.

With group indicators and a common baseline slope, the fitted treatment coefficient can be written in terms of the group means and fitted slope:

This identity reveals why adjustment changes the estimate: the endpoint difference is corrected by the part associated with the baseline difference, according to the fitted baseline slope. It also shows why a change-score contrast and an adjusted contrast are generally different. Their difference is

This is a sample algebraic relationship, not proof that one estimate is always correct and another always biased. Under randomization, both unadjusted endpoint and change-score contrasts can target an average treatment effect over repeated trials when their definitions and assumptions are appropriate. In a particular trial, however, baseline imbalance can make their realized values differ. Adjustment can account for prognostic baseline differences and often improve precision, provided the model is suitable.

Baseline-Adjusted Estimate in a Randomized Example

Suppose a fitted common-slope model estimates that each one-point higher baseline symptom score is associated with 0.8 points higher follow-up score, after accounting for treatment assignment. This positive slope is plausible if participants with more severe baseline symptoms tend to have higher symptoms later. The adjusted group contrast can be calculated without treating the observed baseline means as equal.

The standard error illustration depends on the stated model and variance assumptions. If residual spread changes with baseline or treatment, a model-based standard error using constant variance may be inaccurate; an appropriate heteroscedasticity-robust or design-based method may be preferable. The point estimate also depends on the common-slope specification, not only on randomization.

Baseline and Follow-Up Need Not Share Units

A baseline predictor and follow-up outcome do not have to use the same measurement scale for regression to be mathematically valid. The slope then carries the units of follow-up outcome per unit of baseline predictor. What matters is that baseline was measured before treatment assignment could affect it, and that the model represents the baseline–follow-up relationship adequately. A raw change score, by contrast, requires a meaningful subtraction.

Assumptions, Diagnostics, and Alternatives

A causal interpretation requires more than a regression equation. Treatment must be randomized as described, baseline must precede assignment, and outcomes must correspond to the assigned treatment comparison. The usual interpretation also presumes that each participant’s outcome is not affected by another participant’s assignment and that loss to follow-up or missing outcomes has not created a consequential selection bias. Randomization supports a treatment comparison; it does not guarantee that a particular regression form is correct or that missing outcomes are harmless.

For a constant adjusted contrast \(\tau\), the common-slope model should be reasonable across the observed baseline range. Examine a plot of follow-up against baseline by treatment group and inspect residuals for curvature, outliers, unequal spread, and group-specific patterns. A marked difference in slopes suggests treatment effect modification or a misspecified common-slope model. If a treatment-by-baseline interaction is needed, the treatment coefficient describes the contrast at the reference baseline \(B^*\), rather than a universal effect; an average contrast then requires averaging model predictions over a specified baseline distribution.

The major alternatives have distinct strengths and limitations:

Baseline-adjusted estimates are often more precise when baseline predicts follow-up, but adjustment is not guaranteed to improve precision in every finite sample or under every model. In a small trial with weak baseline prediction, adding a covariate can contribute little. Extrapolation is also a concern when treatment and control groups have little overlap in baseline values: an adjusted contrast may depend heavily on the assumed linear form in regions with sparse data.

Do not choose an analysis because a baseline significance test is below or above a threshold. Such a test does not establish whether a difference is clinically important, whether randomization was properly performed, or whether adjustment is needed. Describe baseline distributions, use the prespecified analysis where available, and consider prognostic relevance and model suitability. This approach is consistent with separating treatment effects from background response trends: the treatment contrast must be distinguished from differences that would occur with baseline-related outcome patterns.

Common Pitfall