Title:	Leaky Instrumental Variables
Version:	0.0.1
Date:	2024-04-09
Maintainer:	David S. Watson <david.s.watson11@gmail.com>
Description:	Instrumental variables (IVs) are a popular and powerful tool for estimating causal effects in the presence of unobserved confounding. However, classical methods rely on strong assumptions such as the exclusion criterion, which states that instrumental effects must be entirely mediated by treatments. In the so-called "leaky" IV setting, candidate instruments are allowed to have some direct influence on outcomes, rendering the average treatment effect (ATE) unidentifiable. But with limits on the amount of information leakage, we may still recover sharp bounds on the ATE, providing partial identification. This package implements methods for ATE bounding in the leaky IV setting with linear structural equations. For details, see Watson et al. (2024) <doi:10.48550/arXiv.2404.04446>.
License:	GPL (≥ 3)
URL:	https://github.com/dswatson/leakyIV
BugReports:	https://github.com/dswatson/leakyIV/issues
Imports:	data.table, corpcor, glasso, Matrix, mvnfast, foreach
Encoding:	UTF-8
RoxygenNote:	7.3.1
NeedsCompilation:	no
Packaged:	2024-04-09 08:34:39 UTC; dswatson
Author:	David S. Watson [aut, cre, cph]
Repository:	CRAN
Date/Publication:	2024-04-09 09:20:03 UTC

Testing Exclusion

Description

Performs a Monte Carlo test against the null hypothesis that minimum leakage is zero, a necessary but insufficient condition for exclusion.

Usage

exclusion_test(
  dat,
  normalize = TRUE,
  method = "mle",
  approx = TRUE,
  n_sim = 1999L,
  parallel = TRUE,
  return_stats = FALSE,
  ...
)

Arguments

Details

The classic linear instrumental variable (IV) model relies on the exclusion criterion, which states that instruments Z have no direct effect on the outcome Y, but can only influence it through the treatment X. This implies a series of tetrad constraints that can be directly tested, given a model for sampling data from the covariance matrix of the observable variables (Watson et al., 2024).

We assume that data are multivariate normal and impose the null hypothesis by modifying the estimated covariance matrix to induce a linear dependence between the vectors for Cov(Z, X) and Cov(Z, Y). Our test statistic is the determinant of the cross product of these vectors, which equals zero if and only if the null hypothesis is true. We generate a null distribution by simulating from the null covariance matrix and compute a p-value by estimating the proportion of statistics that exceed the observed value. Future releases will provide support for a wider range of data generating processes.

Numerous methods exist for estimating covariance matrices. exclusion_test provides support for maximum likelihood estimation (the default), as well as empirical Bayes shrinkage via corpcor::cov.shrink (Schäfer & Strimmer, 2005) and the graphical lasso via glasso::glasso (Friedman et al., 2007). These latter methods are preferable in high-dimensional settings where sample covariance matrices may be unstable or singular. Alternatively, users can pass a pre-computed covariance matrix directly as dat.

Estimated covariance matrices may be singular for some datasets or Monte Carlo samples. Behavior in this case is determined by the approx argument. If TRUE, the test proceeds with the nearest positive definite approximation, computed via Higham's (2002) algorithm (with a warning). If FALSE, the sampler will attempt to use the singular covariance matrix (also with a warning), but results may be invalid.

Value

Either a scalar representing the Monte Carlo p-value of the exclusion test (default) or, if return_stats = TRUE, a named list with three entries: psi, the observed statistic; psi0, a vector of length n_sim with simulated null statistics; and p_value, the resulting p-value.

References

Watson, D., Penn, J., Gunderson, L., Bravo-Hermsdorff, G., Mastouri, A., and Silva, R. (2024). Bounding causal effects with leaky instruments. arXiv preprint, 2404.04446.

Spirtes, P. Calculation of entailed rank constraints in partially non-linear and cyclic models. In Proceedings of the 29th Conference on Uncertainty in Artificial Intelligence, 606–615, 2013.

Friedman, J., Hastie, T., and Tibshirani, R. (2007). Sparse inverse covariance estimation with the lasso. Biostatistics, 9:432-441.

Schäfer, J., and Strimmer, K. (2005). A shrinkage approach to large-scale covariance estimation and implications for functional genomics. Statist. Appl. Genet. Mol. Biol., 4:32.

Higham, N. (2002). Computing the nearest correlation matrix: A problem from finance. IMA J. Numer. Anal., 22:329–343.

Examples

 
set.seed(123)

# Hyperparameters
n <- 200
d_z <- 4
beta <- rep(1, d_z)
theta <- 2
rho <- 0.5

# Simulate correlated residuals
S_eps <- matrix(c(1, rho, rho, 1), ncol = 2)
eps <- matrix(rnorm(n * 2), ncol = 2)
eps <- eps %*% chol(S_eps)

# Simulate observables from the linear IV model
z <- matrix(rnorm(n * d_z), ncol = d_z)
x <- z %*% beta + eps[, 1]
y <- x * theta + eps[, 2]
obs <- cbind(x, y, z)

# Compute p-value of the test
exclusion_test(obs, parallel = FALSE)

Bounding Causal Effects with Leaky Instruments

Description

Estimates bounds on average treatment effects in linear IV models under limited violations of the exclusion criterion.

Usage

leakyIV(
  dat,
  tau,
  p = 2,
  normalize = TRUE,
  method = "mle",
  approx = TRUE,
  n_boot = NULL,
  bayes = FALSE,
  parallel = TRUE,
  ...
)

Arguments

Details

Instrumental variables are defined by three structural assumptions: they must be (A1) relevant, i.e. associated with the treatment; (A2) unconfounded, i.e. independent of common causes between treatment and outcome; and (A3) exclusive, i.e. only affect outcomes through the treatment. The leakyIV algorithm (Watson et al., 2024) relaxes (A3), allowing some information leakage from IVs Z to outcomes Y in linear systems. While the average treatment effect (ATE) is no longer identifiable in this setting, sharp bounds can be computed exactly.

We assume the following structural equation for the treatment: X := Z \beta + \epsilon_X, where the final summand is a noise term that correlates with the additive noise in the structural equation for the outcome: Y := Z \gamma + X \theta + \epsilon_Y. The ATE is given by the parameter \theta. Whereas classical IV models require each \gamma coefficient to be zero, we permit some direct signal from Z to Y. Specifically, leakyIV provides support for two types of information leakage: (a) thresholding the p-norm of linear weights \gamma (scalar tau); and (b) thresholding the absolute value of each \gamma coefficient one by one (vector tau).

Numerous methods exist for estimating covariance matrices. leakyIV provides support for maximum likelihood estimation (the default), as well as empirical Bayes shrinkage via corpcor::cov.shrink (Schäfer & Strimmer, 2005) and the graphical lasso via glasso::glasso (Friedman et al., 2007). These latter methods are preferable in high-dimensional settings where sample covariance matrices may be unstable or singular. Alternatively, users can pass a pre-computed covariance matrix directly as dat.

Estimated covariance matrices may be singular for some datasets or bootstrap samples. Behavior in this case is determined by the approx argument. If TRUE, leakyIV proceeds with the nearest positive definite approximation, computed via Higham's (2002) algorithm (with a warning). If FALSE, bounds are NA (also with a warning).

Uncertainty can be evaluated in leaky IV models using the bootstrap, provided that covariances are estimated internally and not passed directly. Bootstrapping provides a nonparametric sampling distribution for min/max values of the ATE. Set bayes = TRUE to replace the classical bootstrap with a Bayesian bootstrap for approximate posterior inference (Rubin, 1981).

Value

A data frame with columns for ATE_lo and ATE_hi, representing lower and upper bounds of the partial identification interval for the causal effect of X on Y. When bootstrapping, the output data frame contains n_boot rows, one for each bootstrap replicate.

References

Watson, D., Penn, J., Gunderson, L., Bravo-Hermsdorff, G., Mastouri, A., and Silva, R. (2024). Bounding causal effects with leaky instruments. arXiv preprint, 2404.04446.

Friedman, J., Hastie, T., and Tibshirani, R. (2007). Sparse inverse covariance estimation with the lasso. Biostatistics, 9:432-441.

Schäfer, J., and Strimmer, K. (2005). A shrinkage approach to large-scale covariance estimation and implications for functional genomics. Statist. Appl. Genet. Mol. Biol., 4:32.

Higham, N. (2002). Computing the nearest correlation matrix: A problem from finance. IMA J. Numer. Anal., 22:329–343.

Rubin, D.R. (1981). The Bayesian bootstrap. Ann. Statist., 9(1): 130-134.

Examples

 
set.seed(123)

# Hyperparameters
n <- 200
d_z <- 4
beta <- rep(1, d_z)
gamma <- rep(0.1, d_z)
theta <- 2
rho <- 0.5

# Simulate correlated residuals
S_eps <- matrix(c(1, rho, rho, 1), ncol = 2)
eps <- matrix(rnorm(n * 2), ncol = 2)
eps <- eps %*% chol(S_eps)

# Simulate observables from a leaky IV model
z <- matrix(rnorm(n * d_z), ncol = d_z)
x <- z %*% beta + eps[, 1]
y <- z %*% gamma + x * theta + eps[, 2]
obs <- cbind(x, y, z)

# Run the algorithm
leakyIV(obs, tau = 1)

# With bootstrapping
leakyIV(obs, tau = 1, n_boot = 10)

# With covariance matrix input
S <- cov(obs)
leakyIV(S, tau = 1)