Re: st: Interpretation of Oaxaca decomposition results after re-transformation of log scale

[Austin Nichols <austinnichols@gmail.com>]

Vaidyanathan Ganapathy <vaidyang@usc.edu>:

This looks like a mistake--I wonder if -oaxaca- (SSC, SJ) should even
support the eform option, or issue a warning at least. You want a
-glm- with log link, not a regression of ln(y) on X, per

http://blog.stata.com/2011/08/22/use-poisson-rather-than-regress-tell-a-friend/
http://www.stata.com/meeting/boston10/boston10_nichols.pdf

which is not currently supported by -oaxaca- AFAIK ...but could be,
with a decomp of first moments as in the Yun (2008) approach for
logit/probit used by -oaxaca- since 2010:
http://www.stata.com/statalist/archive/2010-01/msg00042.html

Your current decomp has explained 57% and unexplained 70% for a total>100.

On another topic altogether, I wonder if your sample includes
stillbirths, or only infants surviving to some specified age.

On Wed, Mar 12, 2014 at 3:03 AM, Vaidyanathan Ganapathy
<vaidyang@usc.edu> wrote:
> Dear Statalisters,
>
> I am performing an Oaxaca type decomposition to understand the
> healthcare cost differences between two groups -  controls and
> premature infants. Here is my specification:
>
> . oaxaca lnallhccx2 tpcat2-tpcat6 bpdx2 chdx2 asthbrdx2 resinfxdx2
> cnsdx2 motordx2 physdevdx2 nddx2 chrnic1 period, by(premie_cat) pooled
> vce(cluster pcn) eform
>
> The dependent variable is ln(healthcare costs) and the other variables
> are covariates including poverty levels (tpcat2-tpcat6) and certain
> medical diagnoses. Since the dependent variable is in log scale I used
> the -eform option to exponentiate and report the predicted costs and
> the decomposed cost differentials. While I am able to interpret the
> predicted values for the two groups, I have some trouble in
> interpreting the overall, explained and unexplained differences. Here
> is the output -
>
>
>
> Blinder-Oaxaca decomposition                      Number of obs   =     137972
>
>         1: premie_cat = 0  (controls)
>         2: premie_cat = 1   (premature infants)
>
>                               (Std. Err. adjusted for 68994 clusters in pcn)
> -------------------------------------------------------------------------------
>            |               Robust
> lnallhccx2 |     exp(b)   Std. Err.      z    P>|z|     [95% Conf. Interval]
> --------------+----------------------------------------------------------------
> Differential  |
> Prediction_1 |   348.9868   1.737476  1176.03   0.000      345.598    352.4089
> Prediction_2 |    956.743   75.23525    87.28   0.000     820.0862    1116.172
> Difference |   .3647655   .0287414   -12.80   0.000     .3125676    .4256803
> --------------+----------------------------------------------------------------
> Decomposition |
>  Explained |   .5206098    .035001    -9.71   0.000     .4563368    .5939355
> Unexplained |   .7006504   .0489067    -5.10   0.000     .6110629    .8033722
> -------------------------------------------------------------------------------
>
> Using simple math, it could be seen from the results in panel 1
> (Differential) that healthcare costs among controls is only 36.47% of
> that of healthcare costs among premature infants. This led me to the
> following interpretation about the overall cost differential between
> premature and control infants:  The healthcare cost among premature
> infants increases by 174% of that of the costs among controls as
> predicted by the group models. Is it correct to make this
> interpretation?
>
> The interpretation of the decomposition results (panel 2 above) - the
> explained and unexplained components, doesn't seem to be that straight
> forward.
>
> Any help in understanding these difference estimates will be very helpful.
>
> Thanks!/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/