ssp.quantreg:
Subsampling for Quantile RegressionThis vignette introduces the usage of ssp.quantreg. The
statistical theory and algorithms behind this implementation can be
found in the relevant reference papers.
Quantile regression aims to estimate conditional quantiles by minimizing the following loss function:
\[ \min_{\beta} L(\beta) = \frac{1}{N} \sum_{i=1}^{N} \rho_\tau \left( y_i - \beta^\top x_i \right) = \frac{1}{N} \sum_{i=1}^{N} \left( y_i - \beta^\top x_i \right) \left\{ \tau - I \left( y_i < \beta^\top x_i \right) \right\}, \] where \(\tau\) is the quantile of interest, \(y\) is the response variable, \(x\) is covariates vector and \(N\) is the number of observations in full dataset.
The idea of subsampling methods is as follows: instead of fitting the model on the size \(N\) full dataset, a subsampling probability is assigned to each observation and a smaller, informative subsample is drawn. The model is then fitted on the subsample to obtain an estimator with reduced computational cost.
Full dataset: The whole dataset used as input.
Full data estimator: The estimator obtained by fitting the model on the full dataset.
Subsample: A subset of observations drawn from the full dataset.
Subsample estimator: The estimator obtained by fitting the model on the subsample.
Subsampling probability (\(\pi\)): The probability assigned to each observation for inclusion in the subsample.
We introduce ssp.quantreg with simulated data. \(X\) contains \(d=6\) covariates drawn from multinormal
distribution and \(Y\) is the response
variable. The full data size is \(N = 1 \times
10^4\). The interested quantile \(\tau=0.75\).
set.seed(1)
N <- 1e4
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 ~ .
head(data)
#> Y V1 V2 V3 V4 V5
#> 1 3.0813580 -0.1825325 -0.01613791 -0.01852406 1.0672454 0.9353870
#> 2 0.1114953 -0.3829652 -1.20674035 -0.33354934 0.3818526 0.6610612
#> 3 4.0233475 -0.1384141 0.35758454 -0.08962728 0.8591475 0.7554356
#> 4 -7.0116774 -0.7668158 -1.07028901 -2.57374497 -1.4283868 -0.4782146
#> 5 -1.2551700 -0.9557206 -0.82219260 0.47905721 0.1096016 -0.3116279
#> 6 2.6764218 0.8646208 -0.32527175 0.23441106 0.5800169 1.8153229
#> V6
#> 1 0.44382164
#> 2 0.12626628
#> 3 1.63208199
#> 4 1.10717085
#> 5 -0.08180055
#> 6 -0.03612645The function usage is
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,
...
)The core functionality of ssp.quantreg revolves around
three key questions:
How are subsampling probabilities computed? (Controlled by the
criterion argument)
How is the subsample drawn? (Controlled by the
sampling.method argument)
How is the likelihood adjusted to correct for bias? (Controlled
by the likelihood argument)
criterioncriterion stands for the criterion we choose to compute
the sampling probability for each observation. The choices of
criterion include optL(default) and
uniform. In optL, the optimal subsampling
probability is by minimizing a transformation of the asymptotic variance
of subsample estimator. uniform is a baseline method.
sampling.methodThe options for the sampling.method argument include
withReplacement (default) and poisson.
withReplacement stands for drawing \(n.ssp\) subsamples from full dataset with
replacement, using the specified subsampling probabilities.
poisson stands for drawing subsamples one by one by
comparing the subsampling probability with a realization of uniform
random variable \(U(0,1)\). The
expected number of drawn samples are \(n.ssp\).
likelihoodThe available choice for likelihood in
ssp.quantreg is weighted. It takes the inverse
of sampling probabblity as the weights in likelihood function to correct
the bias introduced by unequal subsampling probabilities.
boot and BAn option for drawing \(B\)
subsamples (each with expected size n.ssp) and deriving
subsample estimator and asymptotic covariance matrix based on them.
After getting \(\hat{\beta}_{b}\) on
the \(b\)-th subsample, \(b=1,\dots B\), it calculates
\[ \hat{\beta}_I = \frac{1}{B} \sum_{b=1}^{B} \hat{\beta}_{b} \] as the final subsample estimator and \[ \hat{V}(\hat{\beta}_I) = \frac{1}{r_{ef} B (B - 1)} \sum_{b=1}^{B} \left( \hat{\beta}_{b} - \hat{\beta}_I \right)^{\otimes 2}, \] where \(r_{ef}\) is a correction term for effective subsample size since the observations in each subsample can be replicated. For more details, see Wang and Ma (2021).
After drawing subsample(s), ssp.quantreg utilizes
quantreg::rq to fit the model on the subsample(s).
Arguments accepted by quantreg::rq can be passed through
... in ssp.quantreg.
Below are two examples demonstrating the use of
ssp.quantreg with different configurations.
B <- 5
n.plt <- 200
n.ssp <- 200
ssp.results1 <- 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'
)
ssp.results2 <- ssp.quantreg(formula,
data,
tau = tau,
n.plt = n.plt,
n.ssp = n.ssp,
B = B,
boot = FALSE,
criterion = 'optL',
sampling.method = 'withReplacement',
likelihood = 'weighted'
)The returned object contains estimation results and index of drawn subsample in the full dataset.
names(ssp.results1)
#> [1] "model.call" "coef.plt" "coef"
#> [4] "cov" "index.plt" "index.ssp"
#> [7] "N" "subsample.size.expect" "terms"summary(ssp.results1)
#> Model Summary
#>
#>
#> Call:
#>
#> ssp.quantreg(formula = formula, data = data, tau = tau, n.plt = n.plt,
#> n.ssp = n.ssp, B = B, boot = TRUE, criterion = "optL", sampling.method = "withReplacement",
#> likelihood = "weighted")
#>
#> Subsample Size:
#> [1] 1000
#>
#> Coefficients:
#>
#> Estimate Std. Error z value Pr(>|z|)
#> Intercept 0.9753 0.0324 30.0654 <0.0001
#> V1 0.9701 0.0220 44.1763 <0.0001
#> V2 1.0295 0.0394 26.1369 <0.0001
#> V3 0.9980 0.0209 47.8506 <0.0001
#> V4 0.9834 0.0609 16.1529 <0.0001
#> V5 1.0508 0.0301 34.8848 <0.0001
#> V6 0.9441 0.0327 28.8878 <0.0001summary(ssp.results2)
#> Model Summary
#>
#>
#> Call:
#>
#> ssp.quantreg(formula = formula, data = data, tau = tau, n.plt = n.plt,
#> n.ssp = n.ssp, B = B, boot = FALSE, criterion = "optL", sampling.method = "withReplacement",
#> likelihood = "weighted")
#>
#> Subsample Size:
#> [1] 1000
#>
#> Coefficients:
#>
#> Estimate
#> Intercept 0.9510
#> V1 1.0234
#> V2 1.0562
#> V3 1.0221
#> V4 0.9676
#> V5 0.9698
#> V6 1.0445Some key returned variables:
index.plt and index are the row indices
of drawn pilot subsamples and optimal subsamples in the full
data.
coef.ssp is the subsample estimator for \(\beta\) and coef is the linear
combination of coef.plt (pilot estimator) and
coef.ssp.
cov.ssp and cov are estimated
covariance matrices of coef.ssp and coef. If
boot=FALSE, covariance matrix would not be estimated and a
size n.ssp * B subsample would be drawn.
See the help documentation of ssp.quantreg for
details.