The Gold Standard of Experimental Design | Baseline Imbalance, Covariate Adjustment and Chance

1. Randomisation is a procedure, not a promise of identical tables

Back to contents

Random allocation makes treatment assignment independent of baseline characteristics under the design. Across many repetitions, groups are balanced on average. In one finite trial, age, prior score or disease severity can still differ by chance.

The first audit question is therefore procedural: was the sequence generated and implemented as planned, with no exclusions or prediction? Visual imbalance alone cannot diagnose allocation failure.

2. Describe baseline variables without running a lottery of tests

Back to contents

Baseline tables help readers understand the enrolled sample and the scale of important differences. Significance tests at baseline answer a question already controlled by randomisation and can mislead because small trials have low power while large trials flag trivial differences.

Use descriptive summaries and clinical or substantive judgment. Investigate data errors and implementation failures, but do not let whichever p-value falls below 0.05 choose the analysis.

3. Prespecify prognostic covariates

Back to contents

A baseline covariate measured before treatment can improve precision when it strongly predicts the outcome. Selection should follow prior evidence, measurement quality and the estimand—not observed treatment-group differences or observed outcome associations revealed after unblinding.

State the variables, functional form, treatment interactions if any, missing-covariate handling and variance estimator in the protocol or statistical analysis plan.

4. Match adjustment to the target treatment effect

Back to contents

For a continuous outcome, analysis of covariance can estimate an adjusted mean difference. For binary, count or time-to-event outcomes, conditional model coefficients may not equal the marginal population effect readers expect. Standardisation or other approaches can recover a marginal contrast.

Name the estimand first: population, outcome, handling of intercurrent events, summary measure and time point. Then choose an adjustment method that estimates it.

5. Account for the randomisation scheme

Back to contents

If randomisation used strata, blocks, minimisation or clusters, the analysis should reflect important design features. Ignoring stratification variables can waste precision; ignoring clustering can understate uncertainty.

Adjustment for a prognostic variable is not the same as changing who counts in the analysis. Keep the randomised population and treatment policy clear.

6. Worked example: imbalance changes precision, not the assignment rule

Back to contents

In a 100-participant trial, the treatment group begins with a mean baseline score of 62 and control with 58. At follow-up the means are 71 and 65. The crude follow-up difference is 6 points.

Suppose a prespecified analysis of covariance estimates 5.1 points with a standard error of 1.6, while the unadjusted standard error is 2.0. The adjustment uses baseline prognosis to improve precision. It is not proof that the 4-point starting difference was bias, nor permission to invent the model after seeing it.

7. Use sensitivity analyses to test modelling choices

Back to contents

Report the prespecified primary analysis and a small set of principled sensitivity analyses: unadjusted, alternative functional form, robust standard errors, treatment-by-covariate interaction where scientifically relevant, and documented missing-covariate handling.

Sensitivity analysis should reveal dependence on assumptions. It should not become a search for the most favourable estimate.

8. Changed-context transfer check

Back to contents

Move the example to a randomised classroom study with classes, not pupils, assigned to teaching methods. Chance imbalance in prior attainment still occurs, and prespecification still transfers. The analysis must now account for intraclass correlation and the smaller number of independent units.

A pupil-level regression that ignores class assignment can produce confidence intervals that are too narrow. Transfer the logic while respecting the design unit.

9. Limits and common misconceptions

Back to contents

Adjustment cannot rescue compromised randomisation, selective exclusion or a poorly measured outcome. A perfectly balanced table does not prove allocation was secure. Adding many weak covariates can harm finite-sample performance, and post-treatment variables should not be treated as ordinary baseline adjusters.

Do not choose covariates because their baseline difference is “significant.” Do not interpret a changed point estimate as automatic evidence of confounding in a randomised design.

10. Audit design, data and analysis together

Back to contents

An audit checks sequence generation, allocation implementation, enrolment timing, baseline measurement before treatment, missingness, prespecified covariates, model code and the correspondence between protocol, registry, analysis plan and report.

Reproduce both adjusted and unadjusted results from locked data. Differences should be explainable by prognosis, scale and estimand—not silent population changes or analyst discretion.

11. Frequently asked questions

Back to contents

Does baseline imbalance invalidate randomisation?

Not by itself. First inspect allocation and data procedures; chance differences are expected in finite samples.

Should every baseline variable be adjusted for?

No. Prefer a limited prespecified set of reliable prognostic variables and design factors.

Must the adjusted and unadjusted estimates match?

No. They can differ by chance and by estimand or model. Report both clearly and explain the primary choice.

12. Plan before the table can influence you

Back to contents

The gold standard is a chain of decisions made early enough to resist hindsight: secure randomisation, a declared estimand, justified prognostic covariates, a suitable model and transparent sensitivity checks.

When the baseline table appears, use it to describe the realised sample and detect errors. Do not use it as a menu from which the observed differences select the analysis.