The implication of this is that your grounds for believing
in the log-log model are so strong that you are prepared to throw
out more than half of the data. Presumably this belief arises from
theory, or previous data, or both. This suggests to me a Bayesian
approach in which you build in your previous knowledge in a formal
way. However, Stata is not a good vehicle for that.
Alternatively, any data analysis could well start with the question:
why do the negative values arise? Other tricks and tips are no
more than that without a clear story on this point.
Cordula Stolberg
I have an unbalanced panel with only a total of 80 observations (over 8
years). The problem for me is to find the right specification for the
regression. I have tried to use a semi-log model, as my variables contain
negative values and I would lose even more observations with a log-log
model.
The semi-log model however fails the -ovtest- and then only works when I
include a lag of the dependent variable. When then using the -xtabond2-
command, I'm running into all sorts of problem, because I always get the
warning that the instruments are large in relation to the number of
observations. In addition, I can't really perform a test for endogeneity for
two potentially endogenous regressors (because of the problems caused by the
few observations).
I have then tried to use the log-log model despite the loss of observations.
The model works well, even without using the lag of the dependent variable.
However, I'm then down to only 36 observations.
My question is therefore whether anyone could tell me how to get a log-log
model without losing so many observations. I have read somewhere on the
Internet that one could add a constant number to all variables such that the
biggest negative number is 1. In an earlier posting on the list someone
suggested to add a positive number only to the variable containing negative
values & then to log it. However, I'm not sure whether this distorts the
relation between the variables.
