ceteris paribus , matching, controlling for others conditions

[Chiara Tasselli]

Good afternoon to everyone,
my task is to calculate gender wage differences by controlling for all other conditions/covariates (such as educational qualification, age, seniority, business area, etc..).
I can do this with both regression and Oaxaca decomposition , but I was also looking for alternative methods (for example calculating the pscore on worker characteristics).

Does anyone have any suggestions? Especially IF THE SAMPLE IS TOO SMALL to do a regression or Oaxaca decomposition?
How can one control for characteristics with a simple method?

Many thanks in advance for your time !

[Maarten Buis]

What makes me very worried is your statement that you want to "controlling for all other conditions/covariates" (the emphasis is mine). You always need to be careful that you do not over-control, i.e. control for variables that capture the mechanism through which a variable of interest influences the outcome.

Women earn less in part because:

• they are pushed in and/or choose certain fields of study,
• they are pushed in and/or choose certain occupations,
• they are pushed in and/or choose to work part-time,
• they are less likely to certain job-titles because of glass ceilings,
• they are pushed in and/or choose to leave the labor market temporarily to take care of children,
• ...

It really depends what you exactly want to estimate and how you believe these factors work whether you want to control for these variables or not.

• If you want to know if a society distributes wages fairly across genders, and you think that these "choices" are subject to substantial societal/cultural pressure, then you do not want to control for those variables, as they represent (at least part of) the mechanism through which the gender wage cap comes about.
• If you want to know whether employers fairly distribute wages across genders, and you believe that employers are just passive consumers of whatever society throws at them, then you may want to control for some of these variables. At a minimum, you should in that case be careful not to "control away" the glass ceiling effect.

Regardless, over-controlling is just as much a source of bias as under-controlling, so controlling for all variables is almost certainly wrong.