Suppressor Variables in Multiple Regression
In a multiple regression $y \sim x_1 + x_2$, both $\hat{\beta}_1$ and $\hat{\beta}_2$ have significant t-statistics. But when you run the simple regression $y \sim x_1$ alone, $\hat{\beta}_1$ is no longer significant.
What could cause this? Explain the mechanism and provide a concrete example.
Hints
- Think about what OLS is doing differently in $y \sim x_1$ versus $y \sim x_1 + x_2$: in the multiple regression, each coefficient is estimated *conditional on* the other predictor.
- A variable can inflate the noise in another predictor without itself being related to the outcome. Consider what happens to $\text{Var}(x_1)$ when part of it is explained by $x_2$.
- Write out the partial coefficient formula: $\hat{\beta}_1 = (\text{Cov}(y, x_1) - \hat{\beta}_2 \, \text{Cov}(x_2, x_1)) / (\text{Var}(x_1)(1 - r_{12}^2))$. If $x_2$ is correlated with $x_1$ but not $y$, trace through how the numerator and denominator each change.
Worked Solution
How to Think About It: This is a classic trap in applied regression. Your first instinct might be that something is wrong -- how can a variable become significant only after you add a second predictor? The answer is that $x_2$ is acting as a *suppressor*: it is soaking up variance in $x_1$ that is noise with respect to $y$. Once you strip that noise out by conditioning on $x_2$, the signal in $x_1$ becomes legible. The key is that $x_2$ does not need to correlate with $y$ on its own -- it just needs to correlate with the junk in $x_1$.
Key Insight: OLS partial coefficients measure the effect of each predictor *after removing* the linear influence of the others. When $x_2$ is correlated with $x_1$, including it re-partitions the variance in $x_1$, and if the piece it removes is noise, the signal-to-noise ratio of $\hat{\beta}_1$ improves.
The Mechanism:
- Simple regression: $\hat{\beta}_1 = \text{Cov}(y, x_1) / \text{Var}(x_1)$. This estimate uses all variance in $x_1$ -- signal and noise alike -- in its denominator. If $x_1$ contains noise that is correlated with $x_2$ but not with $y$, that noise inflates $\text{Var}(x_1)$ without helping $\text{Cov}(y, x_1)$. The t-statistic is small.
2. Multiple regression: The partial coefficient is $$\hat{\beta}_1 = \frac{\text{Cov}(y, x_1) - \hat{\beta}_2 \, \text{Cov}(x_2, x_1)}{\text{Var}(x_1)(1 - r_{12}^2)}$$ where $r_{12}$ is the correlation between $x_1$ and $x_2$. Think of this as: project $x_1$ onto the space orthogonal to $x_2$, and regress $y$ on that residual. The $x_2$-correlated noise has been stripped out. If that noise was the main source of variance in $x_1$, both the numerator and denominator shrink -- but the numerator holds its value while the denominator drops faster, so the t-statistic rises.
- Standard error: The SE of $\hat{\beta}_1$ scales with $1/\sqrt{\text{Var}(x_1)(1-r_{12}^2) \cdot n}$. Removing noise variance reduces the effective denominator, tightening the SE and pushing the t-statistic up.
Concrete Example:
Let $y$ = job performance, $x_1$ = aptitude test score, $x_2$ = test-day anxiety level.
- Anxiety is essentially uncorrelated with actual job performance (anxious people are not worse employees).
- But anxiety is negatively correlated with test scores: it depresses scores for high-ability candidates.
So $x_1$ is a noisy signal of true ability -- the noise comes from anxiety. In the simple regression $y \sim x_1$, the noise blurs the ability-performance relationship and the coefficient is weak. Add $x_2$ (anxiety), and the regression holds anxiety constant, isolating true ability. Now $\hat{\beta}_1$ measures the clean ability effect and becomes significant. Meanwhile $\hat{\beta}_2$ picks up the anxiety-score link, which also shows up as significant because conditioning on ability exposes the anxiety channel.
Practical Considerations:
- Suppression is not a pathology to fix -- it reveals that $x_1$ contains mixed signals. The multiple regression is giving you a more precise answer, not a wrong one.
- Watch for this when variables are proxies or composites (test scores, survey indices, PCA components). Any proxy carries noise from its construction process.
- In finance: a factor loading on a portfolio might look insignificant in isolation but become significant once a correlated risk factor is controlled for. For example, a value factor might appear insignificant until you control for a quality factor that shares variance.
- Suppression can also flip signs (not just significance), which is the more alarming case. If $\hat{\beta}_1$ changes sign when you add $x_2$, check your correlations carefully before trusting any interpretation.
Answer: $x_2$ is a suppressor variable -- it correlates with $x_1$ but not with $y$, so it absorbs noise in $x_1$. In the multiple regression, the partial coefficient of $x_1$ isolates the signal component, increasing the t-statistic. This is not a statistical artifact; it reflects a genuine improvement in precision from conditioning on a relevant nuisance variable.
Intuition
The key lesson here is that significance is not a property of a variable in isolation -- it is a property of a variable *given what else is in the model*. A predictor can be swimming in noise when measured alone, but become a sharp signal once you hold the noise source constant. This is why adding controls to a regression can increase (not just decrease) the precision of other coefficients.
This pattern shows up constantly in applied quant work. In factor models, a raw momentum signal might be noisy because it loads on volatility. Once you orthogonalize against a volatility factor, the clean momentum effect snaps into focus. In econometrics, omitting a variable correlated with your regressor does not just bias the coefficient -- it can completely mask whether the effect exists. The practical takeaway: if a t-statistic jumps when you add a control, that control was suppressing noise, not just soaking up explained variance. Understanding *why* requires looking at the correlation structure between your predictors.