Repeated Measures & Change from Baseline

Longitudinal studies measure an outcome more than once for the same participant. A baseline value may be followed by observations at days, weeks, or months. This can reveal the timing, persistence, or reversal of a response, but it also creates correlated observations, multiple possible summaries, and more opportunities to select a favorable result.[1][4]

“The group improved from baseline” is not a complete treatment-effect statement. Readers need to know what was measured, how each participant’s observations became an analysis value, how groups were summarized, which time point mattered, and how the assigned groups were compared.

An outcome has four layers

CONSORT 2025 separates four elements that are often compressed into one label.[1]

Layer	Example question	Why it matters
Measurement variable	Was the recorded value a laboratory measure, scale score, event, or device output?	Defines what was actually observed
Participant-level metric	Was the analysis based on final value, change from baseline, slope, area under a longitudinal curve, or responder status?	Different metrics answer different questions
Aggregation	Were groups summarized by means, medians, proportions, or another statistic?	Determines the group-level quantity being compared
Time point	Was the target 4 weeks, 12 months, or a trajectory across visits?	A result at one time does not automatically extend to another

For example, “change in score at week 12” specifies a measurement variable only if the scale is named, a participant-level change metric, a mean or other aggregation, and a time point. The report should also state the direction of better scores, allowed range, baseline definition, and rules for visits outside their intended windows.

Without all four layers, “improvement” could refer to a raw measurement, an average within one group, a model-based contrast, or a selected responder threshold. Those are not equivalent findings.

Final value and change from baseline are different metrics

A final-value analysis compares group outcomes at the target follow-up time, often adjusting for the baseline value. A change-score analysis first subtracts baseline from follow-up for each participant and then compares changes between groups. Both can preserve the randomized comparison when planned and analyzed appropriately.[3][4]

They should not be casually mixed. The variability of a change score depends on the variability of both measurements and their within-person correlation. Percentage change adds further dependence on the baseline denominator and can behave poorly when baseline values are near zero. ICH E9 treats change, percentage change, and area under a repeated-measures curve as derived variables whose clinical interpretation and derivation should be justified in the protocol.[4]

A baseline-adjusted final-value model and an unadjusted change-score model can produce different precision. That difference does not mean one group “really changed” while the other did not. The scientific target, model, assumptions, and prespecification determine which comparison is primary.[4][6]

Within-group change is not the randomized treatment effect

Suppose the mean score improves significantly from baseline in the intervention group but does not reach a significance threshold in the comparator group. This does not establish a significant between-group effect. The two within-group tests ask separate questions; comparing their labels, “significant” versus “not significant”, is not a test of the difference between assigned groups.[7]

The randomized treatment effect comes from a direct contrast between groups under the prespecified analysis. Both groups may improve because of natural history, regression toward the mean, background care, measurement familiarity, expectations, or calendar effects. Randomization supports comparison of those trajectories; it does not turn a one-arm before-and-after result into the primary effect.[2][3]

Report the between-group estimate and confidence interval at the prespecified time point or for the prespecified trajectory. Group-specific summaries remain useful context, but they should not replace the randomized contrast.

Repeated observations belong to the same person

Measurements from one participant are usually correlated. Treating ten visits from one person as ten independent participants exaggerates the amount of independent information and can make uncertainty appear too small. A longitudinal method should represent the within-participant structure.[3][4]

Common approaches include mixed-effects models, generalized estimating equations, repeated-measures covariance models, and prespecified summary measures. Their names do not guarantee validity. A useful report states:

which repeated values entered the model;
fixed and random effects or working correlation structure;
how time was represented;
whether treatment-by-time interaction was included;
how baseline was handled;
the effect measure at each reported time or across the trajectory;
model assumptions and estimation method; and
sensitivity analyses for important assumptions.

The unit remains the participant even when thousands of observations are recorded. More visits can improve description of a trajectory, but they do not create new randomized people.

Many time points create many chances to find a result

A study may record outcomes at every visit, yet usually only one or a small number are primary. Testing each time point independently, each scale subdomain, several change definitions, and multiple subgroups creates a large family of potential findings. If the most favorable combination is selected after inspection, its apparent strength is overstated.[1][4]

ICH E9 identifies repeated evaluation over time as one source of multiplicity. The protocol and analysis plan should therefore identify the primary outcome and time, any hierarchy among secondary times, and the planned strategy for controlling or interpreting multiple comparisons.[4]

A repeated-measures model is not automatically a multiplicity correction. One model may estimate a global treatment-by-time pattern, a difference averaged across times, or contrasts at several visits. Readers should identify which quantity was designated primary and whether additional contrasts were confirmatory, supportive, or exploratory.

Area under a trajectory and responder status answer other questions

An area-under-the-longitudinal-curve summary combines values across a fixed time window. It can reflect accumulated outcome burden or average response over that window, depending on its definition. Two participants may have the same area despite different early and late patterns, so the time window and interpretation must be explicit.[4]

A responder analysis classifies participants according to a threshold, such as achieving at least a stated change. This may be easier to communicate but discards information about how far values lie above or below the threshold. The threshold, time point, handling of death or missing observations, and rationale must be prespecified. Trying multiple thresholds and highlighting the most favorable one is another form of multiplicity.[1][5]

Final value, change, trajectory, area, and responder status are complementary only when each is clearly labeled. They should not be swapped after results are known or treated as repeated confirmations of the same hypothesis.

Missing longitudinal data remain unobserved

At each time point, report the number of participants contributing observed data, not only the number randomized. A participant may have baseline and early visits but lack the primary endpoint, or miss one visit and return later. These patterns contain different information and should be described by group.[2]

Mixed models and other repeated-measures methods can use incomplete observed records under stated assumptions. They do not recover the actual unobserved values without assumptions, and they do not automatically remove bias caused by differential dropout. If worsening response, adverse events, treatment discontinuation, rescue treatment, or another intercurrent event affects whether an observation is available, the main estimate may depend strongly on the chosen strategy.[2][4][5]

ICH E9(R1) distinguishes the treatment-effect question from the estimator and calls for sensitivity analysis to examine robustness to deviations from assumptions. A credible report defines how intercurrent events are handled, explains the missing-data model, and compares planned analyses that stress important uncertainties.[5]

A longitudinal-results reading checklist

Name the variable, metric, aggregation, and primary time point.
Confirm that baseline precedes randomization or assigned treatment.
Identify whether the target is final value, change, trajectory, area, or response.
Locate the direct between-group contrast and its confidence interval.
Count participants with observations at every reported time.
Check whether the model preserves within-participant correlation.
Record the time structure, interactions, covariates, and assumptions.
Compare primary, secondary, supportive, and exploratory time points.
Trace missingness and intercurrent events by assigned group.
Keep interpretation within the prespecified follow-up window.

Important limitations

A significant within-group change is not evidence of a significant between-group treatment effect.
Repeated observations are correlated and cannot be counted as independent participants.
Change scores, final values, trajectories, areas, and responder outcomes are different metrics.
Multiple time points and thresholds create multiplicity unless their hierarchy is clear.
A longitudinal model uses assumptions; it does not reveal unobserved outcomes.
Differential dropout can bias a repeated-measures analysis despite sophisticated modeling.
Reporting guidance improves transparency but does not approve a method or conclusion.
A longitudinal result cannot establish the identity or comparability of another material.

Frequently asked questions

Is improvement from baseline enough to show a treatment effect?

No. In a randomized trial, the primary treatment effect is a direct comparison between assigned groups. Within-group change can reflect many influences shared or unequally distributed across time.

Is change from baseline always better than comparing final values?

No. Their performance depends on the design, baseline-follow-up correlation, model, and target effect. A baseline-adjusted final-value analysis can often be efficient; the prespecified estimand and assumptions matter.

Why can’t every visit be treated as a separate participant?

Because values from the same person tend to be correlated. Ignoring that correlation overstates independent information and can understate uncertainty.

Does a repeated-measures model solve missing data?

No. It can use available repeated observations under assumptions, but it cannot observe missing outcomes. Differential dropout and intercurrent events still require explicit strategies and sensitivity analysis.

What does a treatment-by-time interaction mean?

It assesses whether the between-group contrast varies across the modeled time points. Its interpretation depends on how time and outcomes were modeled and whether the interaction was prespecified.

Is an area-under-the-curve summary the same as the final outcome?

No. It combines information across a defined window, whereas a final outcome concerns one time. Similar areas can conceal different trajectories, and a final difference can coexist with similar accumulated responses.

Can longitudinal findings be converted into practical-use instructions?

No. A study’s timing and analytic structure describe its protocol and evidence. They do not provide preparation, administration, dosing, or self-use guidance and do not validate another material.

Are you over 21?

Repeated Measures, Change from Baseline, and Time Points