Heteroskedasticity is the violation of the assumption, made in some linear regression models, that all the errors of the regression have the same variance.
Table of contents
Consider the linear regressionwhere:
is the regressand;
are the regressors;
is the vector of regression coefficients;
is the error term;
the observations are indexed by .
Sometimes we assume that all the error terms have the same variance, that is,
When this hypothesis holds, we say that the errors are homoskedastic (or homoscedastic).
On the contrary, when the errors pertaining to different observations do not have the same variance, the errors are said to be heteroskedastic (or heteroscedastic).
In this case, we also say that the regression suffers from (unconditional) heteroskedasticity.
In most cases, we make an hypothesis stronger than homoskedasticity, called conditional homoskedasticity:where is the design matrix (i.e., the matrix whose rows are the vectors of regressors for ).
In other words, we postulate that the variance of the errors is constant conditional on the design matrix.
When this assumption is violated, the regression is said to suffer from conditional heteroskedasticity.
When the errors are assumed to be not only conditionally homoskedastic, but also conditionally uncorrelated, we can write:where is the vector of error terms and is the identity matrix.
In this case, we say that the errors are spherical.
The ordinary least squares (OLS) estimator of the vector of regression coefficients iswhere is the vector of observations of the dependent variable and is design matrix defined above.
Conditional homoskedasticity is one of the assumptions of the Gauss-Markov theorem.
Therefore, OLS is not guaranteed to be BLUE when a regression suffers from conditional heteroskedasticity.
The generalized least squares (GLS) estimator of the vector of regression coefficients iswhere
If all the conditions of the Gauss-Markov theorem except conditional homoskedasticity are met, OLS is not the BLUE estimator, but the GLS estimator is.
For more details and a proof, see the lecture on the generalized least squares estimator.
When is diagonal, is also called Weighted Least Squares (WLS) estimator.
Conditional heteroskedasticity does not per se introduce biases in the OLS estimator.
If conditional homoskedasticity is violated, but the other Gauss-Markov assumptions hold, then the OLS estimator remains unbiased.
Since we can write the OLS estimator asThe conditional expectation of is where we have used the strict exogeneity assumptions of the Gauss-Markov theorem, namely,By the Law of Iterated Expectations, we have
When all the Gauss-Markov hypotheses hold, the conditional covariance matrix of is
We have proved above thatTherefore,where we have used the assumption of sphericity (no correlation + conditional homoskedasticity), that is,
When only the homoskedasticity condition is violated, then
See previous proof.
The latter formula can be used in practice to compute the exact covariance of the OLS estimator only if:
the other assumptions of the Gauss-Markov theorem are satisfied.
Since these two conditions are unlikely to be both met in practice, we usually compute the heteroskedasticity-consistent estimator presented in the next section.
Under fairly weak assumptions, we can prove (see here) that the asymptotic covariance matrix of iswhereis the so-called long-run covariance matrix.
In a sufficiently large sample, the covariance matrix of is approximately equal to the asymptotic one, divided by the sample size:
The matrix is consistently estimated by
If the terms of the sequence are not serially correlated, the long-run covariance matrix becomesand it is consistently estimated bywhere:
are the residuals
is an diagonal matrix having the squared residuals on its main diagonal.
Putting the various pieces together, we obtain
This popular estimator of the covariance matrix has various names:
White's (1980) estimator;
heteroskedasticity-consistent estimator (HCE);
Note that White's estimator can be used only if the terms of the sequence are uncorrelated (a condition we imposed above).
If the terms of are correlated, then we need to use another estimator called heteroskedasticity and autocorrelation consistent (HAC) or Newey-West estimator.
There are numerous statistical tests that can be used to detect heteroskedasticity, for example:
When the linear regression is performed on time-series data, there are popular models that can be used to analyze and predict how the variance of the error terms changes through time:
the Gauss-Markov theorem;
Greene, W.H., 2017. Econometric analysis, 8th edition, Pearson.
Gujarati, D.N., 2017. Basic econometrics, 5th edition, McGraw-Hill.
White, H., 1980. A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica, 48, pp.817-838.
Previous entry: Factorial
Next entry: IID sequence
Please cite as:
Taboga, Marco (2021). "Heteroskedasticity", Lectures on probability theory and mathematical statistics. Kindle Direct Publishing. Online appendix. https://www.statlect.com/glossary/heteroskedasticity.
Most of the learning materials found on this website are now available in a traditional textbook format.