{smcl}
{* 03aug2011}{...}
{hline}
help for {hi:seldum}
{hline}
{title:Various inequality indicators}
{p 4}{cmd:seldum} {varlist} [, {cmdab:rep:lay}]
{title:Description}
{p}{cmd:seldum} is a post-estimation command to be used after the estimation
of a linear regression model with a logarithmic dependent variable. This means
that it has to be used immediately after a call to an estimation command as
e.g. {cmd:regress}.
{p}{cmd:seldum} transforms the coefficients estimated by the model so that
they can be interpreted as percentage effects on the dependent variable. The
calculations are made for coefficients relating to the variables specified
in {it: varlist} .
{p}At the moment, {cmd:seldum} does not work for multiple-equation models,
including {cmd:heckman} and {cmd:treatreg}.
{title:Option}
{p} {cmdab:rep:lay} requests the display of the results of the last estimation model, i.e. the model
to be used in the calculations.
{title:Details}
{p}Consider the following semilogarthmic regression model:
ln(Y) = a + b*X + e
{p}If the independent variable X is continous, b gives the percentage change
in Y implied by a small change in X. In the case where X is a dummy varibale,
the same interpretation is frequently used. However, if X is a dummy variable,
this interpretation is only approximative. Kennedy (1981) has proposed the
following estimator:
p = exp(b - 0.5*V(b)) - 1 (1)
{p}where V(b) is the variance of b. In principle, the variance of p could
be computed by using the delta method. Alternatively, Van Garderen and Shah (2002)
have derived the following estimator for the variance of p, which they claim to
be more accurate than the delta method:
V(p) = exp(2*b)*(exp(-V(b)) - exp(-2*V(b))) (2)
{p}{cmd:seldum} computes Kennedy's estimator as given by equation (1) as well
as its standard error, as given by equation (2).
{p}The paper by Van Garderen and Shah (2002) provides a more detailed discussion
on this topic.
{title:Saved Results}
{p}{cmd:seldum} returns its results in r():
{p 4}r(b_varname) percentage impact of the varibale {it:varname}
{p 4}r(se_varname) standard error of the percentage impact of variable {it:varname}
{p}The advantage of saving the results in r() is that the results saved by the
preceeding regression are not modified. These results remain available after a
call to {cmd:seldum}.
{title:Example}
{p 4 12}{inp:.} {stata "sysuse nlsw88, clear ":sysuse nlsw88, clear}
{p 4 12}{inp:.} {stata "generate lwage = ln(wage) ":generate lwage = ln(wage)}
{p 4 12}{inp:.} {stata "regress lwage collgrad smsa union south race ttl_exp ":regress lwage collgrad smsa union south race ttl_exp}
{p 4 12}{inp:.} {stata "seldum collgrad union ":seldum collgrad union}
{p 4 12}{inp:.} {stata "seldum collgrad union, replay ":seldum collgrad union, replay}
{title:References}
{p 0 4}Kennedy, P. E. (1981). Estimation with Correctly Interpreted Dummy Variables
in Semilogarithmic Equations. American Economic Review, 71(4):801.
{p 0 4}van Garderen, K. and Shah, C. (2002). Exact Interpretation of Dummy
Variables in Semilogarithmic Equations. The Econometrics Journal, 5(1):149-159.
{title:Also see}
{p} {help levpredict} if installed (available via {help ssc})
{title:Author}
Jean Ries
STATEC, Luxembourg
jean.ries@statec.etat.lu