Dear members, need your help regarding system GMM.
I have panel data for 14 years and 70 countries and lagged dependent variable.Most of variables were in log form. I applied System GMM and got expected result.
Later when I averaged data over three and five years and applied system GMM with same log transformed variables results became insignificant.
But when I applied System GMM with same variables in its original metric form, results were again as per expectation.
This is what I suspect is happening: log transforming variables is generally done to put values on different scales onto a more common scale to facilitate analysis. This process then shrinks the range of values towards some central value. Averaging will further shrink the range of values over a smaller support. Both of these processes can then make it tougher to meet the criteria of significance since the values are not as far apart as in their original state.
Thanks, Gitz. I recommend to carefully look at the data before continuing. Five period T does not seem small (but check it).
In principle, it is not feasible to apply the non-linear (in X and Y) function lnY=a*lnX+c on the arithmetic averages of Y and X (the geometric averages would be correct). In practice, the function Y=kX+d would be almost linear, too. Therefore, the coefficients from the regression with the means of the logs are likely to be similar. The statistical problem with the means is, that there is a high correlation between the X's, Y's and the error terms. This means a higher probability of biases and, in any case with several independent, partly lagged variables, higher standard deviations of the estimated coefficients, i.e. less significance.
you only log dependant variabale ? Why ? What kind of function are expected ? U-shape, Linear function, ....In your dependant variable zero exist ? negative value ? What is the dependant variable (GDP, Export, Inflation......).
Many questions
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling