The R package is a port of the Stata command. Both fit the same
models with the same formulas, and on the same data they agree to about
seven significant digits: the R test suite compares the two on 29
specifications, covering every first stage, clustering, interaction
effects, the bivariate probit, the three scales for the violation and
the local-to-zero refits at Stata’s own draws.
The model
apexog uci y x1 x2 (d = z1 z2) |
apexog(y ~ x1 + x2, endog = d ~ z1 + z2, method = "uci", ...) |
apexog ltz ... |
method = "ltz" |
(d1 d2 = z1 z2): shared instruments |
endog = d1 + d2 ~ z1 + z2 |
(d1 = z1 z2) (d2 = z2 z3): own instruments |
endog = list(d1 ~ z1 + z2, d2 ~ z2 + z3) |
i.cat among the controls or instruments |
factor(cat) |
(c.d c.d#c.x = z c.z#c.x): endogenous interaction |
endog = d + d:x ~ z + z:x |
if x > 0 |
subset = x > 0 |
| at least one control is needed |
y ~ 1 for a model without controls |
Options and arguments
gmin(), gmax(), grid()
|
gmin, gmax, grid
|
mu(), omega()
|
mu, omega
|
distribution(uniform a b) |
prior = "uniform", prior_params = list(c(a, b)) |
distribution(normal m s), chi2,
t, gamma, poisson
|
prior = "normal", "chi2",
"t", "gamma", "poisson" with
prior_params
|
distribution(special var) |
prior = "special", prior_params = list(data$var) |
| (none) |
prior = function(n) ..., or
draws = <matrix>
|
iterations(), seed()
|
iterations, seed
|
firststage(linear probit fprobit poisson biprobit) |
firststage = "linear", … (one value or one per
endogenous regressor) |
i.d / c.d prefix in the clause |
firststage = "probit" / "linear"
|
target(d1) |
target = "d1" |
target(c.x#i.w) or aie(x w)
|
aie = c("x", "w"), with
aie_binary = c(FALSE, TRUE) to fix the types |
cfinteract |
cfinteract = TRUE |
gunits(latent pr structural) |
gunits = "latent", "pr",
"structural"
|
level(95) or level(.95)
|
level = 0.95 or 95
|
vce(robust) (default) |
(default) |
vce(cluster id) |
cluster = ~ id |
graphomega(), graphmu(),
graphdelta()
|
sweep_omega, sweep_mu,
sweep_delta
|
graph |
plot(fit) |
noheader |
do not print the result; verbose = FALSE silences the
notes |
Stored results
e(b), e(V)
|
coef(fit), vcov(fit)
|
e(baseline_ape), e(baseline_se),
e(baseline_lo), e(baseline_hi)
|
fit$baseline |
e(lb_d), e(ub_d); e(lb_aie),
e(ub_aie)
|
confint(fit), fit$bounds
|
e(lo_prior_vec), e(hi_prior_vec)
|
fit$prior_band |
e(grid), e(grid_pr)
|
fit$grid, fit$grid_pr
|
e(gamma_sim), e(gamma_pr_sim)
|
fit$sims$gamma, fit$sims$gamma_pr
|
e(ape_sim), e(se_sim),
e(ape_sim_cnv)
|
fit$sims$est, fit$sims$se,
fit$sims$est_cnv
|
e(sweep) |
fit$sweep |
e(mphi), e(stilde),
e(gs_max), e(rho)
|
fit$mphi, fit$stilde,
fit$gs_max, fit$rho
|
e(N), e(N_clust)
|
nobs(fit), fit$N_clust
|
e(firststage), e(estimator),
e(vce), e(instruments)
|
fit$firststage, fit$estimator,
fit$vce, fit$instruments
|
| (not stored) |
fit$model: the main equation’s coefficients and their
covariance at γ = 0 |
Where the two differ
-
Random numbers. R and Stata draw different random
numbers, so their local-to-zero intervals agree only up to simulation
error. To reproduce a Stata run exactly, pass its draws:
draws = <e(gamma_sim) as a matrix> (in latent units,
the default gunits).
-
No controls. The Stata parser needs at least one
exogenous control; R accepts
y ~ 1.
-
Subsets with missing values. Observations where the
subset expression is NA are dropped in R.
Stata’s if treats a missing value as larger than any number
and may keep them.
-
Factor variables. R’s model matrix drops base
levels itself; Stata’s apexog removes base levels and collinear
instrument terms before fitting. Both end with the same full-rank set of
instruments.
-
An interaction already in the formula. With
aie = c("x", "w") R uses a product x:w that
the formula already contains (as in y ~ x * w); in Stata
write the main effects only and let apexog add the product.