Analysis of Covariance (ANCOVA)

Two ways to analyse the same two numbers

A two-arm trial with a baseline visit and a follow-up visit looks like it should have one obvious analysis: subtract baseline from follow-up for every patient, then compare the mean change between groups. That is a change-score analysis, and it makes an assumption that is easy to miss — it treats every point of baseline as worth exactly one point of follow-up, as though the true relationship between the two had a slope of 1. That is rarely known in advance and often is not the case.

ANCOVA asks a related but distinct question. Rather than folding baseline and follow-up into a single difference score, it models follow-up directly, with treatment group as the factor of interest and baseline entered as a continuous covariate. The model estimates the baseline–follow-up relationship from the data itself, then reports the group difference in follow-up scores after adjusting for whatever that relationship turns out to be. That is a more specific mechanism than “controlling for baseline” in the loose sense the phrase often gets used — ANCOVA adjusts the outcome for the covariate’s estimated linear relationship with it, not simply for group membership on that variable. The quantity it tests, the baseline-adjusted mean difference at follow-up, is conceptually closer to “how different are the groups, holding starting point constant” than a raw change score is.

Why ANCOVA usually wins on power

This is not a matter of house style. Vickers and Altman addressed it directly in a widely cited BMJ statistics note comparing three ways of analysing the same trial: a follow-up-only comparison, a change-score comparison, and ANCOVA.¹ Their worked example is concrete: a trial needing 85 patients if analysed on follow-up scores alone would need only 68 with a change-score analysis, and just 54 with ANCOVA, at a baseline-to-follow-up correlation of 0.6.¹ The formal literature on statistical power for these designs backs the same ordering: Oakes and Feldman derived unified power formulas for both the change-score and ANCOVA models across nonequivalent pretest-posttest designs specifically so evaluators could compare sample-size requirements between the two rather than guessing.² A more recent methods paper looking specifically at the baseline–follow-up correlation makes a related recommendation from a different angle: adjust for baseline by ANCOVA regardless of which outcome you ultimately report, since an unadjusted analysis leaves the trial exposed to baseline imbalance and regression to the mean in a way ANCOVA specifically corrects for.³

The intuition is straightforward once the mechanism is visible. A change score inherits variance from both the baseline measurement and the follow-up measurement. ANCOVA, by estimating the baseline–follow-up relationship rather than assuming it, removes only the variance that baseline actually predicts, and leaves the rest. When baseline and follow-up are highly correlated the two approaches converge; when they are not, the change score is carrying variance ANCOVA does not have to.

The other reason: randomisation balances groups on average, not in every trial

Power is only half of the practical difference. Randomisation balances group means in expectation, not in any single trial, and a modest baseline imbalance is common at the sample sizes typical of a single-centre surgical trial. Vickers and Altman worked through exactly what each analysis does with that imbalance: if the treatment arm happened, by chance, to start slightly worse, a follow-up-only comparison underestimates the treatment effect and a change-score comparison overestimates it, while ANCOVA “gives the same answer whether or not there is baseline imbalance.”¹ A change score has no mechanism to correct for that imbalance; it carries it straight through into the result.

One assumption ANCOVA doesn’t get to skip

ANCOVA's efficiency comes with a condition a change-score analysis never has to check: homogeneity of regression slopes. The model assumes the relationship between baseline and follow-up is the same in every treatment group — one regression line, not several lines running at different angles. If the slopes genuinely differ between groups, the group difference in follow-up scores is not constant across the range of baseline values, and a single adjusted mean difference stops being a fair summary of the effect. The standard check is to add a group×baseline interaction term to the model: a non-significant interaction supports a common slope and the adjusted-mean-difference result stands; a significant one means the group difference has to be reported at specific baseline values instead of as one number, which is a materially different result to write up.

How to run it — SPSS vs StatsPlease

In IBM SPSS Statistics (documentation current to version 32), the path is Analyze ▸ General Linear Model ▸ Univariate…. The follow-up score goes into the Dependent Variable box, treatment group into Fixed Factor(s), and the baseline score into Covariate(s). By default the output’s Tests of Between-Subjects Effects table reports the raw model — to get the number that actually matters, the baseline-adjusted group means, open the EM Means button, move the group factor into “Display Means for,” and tick the box to compare them; that produces a separate Estimated Marginal Means table holding the adjusted means, distinct from the unadjusted means SPSS would show under Descriptives. The Tests of Between-Subjects Effects table itself is still worth reading past the Group row: the row for the covariate (baseline) carries its own F-test and P value, and a significant one confirms baseline is doing real work in the model. The common misread is treating that covariate row as evidence of baseline imbalance between groups — it isn’t; it only says baseline predicts follow-up, which is a different claim, and checking the groups for imbalance is a separate step.

Upload the dataset (or the relevant outcome, group, and covariate columns). StatsPlease’s deterministic engine identifies the appropriate test from the variable type and study design, checks the relevant assumptions, computes the result using fixed, non-LLM algorithms, and drafts the Methods/Results sentence in AMA format — the same number a reader would get running the test by hand in SPSS.

References

1. Vickers AJ, Altman DG. Statistics notes: analysing controlled trials with baseline and follow up measurements. BMJ. 2001;323(7321):1123-1124. https://doi.org/10.1136/bmj.323.7321.1123
2. Oakes JM, Feldman HA. Statistical power for nonequivalent pretest-posttest designs: the impact of change-score versus ANCOVA models. Eval Rev. 2001;25(1):3-28.
3. Clifton L, Clifton DA. The correlation between baseline score and post-intervention score, and its implications for statistical analysis. Trials. 2019;20(1):43. https://doi.org/10.1186/s13063-018-3108-3