Baseline Characteristics in Clinical Trials

Overview

A baseline table describes who entered a trial and supports assessment of applicability, data quality and randomisation implementation.

It is not primarily a hypothesis-testing table.

Randomisation permits chance imbalance. A small baseline p-value does not prove failed randomisation, and a large one does not prove correct allocation.

CONSORT recommends reporting baseline demographic and clinical characteristics by randomised group without significance testing [1].

Why this matters

Baseline tests answer a question that is already known to be governed by random allocation.

They also depend strongly on sample size:

• small trials may have clinically important imbalances with large p-values;
• large trials may show tiny imbalances with small p-values.

The analysis should not add or remove covariates solely because baseline p crosses 0.05.

When to use this framework

Use this framework for parallel, factorial, cluster and crossover trials.

In cluster trials, describe both cluster-level and participant-level characteristics.

In crossover trials, describe participants by sequence rather than period treatment alone.

Research question and variables

A useful baseline table reports:

• group counts;
• age and sex;
• disease severity;
• baseline outcome;
• stratification variables;
• important prognostic factors;
• missing values.

Variables should be selected before inspecting group differences.

Data structure

Suggested baseline-table structure

simulated data

Characteristic	Control	Intervention
Participants	120	120
Age, mean (SD), years	Reported	Reported
Female, n/N (%)	Reported	Reported
Baseline outcome, mean (SD)	Reported	Reported
Site, n (%)	Reported	Reported

Statistical assumptions

Descriptive summaries

Use:

• mean and SD for clinically meaningful numerical summaries;
• median and quartiles for markedly skewed variables;
• counts and percentages for categories;
• variable-specific denominators when missing.

Avoid reporting SE at baseline because the table describes participants rather than estimating a population mean as its primary purpose.

Standardised mean differences

For a numerical variable:

For a binary variable, use a pooled binomial variance.

SMDs are descriptive diagnostics, not tests.

Thresholds such as 0.1 are rough guides rather than universal decision rules.

Worked example

Worked analysis in R

trial <- read.csv(
  "data/part-13-trial-data.csv"
)

standardised_mean_difference <- function(
  x,
  group
) {
  x1 <- x[group == 1]
  x0 <- x[group == 0]

  (
    mean(x1) - mean(x0)
  ) / sqrt(
    (
      var(x1) + var(x0)
    ) / 2
  )
}

c(
  age = standardised_mean_difference(
    trial$age,
    trial$group
  ),
  baseline = standardised_mean_difference(
    trial$baseline,
    trial$group
  )
)

aggregate(
  cbind(
    age,
    baseline
  ) ~ group_label,
  data = trial,
  FUN = function(x) {
    c(
      mean = mean(x),
      sd = sd(x)
    )
  }
)

with(
  trial,
  table(group_label, site)
)

Selected analysis output

Age SMD = 0.027
Female SMD = 0.000
Baseline outcome SMD = -0.041
Site 1 SMD = 0.000

(The standardised-difference calculations for female sex and Site 1 are not shown here.)

Visualising baseline characteristics

Covariate adjustment

Prespecified adjustment for strong baseline prognostic variables can improve precision.

Common examples include:

• baseline value of the outcome;
• randomisation strata;
• site;
• key prognostic factors.

Adjustment is justified by prognosis and design, not by observed imbalance.

For continuous outcomes, ANCOVA is generally preferable to analysing change scores alone when baseline is strongly related to follow-up.

Suspicious imbalance

A pattern can justify investigation when:

• allocation counts conflict with the randomisation list;
• site or time patterns are implausible;
• enrolment changes before known assignments;
• exclusions occur after allocation.

The response is to audit trial conduct, not to repair randomisation using baseline p-values.

Interpreting the table

Presenting the results

Suggested reporting

Report group-specific values, missing denominators and relevant cluster characteristics.

CONSORT 2025 retains baseline descriptive reporting and discourages tests of baseline differences [1].

The rationale against routine baseline testing, the use of standardised balance diagnostics and prespecified baseline covariate adjustment are developed in methodological guidance [2], [3], [4].

Common mistakes

Knowledge check

What is the main purpose of the baseline table?

Answer and reasoning

To describe the realised trial population and support assessment of applicability and data quality.

Why are baseline p-values unnecessary?

Answer and reasoning

Random allocation generated the differences; testing them does not validate randomisation.

Should covariates be adjusted only when imbalanced?

Answer and reasoning

No. Prespecify adjustment using prognosis and design.

What does an SMD near zero show?

Answer and reasoning

Little observed difference on that measured variable in standardised units.

What should suspicious imbalance trigger?

Answer and reasoning

An audit of allocation and enrolment procedures.

Key take-home messages

References

1. Hopewell S, Chan AW, Collins GS, Hróbjartsson A, Moher D, Schulz KF, et al. CONSORT 2025 statement. BMJ. 2025;389:e081123. https://doi.org/10.1136/bmj-2024-081123
2. Senn S. Testing for baseline balance in clinical trials. Stat Med. 1994;13(17):1715–1726. https://doi.org/10.1002/sim.4780131703
3. Austin PC. Balance diagnostics for comparing the distribution of baseline covariates. Stat Med. 2009;28(25):3083–3107. https://doi.org/10.1002/sim.3697
4. European Medicines Agency. Guideline on adjustment for baseline covariates in clinical trials. London: EMA; 2015.