What Does “Adjusted for” Mean in Regression? Covariates, Controls, and Confounders Explained

Start With the Regression Question

Suppose a researcher wants to examine whether study time (X) is associated with an examination score (Y).

An unadjusted linear regression might be written as:

Y = b₀ + b₁X + e

Here, b₁ describes the estimated difference in the outcome associated with a one-unit difference in study time.

Now suppose the researcher believes age, sex, and baseline score also need to be considered. A multiple regression could instead be written as:

Y = b₀ + b₁X + b₂(age) + b₃(sex) + b₄(baseline) + e

The meaning of b₁ has now changed. In multiple regression, a coefficient represents the estimated difference in the outcome between cases differing by one unit on that predictor while being equal on the other predictors included in the model. Hayes (2022) describes this as holding the other modeled antecedent variables constant; Meier et al. (2014) similarly develop multiple regression through partial slopes and the logic of statistical controls. (Meier et al., 2014; Hayes, 2022).

What Does “Holding Other Variables Constant” Mean?

Consider the coefficient for study time in the adjusted model.

Its interpretation is approximately:

Among observations that are equal on the other modeled predictors, how much does the predicted outcome differ for a one-unit difference in study time?

The important qualification is modeled predictors. The regression controls for variables that have actually been included in the model; it does not automatically control for every other difference between participants. Hayes (2022) explicitly notes that statistical control for one potential confound does not eliminate other possible sources of confounding. (Hayes, 2022).

This distinction is central to the meaning of control variables in regression. Statistical adjustment is conditional on the specified model and measured covariates. It is not equivalent to making otherwise different observations identical in every relevant respect.

Covariate, Control Variable, and Confounder Are Not Perfect Synonyms

These terms are often used loosely, but they refer to different ideas.

Why Adjustment Can Change a Coefficient

Suppose study time and examination score are positively associated. Students with higher baseline achievement might also tend to study more and score higher later.

In an unadjusted model, part of the study-time association may therefore reflect differences associated with baseline achievement.

Adding baseline score to the regression changes the question. The study-time coefficient now compares observations conditional on baseline score and the other included predictors.

Meier et al. (2014) illustrate this logic with a spurious bivariate relationship that largely disappears after a relevant third variable is introduced into a multiple regression. They also show the reverse possibility: controlling another predictor can reveal a relationship that was obscured in the bivariate analysis. Adjustment therefore does not have a predetermined effect on a coefficient; it can reduce, increase, reveal, or otherwise alter the modeled association. (Meier et al., 2014).

That is why the difference between an unadjusted and adjusted estimate can be substantively informative—but it still requires an explanation grounded in the variables and design.

A Small Synthetic Example

A researcher models final assessment score from hours of weekly study.

The coefficient has fallen from 2.4 to 1.3 after adjustment.

That result is compatible with some of the original association being statistically shared with age, sex, baseline score, or some combination of them. Multiple regression estimates the focal coefficient conditional on the other predictors rather than attributing all of their overlapping association with the outcome to the focal predictor. (Meier et al., 2014; Hayes, 2022).

Continuous and Categorical Covariates

Covariates do not all enter regression in the same form.

So covariates in regression can be continuous or categorical. What matters is how they are represented in the model and what conditional comparison the resulting coefficients describe.

Confounding Is About Alternative Explanations

The purpose of adjustment is often described as “controlling for confounding.” That can be reasonable, but the phrase should not be used mechanically.

Suppose X is associated with Y, but a third variable C helps produce the observed association. Failing to consider C can lead the analyst to attribute to X an association that is partly or wholly explained by another process. Meier et al. (2014) describe such relationships as spurious when a third variable accounts for the apparent relationship. (Meier et al., 2014).

Regression can help address a measured and appropriately selected confounding variable by conditioning the focal association on that variable. But this only addresses the confounding represented by the variables actually measured and appropriately modeled. Hayes (2022) emphasizes that controlling one covariate does not eliminate other potential confounding variables. (Hayes, 2022).

Causal Order Matters Before You Add a Control

Control-variable selection should not be treated as a search for every variable correlated with the outcome.

The International Encyclopedia of Statistical Science emphasizes that identifying potential confounders is fundamentally informed by subject-matter reasoning about causal ordering. Variables affected by the exposure or outcome are not ordinary baseline confounders, and controlling intermediate variables or consequences of exposure or outcome can introduce bias. Even some pretreatment variables can create problems when controlled in particular causal structures. (Lovric, 2011).

If X plausibly causes M, which subsequently influences Y, adding M as if it were an ordinary pre-existing confounder changes what the model is estimating. The analyst is no longer estimating the same X–Y relationship that would be estimated without conditioning on M. Causal ordering therefore needs to be considered before the regression model is specified. (Lovric, 2011; Hayes, 2022).

Hayes (2022) makes the broader temporal problem especially clear for observational data: when variables are measured contemporaneously without experimental manipulation, observed associations alone cannot establish which variable caused which. Statistical equations do not manufacture temporal ordering that the design failed to establish. (Hayes, 2022).

Why Choosing Controls by p-Value Is Not Enough

A tempting workflow is:

1. Test every possible covariate.
2. Keep the statistically significant ones.
3. Call the resulting coefficient “adjusted.”

Lovric (2011) describes selecting confounders according to significance tests of their associations with treatment or outcome as a common but unjustified approach. Confounder selection begins with substantive knowledge and causal ordering rather than simply asking which candidate variables achieve a conventional significance threshold in the current sample. The encyclopedia also cautions against selecting potential confounders solely according to how much their inclusion changes the effect estimate. (Lovric, 2011).

This distinction matters because prediction-variable selection and confounder control are not the same decision problem.

A variable can matter for causal adjustment even when its sample association does not happen to produce a small p-value. Conversely, a variable can strongly predict the outcome without being an appropriate confounder to control for.

Research Design Comes Before Regression Adjustment

Regression cannot compensate automatically for a weak causal design.

Sekaran and Bougie (2016) identify covariation, temporal precedence, control of alternative explanations, and a logical theoretical basis among the requirements involved in establishing cause-and-effect relationships. Their discussion of experimental design emphasizes randomization as a design mechanism for distributing both known and unknown confounding influences across groups. (Sekaran & Bougie, 2016).

Adams and Lawrence (2018) similarly distinguish correlational and experimental designs. Without random assignment, observed group or variable differences can remain open to third-variable explanations; correlational associations do not by themselves demonstrate causation. (Adams & Lawrence, 2018).

This highlights an important difference between design control and statistical control.

They are not interchangeable. (Sekaran & Bougie, 2016).

Association Versus Causation

Return to the original sentence:

“The association remained significant after adjusting for age, sex, and baseline score.”

A defensible interpretation is:

Depending on the model, the focal coefficient represents the estimated association conditional on those included predictors. (Meier et al., 2014; Hayes, 2022).

Causal interpretation requires evidence beyond the existence of an adjusted regression coefficient, particularly evidence supplied by research design, temporal ordering, substantive theory, and the handling of plausible competing explanations. (Meier et al., 2014; Lovric, 2011; Hayes, 2022; Sekaran & Bougie, 2016; Adams & Lawrence, 2018).

“It Stayed Significant” Is Not the Main Result

Researchers often compare models only by asking whether the p-value stayed below a significance threshold.

That misses important information.

Suppose the unadjusted coefficient is 2.4 and the adjusted coefficient is 1.3. Both might be statistically significant, yet adjustment has changed the estimated magnitude substantially.

Multiple regression is valuable because it changes the conditional comparison represented by a coefficient. Reporting only that an association “remained significant” hides that substantive information. (Meier et al., 2014; Hayes, 2022).

Why “Adjusted” Does Not Automatically Mean “Unbiased Causal Estimate”

An adjusted coefficient can still have a problematic causal interpretation for several reasons.

“Adjusted” should therefore be read as:

conditional on the covariates and model specification used in this analysis

—not as:

guaranteed free of bias and causally identified.

Five Questions to Ask Before Adding a Control Variable

1. Why is this variable being controlled?

   Is there a substantive reason to believe it represents an alternative explanation for the focal association, or is it being included merely because it was measured or happens to have a small p-value? Confounder selection should be informed by causal and subject-matter reasoning rather than automatic significance screening. (Lovric, 2011).

2. Where does the variable occur in the causal order?

   Ask whether the proposed control existed before the focal exposure or predictor, could have been caused by it, could have been caused by the outcome, or might lie on a pathway connecting predictor and outcome. Controlling an intermediate variable or a consequence of exposure or outcome can introduce bias or change the target relationship. (Lovric, 2011; Hayes, 2022).

3. What comparison will the adjusted coefficient represent?

   After adding the control, interpret the focal coefficient explicitly: it compares predicted outcomes for observations differing on the focal predictor but equal on the other modeled predictors. If that conditional comparison is scientifically unclear or irrelevant, simply adding another covariate has not improved the analysis. (Hayes, 2022).

4. Does the research design support the interpretation I want to make?

   Cross-sectional association, longitudinal observation, quasi-experimental comparison, and randomized experimentation provide different foundations for causal reasoning. Statistical adjustment should be interpreted within—not instead of—the study design. (Hayes, 2022; Sekaran & Bougie, 2016; Adams & Lawrence, 2018).

5. What important alternative explanations remain after adjustment?

   A model may control age, sex, and baseline score while omitting other relevant causes or proxies. Controlling measured covariates does not prove that no residual confounding remains. (Hayes, 2022).

How to Read an “Adjusted for” Sentence

When you encounter:

“The association remained significant after adjusting for age, sex, and baseline score.”

translate it mentally into something closer to:

“In a regression model containing the focal predictor, age, sex, and baseline score, the estimated coefficient for the focal predictor remained statistically distinguishable from zero under the model’s inferential procedure.”

Then ask what the paper still needs to establish:

Those questions turn “adjusted for” from a reassuring phrase into an interpretable methodological decision.

The Bottom Line

What does controlling for variables mean? In multiple regression, it means estimating a predictor’s coefficient conditional on the other variables included in the model—often described as holding those modeled predictors constant. (Meier et al., 2014; Hayes, 2022).

Adjustment can be valuable. It can show whether an association persists after accounting statistically for specified covariates, help investigate third-variable explanations, and substantially change the estimated coefficient. (Meier et al., 2014; Hayes, 2022).

But adjustment is only as meaningful as the reasoning behind the model. A covariate is not automatically a confounder, more controls are not automatically better, and statistical significance after adjustment does not establish causation. Causal order, appropriate variable selection, temporal information, research design, and remaining alternative explanations still determine how far interpretation can go. (Lovric, 2011; Hayes, 2022; Sekaran & Bougie, 2016; Adams & Lawrence, 2018).

References

Adams, K. A., & Lawrence, E. K. (2018). Research methods, statistics, and applications (2nd ed.). SAGE Publications.

Hayes, A. F. (2022). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach (3rd ed.). The Guilford Press.

Lovric, M. (Ed.). (2011). International encyclopedia of statistical science. Springer. https://doi.org/10.1007/978-3-642-04898-2

Meier, K. J., Brudney, J. L., & Bohte, J. (2014). Applied statistics for public and nonprofit administration (9th ed.). Cengage Learning.

Sekaran, U., & Bougie, R. (2016). Research methods for business: A skill-building approach (7th ed.). John Wiley & Sons.