KHB Decomposition

[Bai Yang]

Dear all！
I had some problems with KHB decomposition，My Z-Vars is Categorical variables and its not significant，but when further analysed，I found some of them became significant，How can i explain it？
this is my example.

diff of fedu is not significant.
------------------------------------------------------------------------------
zddx | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
fedu |
Reduced | .0743868 .0239341 3.11 0.002 .0274768 .1212969
Full | .0704145 .0239126 2.94 0.003 .0235468 .1172823
Diff | .0039723 .0026069 1.52 0.128 -.0011372 .0090818
-------------+----------------------------------------------------------------

but when I disentangled it，one of them became significant

Z-Variable | Coef Std_Err P_Diff P_Reduced
-------------+---------------------------------------------
fedu |
_i_zdgz2 | -.0020702 .0021099 -52.12 -2.78
_i_zdgz3 | .0060425 .0026391 152.12 8.12

i_zdgz3 is significant while i_zdgz2 not significant。

How can I explain it？ should I stop further explain when diff not significant？

[Bai Yang]

Hello

[Richard Williams]

Welcome to Statalist.

Try showing the exact commands you gave along with how Stata responded. Use code tags to make the output easier to read. See the Statalist FAQ for more tails on asking questions effectively.

[Bai Yang]

Welcome to Statalist.

Try showing the exact commands you gave along with how Stata responded. Use code tags to make the output easier to read. See the Statalist FAQ for more tails on asking questions effectively.

thanks .
My commands is “khb mlogit Y K-vars ||i.Z-var ,concomitant(X) outcome(1) baseoutcome(0) summary disentangle"

[Bai Yang]

Hello？

[Richard Williams]

Hi Bai. Like I said before, it might be helpful if you showed the full output, using code tags so the output is easily readable. I also find the output confusing because it refers to fedu and zddx which I don't see in the command, but maybe these are values of the Z vars or something.

Also, you say "when further analysed" which makes me wonder if you have omitted one or more commands.

In any event, if I understand what you are doing, such seeming inconsistencies do not surprise me. Significance tests need not be terribly consistent. When you do something different, something could easily switch from being significant to insignificant, and vice versa.

Further you are only showing one outcome. When you look at all of them, you may find a clearer pattern about a variable's effects.

In short, I don't really understand the concern. But I also have no idea what your variables are, and it is very difficult to read the original output. Perhaps if you explain more what you are doing and show results more clearly, somebody could help you more. But just on the surface of it, I don't see a major cause for concern.