Dear Neelam Raghav

You can use other transformations such as square root and inverse value, and you can also use the Johansen transform

HETEROSCADASTICITY is a problem if the underlying model assumes unifkr error. The lack of uniform errors is called heteroscadasticity. The requirement of homoscadastcjty or uniform modelling error is common in the regression family which calls for the Gauss-Markov theorem. But the existence of heteroscadasticity may be useful in telling you that your proposed model type is not appropriate for your data set. You need to explore alternative modeling options.

https://en.m.wikipedia.org/wiki/Heteroscedasticity

DATA MANIPULATION is unprofessional and unethical. If one of the variable produces hetroscadasticiy and is not what is required in the model, you must remove that variable; to keep it otherwise would result in Type 1 error: insisting of being correct despite the data is telling you that it is wrong. To manipulated the data in order to remove heteroscadasticity is unethical.

REMOVAL here is to remove the variable column itself, not removing the heteroscadasticity characteristics of the data means means of manipulating the data. Such conduct is not ethical. In such a ase, the data itself, after manipulation by whatever method under whatever sophisticated name it may be, lacks data integrity, and thus, falls short of scientific standard.

Dear Neelam Raghav

You can use other transformations such as square root and inverse value, and you can also use the Johansen transform

You can try different power transformations; if it is not removed, you can try VAR-GARCH models

Clearly, more details on your problem are needed

Hi,

I'm in complete agree with Paul about data manipulation but if you really need to remove heteroscedasticity, you may model this variable using volatility models (GARCH, Stochastic volatility or other models) then you compute the conditional variance and finally standardize your variable using these computed conditional variances.

Hi Neelam,

Using logs is just one possibility for stabilizing the variance of a time-series variable. If instead you use the family of power transformations, you will have access to an infinite number of possibilities, by just changing the index of the power transformation family.

The following paper deals with this topic: Guerrero, V.M. (1993) Time series analysis supported by power transformations. Journal of Forecasting, Vol. 12, pp. 37-48.

Regards

Neelam -

My experience is primarily with finite population sampling and inference, where heteroscedasticity naturally ocurs due to different sizes of predictions, but can be made of larger or smaller impact due to model specification or data issues. However, if you only need to go from a one level OLS model to a weighted least squares (WLS) regression model, that is, just handle heteroscedasticity at the basic level, then you could try the following:

.....................

A spreadsheet tool for estimating or considering a default value for the coefficient of heteroscedasticity, developed for linear regression, is found here (with references):

https://www.researchgate.net/publication/333659087_Tool_for_estimating_coefficient_of_heteroscedasticityxlsx

In order to move from OLS to WLS regression, one may pick a coefficient of heteroscedasticity from the above, and use that to find the regression weight expression for this purpose.  The regression weight would be "w" entered into SAS PROC REG, for example. 

......................

Heteroscedasticity is a natural phenomenon, not an anomaly. OLS regression is a special case of WLS regression where the regression weights are all equal (i.e., w is a constant such as w = 1). OLS is often not a good default. See the above link, and references given in there.

Best wishes - Jim

Several people have already suggested above what is also known as Tukey's transformation,

Dear Neelam Raghav , Dear All,

Various existing methods for the treatment of the problem of heterocedasticity. Once again, the heterogeneity of the observation residue variations allows the instability of the estimator modeled by the equation in question.

How to get rid of it in order finally to propose a reliable, coherent and robust estimator on which decision-making can be claimed.

Some propose the manipulation of data: No, the analysis of samples is predisposed as being fixed on a period of observation.

The option of deleting the variables may omit explanatory variables which, in turn, will generate another so-called simultaneity problem in the model.

The model transformation solution in GARCH assumes that the model relies on certain conditions such as stationarity.

Johansen's transference is applied in the field of spatial econometrics.

The simplest measure is to use the sandwich estimator, which offers a reliable estimator without the problem of heterocedasticity.

Certainly, the availability of MULTIPLE solutions exists. Which method must be approved, knowing that the choice of another model is based on certain criteria.

This returns to an iterative correction, in which case the redefinition of the model becomes necessary.

Kind Regards

As I noted above, going to WLS regression from OLS (which is a special case of WLS, but often it is not the one you want) is quite manageable. Please see https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression