margins using a khb-method

[Floriane Bolazzi]

Hi,
I am running a logit model using the KHB-method (program by Kohler and Karlson) to estimate the confounding percentage of jati (z-var) and class3_f (key-var) as it follows:

khb logit cd31 i.class3_f || i.jati, c(i.agec i.birthCO $char)

Here what I get:

Decomposition using the KHB-Method

Model-Type: logit Number of obs = 1287
Variables of Interest: i.class3_f Pseudo R2 = 0.29
Z-variable(s): i.jati
Concomitant: i.agec i.birthCO highedu married eldest_son hhsize
------------------------------------------------------------------------------
cd31 | Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
1.class3_f | (base outcome)
-------------+----------------------------------------------------------------
2.class3_f |
Reduced | -1.719203 .2038593 -8.43 0.000 -2.11876 -1.319646
Full | -.9050135 .2036712 -4.44 0.000 -1.304202 -.5058254
Diff | -.8141896 .1614067 -5.04 0.000 -1.130541 -.4978383
-------------+----------------------------------------------------------------
3.class3_f |
Reduced | -1.469628 .341212 -4.31 0.000 -2.138392 -.8008653
Full | -.0675756 .3705604 -0.18 0.855 -.7938607 .6587095
Diff | -1.402053 .2174421 -6.45 0.000 -1.828231 -.9758742
------------------------------------------------------------------------------

Summary of confounding

Variable | Conf_ratio Conf_Pct Resc_Fact
-------------+-------------------------------------
1b.class3_f | . . .
2.class3_f | 1.8996435 47.36 1.1349222
3.class3_f | 21.747914 95.40 1.0235198
---------------------------------------------------

Now, I would like to know how the conf_pct varies over birthCO (control variable). The idea is to explore whether the indirect effect of class3_f due to the jati (diff between reduced and full models) varies across different birth cohorts. I don't want to split the sample by birthCO and run the model for each birthCO because I imagine that this strategy might raise problems of comparability. Instead, I would like to apply something like a margins analysis as in logit models. For instance: margins jati, over(birthCO).
But post-estimation seems not to be supported by khb. Any suggestions?
Thanks in advance
Floriane Bolazzi

[Sofiya Volvakova]

Hi,
I am running a logit model using the KHB-method (program by Kohler and Karlson) to estimate the confounding percentage of jati (z-var) and class3_f (key-var) as it follows:

khb logit cd31 i.class3_f || i.jati, c(i.agec i.birthCO $char)

Here what I get:

Decomposition using the KHB-Method

Model-Type: logit Number of obs = 1287
Variables of Interest: i.class3_f Pseudo R2 = 0.29
Z-variable(s): i.jati
Concomitant: i.agec i.birthCO highedu married eldest_son hhsize
------------------------------------------------------------------------------
cd31 | Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
1.class3_f | (base outcome)
-------------+----------------------------------------------------------------
2.class3_f |
Reduced | -1.719203 .2038593 -8.43 0.000 -2.11876 -1.319646
Full | -.9050135 .2036712 -4.44 0.000 -1.304202 -.5058254
Diff | -.8141896 .1614067 -5.04 0.000 -1.130541 -.4978383
-------------+----------------------------------------------------------------
3.class3_f |
Reduced | -1.469628 .341212 -4.31 0.000 -2.138392 -.8008653
Full | -.0675756 .3705604 -0.18 0.855 -.7938607 .6587095
Diff | -1.402053 .2174421 -6.45 0.000 -1.828231 -.9758742
------------------------------------------------------------------------------

Summary of confounding

Variable | Conf_ratio Conf_Pct Resc_Fact
-------------+-------------------------------------
1b.class3_f | . . .
2.class3_f | 1.8996435 47.36 1.1349222
3.class3_f | 21.747914 95.40 1.0235198
---------------------------------------------------

Now, I would like to know how the conf_pct varies over birthCO (control variable). The idea is to explore whether the indirect effect of class3_f due to the jati (diff between reduced and full models) varies across different birth cohorts. I don't want to split the sample by birthCO and run the model for each birthCO because I imagine that this strategy might raise problems of comparability. Instead, I would like to apply something like a margins analysis as in logit models. For instance: margins jati, over(birthCO).
But post-estimation seems not to be supported by khb. Any suggestions?
Thanks in advance
Floriane Bolazzi

Hi Floriane. Did you figure the code out? I am having the same problem..