CVtreeMLE
:deciduous_tree: :dart: Cross Validated Decision Trees with Targeted Maximum Likelihood Estimation
Install / Use
/learn @blind-contours/CVtreeMLEREADME
CVtreeMLE <img src="man/figures/CVtreeMLE_sticker.png" style="float:right; height:200px;">
<!-- badges: start -->
Discovery of Critical Thresholds in Mixed Exposures and Estimation of Policy Intervention Effects using Targeted Learning
Author: David McCoy
What is CVtreeMLE?
This package operationalizes the methodology presented here:
https://arxiv.org/abs/2302.07976
People often encounter multiple simultaneous exposures (e.g. several drugs or pollutants). Policymakers are interested in setting safe limits, interdictions, or recommended dosage combinations based on a combination of thresholds, one per exposure. Setting these thresholds is difficult because all relevant interactions between exposures must be accounted for. Previous statistical methods have used parametric estimators which do not directly address the question of safe exposure limits, rely on unrealistic assumptions, and do not result in a threshold based statistical quantity that is directly relevant to policy regulators.
Here we present an estimator that a) identifies thresholds that minimize/maximize the expected outcome controlling for covariates and other exposures; and which b) efficiently estimates a policy intervention which compares the expected outcome if everyone was forced to these safe levels compared to the observed outcome under observed exposure distribution.
This is done by using cross-validation where in training folds of the data, a custom g-computation tree-based search algorithm finds the minimizing region, and an estimation sample is used to estimate the policy intervention using targeted maximum likelihood estimation.
Inputs and Outputs
This package takes in a mixed exposure, covariates, outcome, super learner stacks of learners if determined (if not default are used), number of folds, minimum observations in a region, if the desired region is minimizer or maximizer and parallelization parameters.
The output are k-fold specific results for the region found in each fold with valid inference, a pooled estimate of the overall oracle parameter across all folds and pooled exposure sets if the region has some inconsistency across the folds.
Installation
Note: Because CVtreeMLE package (currently) depends on sl3 that
allows ensemble machine learning to be used for nuisance parameter
estimation and sl3 is not on CRAN the CVtreeMLE package is not
available on CRAN and must be downloaded here.
There are many dependencies for CVtreeMLE so it’s easier to break up
installation of the various packages to ensure proper installation.
CVtreeMLE uses the sl3 package to build ensemble machine learners
for each nuisance parameter.
Install sl3 on devel:
remotes::install_github("tlverse/sl3@devel")
Make sure sl3 installs correctly then install CVtreeMLE
remotes::install_github("blind-contours/CVtreeMLE@main")
Example
First load the package and other packages needed
library(CVtreeMLE)
library(sl3)
library(dplyr)
library(kableExtra)
library(ggplot2)
seed <- 98484
set.seed(seed)
To illustrate how CVtreeMLE may be used to find and estimate a region
that, if intervened on would lead to the biggest reduction in an outcome
we use synthetic data from the National Institute of Environmental
Health:
National Institute of Environmental Health Data
The 2015 NIEHS Mixtures Workshop was developed to determine if new
mixture methods detect ground-truth interactions built into the
simulated data. In this way we can simultaneously show CVtreeMLE
output, interpretation and validity.
For detailed information on this simulated data please see:
https://github.com/niehs-prime/2015-NIEHS-MIxtures-Workshop
niehs_data <- NIEHS_data_1
head(niehs_data) %>%
kableExtra::kbl(caption = "NIEHS Data") %>%
kableExtra::kable_classic(full_width = FALSE, html_font = "Cambria")
Briefly, this synthetic data can be considered the results of a prospective cohort epidemiologic study. The outcome cannot cause the exposures (as might occur in a cross-sectional study). Correlations between exposure variables can be thought of as caused by common sources or modes of exposure. The nuisance variable Z can be assumed to be a potential confounder and not a collider. There are 7 exposures which have a complicated dependency structure. $X_3$ and $X_6$ do not have an impact on the outcome.
One issue is that many machine learning algorithms will fail given only 1 variable passed as a feature so let’s add some other covariates.
niehs_data$Z2 <- rbinom(nrow(niehs_data),
size = 1,
prob = 0.3
)
niehs_data$Z3 <- rbinom(nrow(niehs_data),
size = 1,
prob = 0.1
)
Run CVtreeMLE
ptm <- proc.time()
# Convert continuous X variables to their corresponding deciles for example
niehs_data <- niehs_data %>%
mutate(across(starts_wit
Related Skills
proje
Interactive vocabulary learning platform with smart flashcards and spaced repetition for effective language acquisition.
YC-Killer
2.7kA library of enterprise-grade AI agents designed to democratize artificial intelligence and provide free, open-source alternatives to overvalued Y Combinator startups. If you are excited about democratizing AI access & AI agents, please star ⭐️ this repository and use the link in the readme to join our open source AI research team.
best-practices-researcher
The most comprehensive Claude Code skills registry | Web Search: https://skills-registry-web.vercel.app
groundhog
400Groundhog's primary purpose is to teach people how Cursor and all these other coding agents work under the hood. If you understand how these coding assistants work from first principles, then you can drive these tools harder (or perhaps make your own!).
