Provides tools for designing and analyzing Acceptance Sampling plans. Supports both Attributes Sampling (Binomial and Poisson distributions) and Variables Sampling (Normal and Beta distributions), enabling quality control for fractional and compositional data. Uses nonlinear programming for sampling plan optimization, minimizing sample size while controlling producer's and consumer's risks. Operating Characteristic curves are available for plan visualization.
An R package for designing and analyzing acceptance sampling plans
π¦ Now available on
CRAN! π β
See NEWS.md for release notes and user-visible changes.
The package is described in Truong, Miranda, and Kissling (2026), AccSamplingDesign: An R Package for Optimizing Acceptance Sampling Plans, published in The R Journal.
The AccSamplingDesign package provides flexible tools to create and evaluate acceptance sampling plans in quality control, for both attributes (pass/fail) and variables (measurable) data. It supports optimization using nonlinear programming (NLP) to meet specified risks while minimizing the required sample size.
alpha and
beta constraints# Install from CRAN
install.packages("AccSamplingDesign")
# Or install development version from GitHub
devtools::install_github("vietha/AccSamplingDesign")
# Load the package
library(AccSamplingDesign)
plan_attr <- optPlan(
PRQ = 0.01, # Acceptable quality
CRQ = 0.05, # Rejectable quality
alpha = 0.02, # Producer's risk
beta = 0.15, # Consumer's risk
distribution = "binomial"
)
summary(plan_attr)
accProb(plan_attr, 0.03) # P(accept) if 3% defective
plot(plan_attr) # OC curve
plan_var <- optPlan(
PRQ = 0.025,
CRQ = 0.1,
alpha = 0.05,
beta = 0.10,
distribution = "normal",
sigma_type = "known"
)
summary(plan_var)
plot(plan_var)
plan_var2 <- optPlan(
PRQ = 0.025,
CRQ = 0.1,
alpha = 0.05,
beta = 0.10,
distribution = "normal",
sigma_type = "unknown"
)
summary(plan_var2)
plan_beta <- optPlan(
PRQ = 0.05,
CRQ = 0.2,
alpha = 0.05,
beta = 0.10,
distribution = "beta",
theta = 44000000,
theta_type = "known",
LSL = 0.00001 # Lower Specification Limit
)
summary(plan_beta)
plot(plan_beta) # By defect level
plot(plan_beta, by = "mean") # By mean value
Unknown-theta plans support analytical Delta--MLE ("delta_mle"), analytical
Delta--MoM ("delta_mom"), and the earlier Govindaraju--Kissling sample-size
adjustment ("gk_adjustment"). Version 0.1.1 also provides the experimental
Saddlepoint MLE OC ("saddlepoint_mle"). Delta--MLE is used when method is omitted.
saddlepoint_mle retains the MLE inspection rule and numerically integrates a
normalized joint saddlepoint density. It uses no simulation during optimization.
It has a substantially larger initial computation cost than the Delta methods;
repeated calculations over k reuse the joint distribution. Its probabilities
are bounded and monotone in k, but its finite-sample risk accuracy must still
be assessed. Nonconvergent integrations raise an error, without a Delta fallback.
New saddlepoint plans store their delivered integer size in both n and
sample_size, and expose numerical diagnostics in integration. Existing Delta
reporting conventions are preserved.
plan_sp <- optPlan(
PRQ = 0.01, CRQ = 0.03, alpha = 0.05, beta = 0.10,
distribution = "beta", theta_type = "unknown", theta = 500,
USL = 0.05, method = "saddlepoint_mle"
)
summary(plan_sp)
accProb(plan_sp, 0.03)
The construction uses the saddlepoint framework in McCullagh, Tensor Methods in Statistics, Chapter 6. It does not establish universal finite-sample risk control or a global sample-size optimum. Numerical convergence diagnostics measure integration error, not statistical approximation error.
plan_beta <- optPlan(
PRQ = 0.05,
CRQ = 0.2,
alpha = 0.05,
beta = 0.10,
distribution = "beta",
theta = 44000000,
theta_type = "unknown",
method = "delta_mle", # Default; may be omitted
LSL = 0.00001
)
summary(plan_beta)
plot(plan_beta) # By defect level
plot(plan_beta, by = "mean") # By mean value
pd <- seq(0, 0.15, by = 0.001)
oc_opt <- OCdata(plan = plan_attr, pd = pd)
mplan1 <- manualPlan(n = plan_attr$n, c = plan_attr$c - 1, distribution = "binomial")
oc_alt1 <- OCdata(plan = mplan1, pd = pd)
plot(pd, oc_opt$paccept, type = "l", col = "blue", lwd = 2,
xlab = "Proportion Defective", ylab = "Probability of Acceptance",
main = "OC Curves Comparison for Attributes Sampling Plan")
lines(pd, oc_alt1$paccept, col = "red", lwd = 2, lty = 2)
legend("topright", legend = c("Optimal Plan", "Manual Plan (c - 1)"),
col = c("blue", "red"), lty = c(1, 2), lwd = 2)
This README provides a quick start for using the AccSamplingDesign package. For a full discussion of the statistical foundations, models, and optimization methods used, please refer to the foundation sources such as:
Contributions, suggestions, and bug reports are welcome!
Please use GitHub
Issues or submit a
pull request.
method = "saddlepoint_mom" evaluates the unchanged MOM inspection rule through
its joint raw-moment distribution. It enforces the finite-sample moment support
and uses deterministic, positive quadrature. The numerical and finite-sample
validation is incomplete: some ordinary calculations fail convergence, and an
unsupported plan or probability raises an error. This method is an experimental
interface, not a validated replacement for either Delta MOM or saddlepoint MLE.
delta_mle remains the default. There is no simulation-fitted correction.
plan <- manualPlan(distribution = "beta", theta_type = "unknown",
theta = 500, USL = 0.05, n = 121, k = 2.333335566,
method = "saddlepoint_mom")
# This development case is unsupported; retain the error as a diagnostic.
tryCatch(accProb(plan, 0.03), error = function(e) conditionMessage(e))
An integer-only experimental plan search is available through optPlan().
Numerical support is required at every search point; a failed evaluation stops
that search. Numerical convergence alone does not guarantee finite-sample risks.
The integrated MOM interface also rejects tilts outside a sufficient log-concavity regime for its numerical integration kernel. This conservative support restriction can reject a mathematically valid Beta probability; it does not discard nodes and renormalize the remaining mass.