The variance of the error term is not constant, which makes the model inefficient and masks the statistical importance of features.

Get ready for the GARP Risk and AI Exam with flashcards and multiple choice questions. Each question comes with hints and explanations. Prepare for success!

Multiple Choice

The variance of the error term is not constant, which makes the model inefficient and masks the statistical importance of features.

In regression, having a constant spread of errors across all observations is a basic assumption. When that spread changes with the level of the predicted value or over the data, it’s called heteroskedasticity. The key problem it creates is not that the coefficient estimates become biased, but that the standard errors used to judge their significance become biased too. That means the usual t-tests and confidence intervals can be unreliable, so you might misjudge which features truly matter. In other words, heteroskedasticity lowers the efficiency of the estimates and can mask the true statistical importance of predictors. Residual plots are a handy way to spot potential heteroskedasticity—look for a pattern in the spread of residuals as fitted values grow. There are formal tests as well, like Breusch-Pagan or White tests. Remedies focus on adjusting standard errors to be robust (robust standard errors), transforming the dependent variable (for example, log or Box-Cox transformations), or using methods that accommodate non-constant error variance (such as weighted least squares) if the form is known. This phenomenon is different from multicollinearity, which involves high correlations among predictors, not the variance of the error term. Ridge is a technique to address multicollinearity and overfitting, and residual plots are diagnostic tools rather than the phenomenon itself.

In regression, having a constant spread of errors across all observations is a basic assumption. When that spread changes with the level of the predicted value or over the data, it’s called heteroskedasticity. The key problem it creates is not that the coefficient estimates become biased, but that the standard errors used to judge their significance become biased too. That means the usual t-tests and confidence intervals can be unreliable, so you might misjudge which features truly matter. In other words, heteroskedasticity lowers the efficiency of the estimates and can mask the true statistical importance of predictors.

Residual plots are a handy way to spot potential heteroskedasticity—look for a pattern in the spread of residuals as fitted values grow. There are formal tests as well, like Breusch-Pagan or White tests. Remedies focus on adjusting standard errors to be robust (robust standard errors), transforming the dependent variable (for example, log or Box-Cox transformations), or using methods that accommodate non-constant error variance (such as weighted least squares) if the form is known.

This phenomenon is different from multicollinearity, which involves high correlations among predictors, not the variance of the error term. Ridge is a technique to address multicollinearity and overfitting, and residual plots are diagnostic tools rather than the phenomenon itself.

More practice questions