Optimal Subsampling Methods for Statistical Models

Description

Subsampling methods are utilized in statistical modeling for massive datasets. These methods aim to draw representative subsamples from the full dataset based on specific sampling probabilities, with the goal of maintaining inference efficiency. The sampling probabilities are tailored to particular objectives, such as minimizing the variance of the estimated coefficients or reducing prediction error. By using subsampling techniques, the package balances the trade-off between computational efficiency and statistical efficiency, making it a practical tool for massive data analysis.

Models Supported

• Generalized Linear Models (GLMs)

• Softmax (Multinomial) Regression

• Rare Event Logistic Regression

• Quantile Regression

Author(s)

Maintainer: Qingkai Dong qingkai.dong@uconn.edu [copyright holder]

Authors:

• Yaqiong Yao

• Haiying Wang

Other contributors:

• Qiang Zhang [contributor]

• Jun Yan [contributor]

See Also

Useful links:

• https://github.com/dqksnow/Subsampling

• Report bugs at https://github.com/dqksnow/Subsampling/issues

Optimal Subsampling Methods for Generalized Linear Models

Description

Draw subsample from full dataset and fit a generalized linear model (GLM) on the subsample. For a quick start, refer to the vignette.

Usage

ssp.glm(
  formula,
  data,
  subset = NULL,
  n.plt,
  n.ssp,
  family = "binomial",
  criterion = "optL",
  sampling.method = "poisson",
  likelihood = "weighted",
  control = list(...),
  contrasts = NULL,
  ...
)

Arguments

Details

A pilot estimator for the unknown parameter \beta is required because both optA and optL subsampling probabilities depend on \beta. There is no "free lunch" when determining optimal subsampling probabilities. Fortunately the pilot estimator only needs to satisfy mild conditions. For logistic regression, this is achieved by drawing a size n.plt subsample with replacement from full dataset. The case-control subsample probability is applied, that is, \pi_i = \frac{1}{2N_1} for Y_i=1 and \pi_i = \frac{1}{2N_0} for Y_i=0, i=1,...,N, whereN_0 and N_1 are the counts of observations with Y = 0 and Y = 1, respectively. For other families, uniform subsampling probabilities are applied. Typically, n.plt is relatively small compared to n.ssp.

When criterion = 'uniform', there is no need to compute the pilot estimator. In this case, a size n.plt + n.ssp subsample will be drawn with uniform sampling probability and coef is the corresponding estimator.

As suggested by survey::svyglm(), for binomial and poisson families, use family=quasibinomial() and family=quasipoisson() to avoid a warning "In eval(family$initialize) : non-integer #successes in a binomial glm!". The quasi versions of the family objects give the same point estimates and suppress the warning. Since subsampling methods only rely on point estimates from svyglm() for further computation, using the quasi families does not introduce any issues.

For Gamma family, ssp.glm returns only the estimation of coefficients, as the dispersion parameter is not estimated.

Value

ssp.glm returns an object of class "ssp.glm" containing the following components (some are optional):

model.call

The original function call.

coef.plt

The pilot estimator. See Details for more information.

coef.ssp

The estimator obtained from the optimal subsample.

coef

The weighted linear combination of coef.plt and coef.ssp. The combination weights depend on the relative size of n.plt and n.ssp and the estimated covariance matrices of coef.plt and coef.ssp. We blend the pilot subsample information into optimal subsample estimator since the pilot subsample has already been drawn. The coefficients and standard errors reported by summary are coef and the square root of diag(cov).

cov.ssp

The covariance matrix of coef.ssp.

cov

The covariance matrix of coef.

index.plt

Row indices of pilot subsample in the full dataset.

index.ssp

Row indices of of optimal subsample in the full dataset.

N

The number of observations in the full dataset.

subsample.size.expect

The expected subsample size, equals to n.ssp for ssp.glm. Note that for other functions, such as ssp.relogit, this value may differ.

terms

The terms object for the fitted model.

References

Wang, H. (2019). More efficient estimation for logistic regression with optimal subsamples. Journal of machine learning research, 20(132), 1-59.

Ai, M., Yu, J., Zhang, H., & Wang, H. (2021). Optimal subsampling algorithms for big data regressions. Statistica Sinica, 31(2), 749-772.

Wang, H., & Kim, J. K. (2022). Maximum sampled conditional likelihood for informative subsampling. Journal of machine learning research, 23(332), 1-50.

Examples

# logistic regression
set.seed(2)
N <- 1e4
beta0 <- rep(-0.5, 7)
d <- length(beta0) - 1
corr <- 0.5
sigmax  <- matrix(corr, d, d) + diag(1-corr, d)
X <- MASS::mvrnorm(N, rep(0, d), sigmax)
Y <- rbinom(N, 1, 1 - 1 / (1 + exp(beta0[1] + X %*% beta0[-1])))
data <- as.data.frame(cbind(Y, X))
formula <- Y ~ .
n.plt <- 500
n.ssp <- 1000
subsampling.results <- ssp.glm(formula = formula, 
data = data, 
n.plt = n.plt,
n.ssp = n.ssp,
family = 'quasibinomial',
criterion = "optL",
sampling.method = 'poisson',
likelihood = "logOddsCorrection")
summary(subsampling.results)
subsampling.results <- ssp.glm(formula = formula, 
data = data, 
n.plt = n.plt,
n.ssp = n.ssp,
family = 'quasibinomial', 
criterion = "optL",
sampling.method = 'withReplacement', 
likelihood = "weighted")
summary(subsampling.results)
Uni.subsampling.results <- ssp.glm(formula = formula, 
data = data, 
n.plt = n.plt,
n.ssp = n.ssp,
family = 'quasibinomial', 
criterion = 'uniform')
summary(Uni.subsampling.results)
####################
# poisson regression
set.seed(1)
N <-  1e4
beta0 <- rep(0.5, 7)
d <- length(beta0) - 1
X <- matrix(runif(N * d), N, d)
epsilon <- runif(N)
lambda <- exp(beta0[1] + X %*% beta0[-1])
Y <- rpois(N, lambda)
data <- as.data.frame(cbind(Y, X))
formula <- Y ~ .
n.plt <- 200
n.ssp <- 600
subsampling.results <- ssp.glm(formula = formula, 
data = data, 
n.plt = n.plt,
n.ssp = n.ssp,
family = 'poisson',
criterion = "optL", 
sampling.method = 'poisson',
likelihood = "weighted")
summary(subsampling.results)
subsampling.results <- ssp.glm(formula = formula, 
data = data, 
n.plt = n.plt,
n.ssp = n.ssp,
family = 'poisson', 
criterion = "optL", 
sampling.method = 'withReplacement',
likelihood = "weighted")
summary(subsampling.results)
Uni.subsampling.results <- ssp.glm(formula = formula, 
data = data, 
n.plt = n.plt,
n.ssp = n.ssp,
family = 'poisson', 
criterion = 'uniform')
summary(Uni.subsampling.results)
##################
# gamma regression
set.seed(1)
N <- 1e4
p <- 3
beta0 <- rep(0.5, p + 1)
d <- length(beta0) - 1
shape <- 2
X <- matrix(runif(N * d), N, d)
link_function <- function(X, beta0) 1 / (beta0[1] + X %*% beta0[-1])
scale <- link_function(X, beta0) / shape
Y <- rgamma(N, shape = shape, scale = scale)
data <- as.data.frame(cbind(Y, X))
formula <- Y ~ .
n.plt <- 200
n.ssp <- 1000
subsampling.results <- ssp.glm(formula = formula, 
data = data, 
n.plt = n.plt,
n.ssp = n.ssp,
family = 'Gamma',
criterion = "optL", 
sampling.method = 'poisson',
likelihood = "weighted")
summary(subsampling.results)

Optimal Subsampling Methods for Quantile Regression Model

Description

Draw subsample from full dataset and fit quantile regression model. For a quick start, refer to the vignette.

Usage

ssp.quantreg(
  formula,
  data,
  subset = NULL,
  tau = 0.5,
  n.plt,
  n.ssp,
  B = 5,
  boot = TRUE,
  criterion = "optL",
  sampling.method = "withReplacement",
  likelihood = c("weighted"),
  control = list(...),
  contrasts = NULL,
  ...
)

Arguments

Details

Most of the arguments and returned variables have the same meaning with ssp.glm. Refer to vignette

A pilot estimator for the unknown parameter \beta is required because optL subsampling probabilities depend on \beta. There is no "free lunch" when determining optimal subsampling probabilities. For quantile regression, this is achieved by drawing a size n.plt subsample with replacement from full dataset, using uniform sampling probability.

If boot=TRUE, the returned value subsample.size.expect equals to B*n.ssp, and the covariance matrix for coef would be calculated. If boot=FALSE, the returned value subsample.size.expect equals to B*n.ssp, but the covariance matrix won't be estimated.

Value

ssp.quantreg returns an object of class "ssp.quantreg" containing the following components (some are optional):

model.call

The original function call.

coef.plt

The pilot estimator. See Details for more information.

coef

The estimator obtained from the optimal subsample.

cov

The covariance matrix of coef

index.plt

Row indices of pilot subsample in the full dataset.

index.ssp

Row indices of of optimal subsample in the full dataset.

N

The number of observations in the full dataset.

subsample.size.expect

The expected subsample size

terms

The terms object for the fitted model.

References

Wang, H., & Ma, Y. (2021). Optimal subsampling for quantile regression in big data. Biometrika, 108(1), 99-112.

Examples

#quantile regression
set.seed(1)
N <- 1e4
B <- 5
tau <- 0.75
beta.true <- rep(1, 7)
d <- length(beta.true) - 1
corr  <- 0.5
sigmax  <- matrix(0, d, d)
for (i in 1:d) for (j in 1:d) sigmax[i, j] <- corr^(abs(i-j))
X <- MASS::mvrnorm(N, rep(0, d), sigmax)
err <- rnorm(N, 0, 1) - qnorm(tau)
Y <- beta.true[1] + X %*% beta.true[-1] + 
err * rowMeans(abs(X))
data <- as.data.frame(cbind(Y, X))
colnames(data) <- c("Y", paste("V", 1:ncol(X), sep=""))
formula <- Y ~ .
n.plt <- 200
n.ssp <- 100
optL.results <- ssp.quantreg(formula,data,tau = tau,n.plt = n.plt,
n.ssp = n.ssp,B = B,boot = TRUE,criterion = 'optL',
sampling.method = 'withReplacement',likelihood = 'weighted')
summary(optL.results)
uni.results <- ssp.quantreg(formula,data,tau = tau,n.plt = n.plt,
n.ssp = n.ssp,B = B,boot = TRUE,criterion = 'uniform',
sampling.method = 'withReplacement', likelihood = 'weighted')
summary(uni.results)

Optimal Subsampling for Logistic Regression Model with Rare Events Data

Description

Draw subsample from full dataset and fit logistic regression model on subsample. For a quick start, refer to the vignette.

Usage

ssp.relogit(
  formula,
  data,
  subset = NULL,
  n.plt,
  n.ssp,
  criterion = "optL",
  likelihood = "logOddsCorrection",
  control = list(...),
  contrasts = NULL,
  ...
)

Arguments

Details

'Rare event' stands for the number of observations where Y=1 is rare compare to the number of Y=0 in the full data. In the face of logistic regression with rare events, @wang2021nonuniform shows that the available information ties to the number of positive instances instead of the full data size. Based on this insight, one can keep all the rare instances and perform subsampling on the non-rare instances to reduce the computational cost. When criterion = optA, optL or LCC, all observations with Y=1 are preserved and it draw n.ssp subsmples from observations with Y=0. When criterion = uniform, it draws (n.plt+n.ssp) subsmples from the full sample with equal sampling probability.

A pilot estimator for the unknown parameter \beta is required because both optA and optL subsampling probabilities depend on \beta. This is achieved by drawing half size subsample from rare observations and half from non-rare observations.

Most of the arguments and returned variables have similar meaning with ssp.glm. Refer to vignette

Value

ssp.relogit returns an object of class "ssp.relogit" containing the following components (some are optional):

model.call

The original function call.

coef.plt

The pilot estimator. See Details for more information.

coef.ssp

The estimator obtained from the optimal subsample.

coef

The weighted linear combination of coef.plt and coef.ssp. The combination weights depend on the relative size of n.plt and n.ssp and the estimated covariance matrices of coef.plt and coef.ssp. We blend the pilot subsample information into optimal subsample estimator since the pilot subsample has already been drawn. The coefficients and standard errors reported by summary are coef and the square root of diag(cov).

cov.ssp

The covariance matrix of coef.ssp.

cov

The covariance matrix of beta.cmb.

index.plt

Row indices of pilot subsample in the full dataset.

index.ssp

Row indices of of optimal subsample in the full dataset.

N

The number of observations in the full dataset.

subsample.size.expect

The expected subsample size.

terms

The terms object for the fitted model.

References

Wang, H., Zhang, A., & Wang, C. (2021). Nonuniform negative sampling and log odds correction with rare events data. Advances in Neural Information Processing Systems, 34, 19847-19859.

Examples

set.seed(1)
N <- 2 * 1e4
beta0 <- c(-5, -rep(0.7, 6))
d <- length(beta0) - 1
X <- matrix(0, N, d)
corr <- 0.5
sigmax <- corr ^ abs(outer(1:d, 1:d, "-"))
sigmax <- sigmax / 4
X <- MASS::mvrnorm(n = N, mu = rep(0, d), Sigma = sigmax)
Y <- rbinom(N, 1, 1 - 1 / (1 + exp(beta0[1] + X %*% beta0[-1])))
print(paste('N: ', N))
print(paste('sum(Y): ', sum(Y)))
n.plt <- 200
n.ssp <- 1000
data <- as.data.frame(cbind(Y, X))
colnames(data) <- c("Y", paste("V", 1:ncol(X), sep=""))
formula <- Y ~ .
subsampling.results <- ssp.relogit(formula = formula,
                                     data = data,
                                     n.plt = n.plt,
                                     n.ssp = n.ssp,
                                     criterion = 'optA',
                                     likelihood = 'logOddsCorrection')
summary(subsampling.results)

Optimal Subsampling Method for Softmax (multinomial logistic) Regression Model

Description

Draw subsample from full dataset and fit softmax(multinomial logistic) regression model on the subsample. Refer to vignette for a quick start.

Usage

ssp.softmax(
  formula,
  data,
  subset,
  n.plt,
  n.ssp,
  criterion = "MSPE",
  sampling.method = "poisson",
  likelihood = "MSCLE",
  constraint = "summation",
  control = list(...),
  contrasts = NULL,
  ...
)

Arguments

Details

A pilot estimator for the unknown parameter \beta is required because MSPE, optA and optL subsampling probabilities depend on \beta. There is no "free lunch" when determining optimal subsampling probabilities. For softmax regression, this is achieved by drawing a size n.plt subsample with replacement from full dataset with uniform sampling probability.

Value

ssp.softmax returns an object of class "ssp.softmax" containing the following components (some are optional):

model.call

The original function call.

coef.plt

The pilot estimator. See Details for more information.

coef.ssp

The estimator obtained from the optimal subsample.

coef

The weighted linear combination of coef.plt and coef.ssp, under baseline constraint. The combination weights depend on the relative size of n.plt and n.ssp and the estimated covariance matrices of coef.plt and coef.ssp. We blend the pilot subsample information into optimal subsample estimator since the pilot subsample has already been drawn. The coefficients and standard errors reported by summary are coef and the square root of diag(cov).

coef.plt.sum

The pilot estimator under summation constrraint. coef.plt.sum = G %*% as.vector(coef.plt).

coef.ssp.sum

The estimator obtained from the optimal subsample under summation constrraint. coef.ssp.sum = G %*% as.vector(coef.ssp).

coef.sum

The weighted linear combination of coef.plt and coef.ssp, under summation constrraint. coef.sum = G %*% as.vector(coef).

cov.plt

The covariance matrix of coef.plt.

cov.ssp

The covariance matrix of coef.ssp.

cov

The covariance matrix of coef.cmb.

cov.plt.sum

The covariance matrix of coef.plt.sum.

cov.ssp.sum

The covariance matrix of coef.ssp.sum.

cov.sum

The covariance matrix of coef.sum.

index.plt

Row indices of pilot subsample in the full dataset.

index.ssp

Row indices of of optimal subsample in the full dataset.

N

The number of observations in the full dataset.

subsample.size.expect

The expected subsample size.

terms

The terms object for the fitted model.

References

Yao, Y., & Wang, H. (2019). Optimal subsampling for softmax regression. Statistical Papers, 60, 585-599.

Han, L., Tan, K. M., Yang, T., & Zhang, T. (2020). Local uncertainty sampling for large-scale multiclass logistic regression. Annals of Statistics, 48(3), 1770-1788.

Wang, H., & Kim, J. K. (2022). Maximum sampled conditional likelihood for informative subsampling. Journal of machine learning research, 23(332), 1-50.

Yao, Y., Zou, J., & Wang, H. (2023). Optimal poisson subsampling for softmax regression. Journal of Systems Science and Complexity, 36(4), 1609-1625.

Yao, Y., Zou, J., & Wang, H. (2023). Model constraints independent optimal subsampling probabilities for softmax regression. Journal of Statistical Planning and Inference, 225, 188-201.

Examples

# softmax regression
d <- 3 # dim of covariates
K <- 2 # K + 1 classes
G <- rbind(rep(-1/(K+1), K), diag(K) - 1/(K+1)) %x% diag(d)
N <- 1e4
beta.true.baseline <- cbind(rep(0, d), matrix(-1.5, d, K))
beta.true.summation <- cbind(rep(1, d), 0.5 * matrix(-1, d, K))
set.seed(1)
mu <- rep(0, d)
sigma <- matrix(0.5, nrow = d, ncol = d)
diag(sigma) <- rep(1, d)
X <- MASS::mvrnorm(N, mu, sigma)
prob <- exp(X %*% beta.true.summation)
prob <- prob / rowSums(prob)
Y <- apply(prob, 1, function(row) sample(0:K, size = 1, prob = row))
n.plt <- 500
n.ssp <- 1000
data <- as.data.frame(cbind(Y, X))
colnames(data) <- c("Y", paste("V", 1:ncol(X), sep=""))
head(data)
formula <- Y ~ . -1
WithRep.MSPE <- ssp.softmax(formula = formula,
 data = data, 
 n.plt = n.plt,
 n.ssp = n.ssp,
 criterion = 'MSPE', 
 sampling.method = 'withReplacement',
 likelihood = 'weighted',
 constraint = 'baseline')
summary(WithRep.MSPE)