The RATS Software Forum

Dear Sir

I am estimating an OLS regression, dependent variable is inflation, explanatory variables include usual controls and exchange rate, data is monthly, from 1999 to 2009, all variables are stationary, coefficients are signigicant, regression passes usual Bruesch-Godfrey LM test, ARCH Test, Ramsey RESET, but fails to pass Jarque-Bera test (so the regression residuals are non-normal). My question is:
(i) Should I rely on the results in the absence of residual non-normaltiy?

secondly, I guess the residual non-normality is due to a unusually high inflation values for the year 2003 (12 values in our case as we have monthly data, these look like outliers), in this context my question is:
(ii) what could be the solution to get normal residuals? Can we use dummy variable to get regression residuals normality?

Thanking you in advance.

Tahir

When only the assumption of normality is not fulfilled, the LS estimators continue to have the smallest variance among all linear unbiased estimators. Note that the normality assumption is not necessary for the partitioning of total variation or for variance estimation. But, it is necessary for tests of significance (p-values). In such cases, transformations of the dependent variable and alternative estimation procedures (weighted least squares and FGLS) could be used to improve the non-normality. Examples for the transformations include the square root, the logarithmic, and the logistic transformations.