In presence of heteroskedasticity you can go with Weighted least squares based estimation of the model parameters and associated appropriate regression analysis.

We may use also Generalized LS or GLM. You can use also HAC (white) standard errors and analyze differences. Good luck!

Data transformations can also be used. These are equivalent to the above recommended methods. If you don't make the effort you are going to have an incorrect result. D Booth

In regression analysis, heteroscedasticity is defined in the context of residuals or error term. The OLS estimators remains unbiased and consistent in the presence of heteroscedasticity, but they are no longer efficient not even asymptotically. To overcome this probem of inefficiency, one can do the weighted least square (WLS) if the heteroscadastic variances are known, otherwise can apply appropriate transformation of dependent or independent variables. You can read the discussion provided in the link below:

http://itfeature.com/heteroscedasticity/remedial-measures-for-heteroscedasticity

Aleksandr -

I'd go with weighted least squares (WLS) regression.  To do that, you need a regression weight.  For that an estimate of the coefficient of heteroscedasticity, gamma, would be nice.  I just wrote a(nother) paper on that:

 https://www.researchgate.net/publication/333642828_Estimating_the_Coefficient_of_Heteroscedasticity

I'll soon add an excel file to this which will do the quick methods illustrated. 

Default values might include 0.75, preferred by Ken Brewer for business surveys.

Cheers - Jim

PS - Aleksandar -

Here is that spreadsheet:

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

So the problem with an hypothesis test - one of the problems with hypothesis tests - in the case of considering heteroscedasticity, is that if you decide (1) to use OLS, then you won't know if there was really an impact on results, and if you decide (2) that you do have heteroscedasticity, then you are still concerned about the practical problem: "How much?"

However, if you estimate the coefficient of heteroscedasticity, or use a default based on other knowledge, you can try that in a regression weight and see what the impact is on results.

Please note the following:

https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression