Provides a toolkit for analytical variance estimation in survey sampling. Apart from the implementation of standard variance estimators, its main feature is to help the sampling expert produce easy-to-use variance estimation "wrappers", where systematic operations (linearization, domain estimation) are handled in a consistent and transparent way.
Gustave (Gustave: a User-oriented Statistical Toolkit for Analytical Variance Estimation) is an R package that provides a toolkit for analytical variance estimation in survey sampling.
Apart from the implementation of standard variance estimators (Sen-Yates-Grundy, Deville-TillĂ©), its main feature is to help he methodologist produce easy-to-use variance estimation wrappers, where systematic operations (statistic linearization, domain estimation) are handled in a consistent and transparent way.
The ready-to-use variance estimation wrapper qvar()
, adapted for common cases (e.g. stratified simple random sampling, non-response correction through reweighting in homogeneous response groups, calibration), is also included. The core functions of the package (e.g. define_variance_wrapper()
) are to be used for more complex cases.
gustave is available on CRAN and can therefore be installed with the install.packages()
function:
install.packages("gustave")
However, if you wish to install the latest version of gustave, you can use devtools::install_github()
to install it directly from the github.com repository:
install.packages("devtools")
devtools::install_github("martinchevalier/gustave")
In this example, we aim at estimating the variance of estimators computed using simulated data inspired from the Information and communication technology (ICT) survey. This survey has the following characteristics:
The ICT simulated data files are shipped with the gustave package:
library(gustave)
data(package = "gustave")
? ict_survey
A variance estimation can be perform in a single call of qvar()
:
qvar(
# Sample file
data = ict_sample,
# Dissemination and identification information
dissemination_dummy = "dissemination",
dissemination_weight = "w_calib",
id = "firm_id",
# Scope
scope_dummy = "scope",
# Sampling design
sampling_weight = "w_sample",
strata = "strata",
# Non-response correction
nrc_weight = "w_nrc",
response_dummy = "resp",
hrg = "hrg",
# Calibration
calibration_weight = "w_calib",
calibration_var = c(paste0("N_", 58:63), paste0("turnover_", 58:63)),
# Statistic(s) and variable(s) of interest
mean(employees)
)
The survey methodology description is however cumbersome when several variance estimations are to be conducted. As it does not change from one estimation to another, it could be defined once and for all and then re-used for all variance estimations. qvar()
allows for this by defining a so-called variance wrapper, that is an easy-to-use function where the variance estimation methodology for the given survey is implemented and all the technical data used to do so included.
# Definition of the variance estimation wrapper precision_ict
precision_ict <- qvar(
# As before
data = ict_sample,
dissemination_dummy = "dissemination",
dissemination_weight = "w_calib",
id = "firm_id",
scope_dummy = "scope",
sampling_weight = "w_sample",
strata = "strata",
nrc_weight = "w_nrc",
response_dummy = "resp",
hrg = "hrg",
calibration_weight = "w_calib",
calibration_var = c(paste0("N_", 58:63), paste0("turnover_", 58:63)),
# Replacing the variables of interest by define = TRUE
define = TRUE
)
# Use of the variance estimation wrapper
precision_ict(ict_sample, mean(employees))
# The variance estimation wrapper can also be used on the survey file
precision_ict(ict_survey, mean(speed_quanti))
The variance estimation wrapper is much easier-to-use than a standard variance estimation function:
several statistics in one call (with optional labels):
precision_ict(ict_survey,
"Mean internet speed in Mbps" = mean(speed_quanti),
"Turnover per employee" = ratio(turnover, employees)
)
domain estimation with where and by arguments
precision_ict(ict_survey,
mean(speed_quanti),
where = employees >= 50
)
precision_ict(ict_survey,
mean(speed_quanti),
by = division
)
# Domain may differ from one estimator to another
precision_ict(ict_survey,
"Mean turnover, firms with 50 employees or more" = mean(turnover, where = employees >= 50),
"Mean turnover, firms with 100 employees or more" = mean(turnover, where = employees >= 100)
)
handy variable evaluation
# On-the-fly evaluation (e.g. discretization)
precision_ict(ict_survey, mean(speed_quanti > 100))
# Automatic discretization for qualitative (character or factor) variables
precision_ict(ict_survey, mean(speed_quali))
# Standard evaluation capabilities
variables_of_interest <- c("speed_quanti", "speed_quali")
precision_ict(ict_survey, mean(variables_of_interest))
Integration with %>% and dplyr
library(dplyr)
ict_survey %>%
precision_ict("Internet speed above 100 Mbps" = mean(speed_quanti > 100)) %>%
select(label, est, lower, upper)
This software is an R package developed with the RStudio IDE and the devtools, roxygen2 and testthat packages. Much help was found in R packages and Advanced R both written by Hadley Wickham.
From the methodological point of view, this package is related to the Poulpe SAS macro (in French) developed at the French statistical institute. From the implementation point of view, some inspiration was found in the ggplot2 package. The idea of developing an R package on this specific topic was stimulated by the icarus package and its author.
Breaking: Heavy remanufacturing of define_variance_wrapper
technical_data
argument offers a more consistent way to include technical data within the enclosing environment of the wrapper. objects_to_include
is kept for non-data objects (such as additional statistic wrappers) or advanced customization.technical_param
argument offers a more convenient way to specify default values for parameters used by the variance function.reference_weight
replaces default$weight
. This means that the reference weight used for point estimation and linearization is set while defining the variance wrapper and not at run-time.stat
, which was a remain of an early implementation of linearization functions, is not a parameter of the variance wrappers anymore. Its purpose (to apply a given variance wrapper to several variables without having to type the name of the linearization wrapper) is now covered by the standard evaluation capabilities of statistic wrappers (see below).default
is replaced by default_id
, as default$weight
and default$stat
are no longer needed. As for default$alpha
, its value is set to 0.05 and cannot be changed anymore while defining the variance wrapper (as this can easily be done afterwards using formals<-
).objects_to_include_from
Breaking: Rebranding and heavy remanufacturing of define_statistic_wrapper
(previously known as define_linearization_wrapper
), added support for standard evaluation (see define_variance_wrapper
examples).
New: the qvar
function allows for a straigthforward variance estimation in common cases (stratified simple random sampling with non-response through reweighting and calibration) and performs both technical and methodological checks.
Some normalization in function names: add0
becomes add_zero
, sumby
becomes sum_by
, rescal
becomes res_cal
Example data: calibration variables in ict_sample instead of ict_survey, new LFS example data
Significant increase of unit tests
define_variance_wrapper
example.varDT