Confidence bands and multiplier bootstrap for valid simultaneous inference#

DoubleML provides methods to perform valid simultaneous inference for multiple treatment variables. As an example, consider a PLR with \(p_1\) causal parameters of interest \(\theta_{0,1}, \ldots, \theta_{0,p_1}\) associated with treatment variables \(D_1, \ldots, D_{p_1}\). Inference on multiple target coefficients can be performed by iteratively applying the DML inference procedure over the target variables of interests: Each of the coefficients of interest, \(\theta_{0,j}\), with \(j \in \lbrace 1, \ldots, p_1 \rbrace\), solves a corresponding moment condition

\[\mathbb{E}[ \psi_j(W; \theta_{0,j}, \eta_{0,j})] = 0.\]

Analogously to the case with a single parameter of interest, the PLR model with multiple treatment variables includes two regression steps to achieve orthogonality. First, the main regression is given by

\[Y = D_j \theta_{0,j} + g_{0,j}([D_k, X]) + \zeta_j, \quad \mathbb{E}(\zeta_j | D, X) = 0,\]

with \([D_k, X]\) being a matrix comprising the confounders, \(X\), and all remaining treatment variables \(D_k\) with \(k \in \lbrace 1, \ldots, p_1\rbrace \setminus j\), by default. Second, the relationship between the treatment variable \(D_j\) and the remaining explanatory variables is determined by the equation

\[D_j = m_{0,j}([D_k, X]) + V_j, \quad \mathbb{E}(V_j | D_k, X) = 0,\]

For further details, we refer to Belloni et al. (2018). Simultaneous inference can be based on a multiplier bootstrap procedure introduced in Chernozhukov et al. (2013, 2014). Alternatively, traditional correction approaches, for example the Bonferroni correction, can be used to adjust p-values.

Multiplier bootstrap and joint confidence intervals#

The bootstrap() method provides an implementation of a multiplier bootstrap for double machine learning models. For \(b=1, \ldots, B\) weights \(\xi_{i, b}\) are generated according to a normal (Gaussian) bootstrap, wild bootstrap or exponential bootstrap. The number of bootstrap samples is provided as input n_rep_boot and for method one can choose 'Bayes', 'normal' or 'wild'. Based on the estimates of the standard errors \(\hat{\sigma}_j\) and \(\hat{J}_{0,j} = \mathbb{E}_N(\psi_{a,j}(W; \eta_{0,j}))\) that are obtained from DML, we construct bootstrap coefficients \(\theta^{*,b}_j\) and bootstrap t-statistics \(t^{*,b}_j\) for \(j=1, \ldots, p_1\)

\[ \begin{align}\begin{aligned}\theta^{*,b}_{j} &= \frac{1}{\sqrt{N} \hat{J}_{0,j}}\sum_{k=1}^{K} \sum_{i \in I_k} \xi_{i}^b \cdot \psi_j(W_i; \tilde{\theta}_{0,j}, \hat{\eta}_{0,j;k}),\\t^{*,b}_{j} &= \frac{1}{\sqrt{N} \hat{J}_{0,j} \hat{\sigma}_{j}} \sum_{k=1}^{K} \sum_{i \in I_k} \xi_{i}^b \cdot \psi_j(W_i; \tilde{\theta}_{0,j}, \hat{\eta}_{0,j;k}).\end{aligned}\end{align} \]

The output of the multiplier bootstrap can be used to determine the constant, \(c_{1-\alpha}\) that is required for the construction of a simultaneous \((1-\alpha)\) confidence band

\[\left[\tilde\theta_{0,j} \pm c_{1-\alpha} \cdot \hat\sigma_j/\sqrt{N} \right].\]

To demonstrate the bootstrap, we simulate data from a sparse partially linear regression model. Then we estimate the PLR model and perform the multiplier bootstrap. Joint confidence intervals based on the multiplier bootstrap are then obtained by setting the option joint when calling the method confint.

Moreover, a multiple hypotheses testing adjustment of p-values from a high-dimensional model can be obtained with the method p_adjust. DoubleML performs a version of the Romano-Wolf stepdown adjustment, which is based on the multiplier bootstrap, by default. Alternatively, p_adjust allows users to apply traditional corrections via the option method.

Python

In [1]: import doubleml as dml

In [2]: import numpy as np

In [3]: from sklearn.base import clone

In [4]: from sklearn.linear_model import LassoCV

# Simulate data
In [5]: np.random.seed(1234)

In [6]: n_obs = 500

In [7]: n_vars = 100

In [8]: X = np.random.normal(size=(n_obs, n_vars))

In [9]: theta = np.array([3., 3., 3.])

In [10]: y = np.dot(X[:, :3], theta) + np.random.standard_normal(size=(n_obs,))

In [11]: dml_data = dml.DoubleMLData.from_arrays(X[:, 10:], y, X[:, :10])

In [12]: learner = LassoCV()

In [13]: ml_l = clone(learner)

In [14]: ml_m = clone(learner)

In [15]: dml_plr = dml.DoubleMLPLR(dml_data, ml_l, ml_m)

In [16]: print(dml_plr.fit().bootstrap().confint(joint=True))
        2.5 %    97.5 %
d1   2.813342  3.055680
d2   2.815224  3.083258
d3   2.860663  3.109069
d4  -0.141546  0.091391
d5  -0.060845  0.176929
d6  -0.158697  0.078474
d7  -0.172022  0.062964
d8  -0.067721  0.174499
d9  -0.092365  0.139491
d10 -0.110717  0.138698

In [17]: print(dml_plr.p_adjust())
       thetas   pval
d1   2.934511  0.000
d2   2.949241  0.000
d3   2.984866  0.000
d4  -0.025077  0.902
d5   0.058042  0.784
d6  -0.040112  0.808
d7  -0.054529  0.784
d8   0.053389  0.784
d9   0.023563  0.902
d10  0.013990  0.902

In [18]: print(dml_plr.p_adjust(method='bonferroni'))
       thetas  pval
d1   2.934511   0.0
d2   2.949241   0.0
d3   2.984866   0.0
d4  -0.025077   1.0
d5   0.058042   1.0
d6  -0.040112   1.0
d7  -0.054529   1.0
d8   0.053389   1.0
d9   0.023563   1.0
d10  0.013990   1.0

R

library(DoubleML)
library(mlr3)
library(mlr3learners)
library(data.table)
lgr::get_logger("mlr3")$set_threshold("warn")

set.seed(3141)
n_obs = 500
n_vars = 100
theta = rep(3, 3)
X = matrix(stats::rnorm(n_obs * n_vars), nrow = n_obs, ncol = n_vars)
y = X[, 1:3, drop = FALSE] %*% theta  + stats::rnorm(n_obs)
dml_data = double_ml_data_from_matrix(X = X[, 11:n_vars], y = y, d = X[,1:10])

learner = lrn("regr.cv_glmnet", s="lambda.min")
ml_l = learner$clone()
ml_m = learner$clone()
dml_plr = DoubleMLPLR$new(dml_data, ml_l, ml_m)

dml_plr$fit()
dml_plr$bootstrap()
dml_plr$confint(joint=TRUE)
dml_plr$p_adjust()
dml_plr$p_adjust(method="bonferroni")

A matrix: 10 × 2 of type dbl
	2.5 %	97.5 %
d1	2.89027368	3.14532650
d2	2.90794478	3.14368145
d3	2.87430335	3.12752825
d4	-0.14790924	0.07828372
d5	-0.09779675	0.16803512
d6	-0.12105472	0.12539340
d7	-0.16536299	0.09310496
d8	-0.10127930	0.14200098
d9	-0.13868238	0.09980311
d10	-0.04444978	0.19680840

A matrix: 10 × 2 of type dbl
	Estimate.	pval
d1	3.017800092	0.000
d2	3.025813114	0.000
d3	3.000915799	0.000
d4	-0.034812763	0.938
d5	0.035119185	0.938
d6	0.002169338	0.958
d7	-0.036129015	0.938
d8	0.020360838	0.954
d9	-0.019439633	0.954
d10	0.076179312	0.428

A matrix: 10 × 2 of type dbl
	Estimate.	pval
d1	3.017800092	0.0000000
d2	3.025813114	0.0000000
d3	3.000915799	0.0000000
d4	-0.034812763	1.0000000
d5	0.035119185	1.0000000
d6	0.002169338	1.0000000
d7	-0.036129015	1.0000000
d8	0.020360838	1.0000000
d9	-0.019439633	1.0000000
d10	0.076179312	0.8116912

References#

Belloni, A., Chernozhukov, V., Chetverikov, D., Wei, Y. (2018), Uniformly valid post-regularization confidence regions for many functional parameters in z-estimation framework. The Annals of Statistics, 46 (6B): 3643-75, doi: 10.1214/17-AOS1671.
Chernozhukov, V., Chetverikov, D., Kato, K. (2013). Gaussian approximations and multiplier bootstrap for maxima of sums of high-dimensional random vectors. The Annals of Statistics 41 (6): 2786-2819, doi: 10.1214/13-AOS1161.
Chernozhukov, V., Chetverikov, D., Kato, K. (2014), Gaussian approximation of suprema of empirical processes. The Annals of Statistics 42 (4): 1564-97, doi: 10.1214/14-AOS1230.