AI Glossary / Statistics and probability / Intermediate

Adjusted means

Definition

Adjusted means are model-based group averages calculated after accounting for specified covariates or other predictors. In a simple analysis of covariance, they represent predicted group means at a common covariate value. More generally, estimated marginal means average model predictions over a defined reference grid and weighting scheme.

What it means

Raw group means describe the observations actually collected. Adjusted means answer a model-based comparison question, such as how groups would compare at the same baseline score. Their interpretation depends on which variables are included, the fitted relationships and the reference values or distribution used.

Adjustment does not physically change the observations or automatically make groups equivalent. It supplies predictions under a specified model. A useful report therefore identifies both the adjustment variables and the target comparison, rather than presenting an adjusted number as an inherently fairer or more truthful average.

How it works

Fit a model that includes the group indicator and relevant covariates. For a simple parallel-slopes linear model, the adjusted mean for group g at reference value x0 can be calculated as its observed mean minus the estimated common slope times the difference between its covariate mean and x0.

Equivalently, predict the outcome for each group at the same covariate value. With several factors or covariates, define a reference grid, obtain predictions and average them using stated weights. Equal weighting across factor levels and weighting according to observed frequencies can answer different questions.

Check whether the model allows interactions. If a covariate's relationship with the outcome differs by group, the group difference can depend on the reference value, making one overall comparison inadequate. Also inspect data support: predicting one group's response at covariate values it barely or never exhibits may rely on extrapolation. Uncertainty should incorporate estimation of the model rather than treating the adjustment as a known constant.

Three practical examples

1. A training evaluator compares follow-up scores at a common baseline score, reporting adjusted group means alongside the observed baseline differences.

2. A manufacturing analyst compares process settings at a shared temperature when temperature varies across observations and is included in the fitted model.

3. A service analyst averages predicted outcomes over the same defined customer mix for each service channel, making the standardization target explicit.

Think of it this way

Think of comparing two journeys after estimating how much their different distances contribute to travel time. A common-distance prediction can help isolate the comparison of interest, but its usefulness depends on whether the distance-time model is credible and whether that common distance is relevant for both journeys.

Worked example

Consider an original parallel-slopes example with two groups of equal size. Group A has an observed outcome mean of 70 and a covariate mean of 40. Group B has an outcome mean of 80 and a covariate mean of 60. Suppose the fitted common slope is 0.4 outcome units per covariate unit, and the reference value is 50.

A's adjusted mean is 70 - 0.4(40 - 50) = 74. B's adjusted mean is 80 - 0.4(60 - 50) = 76. The raw difference B minus A is 10, while the adjusted difference is 2. The change reflects the specified fitted relationship, not proof that eight units were caused by the covariate.

At reference value 55, the predictions become 76 and 78, preserving the two-unit difference because the slopes are parallel. If the groups had different slopes, moving the reference value could change the difference. Standard errors cannot be recovered from these summary values alone; the model's uncertainty information is also needed.

Illustrative case study

Fictional case study: Skillharbor compares two training groups whose participants began with different experience levels. Its analysts fit a model using baseline experience and group membership, then report predicted scores at a reference experience level supported in both groups.

They inspect whether the experience-score relationship differs between groups and show the observed means as context. The report explains that participation was not randomized, so the adjusted comparison may still reflect unmeasured differences. A later evaluation uses a stronger design to address the causal question.

This fictional case illustrates transparent covariate adjustment. It does not claim that adjusting for one baseline measure removes all selection bias or establishes which training approach caused better outcomes.

Common misconceptions and mistakes

Adjusted means are not simply raw means with arbitrary values added or subtracted. The adjustments come from a model and a defined target. They are also not necessarily averages for the population actually sampled; reference-grid weights may describe a different standardized population.

Including every available variable is not automatically beneficial. In causal analyses, controlling for variables affected by the intervention or inappropriate selection variables can change the question or introduce bias. A statistically polished adjusted mean cannot rescue an unsuitable model, poor overlap or an unclear comparison.

Business use

Adjusted means can make comparisons more informative when groups differ in measured characteristics relevant to the outcome. They are useful in experiments with baseline measures, observational reporting and process comparisons, provided the purpose and assumptions are clear.

Publish the outcome units, covariate set, reference values or weights, uncertainty intervals and group contrast being evaluated. Compare the result with raw summaries and investigate large changes rather than assuming they validate the adjustment. For business decisions, consider whether the standardized population matches the customers, processes or participants to whom the decision will apply. Keep descriptive comparisons separate from claims about what would happen under an intervention.

Check your understanding

A group's raw mean is 50, its covariate mean is 12, the fitted common slope is 2 and the reference covariate value is 10. What is its adjusted mean?

Answer: 50 - 2(12 - 10) = 46, under the stated parallel-slopes linear model.

Related terms

Analysis of covariance
Standardization
estimated marginal means
least-squares means
covariate adjustment
reference grid
interaction and confounding

Learn next

calculate adjusted means at several reference values, then add a group-by-covariate interaction and examine how the comparisons change. State the target population and weights before interpreting any averaged predictions.

Sources

Primary sources: Howell, Statistical Methods for Psychology glossary, Adjusted means: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

Penn State, STAT 502, ANCOVA Part I: https://online.stat.psu.edu/stat502/Lesson09

Lenth, emmeans documentation, Basics of estimated marginal means: https://rvlenth.github.io/emmeans/articles/basics.html

Previous: Adjusted correlation ( r adj ) Next: Adjusting or controlling for a variable