A violation of the exclusion restriction is a number, γ, and
a number needs a unit. In a linear model there is only one: γ
is the direct effect of the instrument on the outcome. In a probit there
are three natural ones, and apexog lets you use any of them through
gunits (R) or gunits() (Stata). Internally the
violation always enters the model the same way, as an offset
z’γ in the probit index.
The three scales
The latent index (gunits = "latent",
the default; synonym "index"). γ is a coefficient
in the index of the main probit. That probit includes the control
function, and it normalizes the variance of the error conditional on
the control function to one. This is the scale of the offset
itself, so no conversion is made. One unit of γ moves the
probability of the outcome by roughly the average density of the index,
mφ, per unit of the instrument; apexog prints mφ
in the first “Scale” line.
Probability points (gunits = "pr").
γ is the direct effect of the instrument on the probability
that y = 1, per unit of the instrument (0.02 is two percentage
points). apexog divides every value you give by mφ from the
baseline fit (a variance by mφ2). This is the
scale on which plausexog states the violation when it is
run on a binary outcome, as a linear probability model, so it is the
scale to use when you compare the two commands for the same belief. It
is also the scale on which most people find it easiest to say how large
a direct effect could plausibly be.
The structural latent scale
(gunits = "structural"). γ is a coefficient in the
structural equation y* = x’β + dδ + z’γ + ε with the
total error variance normalized to one, the way latent-variable
models are usually written and simulated. With an endogenous d,
part of the error is explained by the control function, so the probit
index measures in units of the smaller conditional standard deviation σ̃
= sd(ε | control function). apexog estimates σ̃ = (1 + bcf’
Var(ĉ) bcf)−1/2 from the baseline fit and divides
every value by it; the second “Scale” line reports it.
The structural conversion is derived for a linear first
stage, where the control function estimates the first-stage
error. With a probit, fractional probit or Poisson first stage it does
not apply: in simulations the structural γ enters the index
essentially unscaled. Do not combine gunits = "structural"
with those first stages; state the violation on the index or in
probability points instead. apexog accepts the combination but prints a
warning.
Even with a linear first stage the conversion divides by the baseline
σ̃, which is exact only as γ shrinks to zero. With one
continuous endogenous regressor and one instrument, a structural value
has a unique index-scale counterpart only below
|πz|/σv in absolute value (the first-stage
coefficient of the instrument over the first-stage residual standard
deviation). The output reports that limit (gs_max; in Stata
e(gs_max)) and warns when the range or the prior draws
reach it. The results themselves are not changed.
The same belief, three ways
In the example data mφ = 0.209 and σ̃ = 0.834. A direct effect of ±0.10 on the latent index is therefore about ±0.0209 in probability and ±0.0834 on the structural scale. Stated in those units, the three calls run the same grid and give the same bounds:
library(apexog)
data(promotion)
f <- promoted ~ tenure + manager
e <- training ~ subsidy
index <- apexog(f, endog = e, data = promotion, gmin = -0.10, gmax = 0.10,
grid = 5, verbose = FALSE)
m <- index$mphi
s <- index$stilde
c(m_phi = m, sigma_tilde = s)
#> m_phi sigma_tilde
#> 0.2090548 0.8342901
pr <- apexog(f, endog = e, data = promotion, gunits = "pr",
gmin = -0.10 * m, gmax = 0.10 * m, grid = 5, verbose = FALSE)
st <- apexog(f, endog = e, data = promotion, gunits = "structural",
gmin = -0.10 * s, gmax = 0.10 * s, grid = 5, verbose = FALSE)
rbind(index = confint(index), pr = confint(pr), structural = confint(st))
#> lower upper
#> training 0.09709467 0.1904491
#> training 0.09709467 0.1904491
#> training 0.09709467 0.1904491The grid stored in the result is always on the latent index;
grid_pr holds the exact probability-scale value of each
grid point:
cbind(index$grid["g_subsidy"], index$grid_pr)
#> g_subsidy gpr_subsidy
#> 1 -0.10 -0.02090548
#> 2 -0.05 -0.01045274
#> 3 0.00 0.00000000
#> 4 0.05 0.01045274
#> 5 0.10 0.02090548Why the scale matters
The conversions are not cosmetic. If you have a belief on the structural scale and enter it as if it were on the index, the violation you assume is too small by the factor σ̃. Under the local-to-zero method the interval is then too narrow: with a wide prior its coverage approaches 2Φ(z σ̃) − 1 instead of the nominal level (about 0.88 instead of 0.95 at an endogeneity correlation of 0.6 in simulations). Stating the belief in probability points avoids the question of which latent scale is meant.
Exact probability scale
The conversion gunits = "pr" is a first-order one: it
uses mφ from the model with γ = 0. apexog also
reports, at every refit, the exact probability-scale value of each
instrument’s violation in the model refitted at that violation:
- for a 0/1 instrument, the average change in the probability of the outcome when the instrument switches from 0 to 1;
- for any other instrument, the average derivative, mean φ(η)·γ.
The two agree for small violations. They can differ by several
percent for a large violation of a 0/1 instrument when the outcome is
close to 0 or 1. With a linear first stage, as many instruments as
endogenous regressors and a continuous instrument they coincide exactly,
as in the example above. The values are printed beside each range
(“exact Pr scale”) and stored in grid_pr (UCI) and
sims$gamma_pr (LTZ); in Stata in e(grid_pr)
and e(gamma_pr_sim).