Typically, heteroscedasticity influences the infetential statistics and not the parameter estimation. Therefore, you can use robust estimators like HC3 or HC4, no need to transform the Data.
What statistical package will you use to process the data? For example, with STATA you can use the vce (robust) option to obtain robust errors in the presence of heteroscedasticity
Model and data issues can influence heteroscedasticity, either increasing or decreasing the impact, but using the format above can handle all of this, as shown in examples here:
Attached is the summary from "Estimating the Coefficient of Heteroscedasticity." I also included this in responses to https://www.researchgate.net/post/Is_there_heteroscedasticity_in_my_data_and_if_so_how_can_I_reduce_it
Also, please note in the project link above that one of the updates explains why using an hypothesis test for heteroscedasticity is not useful.
You asked "If homoscedasticity criteria is not being met in regression analysis, what tools can used to make it homosedastic?"
However, the homoscedasticity 'assumption' is not really appropriate because heteroscedasticity is to be expected for finite population applications when your model and data are ideal. That is, one should expect that with larger predictions, the sigma for the estimated residuals will be larger. One can handle that using weighted least squares (WLS) regression, which uses a weight which is a measure of size, say, for practical purposes, preliminary predicted y values, raised to the -2gamma power, i.e.,
w =
(preliminary predicted y)^(-2gamma),
where gamma is the coefficient of heteroscedasticity in papers I referenced above. Now, if gamma = 0, which ideally would not happen, then the weights are equal, and we have OLS regression. Finding a good gamma value is covered in the references I already provided.
If your software will let you enter regression weights, w, then it should also provide the estimated variance of the prediction error, which can be used in prediction intervals, around each prediction.
If the homoscedasticity hypothesis is not fulfilled, the standard errors will not be correct. You must use a robust estimation method against heteroscedasticity. What statistical software will you use? You can see:
Predictions are impacted to some degree, though the main problem is standard errors. Instead of ignoring impact on predictions and patching the standard errors, I think it is a lot cleaner to use weighted least squares (WLS) regression. That more closely models things as they actually are.
The WLS regression that James R Knaub suggests often does the trick if you can guess the weights (and sd rising with mean often works). Another straight forward approach is to model both the mean and the sd of the response variable. The R package lmvar does this in a fairly straight forward manner. This is useful if you aren't thinking of heteroskedaticity as just a nuisance, but you actually are interested in exploring it. More info at: https://rdrr.io/cran/lmvar/#vignettes
As a reminder of my earlier response, I have provided a tool for estimating regression weights.
I expect your software will let you enter w for the regression weights.
w = z^(-2 gamma)
Examples are in here: https://www.researchgate.net/publication/333642828_Estimating_the_Coefficient_of_Heteroscedasticity
A tool for doing this with your data and model is given here: https://www.researchgate.net/publication/333659087_Tool_for_estimating_coefficient_of_heteroscedasticityxlsx