Optimum Sample Allocation in Stratified Sampling with stratallo

The goal of stratallo package is to provide implementations of the efficient algorithms that solve a classical problem in survey methodology - an optimum sample allocation in stratified sampling. In this context, the classical problem of optimum sample allocation is the Tschuprow-Neyman’s sense (Neyman 1934; Tschuprow 1923). It is formulated as determination of a vector of strata sample sizes that minimizes the variance of the stratified \(\pi\) estimator of the population total of a given study variable, under constraint on total sample size. More specifically, the algorithms provided in this package are valid given that the variance of the stratified estimator is of the following generic form:

\[ V_{st}(\mathbf n) = \sum_{h=1}^{H} \frac{A_h^2}{n_h} - A_0, \] where \(H\) denotes total number of strata, \(\mathbf n= (n_h)_{h \in \{1,\ldots,H\}}\) is the allocation vector with strata sample sizes, and population parameters \(A_0,\, A_h > 0,\, h = 1,\ldots,H\), do not depend on the \(x_h,\, h = 1,\ldots,H\). The allocation problem mentioned, can be further complemented by imposing lower or upper bounds on sample sizes is strata.

Among stratified estimators and stratified sampling designs that jointly give rise to a variance of the above form, is the so called stratified \(\pi\) estimator of the population total with stratified simple random sampling without replacement design, which is one of the most basic and commonly used stratified sampling designs. This case yields \(A_0 = \sum_{h = 1}^H N_h S_h^2\), \(A_h = N_h S_h,\, h = 1,\ldots,H\), where \(S_h\) denotes stratum standard deviation of study variable and \(N_h\) is the stratum size (see e.g. Särndal et al. (1993), Result 3.7.2, p.103).

A minor modification of the classical optimum sample allocation problem leads to the minimum cost allocation. This problem lies in the determination of a vector of strata sample sizes that minimizes total cost of the survey, under assumed fixed level of the stratified \(\pi\) estimator’s variance. As in the case of the classical optimum allocation, the problem of minimum cost allocation can be complemented by imposing upper bounds on sample sizes in strata.

Package stratallo provides two user functions:

that solve sample allocation problems briefly characterized above as well as the following helpers functions:

Functions var_st() and var_st_tsi() compute a value of the variance \(V_{st}\). The var_st_tsi() is a simple wrapper of var_st() that is dedicated for the case when \(A_0 = \sum_{h = 1}^H N_h S_h^2\) and \(A_h = N_h S_h,\, h = 1,\ldots,H\). asummary() creates a data.frame object with summary of the allocation. Functions ran_round() and round_oric() are the rounding functions that can be used to round non-integers allocations (see section Rounding, below). The package comes with three predefined, artificial populations with 10, 507 and 969 strata. These are stored under pop10_mM, pop507 and pop969 objects, respectively.

Minimization of the variance with opt() function

The opt() function solves the following three problems of the optimum sample allocation, formulated in the language of mathematical optimization. User of opt() can choose whether the solution computed will be for Problem 1, Problem 2 or Problem 3. This is achieved with the proper use of m and M arguments of the function. Also, if required, the inequality constraints can be removed from the optimization problem. For more details, see the help page for opt() function.

Problem 1 (one-sided upper bounds)

Given numbers \(n > 0,\, A_h > 0,\, M_h > 0\), such that \(M_h \leq N_h,\, h = 1,\ldots,H\), and \(n \leq \sum_{h=1}^H M_h\), \[\begin{align*} \underset{\mathbf x\in {\mathbb R}_+^H}{\mathrm{minimize ~\,}} & \quad f(\mathbf x) = \sum_{h=1}^H \tfrac{A_h^2}{x_h} \\ \mathrm{subject ~ to} & \quad \sum_{h=1}^H x_h = n \\ & \quad x_h \leq M_h, \quad{h = 1,\ldots,H,} \end{align*}\] where \(\mathbf x= (x_h)_{h \in \{1,\ldots,H\}}\).

There are four different algorithms available to use for Problem 1, RNA (default), SGA, SGAPLUS, COMA. All these algorithms, except SGAPLUS, are described in detail in Wesołowski et al. (2021). The SGAPLUS is defined in Wójciak (2019) as Sequential Allocation (version 1) algorithm.

Examples

library(stratallo)

Define example population.

N <- c(3000, 4000, 5000, 2000) # Strata sizes.
S <- c(48, 79, 76, 16) # Standard deviations of a study variable in strata.
A <- N * S
n <- 190 # Total sample size.

Tschuprow-Neyman allocation (no inequality constraints).

xopt <- opt(n = n, A = A)
xopt
#> [1] 31.376147 68.853211 82.798165  6.972477
sum(xopt) == n
#> [1] TRUE
# Variance of the st. estimator that corresponds to the optimum allocation.
var_st_tsi(xopt, N, S)
#> [1] 3940753053

One-sided upper bounds.

M <- c(100, 90, 70, 80) # Upper bounds imposed on the sample sizes in strata.
all(M <= N)
#> [1] TRUE
n <= sum(M)
#> [1] TRUE

# Solution to Problem 1.
xopt <- opt(n = n, A = A, M = M)
xopt
#> [1] 35.121951 77.073171 70.000000  7.804878
sum(xopt) == n
#> [1] TRUE
all(xopt <= M) # Does not violate upper-bounds constraints.
#> [1] TRUE
# Variance of the st. estimator that corresponds to the optimum allocation.
var_st_tsi(xopt, N, S)
#> [1] 4018789143

Problem 2 (one-sided lower bounds)

Given numbers \(n,\, A_h > 0,\, m_h > 0\), such that \(m_h \leq N_h,\, h = 1,\ldots,H\), and \(n \geq \sum_{h=1}^H m_h\), \[\begin{align*} \underset{\mathbf x\in {\mathbb R}_+^H}{\mathrm{minimize ~\,}} & \quad f(\mathbf x) = \sum_{h=1}^H \tfrac{A_h^2}{x_h} \\ \mathrm{subject ~ to} & \quad \sum_{h=1}^H x_h = n \\ & \quad x_h \geq m_h, \quad{h = 1,\ldots,H,} \end{align*}\] where \(\mathbf x= (x_h)_{h \in \{1,\ldots,H\}}\).

The optimization Problem 2 is solved by the LRNA that in principle is based on the RNA and it is introduced in Wójciak (2023).

Examples

m <- c(50, 120, 1, 2) # Lower bounds imposed on the sample sizes in strata.
n >= sum(m)
#> [1] TRUE

# Solution to Problem 2.
xopt <- opt(n = n, A = A, m = m)
xopt
#> [1]  50 120  18   2
sum(xopt) == n
#> [1] TRUE
all(xopt >= m) # Does not violate lower-bounds constraints.
#> [1] TRUE
# Variance of the st. estimator that corresponds to the optimum allocation.
var_st_tsi(xopt, N, S)
#> [1] 9719807556

Problem 3 (box constraints)

Given numbers \(n,\, A_h > 0,\, m_h > 0,\, M_h > 0\), such that \(m_h < M_h \leq N_h,\, h = 1,\ldots,H\), and \(\sum_{h=1}^H m_h \leq n \leq \sum_{h=1}^H M_h\), \[\begin{align*} \underset{\mathbf x\in {\mathbb R}_+^H}{\mathrm{minimize ~\,}} & \quad f(\mathbf x) = \sum_{h=1}^H \tfrac{A_h^2}{x_h} \\ \mathrm{subject ~ to} & \quad \sum_{h=1}^H x_h = n \\ & \quad m_h \leq x_h \leq M_h, \quad{h = 1,\ldots,H,} \end{align*}\] where \(\mathbf x= (x_h)_{h \in \{1,\ldots,H\}}\).

The optimization Problem 3 is solved by the RNABOX which is a new algorithm proposed by the authors of this package and it will be published soon.

Examples

m <- c(100, 90, 500, 50) # Lower bounds imposed on sample sizes in strata.
M <- c(300, 400, 800, 90) # Upper bounds imposed on sample sizes in strata.
n <- 1284
n >= sum(m) && n <= sum(M)
#> [1] TRUE

# Optimum allocation under box constraints.
xopt <- opt(n = n, A = A, m = m, M = M)
xopt
#> [1] 228.9496 400.0000 604.1727  50.8777
sum(xopt) == n
#> [1] TRUE
all(xopt >= m & xopt <= M) # Does not violate any lower or upper bounds constraints.
#> [1] TRUE
# Variance of the st. estimator that corresponds to the optimum allocation.
var_st_tsi(xopt, N, S)
#> [1] 538073357

Minimization of the total cost with optcost() function

The optcost() function solves the following minimum total cost allocation problem, formulated in the language of mathematical optimization.

Problem 4

Given numbers \(A_h > 0,\, c_h > 0,\, M_h > 0\), such that \(M_h \leq N_h,\, h = 1,\ldots,H\), and \(V \geq \sum_{h=1}^H \tfrac{A_h^2}{M_h} - A_0\), \[\begin{align*} \underset{\mathbf x\in {\mathbb R}_+^H}{\mathrm{minimize ~\,}} & \quad c(\mathbf x) = \sum_{h=1}^H c_h x_h \\ \mathrm{subject ~ to} & \quad \sum_{h=1}^H \tfrac{A_h^2}{x_h} - A_0 = V \\ & \quad x_h \leq M_h, \quad{h = 1,\ldots,H,} \end{align*}\] where \(\mathbf x= (x_h)_{h \in \{1,\ldots,H\}}\).

The algorithm that solves Problem 4 is based on the LRNA and it is described in Wójciak (2023).

Examples

A <- c(3000, 4000, 5000, 2000)
A0 <- 70000
unit_costs <- c(0.5, 0.6, 0.6, 0.3) # c_h, h = 1,...4.
M <- c(100, 90, 70, 80)
V <- 1e6 # Variance constraint.
V >= sum(A^2 / M) - A0
#> [1] TRUE

xopt <- optcost(V = V, A = A, A0 = A0, M = M, unit_costs = unit_costs)
xopt
#> [1] 40.39682 49.16944 61.46181 34.76805
sum(A^2 / xopt) - A0 == V
#> [1] TRUE
all(xopt <= M)
#> [1] TRUE

Rounding

stratallo comes with 2 functions: ran_round() and round_oric() that can be used to round non-integer allocations.

Examples

m <- c(100, 90, 500, 50)
M <- c(300, 400, 800, 90)
n <- 1284

# Optimum, non-integer allocation under box constraints.
xopt <- opt(n = n, A = A, m = m, M = M)
xopt
#> [1] 297.4286 396.5714 500.0000  90.0000

xopt_int <- round_oric(xopt)
xopt_int
#> [1] 297 397 500  90

Installation

You can install the released version of stratallo package from CRAN with:

install.packages("stratallo")

Note on finite precision arithmetic

Consider the following example

N <- c(3000, 4000, 5000, 2000)
S <- c(48, 79, 76, 17)
a <- N * S
n <- 190
xopt <- opt(n = n, A = A) # which after simplification is (n / sum(a)) * a
xopt
#> [1] 40.71429 54.28571 67.85714 27.14286

and note that

sum(xopt) == n
#> [1] TRUE

which results from the fact that

options(digits = 22)
sum(xopt)
#> [1] 190

sum((n / sum(A)) * A) == n # mathematically, it should be TRUE!
#> [1] TRUE

References

Neyman, J. (1934), “On the Two Different Aspects of the Representative Method: The Method of Stratified Sampling and the Method of Purposive Selection,” Journal of the Royal Statistical Society, 97, 558–606.
Särndal, C.-E., Swensson, B., and Wretman, J. (1993), Model Assisted Survey Sampling, Springer.
Tschuprow, A. A. (1923), “On the Mathematical Expectation of the Moments of Frequency Distributions in the Case of Correlated Observations,” Metron, 2, 461–493, 636–680.
Wesołowski, J., Wieczorkowski, R., and Wójciak, W. (2021), “Optimality of the Recursive Neyman Allocation,” Journal of Survey Statistics and Methodology. https://doi.org/10.1093/jssam/smab018. https://arxiv.org/abs/2105.14486.
Wójciak, W. (2019), “Optimal Allocation in Stratified Sampling Schemes,” MSc Thesis, Warsaw University of Technology. http://home.elka.pw.edu.pl/~wwojciak/msc_optimal_allocation.pdf.
Wójciak, W. (2023), “Another Solution of Some Optimum Allocation Problem,” Statistics in Transition new series. https://arxiv.org/abs/2204.04035.

mirror server hosted at Truenetwork, Russian Federation.