Package 'dmlrs'

Confidence intervals for DML estimates

Description

Pointwise intervals, or simultaneous intervals for all treatments from the multiplier bootstrap (joint = TRUE), as DoubleML's confint().

Usage

## S3 method for class 'dmlrs_fit'
confint(
  object,
  parm,
  level = 0.95,
  joint = FALSE,
  method = c("normal", "Bayes", "wild"),
  n_boot = 500,
  seed = 0,
  ...
)

Arguments

object	A fitted model.
parm	Treatments (names or indices); all by default.
level	Confidence level.
joint	Simultaneous intervals from the multiplier bootstrap.
method	Bootstrap weights: "normal", "Bayes" or "wild".
n_boot	Number of bootstrap draws.
seed	Seed of the bootstrap.
...	Not used.

Value

A matrix with one row per treatment.

Examples

set.seed(3)
n <- 400
x <- matrix(rnorm(n * 3), n)
d <- cbind(d1 = x[, 1] + rnorm(n), d2 = x[, 2] + rnorm(n))
y <- 0.5 * d[, 1] - 0.2 * d[, 2] + x[, 3] + rnorm(n)
fit <- dml_plr(y, d, x, ml_l = lrn_ols())
confint(fit, joint = TRUE, n_boot = 1000)

---

Difference in differences

Description

The average treatment effect on the treated with two periods (Sant'Anna and Zhao, 2020), as DoubleMLDID and DoubleMLDIDCS. With panel data, y is the change of the outcome between the periods; with repeated cross-sections, y is the outcome and t the period (0/1).

Usage

dml_did(
  y,
  d,
  x,
  t = NULL,
  ml_g = lrn_lasso_cv(),
  ml_m = lrn_logistic_cv(),
  score = c("observational", "experimental"),
  in_sample_normalization = TRUE,
  n_folds = 5,
  n_rep = 1,
  seed = 0
)

Arguments

y	Outcome change (panel) or outcome (repeated cross-sections).
d	Binary treatment group indicator (0/1).
x	Covariates: a numeric matrix or data frame.
t	NULL for panel data, or the period indicator (0/1) of repeated cross-sections.
ml_g	Learner for the outcome regressions.
ml_m	Classifier for the propensity score (not used by the "experimental" score).
score	"observational" or "experimental".
in_sample_normalization	Normalise the weights in sample, as DoubleML.
n_folds	Number of cross-fitting folds.
n_rep	Number of repetitions of the sample split; estimates are aggregated with the median, as in DoubleML.
seed	Seed of the sample splits.

Value

A dmlrs_fit object; see stats::coef(), confint.dmlrs_fit(), base::summary() and sensitivity_analysis().

---

Interactive IV model

Description

Local average treatment effect of a binary treatment with a binary instrument, as DoubleMLIIVM.

Usage

dml_iivm(
  y,
  d,
  z,
  x,
  ml_g = lrn_lasso_cv(),
  ml_m = lrn_logistic_cv(),
  ml_r = ml_m,
  subgroups = c(always_takers = TRUE, never_takers = TRUE),
  n_folds = 5,
  n_rep = 1,
  seed = 0
)

Arguments

y	Outcome (numeric vector).
d	Binary treatment (0/1).
z	Binary instrument (0/1).
x	Covariates: a numeric matrix or data frame.
ml_g	Learner for $E[Y \mid Z, X]$.
ml_m	Classifier for $P(Z = 1 \mid X)$.
ml_r	Classifier for $P(D = 1 \mid Z, X)$.
subgroups	Named logical vector: whether there can be always_takers and never_takers (as DoubleML's subgroups).
n_folds	Number of cross-fitting folds.
n_rep	Number of repetitions of the sample split; estimates are aggregated with the median, as in DoubleML.
seed	Seed of the sample splits.

Value

A dmlrs_fit object; see stats::coef(), confint.dmlrs_fit(), base::summary() and sensitivity_analysis().

---

Interactive regression model

Description

Average treatment effect ("ATE") or average treatment effect on the treated ("ATTE") of a binary treatment, as DoubleMLIRM.

Usage

dml_irm(
  y,
  d,
  x,
  ml_g = lrn_lasso_cv(),
  ml_m = lrn_logistic_cv(),
  score = c("ATE", "ATTE"),
  clipping_threshold = 0.01,
  normalize_ipw = FALSE,
  n_folds = 5,
  n_rep = 1,
  seed = 0
)

Arguments

y	Outcome (numeric vector).
d	Binary treatment (0/1).
x	Covariates: a numeric matrix or data frame.
ml_g	Learner for $E[Y \mid D, X]$ (fitted on each treatment arm).
ml_m	Classifier for the propensity score $P(D = 1 \mid X)$.
score	"ATE" or "ATTE".
clipping_threshold	Propensity scores are clipped to $[c, 1 - c]$.
normalize_ipw	Normalise the inverse-probability weights.
n_folds	Number of cross-fitting folds.
n_rep	Number of repetitions of the sample split; estimates are aggregated with the median, as in DoubleML.
seed	Seed of the sample splits.

Value

A dmlrs_fit object; see stats::coef(), confint.dmlrs_fit(), base::summary() and sensitivity_analysis().

Examples

set.seed(2)
n <- 1000
x <- matrix(rnorm(n * 4), n)
d <- rbinom(n, 1, plogis(x[, 1]))
y <- d + x[, 1] + rnorm(n)
dml_irm(y, d, x, ml_g = lrn_ols(), ml_m = lrn_logistic())

---

Partially linear IV regression

Description

The partially linear model with an endogenous treatment and one instrument $Z$, as DoubleMLPLIV ("partialling out" score).

Usage

dml_pliv(
  y,
  d,
  z,
  x,
  ml_l = lrn_lasso_cv(),
  ml_m = ml_l,
  ml_r = ml_l,
  n_folds = 5,
  n_rep = 1,
  seed = 0
)

Arguments

y	Outcome (numeric vector).
d	Treatment: a numeric vector, or a matrix with one column per treatment (column names become the treatment names).
z	Instrument (numeric vector).
x	Covariates: a numeric matrix or data frame.
ml_l	Learner for $E[Y \mid X]$.
ml_m	Learner for $E[Z \mid X]$.
ml_r	Learner for $E[D \mid X]$.
n_folds	Number of cross-fitting folds.
n_rep	Number of repetitions of the sample split; estimates are aggregated with the median, as in DoubleML.
seed	Seed of the sample splits.

Value

A dmlrs_fit object; see stats::coef(), confint.dmlrs_fit(), base::summary() and sensitivity_analysis().

---

Partially linear regression

Description

Double / debiased machine learning for the partially linear model $Y = D\theta + g(X) + \varepsilon$, as DoubleMLPLR. The nuisance functions are cross-fitted in Rust.

Usage

dml_plr(
  y,
  d,
  x,
  ml_l = lrn_lasso_cv(),
  ml_m = ml_l,
  ml_g = NULL,
  score = c("partialling out", "IV-type"),
  n_folds = 5,
  n_rep = 1,
  seed = 0
)

Arguments

y	Outcome (numeric vector).
d	Treatment: a numeric vector, or a matrix with one column per treatment (column names become the treatment names).
x	Covariates: a numeric matrix or data frame.
ml_l	Learner for $E[Y \mid X]$ (see learners).
ml_m	Learner for $E[D \mid X]$.
ml_g	Learner for $g(X)$, used by the "IV-type" score (default: ml_l).
score	"partialling out" or "IV-type".
n_folds	Number of cross-fitting folds.
n_rep	Number of repetitions of the sample split; estimates are aggregated with the median, as in DoubleML.
seed	Seed of the sample splits.

Value

A dmlrs_fit object; see stats::coef(), confint.dmlrs_fit(), base::summary() and sensitivity_analysis().

Examples

set.seed(1)
n <- 500
x <- matrix(rnorm(n * 5), n)
d <- x[, 1] + rnorm(n)
y <- 0.5 * d + x[, 2] + rnorm(n)
fit <- dml_plr(y, d, x, ml_l = lrn_lasso_cv())
fit
confint(fit)

---

Built-in nuisance learners

Description

Learners for the nuisance functions of the DML models, implemented in Rust. Parameters follow scikit-learn, and results match the learners of the Python package dml-rs. Regressors learn conditional means; classifiers (lrn_logistic(), lrn_logistic_cv(), and the tree ensembles) learn probabilities of a binary variable.

Usage

lrn_ols(ridge = 0)

lrn_lasso(alpha = 1, tol = 1e-04, max_iter = 1e+05)

lrn_lasso_cv(n_alphas = 100, eps = 0.001, cv = 5)

lrn_logistic(C = 1)

lrn_logistic_cv(n_cs = 10, cv = 5)

lrn_random_forest(
  n_estimators = 100,
  max_depth = NULL,
  min_samples_leaf = 1,
  max_features = NULL,
  random_state = 0
)

lrn_gradient_boosting(
  n_estimators = 100,
  learning_rate = 0.1,
  max_depth = 3,
  min_samples_leaf = 1,
  subsample = 1,
  random_state = 0
)

lrn_hist_gradient_boosting(
  max_iter = 100,
  learning_rate = 0.1,
  max_leaf_nodes = 31,
  min_samples_leaf = 20,
  l2_regularization = 0,
  random_state = 0
)

lrn_regression_forest(n_estimators = 2000, min_node_size = 5, random_state = 0)

Arguments

ridge, alpha, C	Penalties (see above).
tol, max_iter	Convergence tolerance and maximum iterations (sweeps of coordinate descent, or boosting iterations for lrn_hist_gradient_boosting()).
n_alphas, eps, n_cs, cv	Grid size, smallest-to-largest penalty ratio and number of folds of the cross-validated learners.
n_estimators	Number of trees.
max_depth	Maximum tree depth (NULL: unlimited).
min_samples_leaf	Minimum observations per leaf.
max_features	Features tried per split: NULL (all for regression, "sqrt" for classification), "sqrt", "log2", a fraction or a number.
random_state	Seed.
learning_rate	Shrinkage of each tree.
subsample	Share of observations used for each tree.
max_leaf_nodes	Maximum leaves per tree (NULL: unlimited).
l2_regularization	L2 penalty on leaf values.
min_node_size	Minimum observations per leaf (grf's min.node.size).

Details

• lrn_ols(): least squares with intercept, optionally ridge-penalised (ridge is scikit-learn's Ridge(alpha)).

• lrn_lasso(): lasso with penalty alpha (scikit-learn's Lasso).

• lrn_lasso_cv(): lasso with the penalty chosen by cv-fold cross-validation over n_alphas values (scikit-learn's LassoCV).

• lrn_logistic(): L2-penalised logistic regression with inverse penalty C.

• lrn_logistic_cv(): logistic regression with C chosen by stratified cross-validation (log loss) over n_cs values from 1e-4 to 1e4.

• lrn_random_forest(): random forest (regressor or classifier).

• lrn_gradient_boosting(): gradient boosting with the trees of scikit-learn's ⁠GradientBoosting*⁠.

• lrn_hist_gradient_boosting(): histogram gradient boosting, identical to scikit-learn's ⁠HistGradientBoosting*⁠; the learner for large data.

• lrn_regression_forest(): honest regression forest (R's grf::regression_forest), regression only.

Value

A learner specification (class dmlrs_learner).

Examples

lrn_lasso_cv()
lrn_random_forest(n_estimators = 200, min_samples_leaf = 5)

---

Sensitivity analysis for omitted confounding

Description

Bounds on the effect and robustness values when the confounding is as strong as cf_y (share of the residual outcome variance) and cf_d (share of the residual Riesz representer variance), with correlation rho (Chernozhukov et al., 2022), as DoubleML's sensitivity_analysis(). Available for PLR, IRM and DID fits.

Usage

sensitivity_analysis(
  object,
  cf_y = 0.03,
  cf_d = 0.03,
  rho = 1,
  level = 0.95,
  null_hypothesis = 0
)

Arguments

object	A fitted model.
cf_y, cf_d	Confounding strengths.
rho	Correlation of the confounding in the outcome and treatment.
level	Confidence level of the bounds.
null_hypothesis	Value of the effect under the null hypothesis.

Value

A data frame with one row per treatment: the bounds on the effect (theta_lower, theta_upper), their confidence bounds (ci_lower, ci_upper) and the robustness values rv and rva (in percent).

Examples

set.seed(4)
n <- 500
x <- matrix(rnorm(n * 3), n)
d <- x[, 1] + rnorm(n)
y <- 0.5 * d + x[, 2] + rnorm(n)
sensitivity_analysis(dml_plr(y, d, x, ml_l = lrn_ols()))

---