In regression diagnostics, a model is efficient if the dispersion of the residuals is homogeneous (one of the 9 criteria). In the absence of homoscedasticity, the problem is one of heteroscedasticity.
The heteroscedasticity may be caused by a variable (by transformation or omission of an important variable) or by the sample itself (model function error, variable heterogeneity, etc.).
Several tests are available to detect this problem: the Breusch-Pagen and White tests, the Goldfeld-Quant test, etc.
The use of logarithmic or Box-Cox transformations to homogenize model variables or normalize distributions is recommended to correct heterocodasticity. However, when the variables are negative, this function is unusable.
The second solution is to weight the model. In Eviews, you need to click on the Weighted option after "Quick", then "Estimate equation". In Stata, this is done with the "robust" function in the equation. In R, you need to install the "MASS" library to use the rlm #(it robust regression model) function.
Abdi-Basid Adan thank you very much for the comprehensive answer , i have one more question in eviews 13 after i click on quick i get GLS weights the i choose cross-section SUR , is this correct in my case or may i have to choose another one
There's a full tutorial on the subject, just click on the link below. I think the differences between versions are minimal and don't represent an obstacle. I wish you the best of luck.
,how to remove Heteroskedasticity in EVIEWS 13, trying by using LOG values of the model but i am encountering an issue of negative variable values?
I presume that you are dealing with a problem of heteroskedasticity in economics or social science of similar.
The minimum assumptions underlying OLS estimation in such circumstances can be set out as follows. You have a population and some variables y, x1, x2,...,x_k that can be observed. The linear relationship
y = b0 + b1 x1 + b2 x2 + bk xk +u
between y and the x's hold. The purpose of the exercise is to estimate the b's. u is a random variable that has the following properties
1. E[e]=0
2. E[u|x1,x2,...xk]=0
3. The covariance matrix of the x's is positive definite
Note that I have made no assumption about heteroskedasticity in the population.
Now draw a random sample {x1, x2,...,x_n} from thae population.
The standard OLS estimates of the coefficients are consistent. If one also assumes that the variance of u is constant the standard OLS estimates of the variances of these estimates are correct. If you can not assume that they are correct you must use an alternative method of estimating these standard errors.
1. The method of adjustment most used is a robust method due to Eicker, Huber, and White. These may be called HCSE or robust standard errors or some reference to one of the originators in your software.
2 If you know the standard errors (or know they are proportional to some variable) then you can use GLS. However, GLS with some guesses of the standard errors may be inferior to OLS and robust errors. These guesses may not improve the estimated standard errors.
you
3. The third cause of heteroskedasticity is misspecification in your model. To look at this you must return to the original population equation. If there are likely to be negative values in this original equation you can not use log transforms of such variables to improve the diagnostics. Heteroskedasticity may be an indication of missing variables. It may also be an indication of some discontinuity in your model. Were there industrial disputes, or changes in government policy, changes in definitions of some of your series, etc? It is more likely that you will find the solutions to your problem in your economic analysis of the problem.
It is difficult to consider a regression in economics where the explanatory variables are non-stochastic (values that can be designed into the sampling). Economic variables are usually observed and the traditional Gauss-Markov results must be revised. Many introductory economics tests fudge this point. In the sciences, this is often not the case.
I think the key question is what kind of panel regression is tested (pooled, FE, ...?). Heteroscedasticity may be related to model specification problems: for example, the absence of a variable or important dummies. Technically, the problem of heteroskedasticity in panel regressions is solved by White robust errors.
For a linear model (see John's function for y), it is not allowed to take log-transformations of the variables, even if all the values were positive, because such a transformation (like other non-linear transformations) changes the specification of the model. Without any information about your specifidation and the variables (smples, time series?) you use, it is hardly possible to give advice. The main problem could be that you have not carefully specified the model/equation you want to estimate.
One way to remove heteroskedasticity in EViews is to use the White test. The White test is a statistical test that can be used to detect heteroskedasticity in a regression model. If heteroskedasticity is detected, the White test can be used to correct it by estimating the variance-covariance matrix using a robust estimator.
Regarding the issue of negative variable values, one way to deal with this issue is to use Box-Cox transformations of Y data with negative values. First, add a constant to make them all positive.
I would prefer not to remove heteroskedasticity (HS), because mostly this would reduce the information content of the data. As HS leaves estimates of coefficients unbiased (consistent), one should first estimate with the original data and analyse the calculated residuals. This procedure will ikely give you more insight about the causes and form of the HS than a standardized White test.