Non linear decomposition

[Alejandra Armesto]

Hi everyone!

I have been working on the effects that couples working together have on the gross income of their businesses and their perceived success. Particularly, I defined my sample as for the couples as 1 if they make joint decision and 0 otherwise. For the gross income analysis I ran an OLS model, therefore the B-O decomposition I ran in Stata was under the command Oaxaca. Yet, for the perceived success I coded 1 as successful and 0 otherwise, and ran it as a logit model. The decomposition I ran in this case was a nldecompose.

I had some results that make me think that I'm not performing it correctly. I would like some guidance or feedback on it.


I already went over: https://journals.sagepub.com/doi/pdf...867X0800800402, and https://www.stata.com/meeting/5germa...esentation.pdf.

Many thanks for your help! ☺️

[Andrew Musau]

For decomposition of a binary outcome, I'd stick to oaxaca or install fairlie from SSC.

ssc describe fairlie

These will give you the contribution of individual variables or a group of variables separately, unlike nldecompose (from the Stata Journal) that only gives you the contribution of all variables. The interpretation is analogous to a continuous outcome except this time you are dealing with proportions.

[Alejandra Armesto]

Many thanks Andrew Musau ! I did what you suggested and this is what I got. So I could conclude that I have only two variables that are contributing to the difference between groups. phigher and hours_p. Each of them contribute negatively and positively (based on the coef.) ?

[Alejandra Armesto]

I have been thinking about your response. For the Oaxaca command, isn't there any problem if the variable is a binary variable? Andrew Musau

Thanks again for your help.

[Andrew Musau]

So I could conclude that I have only two variables that are contributing to the difference between groups. phigher and hours_p. Each of them contribute negatively and positively (based on the coef.)?

Correct. For the negative coefficient (on hours), equalizing this variable between the two groups is associated with a reduction in the overall disparity in the outcome between the groups. For the positive coefficient, vice-versa.

For the Oaxaca command, isn't there any problem if the variable is a binary variable?

Not if you choose the correct option for binary outcomes. This is explained in

help oaxaca

oaxaca also supports the non-linear decomposition for binary dependent variables proposed by Yun (2004). See the logit and probit options.