I am trying to do multiple regression with a dependent variable (in which have half of the values are 0 and others are spread out). Transforming the dependent variable did not solve the problem of heteroscedasticity.
What can be done to solve heteroscedasticity.
Rida -
Depending on the application, you might want to do something special about all those zeroes. But maybe you just need to stratify.
Regarding heteroscedasticity, it is a natural phenomenon which is best characterized as "a feature, not a bug." For regression of form y = y* + e, you might want to consider the following:
https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression
In that project, please particularly note the following:
https://www.researchgate.net/publication/320853387_Essential_Heteroscedasticity
https://www.researchgate.net/publication/333642828_Estimating_the_Coefficient_of_Heteroscedasticity
and.
https://www.researchgate.net/publication/333659087_Tool_for_estimating_coefficient_of_heteroscedasticityxlsx.
Best wishes - Jim
Hello, Rida.
If you have a lot of response's values equal to 0, it could corrupt all your efforts in model's construction because all regression models are going to explain _differences_ in dependent variables by changes in independent ones. That differences in response are absent if you consider the case y = 0, and for the regression is nothing to explain. So one of the ideas how to eliminate heteroscedasticity's problem is to consider two cases separately: y = 0, or then y not equal to 0. logit-, or probit-approaches could help you not to drop all 0-cases.
The other drastic suggestion is to replace your response by some variable which is closely connected to yours, but which is not equal to 0 in numerous cases.
- Badges
- Science topic
- Similar topics
- Correction