16  Baseline Covariate Adjustment

Prerequisites: Chapters 3, 5, and 8.

16.1 Learning objectives

By the end of this chapter you should be able to:

  • State the criteria determining when to adjust, which covariates to adjust for, and how to specify the adjustment.
  • Explain why the decision to adjust must never depend on observed baseline balance.
  • Compute the efficiency gain from adjustment and relate it to the covariate-outcome correlation.
  • Handle adjustment correctly for binary and time-to-event outcomes, where the estimand changes.
  • Recognize when adjustment can hurt.

16.2 Orientation

Covariate adjustment in randomized trials has a mature theoretical literature, consistent simulation evidence, and regulatory guidance from ICH, the FDA, and the EMA, all pointing the same direction. It remains contested in applied practice. The gap is not one of evidence but of translation: an investigator facing a specific trial has had no compact set of criteria to apply without re-deriving the theory.

This chapter supplies those criteria. Its content is not new; it is a consolidation of dispersed guidance into a form a study team can use, with the reasoning behind each rule stated so that it can be defended.

The central claim is simple: adjust, always, for pre-specified prognostic covariates, and the design questions are which ones and how, not whether.

16.3 Provenance

This chapter follows the research compendium 12-baseline-covariate-dist (project baselinedist), which synthesizes the statistical literature, the regulatory guidance, and simulation evidence on adjusted versus unadjusted analyses across covariate-outcome correlations, sample sizes, and imbalance levels, and distils the result into criteria along three axes: when to adjust, which covariates, and how to specify the adjustment. Performance measures are bias, coverage, and relative variance reduction.

16.4 The statistician’s contribution

(Judgment 1.) The covariate list, fixed before unblinding. The list should be short, prognostic, and justified by prior evidence. Choosing it from the trial’s own data invalidates the inference, whether the selection is on observed imbalance or on observed association with the outcome.

(Judgment 2.) The estimand implied by the model. In linear models, adjustment changes precision and not the target. In logistic and Cox models it changes the target, from a marginal to a conditional effect. This is not a technicality; it changes what the number means and whether it can be compared to other trials.

(Judgment 3.) Resisting the balance test. Someone will propose testing baseline characteristics for imbalance and adjusting for those that differ. This is incoherent and should be refused. Chapter 8 gave the reason; this chapter gives the criterion.

16.5 When to adjust

Criterion 1. Always, when prognostic covariates are available and pre-specified. The efficiency gain is real and the cost is nil when the model is correctly specified.

Criterion 2. Always include the stratification variables used in the randomization. This is not merely an efficiency recommendation but a validity requirement, as Chapter 3 established: omitting them leaves the standard error too large and the inference conservative, and under covariate-adaptive schemes the test is not correctly calibrated at all.

Criterion 3. Always include the baseline measurement of the outcome, when it exists. It is nearly always the single most prognostic covariate available, and the efficiency gain from including it is larger than from any other single variable.

And the negative criterion: the decision to adjust must not depend on observed baseline balance. Senn’s argument is decisive. In a randomized trial, any observed imbalance is by construction a chance event; a hypothesis test of baseline balance therefore tests a null hypothesis known to be true, and its rejection carries no information about whether adjustment is needed. Worse, conditioning the model choice on the data makes the type I error of the final test depend on a data-dependent selection, which is not accounted for in the reported \(p\)-value.

16.6 Which covariates

Criterion 4. Select on prior knowledge of prognostic value, from previous trials, meta-analyses, or established clinical understanding.

Criterion 5. Keep the number small relative to the sample size. The efficiency gain from an additional covariate is offset by the degree of freedom it consumes, and the trade-off worsens as the sample size falls. A useful rule is one covariate per 15 to 20 patients as an upper bound, and in practice the useful list is three to five variables. Chapter 14 supplies an additional reason for restraint: the covariate-adjusted Wald test becomes anti-conservative as the list lengthens in small samples.

Criterion 6. Never select on the trial’s own data, and this includes both selecting on observed imbalance and selecting on observed covariate-outcome association.

16.7 The efficiency gain

For a continuous outcome analyzed by ANCOVA with a single covariate of correlation \(\rho\) with the outcome, the residual variance is reduced by the factor \(1 - \rho^2\), so the required sample size for fixed power is reduced by the same factor.

\(\rho\) Variance reduction Equivalent \(N\) saving
0.2 4% 4%
0.4 16% 16%
0.5 25% 25%
0.6 36% 36%
0.7 49% 49%

Baseline measurements of the outcome typically correlate 0.5 to 0.7 with the follow-up value, so adjustment for baseline alone commonly saves a quarter to a half of the sample size. This is the largest free gain available in trial design and it costs one line in the model specification.

set.seed(2026)
n <- 100; rho <- 0.6
x <- rnorm(2 * n)
z <- rep(0:1, each = n)
y <- 0.5 * z + rho * x + sqrt(1 - rho^2) * rnorm(2 * n)

summary(lm(y ~ z))$coefficients['z', 'Std. Error']
#> [1] 0.1418
summary(lm(y ~ z + x))$coefficients['z', 'Std. Error']
#> [1] 0.1131
(0.1131 / 0.1418)^2
#> [1] 0.6363    # variance ratio, close to 1 - rho^2 = 0.64

Question. A reviewer objects that the trial adjusted for five covariates but the baseline table shows the arms were well balanced on all five, so the adjustment was unnecessary and looks like fishing. How do you respond?

Answer.

The objection confuses two roles that adjustment plays.

Adjustment corrects for imbalance, which is what the reviewer has in mind, and in a well-balanced trial that correction is indeed small. But adjustment also reduces residual variance, and that gain is present whether or not the covariates are balanced. It comes from removing explained variation from the error term, not from correcting a difference between arms. A perfectly balanced trial still gains the full \(1 - \rho^2\) precision benefit.

The ‘fishing’ concern is answered by pre-specification: the five covariates were named in the SAP before unblinding, so no selection occurred. Point the reviewer to the SAP version and date.

Finally, note that the reviewer’s implicit alternative, adjusting only when the baseline table shows imbalance, is exactly the procedure that would constitute fishing, since it conditions the model on the data.

16.8 How to adjust

Criterion 7. Continuous outcomes: ANCOVA. Regress the follow-up value on treatment and the covariates, including the baseline value of the outcome. Ordinary or robust standard errors are both acceptable in balanced designs. Lin’s analysis shows that including treatment-by-covariate interactions, with covariates centered, guarantees the adjusted estimator is no less efficient than the unadjusted one asymptotically, which answers Freedman’s critique of adjustment in randomized experiments (Freedman, 2008; Lin, 2013).

Criterion 8. Binary outcomes: fit the logistic model and then standardize. The raw coefficient from an adjusted logistic regression is a conditional odds ratio, which differs from the marginal odds ratio by non-collapsibility even in the absence of confounding. To recover the marginal estimand, use g-computation: predict each patient’s outcome probability under both arms, average, and difference.

fit <- glm(y ~ z + x1 + x2, family = binomial, data = d)
p1  <- predict(fit, transform(d, z = 1), type = 'response')
p0  <- predict(fit, transform(d, z = 0), type = 'response')
c(risk_difference = mean(p1) - mean(p0),
  risk_ratio      = mean(p1) / mean(p0))
#> risk_difference      risk_ratio
#>          0.0912          1.4380

Standard errors by the delta method or the bootstrap; the marginaleffects and RobinCar packages implement this, the latter written specifically for randomized trials. Alternatives are a linear probability model or modified Poisson regression with robust standard errors, both of which target marginal quantities directly.

Criterion 9. Time-to-event outcomes. Use a covariate-adjusted log-rank test, or estimate an adjusted hazard ratio recognizing that it too is conditional and non-collapsible, or target a collapsible summary such as the difference in restricted mean survival time. The choice must be pre-specified because the three answer different questions.

16.9 When adjustment can hurt

Three cases.

Post-baseline variables. Adjusting for anything measured after randomization can be affected by treatment. Conditioning on it removes part of the treatment effect and can induce collider bias. This includes adherence, concomitant medication, and intermediate biomarkers.

Too many covariates relative to \(n\). The degrees of freedom cost, and in small samples the calibration cost documented in Chapter 14, can exceed the precision gain.

Mis-specified functional form. Adjustment for a continuous covariate entered linearly when the true relationship is strongly nonlinear leaves residual confounding of the imbalance and does not deliver the full precision gain. Splines or pre-specified categorization address this; the specification should be in the SAP.

16.10 Worked example: the covariate list for a stroke trial

A trial of an acute intervention, 400 patients, primary endpoint the modified Rankin score at 90 days analyzed as an ordinal outcome, randomization stratified by center and by stroke severity category.

Criterion 2 forces: center and severity category.

Criterion 3 forces: there is no baseline modified Rankin in an acute stroke trial, since all patients are at their pre-morbid state, so the analogue is the pre-stroke functional status, which is included.

Criterion 4 supplies: age and NIHSS at presentation, both established as the dominant prognostic factors from multiple prior trials and meta-analyses.

Criterion 5 checks: five covariates for 400 patients, 80 patients per covariate, comfortably within bounds.

Criterion 9 governs the specification: the outcome is ordinal, so the primary analysis is a proportional-odds model with the five covariates, and because the proportional-odds coefficient is conditional and non-collapsible, the SAP additionally pre-specifies a standardized marginal summary, the difference in mean utility-weighted Rankin, computed by g-computation from the same model.

The negative criterion applies: the SAP states explicitly that no covariate will be added or removed on the basis of the observed baseline table, and that no baseline balance tests will be performed or reported.

Sensitivity: an unadjusted analysis, reported alongside. If adjusted and unadjusted differ materially in a randomized trial of this size, the explanation is either a strongly prognostic covariate that happened to be imbalanced, which is benign and expected, or a specification problem, which is not. Both should be examined before the manuscript.

16.11 Collaborating with an LLM on covariate adjustment

Prompt 1: ‘Which covariates should we adjust for?’

What to watch for. Models produce reasonable lists based on the disease area and often include variables the trial does not collect, or omit the stratification factors, which are mandatory.

Verification. Check the list against the case report form and against the randomization specification.

Prompt 2: ‘Compute the marginal risk difference from this adjusted logistic model.’

What to watch for. Frequently returns the coefficient or its exponential rather than a standardized marginal effect, and frequently attaches a naive standard error to a standardized estimate.

Verification. Confirm the prediction step evaluates every patient under both arms, and that the standard error comes from the delta method or a bootstrap.

Prompt 3: ‘The baseline table shows an imbalance in age. Should we adjust?’

What to watch for. Models often say yes, adopt the imbalance-triggered logic, and reproduce exactly the error this chapter warns against.

Verification. The answer is that age should have been in the pre-specified list if it is prognostic, and the observed imbalance is irrelevant to the decision.

16.12 Principle in use

  1. Fix the covariate list in the SAP and never revisit it. Three to five prognostic variables, including the stratification factors and the baseline outcome.

  2. Never test baseline balance. It tests a null known to be true and invites a data-dependent model choice.

  3. For nonlinear models, say which estimand you are reporting. Conditional and marginal effects differ, and only one of them was pre-specified.

16.13 Exercises

  1. Verify by simulation that the ANCOVA variance ratio is \(1 - \rho^2\) for \(\rho = 0.3, 0.5, 0.7\), and confirm the estimator is unbiased in each case.

  2. Simulate a binary-outcome trial with a strong prognostic covariate. Compare the unadjusted odds ratio, the adjusted conditional odds ratio, and the standardized marginal odds ratio. Explain the ordering.

  3. Implement the imbalance-triggered adjustment rule (adjust only if the baseline test has \(p < 0.05\)) and estimate the resulting type I error of the treatment comparison by simulation.

  4. Show by simulation that adjusting for a post-randomization variable affected by treatment biases the estimated treatment effect, and characterize the direction.

  5. For a trial with \(n = 60\), plot the empirical power of ANCOVA against the number of adjusted covariates from 0 to 10, with each covariate correlated 0.3 with the outcome. Where is the optimum?

16.14 Further reading

  • The compendium 12-baseline-covariate-dist for the full criteria with primary-source citations.
  • Senn (1994), on baseline balance testing.
  • Lin (2013) and Freedman (2008), the exchange on adjustment in randomized experiments, and Wang et al. (2019), which establishes valid intervals without model assumptions.
  • Raab et al. (2000), the practical criteria for selecting covariates, including the subjects-per-covariate rule.
  • Van Lancker et al. (2024) and Ye et al. (2023), the two most useful recent syntheses of adjustment practice.
  • Tsiatis et al. (2008), for the semiparametric framing, and Moore & Laan (2009), for targeted maximum likelihood with binary outcomes.
  • Kahan & Morris (2012), on the requirement to include stratification variables, and Kahan et al. (2014), an empirical assessment across trials.
  • US Food and Drug Administration (2021) and the EMA covariate guideline.
  • The RobinCar and marginaleffects R packages.