what is multicollinearity in regression

June 28, 2026

Multicollinearity in regression is when two or more predictor (independent) variables are highly linearly related to each other, instead of being largely independent.

Quick Scoop: What it is

In a multiple regression model, you ideally want each predictor to contribute unique information about the outcome.

With multicollinearity, one predictor can be (almost) predicted from a combination of others, because they share a strong linear relationship.

In the extreme “perfect multicollinearity” case, one predictor is an exact linear combination of others (for example, X3=2X1−X2X_3=2X_1-X_2X3=2X1−X2), and the regression coefficients cannot be uniquely estimated.

In short: the predictors start “telling the same story,” which confuses the model about how much credit to give each variable.

Why it’s a problem

Multicollinearity does not necessarily ruin predictions, but it makes the coefficients unstable and hard to interpret.

Key issues:

Large standard errors for coefficients, leading to wide confidence intervals and non‑significant t‑tests, even when variables truly matter.

Coefficient estimates become very sensitive to small changes in the data: add or remove a few observations, and signs/magnitudes can flip.

It becomes hard to say which predictor is really associated with the outcome, since several move together.

Example: Suppose you include both “total hours online” and “hours on social media” as predictors of stress. Since those two are highly correlated, the model struggles to decide which one gets the effect, and both coefficients may look weak or unstable.

How you detect it (high level)

Common diagnostics used in practice include:

Correlation matrix of predictors (look for very high correlations).

Variance Inflation Factor (VIF): large VIF values indicate a predictor’s variance is inflated by collinearity.

Condition indices and related measures, which flag linear dependencies among multiple predictors at once.

What you typically do about it

When multicollinearity is severe and harms interpretation, common remedies include:

Dropping or combining strongly overlapping variables (e.g., keep either “total hours online” or the breakdown, not both).

Collecting more data, if feasible, so coefficients are estimated more stably.

Using regularization methods (like ridge regression) that shrink coefficients and handle correlated predictors better.

Re‑expressing predictors (e.g., using orthogonal transformations or orthogonal polynomials for polynomial terms).

Tiny HTML summary table

Here’s a compact HTML table you can reuse:

html

<table>
  <thead>
    <tr>
      <th>Aspect</th>
      <th>Explanation</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Basic idea</td>
      <td>Predictors in a regression model are highly linearly related to each other rather than independent. [web:1][web:3][web:9]</td>
    </tr>
    <tr>
      <td>Main consequence</td>
      <td>Coefficient estimates become unstable, with large standard errors and unreliable significance tests, even if overall prediction can still look good. [web:4][web:5][web:10]</td>
    </tr>
    <tr>
      <td>Detection</td>
      <td>Check predictor correlations, Variance Inflation Factors (VIF), and condition indices or related diagnostics. [web:3][web:5][web:10]</td>
    </tr>
    <tr>
      <td>Remedies</td>
      <td>Remove or combine overlapping variables, collect more data, use regularization methods, or transform predictors into less correlated forms. [web:1][web:2][web:7][web:10]</td>
    </tr>
  </tbody>
</table>

Bottom note: Information gathered from public forums or data available on the internet and portrayed here.