Choose change scores or follow-up analysis deliberately

Define what the outcome means

A change score is follow-up minus baseline. A follow-up outcome is the measurement at the planned endpoint. These descriptions can answer related questions, but the analysis must still specify baseline adjustment, missingness and the treatment contrast. Do not choose whichever version produces the smaller p-value.

The FDA guidance on covariate adjustment addresses change-from-baseline analyses and baseline covariates in randomised trials. Before selecting an analysis, decide whether the scientific quantity is an endpoint level, a within-person change or another longitudinal summary. Record the direction and units of the scale.

Distinguish adjusted and unadjusted change

Regressing change on treatment alone is not the same specification as regressing follow-up on treatment and baseline. However, in an ordinary linear model using the same complete observations, a change outcome adjusted for baseline gives the same treatment coefficient as a follow-up outcome adjusted for baseline. Subtracting baseline changes its coefficient by one while leaving the treatment coefficient unchanged.

That algebra does not establish equivalence for every nonlinear model, missing-data procedure or differently selected sample. Check the actual data rows and model specification before using it to reconcile outputs. An apparent disagreement may reflect different inclusion rules rather than an arithmetic error.

Hypothetical example

A fictional follow-up model is Y1 = 10 + 2T + 0.6Y0 + error. Subtracting Y0 from both sides gives change = 10 + 2T − 0.4Y0 + error. The treatment coefficient remains two. The baseline coefficient changes from 0.6 to −0.4.

If an analyst instead removes baseline from the change model, the equivalence no longer follows. Nor should an analyst compare within-group significance tests and infer a treatment difference because only one group has a significant change.

Make the comparison auditable

Prepare a specification table with the outcome formula, covariates, analysis population, missing-data method and reported contrast. Reconcile the observation counts before comparing estimates. Check for impossible baseline dates, scale reversals and follow-up measurements taken outside the intended window.

Report the selected analysis with a clear rationale and retain alternatives as labelled sensitivity analyses when scientifically useful. Use the paired-measurement audit to check pairing, but remember that a paired within-group analysis does not replace a between-treatment contrast. The final interpretation should describe the actual modelled comparison, not just the name of the variable used on the left side of the equation.

Source verification: 6 November 2026. Examples are hypothetical.