Analyzing Change from Baseline in Clinical Trials

Introduction

Many clinical trials measure a continuous outcome at baseline and again after treatment. Examples include blood pressure, cholesterol, tumor measurements, pulmonary function, pain scores, laboratory values, and many patient-reported outcomes.

The central question is often not simply: What was the outcome at Week 12?

Instead, investigators want to know: How much did the outcome change from baseline, and was the change different between treatment groups?

This is the basic idea behind a change-from-baseline analysis.

What Is Change from Baseline?

Let:

• \(Y_{i0}\) = baseline measurement for patient \(i\)
• \(Y_{ij}\) = post-baseline measurement at visit \(j\)

The change from baseline at visit \(j\) is:

For example, suppose a patient's systolic blood pressure is 150 mmHg at baseline and 135 mmHg at Week 12.

The patient has therefore experienced a 15 mmHg reduction.

Change Is a Derived Endpoint

Change from baseline is not a directly observed variable.

It is calculated from two measurements:

This distinction becomes important because the statistical properties of the change score depend on both the baseline and post-baseline measurements.

If:

and:

then:

Thus, the correlation between baseline and follow-up measurements can have a major effect on the precision of a change score.

Within-Group Change vs. Between-Group Difference

This is one of the most important concepts in clinical trial analysis. Suppose:

The treatment group changed by -15 mmHg and the placebo group changed by -7 mmHg.

The treatment effect on the change scale is therefore:

Thus, the treatment reduced systolic blood pressure by an estimated 8 mmHg more than placebo.

Why "Significant in One Group but Not the Other" Is Not Enough

A common but incorrect argument is:

"The treatment group significantly improved, while placebo did not, so the treatment works."

This reasoning is statistically invalid.

Suppose the treatment-group estimate has:

while the placebo-group estimate has:

The fact that one p-value is below 0.05 and the other is above 0.05 does not establish that the treatment effects differ.

The appropriate hypothesis concerns the treatment contrast:

versus:

Simple Analysis of Change Scores

One straightforward approach is to calculate each patient's change from baseline and compare the mean changes between randomized treatment groups.

For two treatment groups, the estimated treatment effect is:

A two-sample t-test can then be used to compare the change scores when its assumptions are appropriate.

Example

The estimated treatment difference is:

The estimated treatment effect is therefore a 7.8-unit greater reduction with treatment.

Advantages of the Simple Change-Score Approach

• Easy to understand and explain.
• Each patient's change is explicitly calculated.
• The treatment contrast has a straightforward clinical interpretation.
• It is appropriate in many simple randomized-trial settings.

Limitations of the Simple Change-Score Approach

• It does not necessarily make the most efficient use of baseline information.
• It can be less precise than baseline-adjusted ANCOVA.
• It is inconvenient when there are multiple post-baseline visits.
• Patients missing the selected follow-up visit may be excluded from the simple analysis unless an appropriate missing-data strategy is used.
• It does not automatically address baseline imbalance through covariate adjustment.

Baseline Adjustment with ANCOVA

A commonly preferred approach for a single post-baseline endpoint is analysis of covariance (ANCOVA).

Instead of analyzing the change score directly, ANCOVA models the post-baseline outcome while adjusting for baseline.

A simple model is:

where:

• \(Y_i\) is the post-baseline outcome.
• \(\operatorname{Treatment}_i\) indicates treatment assignment.
• \(Y_{i0}\) is the baseline value.
• \(\beta_1\) represents the treatment effect after baseline adjustment.

ANCOVA vs. Analysis of Change

Why ANCOVA Can Be More Efficient

Suppose baseline and follow-up measurements are strongly correlated.

The correlation allows baseline to explain some of the variability in the follow-up outcome.

The ANCOVA treatment estimate can therefore have a smaller standard error than an unadjusted comparison of change scores.

Conceptually:

If baseline adjustment reduces residual variability, the treatment comparison becomes more precise.

The Role of Baseline Imbalance

Randomization is intended to balance baseline characteristics in expectation, but finite samples can still exhibit differences.

For example:

Both groups have the same Week 12 mean, but the treatment group started with a substantially higher baseline value.

A raw comparison of Week 12 outcomes alone may therefore obscure the clinical context.

Baseline adjustment provides a principled way to compare treatment groups while accounting for baseline values.

Baseline Is Not Usually a Post-Randomization Covariate

Baseline measurements are collected before treatment is initiated.

Therefore, baseline is generally suitable for adjustment in randomized treatment comparisons.

This differs from variables measured after treatment begins.

Choosing the Analysis Time Point

Many trials have several post-baseline assessments. For example:

If the primary objective concerns Week 12, a prespecified Week 12 comparison may be appropriate.

If the scientific question concerns the entire trajectory over time, a repeated-measures analysis may be more informative.

Repeated-Measures Analysis

When patients have multiple post-baseline measurements, repeatedly analyzing each visit separately can create several problems.

• Multiple statistical tests may inflate the type I error.
• Measurements within a patient are correlated.
• Patients may have incomplete follow-up.
• The treatment effect may change over time.

A repeated-measures model explicitly accounts for the longitudinal structure of the data.

Mixed Models for Repeated Measures

A commonly used approach in clinical trials is the mixed model for repeated measures (MMRM).

A conceptual model is:

The model can estimate treatment differences at individual visits while accounting for within-patient correlation.

What Does the Treatment-by-Visit Interaction Mean?

The treatment-by-visit interaction allows the treatment effect to vary across time.

For example, suppose the estimated treatment effects are:

The treatment effect appears to increase over time.

A treatment-by-visit interaction provides a framework for formally modeling this pattern.

Change from Baseline in an MMRM

An MMRM may model either the observed post-baseline outcome with baseline included as a covariate or a change-from-baseline outcome directly.

For a change-score formulation:

and the model can be written conceptually as:

The exact model specification should follow the prespecified statistical analysis plan.

Least-Squares Means

Clinical trial analyses frequently report least-squares means (LS means), also called estimated marginal means.

These are model-based estimates of the mean outcome or mean change after accounting for covariates and other model terms.

Suppose an ANCOVA produces:

The estimated treatment difference is:

The clinical interpretation is: The adjusted mean reduction was 7.7 units greater with treatment than with placebo.

Confidence Intervals Are Essential

A treatment effect should not be reported only with a p-value.

A confidence interval communicates the precision and range of effects compatible with the data.

For example:

The confidence interval indicates that the data are compatible with a treatment effect ranging from a 4.4-unit to an 11.2-unit greater reduction compared with placebo.

The interval is often more informative than the p-value alone.

Direction of the Treatment Effect

The sign of a treatment effect must always be interpreted relative to the clinical endpoint.

For an endpoint where lower is better:

For an endpoint where higher is better:

Percentage Change from Baseline

Sometimes absolute change is not directly comparable across patients because baseline values differ substantially.

A percentage change can be defined as:

For a patient whose value decreases from 100 to 80:

The patient has experienced a 20% reduction from baseline.

Percentage Change Requires Caution

Percentage change can be useful, but it is not automatically preferable to absolute change.

Problems can arise when baseline values are close to zero.

For example:

A relatively small absolute difference can produce a very large percentage change when the denominator is small.

Percentage change may also have undesirable statistical properties when baseline values vary widely.

Example: Blood Pressure Trial

Consider a randomized trial comparing an antihypertensive drug with placebo. The primary endpoint is change in systolic blood pressure from baseline to Week 12.

The unadjusted difference in mean change is:

Thus, treatment produced an additional 7.6 mmHg reduction relative to placebo.

ANCOVA Version of the Example

Now suppose baseline systolic blood pressure is included as a covariate. The ANCOVA produces the following model-adjusted estimates:

The adjusted treatment difference is:

The interpretation is: The adjusted mean reduction in systolic blood pressure was 7.7 mmHg greater with the drug than with placebo.

A More Complete Clinical Trial Table

Interpreting the Example Clinically

A statistical interpretation might be:

At Week 12, the adjusted mean reduction in systolic blood pressure was 7.7 mmHg greater with the investigational drug than with placebo (95% CI: −10.4 to −5.0; p<0.0001).

This is preferable to simply stating:

"The drug group had a significant reduction in blood pressure."

The latter statement does not directly establish a treatment effect.

Paired t-Test: When Is It Appropriate?

A paired t-test is designed to test whether the mean change within a group differs from zero.

For example:

versus:

This can be useful for descriptive or single-group studies.

However, in a randomized controlled trial, the primary efficacy question is usually not whether the treatment group's change differs from zero.

It is whether the treatment group's change differs from the control group's change.

Independent vs. Paired Data

Why Randomization Matters

Randomization provides the foundation for interpreting treatment-group differences causally.

When patients are randomized, differences in post-baseline outcomes or changes can be attributed to treatment, subject to the usual assumptions and validity of the trial.

This is why comparing treatment groups is fundamentally different from examining pre-post changes within a single group.

Missing Baseline Measurements

A patient without a valid baseline measurement creates a special problem for a change-from-baseline analysis.

The change:

cannot be calculated if \(Y_0\) is missing.

The protocol and statistical analysis plan should therefore define:

• What constitutes a valid baseline measurement.
• Which measurement is selected if multiple predose measurements exist.
• How baseline measurements outside the assessment window are handled.
• How missing baseline values affect analysis populations.

Missing Post-Baseline Measurements

Missing follow-up measurements are common in clinical trials. Patients may:

• Withdraw from treatment.
• Withdraw consent.
• Be lost to follow-up.
• Miss a study visit.
• Require rescue medication.
• Discontinue because of an adverse event.

A simple change-score analysis at Week 12 may therefore analyze only patients with both baseline and Week 12 measurements.

This can reduce sample size and potentially introduce bias if missingness is related to treatment response.

MMRM and Missing Data

One reason MMRM is frequently used for longitudinal clinical-trial outcomes is that it can use available repeated measurements without requiring every patient to have an observed value at every visit.

Under appropriate assumptions, likelihood-based mixed models can provide valid inferences under a missing at random (MAR) framework.

Last Observation Carried Forward

Historically, some clinical trials used last observation carried forward (LOCF) to replace missing post-baseline measurements.

For example, if a patient's Week 8 measurement is the last available value, that value might be carried forward to Week 12.

LOCF is generally not a universally appropriate solution because it imposes a strong and often unrealistic assumption about what would have happened after withdrawal.

Baseline Observation Carried Forward

Another historically used approach is baseline observation carried forward (BOCF).

This effectively assumes that a patient who drops out returns to their baseline value.

This can also impose an implausible assumption and should not be used automatically.

The Estimand Perspective

Modern clinical-trial analysis distinguishes the statistical model from the scientific question being answered.

Before deciding how to analyze change from baseline, investigators should clarify the target treatment effect.

For example:

• What would the treatment difference be regardless of treatment discontinuation?
• What is the effect while patients remain on treatment?
• Should rescue medication affect the endpoint?
• Should treatment discontinuation be considered an intercurrent event?
• What population is the treatment effect intended to describe?

These questions affect the estimand and consequently the analysis strategy.

Change from Baseline and Intercurrent Events

Suppose a patient discontinues treatment and subsequently receives rescue therapy.

Their later outcome may no longer represent the same treatment condition as patients who remain on randomized therapy.

The analysis therefore needs a prespecified strategy for such events.

Possible approaches include:

• Treatment-policy strategies.
• Hypothetical strategies.
• Composite strategies.
• While-on-treatment strategies.
• Principal-stratum strategies.

The appropriate strategy depends on the scientific question.

Change from Baseline and Covariate Adjustment

In addition to baseline outcome, an ANCOVA or repeated-measures model may include prespecified prognostic covariates.

For example:

However, covariates should be selected prospectively and justified.

Baseline-by-Treatment Interaction

An advanced issue is whether the treatment effect depends on baseline severity. This can be investigated by including a treatment-by-baseline interaction:

If \(\beta_3\) is meaningfully different from zero, the treatment effect may vary according to baseline level.

This can be scientifically interesting but should generally be prespecified when it is intended to support confirmatory inference.

Example of Baseline-Dependent Treatment Effect

Suppose the treatment effect is:

A patient with baseline 100 has an estimated treatment effect of:

while a patient with baseline 140 has:

The treatment appears more effective among patients with higher baseline values.

Such findings require careful interpretation and should not be inferred merely from subgroup p-values.

Analysis of Change When the Outcome Is Skewed

A change-score t-test or ANCOVA assumes that the relevant residual structure is adequately modeled.

For strongly skewed outcomes, investigators may consider:

• Transforming the outcome.
• Using a generalized linear model.
• Using robust methods.
• Using an appropriate nonparametric approach.
• Modeling ratios or logarithmic changes when scientifically justified.

The appropriate method depends on the measurement scale and scientific meaning of the endpoint.

Log-Transformed Outcomes

For biomarkers measured on a multiplicative scale, a logarithmic transformation may be appropriate.

Suppose:

Then a difference on the log scale can be interpreted approximately as a ratio on the original scale.

For example:

can represent a ratio of geometric means.

Change Scores and Regression to the Mean

Regression to the mean is an important phenomenon when measurements vary over time.

Patients with unusually high baseline measurements may tend to have lower measurements at subsequent assessments even without effective treatment.

Likewise, unusually low measurements may tend to increase.

Randomized controlled trials reduce the concern about regression to the mean creating a spurious treatment effect because both treatment groups are subject to the phenomenon.

However, baseline adjustment can improve precision and account for baseline differences.

Do Not Compare Raw Percentage Improvements Without Context

Suppose:

The percentage changes appear to differ by 10 percentage points.

But the absolute reductions differ by 30 units.

The appropriate analysis depends on the clinical scale and the prespecified estimand.

Clinical Meaning vs. Statistical Significance

A statistically significant treatment difference is not necessarily clinically important.

Suppose:

with:

A very large trial can detect a small difference with high statistical significance.

The investigator must therefore ask whether a 1.2-unit difference is clinically meaningful.

Minimal Clinically Important Difference

Suppose the clinically meaningful difference is 5 units. Then:

might be considered clinically important.

The confidence interval can then be compared with this threshold.

For example:

The point estimate exceeds the clinically important threshold, but the confidence interval includes effects smaller than 5 units.

This provides more nuanced evidence than the p-value alone.

Example: Pain Score

Suppose a trial measures pain on a 0–10 scale.

The difference in mean change is:

Thus, treatment improved pain by an additional 1.2 points relative to placebo.

Whether that is clinically important depends on the prespecified interpretation of the pain scale.

Change from Baseline in a Safety Analysis

Change from baseline is also widely used for safety laboratory parameters. Examples include:

• ALT
• AST
• Creatinine
• Hemoglobin
• Platelet count
• Blood glucose

However, safety analysis often requires more than mean change.

Investigators may also examine:

• Shift from baseline categories.
• Worst post-baseline value.
• Clinically significant abnormalities.
• CTCAE grades in oncology.
• Laboratory values crossing predefined thresholds.

Shift Tables vs. Change Analysis

A continuous laboratory measurement can be summarized both by change and by categorical shifts.

A shift table can reveal clinically relevant deterioration that might not be obvious from a mean change alone.

Change from Baseline in Oncology

In oncology trials, change from baseline may be used for tumor burden, laboratory values, biomarkers, symptoms, or quality-of-life measures.

However, tumor response itself is often defined using categorical criteria such as RECIST rather than simply using the mean percentage change in tumor size.

Therefore:

Longitudinal Clinical Trial Example

Consider a 12-week randomized trial with measurements at Weeks 0, 4, 8, and 12.

The treatment effect appears to increase over time.

An MMRM can estimate each visit-specific treatment contrast while accounting for the correlation among repeated measurements.

Covariance Structure in Repeated-Measures Models

Repeated measurements within a patient are correlated.

For example, a patient's Week 4 and Week 8 measurements are generally more similar than measurements from two unrelated patients.

The covariance structure describes these relationships. Common choices include:

• Compound symmetry.
• Unstructured covariance.
• Autoregressive covariance.
• Toeplitz covariance.
• Variance components.

The choice should be justified based on the longitudinal design and prespecified model strategy.

Why the Unstructured Covariance Is Common

An unstructured covariance matrix allows each visit variance and covariance to be estimated separately.

For four post-baseline visits, the covariance matrix has the form:

This is flexible but can require many parameters, particularly with many visits.

Model-Based Treatment Comparisons

Suppose the MMRM produces the following Week 12 estimates:

The estimated treatment effect is therefore a 6.4-unit greater reduction at Week 12.

R Implementation: Simple Change Score

Suppose a dataset contains:

• id
• trt
• baseline
• week12

The change score can be created in R:

dat$change <- dat$week12 - dat$baseline

A simple treatment comparison can then be performed using a linear model:

fit_change <- lm(
  change ~ trt,
  data = dat
)

summary(fit_change)

The coefficient for trt represents the estimated difference in mean change between treatment groups.

R Implementation: ANCOVA

For a single primary post-baseline assessment, an ANCOVA can be fit as:

fit_ancova <- lm(
  week12 ~ trt + baseline,
  data = dat
)

summary(fit_ancova)

The treatment coefficient represents the adjusted treatment difference at Week 12.

R Implementation: Multiple Visits

For longitudinal data, the data are often stored in long format:

id    trt       visit       change
101   Drug      Week4       -4.2
101   Drug      Week8       -8.1
101   Drug      Week12      -11.5
102   Placebo   Week4       -2.0
102   Placebo   Week8       -3.9
102   Placebo   Week12      -5.2

A mixed model can then be specified using an appropriate R package and covariance structure.

library(nlme)

fit_mmrm <- lme(
  change ~ trt * visit,
  random = ~ 1 | id,
  data = dat_long
)

summary(fit_mmrm)

The exact covariance specification should be selected and documented according to the trial's statistical analysis plan.

SAS Implementation: Change Score

A simple change score can be generated using:

data analysis;
    set adsl;
    change = week12 - baseline;
run;

proc ttest data=analysis;
    class trt;
    var change;
run;

This provides a basic comparison of mean change between treatment groups.

SAS Implementation: ANCOVA

proc glm data=analysis;
    class trt;
    model week12 = trt baseline;
    lsmeans trt / pdiff cl;
run;
quit;

The treatment contrast is adjusted for baseline.

SAS Implementation: MMRM

A conceptual repeated-measures model can be implemented using PROC MIXED.

proc mixed data=analysis_long;
    class id trt visit;
    
    model change =
        trt
        visit
        trt*visit
        / ddfm=kr solution;
    
    repeated visit /
        subject=id
        type=un;
    
    lsmeans trt*visit /
        diff
        cl;
run;

Here:

• trt represents treatment.
• visit represents assessment time.
• trt*visit allows treatment effects to vary by visit.
• type=un specifies an unstructured covariance.
• ddfm=kr requests a Kenward-Roger denominator degrees-of-freedom method.

Change from Baseline and the Analysis Population

The analysis population must be clearly defined. Common populations include:

• Full analysis set.
• Intent-to-treat population.
• Modified intent-to-treat population.
• Per-protocol population.
• Safety population.

A change-from-baseline analysis should not silently switch from the primary analysis population to only patients with complete data unless that approach has been prespecified.

Baseline Definition Must Be Prespecified

A trial may collect several measurements before treatment:

The protocol should define which value constitutes baseline.

For example, baseline might be the last valid assessment before the first dose.

Without a prespecified rule, seemingly small programming decisions can change the derived change-from-baseline endpoint.

Visit Windows Matter

Clinical trials rarely collect every assessment at exactly the nominal study day.

A Week 12 visit might occur on Day 80, Day 84, or Day 90.

The analysis dataset therefore needs a prespecified method for assigning observations to analysis visits.

Common Mistakes

1. Testing within-group change instead of the treatment contrast. A significant change in one group does not prove that the treatment groups differ.
2. Ignoring baseline adjustment. For many continuous endpoints, ANCOVA can provide a more efficient treatment comparison than an unadjusted change analysis.
3. Using the wrong baseline. Baseline must be defined prospectively and consistently.
4. Reporting only p-values. The estimated treatment difference and confidence interval should also be reported.
5. Using percentage change automatically. Percentage change is not universally superior to absolute change.
6. Ignoring repeated measurements. Analyzing each visit separately can create multiplicity and correlation problems.
7. Assuming MMRM solves all missing-data problems. MMRM relies on modeling assumptions.
8. Using LOCF without justification. Carrying observations forward imposes assumptions about post-withdrawal outcomes.
9. Confusing statistical significance with clinical importance. A very small effect can be highly statistically significant in a large study.
10. Changing the analysis after seeing the results. The primary analysis should be prespecified before unblinding the comparative treatment results.

A Practical Analysis Workflow

What Should Be Reported?

A clinical trial report should provide enough information to understand both the magnitude and uncertainty of the treatment effect.

For a continuous endpoint analyzed as change from baseline, a useful summary may include:

• Number of patients analyzed.
• Baseline descriptive statistics.
• Post-baseline descriptive statistics.
• Mean or LS mean change.
• Between-group treatment difference.
• 95% confidence interval.
• p-value when appropriate.
• Analysis model.
• Covariates included in the model.
• Handling of missing observations.
• Relevant sensitivity analyses.

Example Statistical Reporting

A concise clinical-trial result could be reported as:

This communicates:

• What was measured.
• When it was measured.
• How much each group changed.
• The treatment contrast.
• The uncertainty around the treatment contrast.
• The statistical evidence.

Change from Baseline vs. Endpoint Analysis

Three common ways of expressing a treatment comparison are:

These approaches can produce similar conclusions, but they are not identical statistical models.

In many randomized trials with continuous endpoints, baseline-adjusted ANCOVA is an efficient and well-established approach for a single primary follow-up assessment.

Why Change Scores and ANCOVA Can Produce Different Results

A change-score analysis effectively models:

while ANCOVA models:

conditional on:

The two methods therefore use baseline information differently.

When baseline strongly predicts the follow-up outcome, ANCOVA may provide greater statistical efficiency.

What If Baseline Values Are Very Different?

A substantial baseline imbalance should not simply be "fixed" by analyzing percentage change or by changing the primary analysis after the fact.

In a randomized trial, baseline adjustment can provide an appropriate model for accounting for baseline differences.

However, extreme baseline imbalance may also indicate:

• Randomization imbalance.
• Small sample size.
• Eligibility problems.
• Measurement differences.
• Potential data-quality issues.

The source of the imbalance should be investigated.

Change from Baseline and Subgroup Analysis

Suppose investigators want to compare treatment effects according to baseline disease severity.

The preferred approach is generally to test or estimate a treatment-by-subgroup interaction rather than simply comparing p-values within each subgroup.

Here \(\beta_3\) addresses whether the treatment effect differs by subgroup.

Change from Baseline and Multiplicity

Suppose a trial reports treatment differences at:

• Week 4
• Week 8
• Week 12
• Week 16

Testing all four visits independently at the 5% level can increase the probability of a false-positive finding.

Multiplicity becomes particularly important when several visits, endpoints, subgroups, or treatment contrasts are all considered confirmatory.

A prespecified testing hierarchy, multiplicity adjustment, or other strategy may be required depending on the confirmatory objective.

Primary Endpoint vs. Supporting Visits

If Week 12 is the primary endpoint, the trial should not casually switch the primary conclusion to Week 8 because the Week 8 result happens to be more favorable.

A clean analysis distinguishes:

Change from Baseline and Responder Analyses

Continuous change and responder endpoints answer different questions.

A continuous endpoint asks: How much did the outcome differ between treatment groups?

A responder endpoint asks: What proportion of patients achieved a clinically meaningful improvement?

For example, if a 5-point reduction is clinically meaningful:

The responder rate can then be compared between treatment groups.

A Complete Decision Framework

The Most Important Concept

The most important point in change-from-baseline analysis is simple: clinical trials compare treatments, not merely changes within treatments.

If a treatment group decreases by 10 units and placebo decreases by 8 units, the treatment effect is not 10 units.

It is:

The treatment produced an additional 2-unit reduction relative to placebo.

For a single post-baseline endpoint, baseline-adjusted ANCOVA is often an efficient choice.

For multiple post-baseline measurements, a repeated-measures model such as MMRM can account for the longitudinal structure and provide visit-specific treatment contrasts.

References

Vickers, A.J. & Altman, D.G. (2001). Analysing controlled trials with baseline and follow up measurements. BMJ, 323, 1123–1124.
Frison, L. & Pocock, S.J. (1992). Repeated measures in clinical trials: analysis using mean summary statistics and its implications for design. Statistics in Medicine, 11, 1685–1704.
Fitzmaurice, G.M., Laird, N.M. & Ware, J.H. Applied Longitudinal Analysis. Wiley.
Little, R.J.A. & Rubin, D.B. Statistical Analysis with Missing Data. Wiley.
ICH E9 (R1). Addendum on Estimands and Sensitivity Analysis in Clinical Trials.
EMA. Guideline on adjustment for baseline covariates in clinical trials.