What are the typical consequences of severe multicollinearity in a regression model?

What are the typical consequences of severe multicollinearity in a regression model?

• A. Inflated standard errors, unstable coefficient estimates, and difficulty identifying the individual effect of correlated predictors
• B. Decreased standard errors and more stable estimates
• C. No impact on inference
• D. Immediately invalidates the model but coefficients stay stable

Answer: (A)

When predictors are highly correlated, it becomes impossible to separate their individual effects. This makes the design matrix nearly singular, so the inverse used in estimating coefficients is unstable. As a result, the standard errors of the estimates inflate, confidence intervals widen, and the t-tests lose power. The coefficients themselves become highly sensitive to small changes in the data or in which predictors are included, so their exact values can swing even when the overall fit looks similar. Inference about which predictor truly drives the outcome becomes unreliable because the effects are entangled. Note that this situation typically does not bias the coefficient estimates in ordinary least squares, but it greatly reduces precision; in the extreme case of perfect multicollinearity, unique coefficient estimates cannot be obtained at all.

When predictors are highly correlated, it becomes impossible to separate their individual effects. This makes the design matrix nearly singular, so the inverse used in estimating coefficients is unstable. As a result, the standard errors of the estimates inflate, confidence intervals widen, and the t-tests lose power. The coefficients themselves become highly sensitive to small changes in the data or in which predictors are included, so their exact values can swing even when the overall fit looks similar. Inference about which predictor truly drives the outcome becomes unreliable because the effects are entangled. Note that this situation typically does not bias the coefficient estimates in ordinary least squares, but it greatly reduces precision; in the extreme case of perfect multicollinearity, unique coefficient estimates cannot be obtained at all.