Employees obtain a certification partly on an unobservable that is
correlated (0.5) with the promotion error, so certification is endogenous.
A randomly assigned certification subsidy shifts certification and has no
direct effect on promotion. The errors are jointly normal, the case in
which the recursive bivariate probit (firststage = "biprobit") is
exact and the generalized-residual control function only approximate. The
same file is shipped as inst/extdata/certification.dta.
Format
A data frame with 4,000 rows and 5 variables:
- promoted
Promoted (0/1); the outcome.
- certified
Certified (0/1); the endogenous treatment.
- subsidy
Certification subsidy (standardized); the instrument.
- tenure
Tenure (standardized); a control.
- manager
Supportive manager (0/1); a control.
Details
Data-generating process (data-raw/example_data.do, seed
90210): certified = 1[-0.2 + 0.9 subsidy + 0.3 tenure + 0.2 manager
+ v > 0], promoted = 1[-0.2 + 0.6 certified + 0.4 tenure +
0.3 manager + e > 0], with (e, v) standard bivariate normal with
correlation 0.5. The true average effect of certification on the
probability of promotion is about 0.21.