Oaxaca Blinder Decomposition Insignificant

[Theresia Verena]

Hello,
I have the results of oaxaca blinder decomposition like this.

Blinder-Oaxaca decomposition                    Number of obs     =      6,810
                                                  Model           =     linear
Group 1: mental = 0                               N of obs 1      =       3872
Group 2: mental = 1                               N of obs 2      =       2938

--------------------------------------------------------------------------------
      lwage_m1 |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
---------------+----------------------------------------------------------------
overall        |
       group_1 |   14.81284   .2203266    67.23   0.000       14.381    15.24467
       group_2 |   14.68194   .2084602    70.43   0.000     14.27337    15.09052
    difference |   .1308951   .3033141     0.43   0.666    -.4635897    .7253799
     explained |   .0182514   .0155813     1.17   0.241    -.0122874    .0487902
   unexplained |   .1126437   .2983131     0.38   0.706    -.4720392    .6973266

I use this command.

oaxaca lwage_m1 usia married female yeduc agriculture mining manufactur utility construction wholesaledist transport finance_estate social other_sector urban training if mental | n_dsbld, by (mental) model1 (heckman, select (worker = usia yeduc female married child0_5 child6_11 child12_15 child16 krt urban) twostep) model2 (heckman, select (worker=usia yeduc female married child0_5 child6_11 child12_15 child16 krt urban) twostep) weight (0) noisily relax

where:
mental is a dummy variable whose value is equal to 1 if you have a mental illness
n_dsbld = 1 if dummy variable mental !=1 & dummy variable physical disabilities!=1

The differences, unexplained and explained components are not significant. What do I have to do?
Thanks in advance!

[Fei Wang]

The difference between groups 1 and 2 is not statistically significant, which means it's unnecessary to examine what factors explain this difference, i.e., Oaxaca decomposition is not needed from the very beginning. You may try decompositions for other pairs of groups.

[Theresia Verena]

The difference between groups 1 and 2 is not statistically significant, which means it's unnecessary to examine what factors explain this difference, i.e., Oaxaca decomposition is not needed from the very beginning. You may try decompositions for other pairs of groups.

The decomposition between 1 group of other types of disability and non-disability group is also the same. Can you give me suggestions to fix this? because I can't change the data.

[Fei Wang]

I guess your data simply shows no significant difference in average wages between the disabled and the non-disabled. Is that true? If you believe there should be substantial differences but your data fail to show, then you may check if something went wrong with your data. With this data, you may also try distributional decomposition, as Oaxaca only does mean decomposition. May run -net install st0588- to install -oaxaca_rif- and related commands.

[Theresia Verena]

I guess your data simply shows no significant difference in average wages between the disabled and the non-disabled. Is that true? If you believe there should be substantial differences but your data fail to show, then you may check if something went wrong with your data. With this data, you may also try distributional decomposition, as Oaxaca only does mean decomposition. May run -net install st0588- to install -oaxaca_rif- and related commands.

Can it run concurrently with Heckman?

[FernandoRios]

No oaxaca_rif does not add anytype of allow for selection corrections.
But as Fei says. There is nothing WRONG with your results.
Perhaps you may start by looking at the distributions across groups first.

[Theresia Verena]

No oaxaca_rif does not add anytype of allow for selection corrections.
But as Fei says. There is nothing WRONG with your results.
Perhaps you may start by looking at the distributions across groups first.

If I use RIF, there is no problem in significance. Then, what should I do next?

[Fei Wang]

You may read the Stata Journal paper related to the commands: https://journals.sagepub.com/doi/10....36867X20909690. Tables there show what the results of distributional decompositions would look like.

[Weiwen Ng]

...

Blinder-Oaxaca decomposition Number of obs = 6,810
Model = linear
Group 1: mental = 0 N of obs 1 = 3872
Group 2: mental = 1 N of obs 2 = 2938

--------------------------------------------------------------------------------
lwage_m1 | Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------+----------------------------------------------------------------
overall |
group_1 | 14.81284 .2203266 67.23 0.000 14.381 15.24467
group_2 | 14.68194 .2084602 70.43 0.000 14.27337 15.09052
difference | .1308951 .3033141 0.43 0.666 -.4635897 .7253799
explained | .0182514 .0155813 1.17 0.241 -.0122874 .0487902
unexplained | .1126437 .2983131 0.38 0.706 -.4720392 .6973266

I use this command.

oaxaca lwage_m1 ...

where:
mental is a dummy variable whose value is equal to 1 if you have a mental illness
n_dsbld = 1 if dummy variable mental !=1 & dummy variable physical disabilities!=1

The differences, unexplained and explained components are not significant. What do I have to do?
...

Let's go over what the components mean first. If lwage_m1 is log wages (i.e. group 1 is earning over 2.6 million units of whatever currency), then group_1 and group_2 are just the mean values of lwage_m1 by group. Go run the mean command to check for yourself. difference is literally the difference in average wages between the two groups. It's 0.131. That's less than 1% of group 1's mean wage. That means there's no disparity in average wages to explain.

Anyway, explained + unexplained = difference, in case that wasn't obvious. The tests next to those terms are all tests that those quantities are equal to zero.

I'm not familiar with oaxaca_rif, but at first glance it looks like it can decompose the difference in median wages, or x-th percentile wages, etc. You say that "there is no problem in significance", which I guess means that the terms for difference, explained, and unexplained are statistically significantly different from 0. But never mind the p-values. How different are the means/medians/25th percentile wages in terms of actual currency? If the difference is as small as above, that means that isn't a meaningful disparity between the two groups. There is not anything to "do" about it per se.

And I don't mean to belabor the point I made on your other question, but if you were using the CES-D score (a depression screening tool) to designate people with mental illness, that is a big problem! The CES-D measures depressive symptoms. You can have depression that's under control, so you aren't currently symptomatic - and then by the CES-D score you don't have a mental illness. You could have schizophrenia, a personality disorder, etc without current depressive symptoms. And depression is episodic, so if you have a current major depressive episode, that doesn't mean you have a disability for life.

[Theresia Verena]

Let's go over what the components mean first. If lwage_m1 is log wages (i.e. group 1 is earning over 2.6 million units of whatever currency), then group_1 and group_2 are just the mean values of lwage_m1 by group. Go run the mean command to check for yourself. difference is literally the difference in average wages between the two groups. It's 0.131. That's less than 1% of group 1's mean wage. That means there's no disparity in average wages to explain.

Anyway, explained + unexplained = difference, in case that wasn't obvious. The tests next to those terms are all tests that those quantities are equal to zero.

I'm not familiar with oaxaca_rif, but at first glance it looks like it can decompose the difference in median wages, or x-th percentile wages, etc. You say that "there is no problem in significance", which I guess means that the terms for difference, explained, and unexplained are statistically significantly different from 0. But never mind the p-values. How different are the means/medians/25th percentile wages in terms of actual currency? If the difference is as small as above, that means that isn't a meaningful disparity between the two groups. There is not anything to "do" about it per se.

And I don't mean to belabor the point I made on your other question, but if you were using the CES-D score (a depression screening tool) to designate people with mental illness, that is a big problem! The CES-D measures depressive symptoms. You can have depression that's under control, so you aren't currently symptomatic - and then by the CES-D score you don't have a mental illness. You could have schizophrenia, a personality disorder, etc without current depressive symptoms. And depression is episodic, so if you have a current major depressive episode, that doesn't mean you have a disability for life.

okay thank you for your helping.