Gaussian processes for the estimation of causal exposure-response curves (GP-CERF)

Summary

Gaussian Process (GP) and nearest neighbor Gaussian Process (nnGP) approaches for nonparametric modeling.

Installation

r library("devtools") install_github("NSAPH-Software/GPCERF", ref="develop") library("GPCERF")

Usage

Note: The following examples will also need installing ranger R package.

GP

```r library(GPCERF) set.seed(781) simdata <- generatesyntheticdata(samplesize = 500, gps_spec = 1)

n_core <- 1

mxgboost <- function(nthread = ncore, ...) { SuperLearner::SL.xgboost(nthread = nthread, ...) }

mranger <- function(num.threads = ncore, ...){ SuperLearner::SL.ranger(num.threads = num.threads, ...) }

Estimate GPS function

gpsm <- estimategps(covmt = simdata[,-(1:2)], wall = simdata$treat, sllib = c("mxgboost", "mranger"), dnormlog = TRUE)

exposure values

q1 <- stats::quantile(simdata$treat, 0.05) q2 <- stats::quantile(simdata$treat, 0.95)

w_all <- seq(q1, q2, 1)

paramslst <- list(alpha = 10 ^ seq(-2, 2, length.out = 10), beta = 10 ^ seq(-2, 2, length.out = 10), gsigma = c(0.1, 1, 10), tune_app = "all")

cerfgpobj <- estimatecerfgp(simdata, wall, gpsm, params = paramslst, outcomecol = "Y", treatmentcol = "treat", covariatescol = paste0("cf", seq(1,6)), nthread = ncore) summary(cerfgpobj) plot(cerfgpobj) GPCERF standard Gaussian grocess exposure response function object

Optimal hyper parameters(#trial: 300): alpha = 12.9154966501488 beta = 12.9154966501488 g_sigma = 0.1

Optimal covariate balance: cf1 = 0.069 cf2 = 0.082 cf3 = 0.063 cf4 = 0.066 cf5 = 0.056 cf6 = 0.081

Original covariate balance: cf1 = 0.222 cf2 = 0.112 cf3 = 0.175 cf4 = 0.318 cf5 = 0.198 cf6 = 0.257 ----***----
```

nnGP

```r set.seed(781) simdata <- generatesyntheticdata(samplesize = 5000, gps_spec = 1)

m_xgboost <- function(nthread = 12, ...) { SuperLearner::SL.xgboost(nthread = nthread, ...) }

m_ranger <- function(num.threads = 12, ...){ SuperLearner::SL.ranger(num.threads = num.threads, ...) }

Estimate GPS function

gpsm <- estimategps(covmt = simdata[,-(1:2)], wall = simdata$treat, sllib = c("mxgboost", "mranger"), dnormlog = TRUE)

exposure values

q1 <- stats::quantile(simdata$treat, 0.05) q2 <- stats::quantile(simdata$treat, 0.95)

w_all <- seq(q1, q2, 1)

paramslst <- list(alpha = 10 ^ seq(-2, 2, length.out = 10), beta = 10 ^ seq(-2, 2, length.out = 10), gsigma = c(0.1, 1, 10), tuneapp = "all", nneighbor = 50, block_size = 1e3)

cerfnngpobj <- estimatecerfnngp(simdata, wall, gpsm, params = paramslst, outcomecol = "Y", treatmentcol = "treat", covariatescol = paste0("cf", seq(1,6)), nthread = 12) summary(cerfnngpobj) plot(cerfnngp_obj) ```

``` GPCERF nearest neighbore Gaussian process exposure response function object summary

Optimal hyper parameters(#trial: 300): alpha = 0.0278255940220712 beta = 0.215443469003188 g_sigma = 0.1

Optimal covariate balance: cf1 = 0.062 cf2 = 0.070 cf3 = 0.091 cf4 = 0.062 cf5 = 0.076 cf6 = 0.088

Original covariate balance: cf1 = 0.115 cf2 = 0.137 cf3 = 0.145 cf4 = 0.296 cf5 = 0.208 cf6 = 0.225 ----***----
```

Code of Conduct

Please note that the GPCERF project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

Contributing

Contributions to the package are encouraged. For detailed information on how to contribute, please refer to the CONTRIBUTING guidelines.

Reporting Issues & Seeking Support

If you encounter any issues with GPCERF, we kindly ask you to report them on our GitHub by opening a new issue. To expedite resolution, including a reproducible example is highly appreciated. For those seeking assistance or further details about a particular topic, feel free to initiate a Discussion on GitHub or open an issue. Additionally, for more direct inquiries, the package maintainer can be reached via the email address provided in the DESCRIPTION file.

References

Ren, B., Wu, X., Braun, D., Pillai, N. and Dominici, F., 2021. Bayesian modeling for exposure response curve via gaussian processes: Causal effects of exposure to air pollution on health outcomes. arXiv preprint arXiv:2105.03454.