pls correct me if i'm wrong. of the econometrics' literature i've read so far, most mentioned heteroskedasticity is not a major problem empirically but multicollinearity would pose a greater concern to researchers and that data transformation would improve empirical results but not completely remove heteroskedasticity. Is this right?
by above, i've used R to conduct a basic linear regression with one regressor and three dummies and conducted Breusch-Pagan test using package 'car' and VIF test using package 'lmtest'. Thereafter, i conducted a second regression, log-transforming the dependent and independent variable and carried similar tests as aforementioned.
for both models, m1 for former and m2 for latter, and using VIF test, both presented less than 5 and reasonably to say multicollinearity isn't a problem
however, for Breusch-Pagan, m1 failed to reject null but m2 (logged) rejected it. can someone enlighten me why the result is not consistent with the theory above?
some related questions,
would it better not to log-transform the model?
i'm suspecting that the regressor is correlated with the residuals. is there test(s) 'out there' for endogeneity?
if log model is preferred, what can i then do to 'lessen' heteroskedasticity in m2?
Thanks in advance!