Change-From-Baseline Contrasts

What Does Change From Baseline Measure?

In a repeated-measures study, measurements may be taken at baseline and at several follow-up visits. A focused question is often whether the outcome at a particular follow-up differs from baseline. Another is whether the average outcome across follow-up visits differs from baseline. These are distinct questions, and each can be represented by a contrast: a weighted combination of visit means whose coefficients sum to zero.

The direction of the contrast matters. Here, change is defined as follow-up minus baseline. A positive estimate therefore indicates that the outcome is higher at follow-up than at baseline; a negative estimate indicates it is lower. Whether an increase is beneficial depends on what the outcome measures, not on the sign alone.

Two Common Baseline Contrasts

For a comparison of follow-up visit \(k\) with baseline, assign coefficient \(-1\) to baseline, \(+1\) to visit \(k\), and \(0\) to all other visits. The contrast is \(\mu_k-\mu_0\). It measures the mean change from baseline to that particular visit, in the outcome’s units.

For a comparison of the average of all \(m\) follow-up means with baseline, assign coefficient \(-1\) to baseline and \(1/m\) to each follow-up visit. This contrast is \(\frac{1}{m}\sum_{j=1}^{m}\mu_j-\mu_0\). It gives equal weight to each follow-up visit, regardless of the time between visits. It is not a trend contrast: it summarizes the average follow-up level rather than a rate or pattern of change over time.

Coefficients determine the scale and meaning of a contrast. Multiplying every coefficient by a constant also multiplies the contrast by that constant, so a rescaled contrast can have a different interpretation. For example, coefficients \((-m,1,\ldots,1)\) represent the sum of follow-up means minus \(m\) times the baseline mean, which is \(m\) times the average follow-up change. The equal-weight average version is usually easier to interpret in the original outcome units.

Estimating a Contrast and Forming an Interval

Let \(Y_{ij}\) be the outcome for participant \(i\) at visit \(j\), and let \(\bar{Y}_j\) be the sample mean at visit \(j\). The estimated contrast is obtained by replacing each population mean in the contrast with its sample mean: \(\hat{\theta}=\sum_{j=0}^{m}c_j\bar{Y}_j\). Equivalently, calculate a contrast score for each participant and average those scores.

The equality follows by rearranging the sums: \(\sum_jc_j\bar{Y}_j=\sum_jc_j(\frac{1}{n}\sum_iY_{ij})=\frac{1}{n}\sum_i\sum_jc_jY_{ij}=\bar{d}\). Thus, the relevant observations for a particular contrast are the participants’ contrast scores, not the visit means treated as if they came from independent groups.

For a confidence interval based on a paired \(t\) procedure, the model and assumptions are as follows: participant vectors of repeated outcomes are independent across participants; the same participants have observed outcomes at the visits needed to calculate the contrast; and the participant-level contrast scores are independent and identically distributed normal observations, with a common mean and variance, for exact \(t\) inference. With a sufficiently large sample, approximate normality of those scores may be adequate. The variance structure among visits can otherwise be unrestricted for this single contrast. In particular, a single contrast does not require the sphericity assumption used for certain omnibus repeated-measures tests.

The interval describes uncertainty about the population mean change. In repeated use, the specified procedure will cover the fixed population contrast at its stated rate, under its assumptions. A 95% interval does not mean there is a 95% probability that this particular fixed parameter lies inside the calculated interval.

Worked Example: Building an Average-Change Contrast From Visit Means

Why the Contrast Must Match the Question

A contrast for one visit and a contrast for average follow-up answer different questions. If a treatment effect rises at one visit and falls at another, an average can obscure that pattern. Conversely, testing each visit separately does not directly estimate the average follow-up change. Define the scientific target first, then choose coefficients that express it.

Confidence intervals calculated separately for several follow-up visits are pointwise: each addresses its own contrast. A collection of such intervals does not automatically provide a specified overall coverage rate for all visits together. This matters when interpreting several comparisons, but the interval formula here applies to the single contrast specified. Also, an interval that includes zero does not establish that the true change is zero; it indicates that the data and procedure do not distinguish the contrast from zero at the corresponding two-sided level.

• Wrong: “The coefficient for baseline should be \(+1\), because baseline is the reference.”
• Correction: With change defined as follow-up minus baseline, the baseline coefficient is \(-1\) and the follow-up coefficient is \(+1\). Reversing both signs gives baseline minus follow-up and reverses the interpretation.
• Wrong: “The standard error can be calculated by treating the baseline and follow-up means as independent.”
• Correction: The same participants supply both measurements. Form each participant’s contrast score first and use the variation among those scores.