Repository for R and Python packages PSPS that implements Task-Agnostic Machine Learning-Assisted Inference.
PSPS is a simple and task-agnotsic protocol for valid and efficient machine learning (ML)-assited infernece. It can be easily adapted to a variety of statistical tasks.
# install.packages("devtools")
devtools::install_github("qlu-lab/PSPS", subdir = "psps-r")fit_PSPS <- PSPS(est_lab_y, est_lab_yhat, est_unlab_yhat, Sigma)Suppose we are interested in estimating a K-dimensional parameter, then
est_lab_y: a K-dimensional vector of Point estimates usingYin labeled data.est_lab_yhat: a K-dimensional vector of Point estimates usingYhatin labeled data.est_unlab_yhat: a K-dimensional vector of Point estimates usingYhatin unlabeled data.Sigma: a 3K x 3K Variance-covariance matrix for the above three estimators (Note: not the asymptotic variance).
Here is an example of PSPS for logistic regression. For other tasks, simply replace logistic regression with other algorithms to produce summary statistics.
# Load the package
library(PSPS)
# Load the Labeled and unlabelled data
lab <- read.csv("./PSPS/test_data/lab.csv")
unlab <- read.csv("./PSPS/test_data/unlab.csv")
# Point Estimates Preparation
## Fit logistic regression models and extract the second coefficient (typically the slope)
est_lab_y <- coef(glm(Y ~ X, data = lab, family = binomial("logit")))[2]
est_lab_yhat <- coef(glm(Yhat ~ X, data = lab, family = binomial("logit")))[2]
est_unlab_yhat <- coef(glm(Yhat ~ X, data = unlab, family = binomial("logit")))[2]
# Variance-Covariance Matrix Preparation
## Bootstrap for Covariance Calculation between est_lab_y and est_lab_yhat
B <- 1000 # Number of bootstrap iterations
n <- nrow(lab) # Total observations in labeled data
### Initialize matrices to store bootstrap estimates
est_lab_y_boot <- vector("numeric", B)
est_lab_yhat_boot <- vector("numeric", B)
### Perform bootstrap to estimate the variance-covariance of est_lab_y and est_lab_yhat
set.seed(123) # Set seed for reproducibility
for (i in 1:B) {
boot_indices <- sample(1:n, n, replace = TRUE)
boot_lab <- lab[boot_indices, ]
est_lab_y_boot[i] <- coef(glm(Y ~ X, data = boot_lab, family = binomial("logit")))[2]
est_lab_yhat_boot[i] <- coef(glm(Yhat ~ X, data = boot_lab, family = binomial("logit")))[2]
}
### Compute the covariance matrix
Sigma <- matrix(0, nrow = 3, ncol = 3)
Sigma[1:2, 1:2] <- cov(cbind(est_lab_y_boot, est_lab_yhat_boot)) # Covariance of bootstrap estimates
Sigma[3, 3] <- summary(glm(Yhat ~ X, data = unlab, family = binomial("logit")))$coefficients[2, "Std. Error"]^2
## One-step Debiasing with the PSPS Method
fit_PSPS <- PSPS(est_lab_y, est_lab_yhat, est_unlab_yhat, Sigma)
print(fit_PSPS)
### Example output from PSPS function:
# Estimate Std.Error Lower.CI Upper.CI P.value
# 0.7537844 0.09366147 0.5702113 0.9373575 8.418062e-16pip install psps_pyPSPS inputs summary statistics from three separate analyses and returns the ML-assisted estimator.
fit_PSPS = PSPS(est_lab_y, est_lab_yhat, est_unlab_yhat, Sigma)Suppose we are interested in estimating a K-dimensional parameter, then
est_lab_y: a K-dimensional vector of Point estimates usingYin labeled data.est_lab_yhat: a K-dimensional vector of Point estimates usingYhatin labeled data.est_unlab_yhat: a K-dimensional vector of Point estimates usingYhatin unlabeled data.Sigma: a 3K x 3K Variance-covariance matrix for the above three estimators (Note: not the asymptotic variance).
Here is an example of PSPS for logistic regression. For other tasks, simply replace logistic regression with other algorithms to produce summary statistics.
import pandas as pd
import numpy as np
import statsmodels.api as sm
from statsmodels.formula.api import glm
# Load the package
from psps_py import PSPS
# Load data
lab = pd.read_csv("./PSPS/test_data/lab.csv")
unlab = pd.read_csv("./PSPS/test_data/unlab.csv")# Fit logistic regression models and extract the second coefficient
est_lab_y = glm('Y ~ X', data=lab, family=sm.families.Binomial()).fit().params['X']
est_lab_yhat = glm('Yhat ~ X', data=lab, family=sm.families.Binomial()).fit().params['X']
est_unlab_yhat = glm('Yhat ~ X', data=unlab, family=sm.families.Binomial()).fit().params['X']
# Bootstrap for covariance calculation
B = 1000 # Number of bootstrap iterations
n = len(lab) # Total observations in labeled data
np.random.seed(123) # For reproducibility
est_lab_y_boot = np.zeros(B)
est_lab_yhat_boot = np.zeros(B)
for i in range(B):
boot_indices = np.random.choice(lab.index, size=n, replace=True)
boot_lab = lab.loc[boot_indices]
est_lab_y_boot[i] = glm('Y ~ X', data=boot_lab, family=sm.families.Binomial()).fit().params['X']
est_lab_yhat_boot[i] = glm('Yhat ~ X', data=boot_lab, family=sm.families.Binomial()).fit().params['X']
# Compute the covariance matrix
Sigma = np.zeros((3, 3))
Sigma[0:2, 0:2] = np.cov(np.vstack([est_lab_y_boot, est_lab_yhat_boot]))
Sigma[2, 2] = glm('Yhat ~ X', data=unlab, family=sm.families.Binomial()).fit().bse['X']**2
# Run PSPS
fit_PSPS = PSPS(est_lab_y, est_lab_yhat, est_unlab_yhat, Sigma)
print(fit_PSPS)
### Example output from PSPS function:
# Estimate Std.Error Lower.CI Upper.CI P.value
# 0.755438 0.090828 0.577417 0.933458 9.007197e-17We provide the script for analysis in the PSPS paper here.
Please submit an issue or contact Jiacheng ([email protected]) or Qiongshi ([email protected]) for questions.
Task-Agnostic Machine Learning-Assisted Inference
Assumption-Lean and Data-Adaptive Post-Prediction Inference
Valid Inference for Machine Learning-Assisted GWAS
- POP-TOOLS (POst-Prediction TOOLS) is a toolkit for conducting valid and powerful machine learning (ML)-assisted genetic association studies. It currently implements
POP-GWAS, where statistical and computational methods are optimized for GWAS applications.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent