Skip to content

General targeted machine learning for modern causal mediation analysis

License

Notifications You must be signed in to change notification settings

nt-williams/crumble

Repository files navigation

crumble (verb): break or fall apart into small fragments

crumble

Lifecycle: experimental CRAN status License: GPL v3

crumble implements a modern, unified estimation strategy (Liu et al. 2024) for common mediation estimands: natural effects (Pearl 2022), organic effects (Lok 2015), interventional effects (Vansteelandt and Daniel 2017), recanting twins (Vo et al. 2024), in causal inference in combination with modified treatment policies. It makes use of recent advancements in “Riesz-learning” to estimate a set of required nuisance parameters using deep learning. The result is a software package that is capable of estimating mediation effects with binary, categorical, continuous, or multivariate exposures with high-dimensional mediators and mediator-outcome confounders using machine learning.

This work was supported by the National Institute on Drug Abuse [R00DA042127].

Installation

remotes::install_github("nt-williams/crumble")

Features

Feature Status
Recanting twins
Natural effects
Organic effects
Interventional effects
Modified treatment Policy
Static intervention
Dynamic intervention
Continuous treatment
Binary treatment
Categorical treatment
Multivariate treatment
Missingness in treatment
Continuous outcome
Binary outcome
Censored outcome
Survey weights Planned
Super learner
Clustered data Planned
Parallel processing
GPU support
Progress bars

Example(s)

library(crumble)
library(mlr3extralearners)

data(weight_behavior, package = "mma")

weight_behavior <- na.omit(weight_behavior)

set.seed(2345)
Recanting twins
crumble(
    data = weight_behavior,
    trt = "sports", 
    outcome = "bmi",
    covar = c("age", "sex", "tvhours"),
    mediators = c("exercises", "overweigh"),
    moc = "snack",
    d0 = \(data, trt) factor(rep(1, nrow(data)), levels = c("1", "2")), 
    d1 = \(data, trt) factor(rep(2, nrow(data)), levels = c("1", "2")), 
    effect = "RT",
    learners = c("mean", "glm", "earth", "ranger"), 
    nn_module = sequential_module(),
    control = crumble_control(crossfit_folds = 1L, epochs = 20L)
)
#> ✔ Permuting Z-prime variables... 1/1 tasks [2.5s]
#> ✔ Fitting outcome regressions... 1/1 folds [25.6s]             
#> ✔ Computing alpha n density ratios... 1/1 folds [39.7s]        
#> ✔ Computing alpha r density ratios... 1/1 folds [41.6s]        
#> 
#> ══ Results `crumble()` ═════════════════════════════════════════
#> 
#> ── E[Y(d1) - Y(d0)] 
#>       Estimate: 1.0537
#>     Std. error: 0.3009
#>         95% CI: (0.4639, 1.6435)
#> 
#> ── Path: A -> Y 
#>       Estimate: 0.0366
#>     Std. error: 0.1842
#>         95% CI: (-0.3245, 0.3976)
#> 
#> ── Path: A -> Z -> Y 
#>       Estimate: -0.0202
#>     Std. error: 0.0238
#>         95% CI: (-0.0668, 0.0264)
#> 
#> ── Path: A -> Z -> M -> Y 
#>       Estimate: -6e-04
#>     Std. error: 0.0099
#>         95% CI: (-0.02, 0.0189)
#> 
#> ── Path: A -> M -> Y 
#>       Estimate: 1.0506
#>     Std. error: 0.2162
#>         95% CI: (0.627, 1.4743)
#> 
#> ── Intermediate Confounding 
#>       Estimate: -0.0127
#>     Std. error: 0.0261
#>         95% CI: (-0.0638, 0.0384)
Natural effects
crumble(
    data = weight_behavior,
    trt = "sports", 
    outcome = "bmi",
    covar = c("age", "sex", "tvhours"),
    mediators = c("exercises", "overweigh"),
    d0 = \(data, trt) factor(rep(1, nrow(data)), levels = c("1", "2")), 
    d1 = \(data, trt) factor(rep(2, nrow(data)), levels = c("1", "2")), 
    effect = "N",
    learners = c("mean", "glm", "earth", "ranger"), 
    nn_module = sequential_module(),
    control = crumble_control(crossfit_folds = 1L, epochs = 20L)
)
#> ✔ Fitting outcome regressions... 1/1 folds [10.6s]             
#> ✔ Computing alpha n density ratios... 1/1 folds [53.1s]        
#> 
#> ══ Results `crumble()` ═════════════════════════════════════════
#> 
#> ── E[Y(d1) - Y(d0)] 
#>       Estimate: 1.0289
#>     Std. error: 0.28
#>         95% CI: (0.48, 1.5777)
#> 
#> ── Natural Direct Effect 
#>       Estimate: 0.0165
#>     Std. error: 0.1717
#>         95% CI: (-0.3201, 0.3531)
#> 
#> ── Natural Indirect Effect 
#>       Estimate: 1.0124
#>     Std. error: 0.2178
#>         95% CI: (0.5856, 1.4393)
Organic effects
crumble(
    data = weight_behavior,
    trt = "sports", 
    outcome = "bmi",
    covar = c("age", "sex", "tvhours"),
    mediators = c("exercises", "overweigh"),
    d0 = \(data, trt) factor(rep(1, nrow(data)), levels = c("1", "2")), 
    d1 = \(data, trt) factor(rep(2, nrow(data)), levels = c("1", "2")), 
    effect = "O",
    learners = c("mean", "glm", "earth", "ranger"), 
    nn_module = sequential_module(),
    control = crumble_control(crossfit_folds = 1L, epochs = 20L)
)
#> ✔ Fitting outcome regressions... 1/1 folds [10.7s]             
#> ✔ Computing alpha n density ratios... 1/1 folds [48.2s]        
#> 
#> ══ Results `crumble()` ═════════════════════════════════════════
#> 
#> ── Organic Direct Effect 
#>       Estimate: 0.011
#>     Std. error: 0.1772
#>         95% CI: (-0.3364, 0.3584)
#> 
#> ── Organic Indirect Effect 
#>       Estimate: 1.0278
#>     Std. error: 0.2231
#>         95% CI: (0.5904, 1.4651)#> 
Randomized interventional effects
crumble(
    data = weight_behavior,
    trt = "sports", 
    outcome = "bmi",
    covar = c("age", "sex", "tvhours"),
    mediators = c("exercises", "overweigh"),
    moc = "snack",
    d0 = \(data, trt) factor(rep(1, nrow(data)), levels = c("1", "2")), 
    d1 = \(data, trt) factor(rep(2, nrow(data)), levels = c("1", "2")), 
    effect = "RI",
    learners = c("mean", "glm", "earth", "ranger"), 
    nn_module = sequential_module(),
    control = crumble_control(crossfit_folds = 1L, epochs = 20L)
)
#> ✔ Permuting Z-prime variables... 1/1 tasks [2s]
#> ✔ Fitting outcome regressions... 1/1 folds [14.2s]             
#> ✔ Computing alpha r density ratios... 1/1 folds [1m 23.2s]     
#> 
#> ══ Results `crumble()` ═════════════════════════════════════════
#> 
#> ── Randomized Direct Effect 
#>       Estimate: 0.0162
#>     Std. error: 0.1774
#>         95% CI: (-0.3315, 0.364)
#> 
#> ── Randomized Indirect Effect 
#>       Estimate: 1.0304
#>     Std. error: 0.2296
#>         95% CI: (0.5805, 0.4662)

References

Liu, Richard, Nicholas T Williams, Kara E Rudolph, and Iván Dı́az. 2024. “General Targeted Machine Learning for Modern Causal Mediation Analysis.” arXiv Preprint arXiv:2408.14620.

Lok, Judith J. 2015. “Organic Direct and Indirect Effects with Post-Treatment Common Causes of Mediator and Outcome.” https://doi.org/10.48550/ARXIV.1510.02753.

Pearl, Judea. 2022. “Direct and Indirect Effects.” In, 373–92. ACM. https://doi.org/10.1145/3501714.3501736.

Vansteelandt, Stijn, and Rhian M. Daniel. 2017. “Interventional Effects for Mediation Analysis with Multiple Mediators.” Epidemiology 28 (2): 258–65. https://doi.org/10.1097/ede.0000000000000596.

Vo, Tat-Thang, Nicholas Williams, Richard Liu, Kara E. Rudolph, and Ivan Dıaz. 2024. “Recanting Twins: Addressing Intermediate Confounding in Mediation Analysis.” https://doi.org/10.48550/ARXIV.2401.04450.

About

General targeted machine learning for modern causal mediation analysis

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published