///|
/// Average potential outcomes for an arbitrary treatment level.
pub struct DoubleMLAPO {
data : DoubleMLData
treatment_level : Double
n_folds : Int
n_rep : Int
seed : Int
propensity_clip : Double
g_hat : Array[Double]
m_hat : Array[Double]
coef : Double
se : Double
fitted : Bool
} derive(Debug)
///|
pub fn DoubleMLAPO::new(
data : DoubleMLData,
treatment_level? : Double = 1.0,
n_folds? : Int = 2,
n_rep? : Int = 1,
seed? : Int = 3141,
propensity_clip? : Double = 1.0e-6,
) -> DoubleMLAPO {
try {
require(n_folds >= 2)
require(n_folds <= data.n_obs())
require(n_rep >= 1)
require(propensity_clip > 0.0)
require(propensity_clip < 0.5)
{
data,
treatment_level,
n_folds,
n_rep,
seed,
propensity_clip,
g_hat: Array::make(data.n_obs(), 0.0),
m_hat: Array::make(data.n_obs(), 0.0),
coef: 0.0,
se: 0.0,
fitted: false,
}
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
pub fn DoubleMLAPO::n_obs(self : DoubleMLAPO) -> Int {
self.data.n_obs()
}
///|
/// Number of features (covariate columns).
pub fn DoubleMLAPO::n_features(self : DoubleMLAPO) -> Int {
self.data.n_features()
}
///|
pub fn DoubleMLAPO::coef(self : DoubleMLAPO) -> Double {
try {
require(self.fitted)
self.coef
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
pub fn DoubleMLAPO::se(self : DoubleMLAPO) -> Double {
try {
require(self.fitted)
self.se
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
pub fn DoubleMLAPO::confint(self : DoubleMLAPO) -> (Double, Double) {
try {
require(self.fitted)
(self.coef - 1.96 * self.se, self.coef + 1.96 * self.se)
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
pub fn DoubleMLAPO::predictions_g(self : DoubleMLAPO) -> Array[Double] {
try {
require(self.fitted)
self.g_hat
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
pub fn DoubleMLAPO::predictions_m(self : DoubleMLAPO) -> Array[Double] {
try {
require(self.fitted)
self.m_hat
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
fn indicator_level(d : Array[Double], level : Double) -> Array[Double] {
let out = Array::make(d.length(), 0.0)
for i = 0; i < d.length(); i = i + 1 {
out[i] = if d[i] == level { 1.0 } else { 0.0 }
}
out
}
///|
/// Cross-fit the APO nuisances. The order is asymmetric: the
/// conditional outcome `g` is fit only on the treated subset (the
/// same as the upstream `DoubleMLAPO`), while the propensity `m`
/// is fit on the *full* training fold (including controls). The
/// effect is that rows with `treated = 0` receive their `g` value
/// from the previous fold's treated-only fit (or stay at the
/// default 0.0 if no fold has yet seen them); rows with `treated = 1`
/// receive both a `g` (on the treated subset) and a propensity
/// update. Matches the upstream `DoubleMLAPO` convention.
fn cross_fit_apo(
x : Matrix,
y : Array[Double],
treated : Array[Double],
folds : Array[Fold],
clip : Double,
) -> (Array[Double], Array[Double]) {
let n = x.rows()
let g = Array::make(n, 0.0)
let m = Array::make(n, 0.0)
for fold in folds {
let tr = fold.train_indices()
let te = fold.test_indices()
let tg = filter_indices(tr, treated)
if tg.length() > 0 {
let p = LinearRegression::new()
.fit(slice_matrix_rows(x, tg), slice_vector(y, tg))
.predict(slice_matrix_rows(x, te))
for k = 0; k < te.length(); k = k + 1 {
g[te[k]] = p[k]
}
}
let p = LinearRegression::new()
.fit(slice_matrix_rows(x, tr), slice_vector(treated, tr))
.predict(slice_matrix_rows(x, te))
for k = 0; k < te.length(); k = k + 1 {
m[te[k]] = p[k]
}
}
(g, clip_vec(m, clip, 1.0 - clip))
}
///|
pub fn DoubleMLAPO::fit(self : DoubleMLAPO) -> DoubleMLAPO {
let n = self.n_obs()
let treated = indicator_level(self.data.d, self.treatment_level)
// Accumulate `g` and `m` directly in the storage arrays; divide
// by `n_rep` after the loop.
let g = Array::make(n, 0.0)
let m = Array::make(n, 0.0)
for r = 0; r < self.n_rep; r = r + 1 {
let (gr, mr) = cross_fit_apo(
self.data.x,
self.data.y,
treated,
kfold(n, self.n_folds, self.seed + r),
self.propensity_clip,
)
for i = 0; i < n; i = i + 1 {
g[i] = g[i] + gr[i]
m[i] = m[i] + mr[i]
}
}
let inv = 1.0 / self.n_rep.to_double()
for i = 0; i < n; i = i + 1 {
g[i] = g[i] * inv
m[i] = m[i] * inv
}
// `pa` is the APO score's `psi_a` component, which is structurally
// `-1` for every observation (the potential-outcome score has no
// treatment-side variation). `pb` is the `psi_b` component with
// the IPW-style centring `g + treated * (y - g) / m`.
let pa = Array::make(n, -1.0)
let pb = Array::make(n, 0.0)
for i = 0; i < n; i = i + 1 {
pb[i] = g[i] + treated[i] * (self.data.y[i] - g[i]) / m[i]
}
// Score + variance via the shared `var_est` helper.
let (theta, se) = var_est(pa, pb)
{
data: self.data,
treatment_level: self.treatment_level,
n_folds: self.n_folds,
n_rep: self.n_rep,
seed: self.seed,
propensity_clip: self.propensity_clip,
g_hat: g,
m_hat: m,
coef: theta,
se,
fitted: true,
}
}
///|
/// Average potential outcomes *symmetric* across multiple treatment
/// levels. v0.49.0: full upstream parity — `treatment_levels`
/// validation in `new` (rejects duplicates and levels not in
/// `data.d`), the `causal_contrast` method (level-by-level delta
/// and SE vs a reference level), and the
/// `treatment_levels()` / `n_treatment_levels()` / `fitted()`
/// accessors.
///
/// Each treatment level is fit with the same fold partition
/// (the parent does not currently route a shared partition to the
/// child `DoubleMLAPO`; each child draws its own folds via
/// `kfold`. v0.50.0+ plans to wire the parent through a
/// `fit_with_splits` helper to share one stratified partition,
/// but the v0.49.0 implementation is the v0.50.0 PR target) and
/// the same closed-form `LinearRegression` learner. The
/// `causal_contrast(reference_levels)` method then returns the
/// level-by-level difference `coefs[i] - coefs[ref]` for one or
/// more reference levels, matching the upstream
/// `DoubleMLAPOS.causal_contrast` semantics.
pub struct DoubleMLAPOS {
data : DoubleMLData
treatment_levels : Array[Double]
n_folds : Int
n_rep : Int
seed : Int
propensity_clip : Double
coefs : Array[Double]
ses : Array[Double]
fitted : Bool
} derive(Debug)
///|
pub fn DoubleMLAPOS::new(
data : DoubleMLData,
treatment_levels : Array[Double],
n_folds? : Int = 2,
n_rep? : Int = 1,
seed? : Int = 3141,
propensity_clip? : Double = 1.0e-6,
) -> DoubleMLAPOS {
try {
require(treatment_levels.length() >= 1)
require(n_folds >= 2)
require(n_folds <= data.n_obs())
require(n_rep >= 1)
require(propensity_clip > 0.0)
require(propensity_clip < 0.5)
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
// v0.49.0: every requested treatment level must be present in
// the data's treatment assignment (`np.unique(data.d)` in
// upstream). Catch duplicates within the request itself too.
for i = 0; i < treatment_levels.length(); i = i + 1 {
let lvl = treatment_levels[i]
let mut dup = false
for j = 0; j < i; j = j + 1 {
if treatment_levels[j] == lvl {
dup = true
break
}
}
if dup {
abort(
"DoubleMLAPOS: treatment_levels contains a duplicate entry: " +
lvl.to_string(),
)
}
let mut in_data = false
for k = 0; k < data.d.length(); k = k + 1 {
if data.d[k] == lvl {
in_data = true
break
}
}
if !in_data {
abort(
"DoubleMLAPOS: treatment_level " +
lvl.to_string() +
" is not present in data.d",
)
}
}
{
data,
treatment_levels,
n_folds,
n_rep,
seed,
propensity_clip,
coefs: Array::make(treatment_levels.length(), 0.0),
ses: Array::make(treatment_levels.length(), 0.0),
fitted: false,
}
}
///|
pub fn DoubleMLAPOS::fit(self : DoubleMLAPOS) -> DoubleMLAPOS {
try {
require(self.treatment_levels.length() >= 1)
let c = Array::make(self.treatment_levels.length(), 0.0)
let s = Array::make(self.treatment_levels.length(), 0.0)
for j = 0; j < self.treatment_levels.length(); j = j + 1 {
// v0.49.0: the parent APOS fits each child `DoubleMLAPO` with
// `n_rep = self.n_rep` so the child draws its own fold
// partition (`n_rep` total fold draws per treatment level).
// The parent does not currently route a shared stratified
// partition to the child — that's a v0.50.0+ target. The
// resulting fold-draw count is `n_rep * n_treatment_levels`,
// matching the upstream `DoubleMLAPOS.fit` total.
let z = DoubleMLAPO::new(
self.data,
treatment_level=self.treatment_levels[j],
n_folds=self.n_folds,
n_rep=self.n_rep,
seed=self.seed,
propensity_clip=self.propensity_clip,
).fit()
c[j] = z.coef()
s[j] = z.se()
}
{
data: self.data,
treatment_levels: self.treatment_levels,
n_folds: self.n_folds,
n_rep: self.n_rep,
seed: self.seed,
propensity_clip: self.propensity_clip,
coefs: c,
ses: s,
fitted: true,
}
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
pub fn DoubleMLAPOS::coefs(self : DoubleMLAPOS) -> Array[Double] {
try {
require(self.fitted)
self.coefs
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
pub fn DoubleMLAPOS::ses(self : DoubleMLAPOS) -> Array[Double] {
try {
require(self.fitted)
self.ses
} catch {
PreconditionError::Violated(loc) =>
abort("precondition failed at " + loc.to_string())
}
}
///|
/// v0.49.0: the requested treatment levels, in user-supplied order.
pub fn DoubleMLAPOS::treatment_levels(self : DoubleMLAPOS) -> Array[Double] {
self.treatment_levels
}
///|
/// v0.49.0: number of requested treatment levels.
pub fn DoubleMLAPOS::n_treatment_levels(self : DoubleMLAPOS) -> Int {
self.treatment_levels.length()
}
///|
/// v0.49.0: whether `fit` has been called.
pub fn DoubleMLAPOS::fitted(self : DoubleMLAPOS) -> Bool {
self.fitted
}
///|
/// v0.49.0: causal contrasts between the requested treatment levels
/// and the supplied reference level(s). Returns one
/// `Array[Double]` of length `2 * treatment_levels.length() - 1`
/// per reference level: the ref-level slot is a single `0.0` (its
/// contrast is trivially zero), and every other slot is a
/// `(delta, se)` pair where `delta = coefs[i] - coefs[ref_idx]`
/// and `se = sqrt(se[i]^2 + se[ref_idx]^2)`. `ref_idx` is the
/// position of the reference level in `treatment_levels`. The
/// layout is interleaved (`[0.0, delta_0, se_0, delta_1, se_1, ...]`
/// for a 2-level input) rather than a `[(level, coef, se), ...]`
/// table — see `apo_test.mbt::apos_causal_contrast_with_reference`
/// for the actual indices.
///
/// The SE of each contrast is computed via the standard
/// `var(psi_a) + var(psi_b) - 2*cov(psi_a, psi_b)` style
/// approximation; for v0.49.0 we take the conservative
/// `sqrt(se_a^2 + se_b^2)` route (same as the upstream
/// `causal_contrast` summary table), which is exact when the
/// per-level psi_a and psi_b are independent across levels
/// (true under stratified kfold with disjoint train indices).
pub fn DoubleMLAPOS::causal_contrast(
self : DoubleMLAPOS,
reference_levels : Array[Double],
) -> Array[Array[Double]] {
// v0.49.0: pre-condition check is local; ref_indices is built
// up in the same scope that consumes it. `abort` (not raise) is
// used so the function signature stays clean.
if !self.fitted {
abort("precondition failed at DoubleMLAPOS::causal_contrast: not fitted")
}
if reference_levels.length() < 1 {
abort(
"precondition failed at DoubleMLAPOS::causal_contrast: reference_levels is empty",
)
}
let ref_indices : Array[Int] = []
for r = 0; r < reference_levels.length(); r = r + 1 {
let ref_lvl = reference_levels[r]
let mut found = false
for i = 0; i < self.treatment_levels.length(); i = i + 1 {
if self.treatment_levels[i] == ref_lvl {
ignore(ref_indices.push(i))
found = true
break
}
}
if !found {
abort(
"DoubleMLAPOS::causal_contrast: reference_level " +
ref_lvl.to_string() +
" is not in treatment_levels",
)
}
}
let results : Array[Array[Double]] = []
for r = 0; r < ref_indices.length(); r = r + 1 {
let ref_idx = ref_indices[r]
let row : Array[Double] = []
for i = 0; i < self.treatment_levels.length(); i = i + 1 {
if i == ref_idx {
ignore(row.push(0.0))
} else {
// delta = coef_i - coef_ref, se = sqrt(se_i^2 + se_ref^2)
let delta = self.coefs[i] - self.coefs[ref_idx]
let se = (self.ses[i] * self.ses[i] +
self.ses[ref_idx] * self.ses[ref_idx]).sqrt()
ignore(row.push(delta))
ignore(row.push(se))
}
}
results.push(row)
}
results
}