Can a predictor with zero outcome correlation matter?

Yes, through its joint relationship with other predictors.

Is a larger coefficient always better?

No; scale, uncertainty, model purpose, and out-of-sample performance matter.

AI Glossary / Statistics and probability / Intermediate

Suppressor variable

Definition

A suppressor variable is a predictor whose inclusion in a regression improves the predictive contribution of another predictor by accounting for variation that obscures its relationship with the outcome. In classical suppression, the suppressor can have little or no simple correlation with the outcome while still helping the predictors work together.

What it means

A variable can be useful in combination even when it looks unhelpful on its own. Suppose a score contains both useful information and an irrelevant component. A second measurement that captures the irrelevant component can help regression separate the two.
Suppression is a property of the relationships among the variables in a specified model. It is not automatically a causal mechanism, proof of confounding, or evidence that the added variable should always be retained. Definitions of particular suppression patterns vary across the methodological literature.

How it works

Begin with the outcome and candidate predictors, making their measurement scales and intended roles explicit. Examine simple relationships as well as relationships conditional on other predictors. Screening solely by outcome correlation can miss a useful joint contribution.
Fit a baseline model and a model including the proposed suppressor. Compare the relevant coefficients, explained variation, residual behavior, and predictive performance. Use standardized coefficients when comparisons require a common scale, while retaining the original units for interpretation.
In a classical two-predictor example, the suppressor is correlated with the other predictor but uncorrelated with the outcome. Its regression coefficient can nevertheless be nonzero, because it helps separate the outcome-related component of the other predictor.
The coefficient of the main predictor may increase in magnitude when the suppressor is included. The added predictor may receive a sign that seems surprising when viewed only through its simple association with the outcome. Interpret that sign conditionally, holding the other model variables fixed.
Do not identify suppression from an increase in training R squared alone. Adding a predictor to ordinary least squares cannot reduce training R squared, even when the new variable contributes little useful information beyond sampling noise.
Check uncertainty and stability. Correlated predictors can create unstable estimates, especially in small samples. Replication, held-out evaluation, or cross-validation helps assess whether the apparent improvement survives beyond the fitted data.
For causal questions, use knowledge of how variables arise. Adjusting for a mediator, collider, or consequence of the outcome may change coefficients for reasons that do not represent a useful causal adjustment. A statistical suppression pattern does not decide which causal estimand is appropriate.
For prediction, confirm that the proposed variable will be available at prediction time and does not leak information about the outcome. A mathematically impressive fit can be operationally invalid if it uses unavailable or retrospectively constructed measurements.

Three practical examples

Assessment: a measurement of response style may help isolate useful information in another score.
Operations: a nuisance measurement can help separate instrument artifacts from a signal associated with quality.
AI feature review: a weak marginal predictor may improve joint prediction when it captures irrelevant variation in another feature.

Think of it this way

Think of a microphone recording speech plus background noise. A second microphone aimed at the noise may contain little speech, yet help recover the first microphone's useful content. The analogy explains joint usefulness; it does not imply that every regression suppressor literally measures noise.

Worked example

Consider an original population correlation example with standardized variables. Let the correlation between outcome Y and predictor X1 be 0.60, between Y and X2 be zero, and between X1 and X2 be 0.50. This is a valid positive-definite correlation structure.
Using X1 alone gives a standardized slope of 0.60 and R squared of 0.36. With both predictors, the standardized slope for X1 is (0.60 - 0.50 times 0)/(1 - 0.50 squared) = 0.80. The slope for X2 is (0 - 0.50 times 0.60)/0.75 = -0.40.
The multiple R squared is 0.80 times 0.60 + (-0.40) times 0 = 0.48. Thus including X2 raises explained population variation from 36% to 48%, despite its zero simple correlation with Y.
The negative conditional coefficient for X2 does not contradict its zero marginal correlation. These stipulated population relationships illustrate classical suppression. They provide no sample-size-based significance test and no causal interpretation.

Illustrative case study

Fictional case: Spruce Analytics predicts task quality using a composite assessment score. A separate response-style measure has almost no simple correlation with quality, so an initial feature screen removes it. Analysts later find that including it stabilizes the useful component of the composite score in a controlled evaluation. They test the improvement on held-out cases and check measurement consistency before deployment. The team describes a possible suppression pattern rather than claiming that changing response style would cause quality to change.

Common misconceptions and mistakes

Do not discard every variable with low marginal correlation. Do not label every coefficient sign reversal as suppression without investigating the model. A rise in training fit is insufficient evidence of useful prediction. Standardized coefficients are not causal effect sizes. Apparent suppression can reflect sampling instability, measurement choices, or leakage rather than a durable relationship.

Business use

Use the concept when reviewing feature selection, composite measures, and unexpected coefficient changes. It encourages analysts to evaluate predictors jointly and explain why a variable is included. For deployment, prioritize reproducible predictive value and available measurements. For causal decisions, require a defensible causal model rather than relying on the suppression label.

Check your understanding

What is X2's simple correlation with Y in the example?

Answer: Zero.

What happens to X1's standardized coefficient?

Answer: It rises from 0.60 to 0.80.

Why is an increase in training R squared alone insufficient?

Answer: Ordinary least squares can improve training fit simply by adding flexibility.

Related terms

Learn next

Study partial regression plots and standardized regression formulas. Then compare suppression with confounding and mediation, keeping prediction and causal interpretation separate.

Sources

Frees, Regression Modeling with Actuarial and Financial Applications, Variable Selection: https://users.ssc.wisc.edu/~ewfrees/ActuarialRegression2009/C5VarSelect.html Smith, Ager and Williams, Suppressor Variables in Multiple Regression/Correlation: https://journals.sagepub.com/doi/10.1177/001316449205200102 The correlation example is original.

Previous: Supporting hyperplane Next: Supremum (or least upper bound)