Added a kernel-based variance estimator (of Breidt, Opsomer, and Sanchez-Borrego 2016) for finely stratified or systematic samples. This can be used by calling as_fays_gen_rep_design(..., variance_estimator = "BOSB") or as_gen_boot_design(..., variance_estimator = "BOSB"). Or you can directly create the quadratic form matrix for this estimator using the function make_kernel_var_matrix(). Currently only supports a single auxiliary variable and only the Epanechnikov kernel function.
Added the option variance_estimator = "Beaumont-Emond" to the generalized replication methods.
For Fay's generalized replication, the default value of max_replicates is now Inf, meaning that there is no limit on the number of replicates that will be included in the output.
Functions for Fay's generalized replication method are now noticeably faster.
Added more detailed documentation to the Rao-Wu-Yue-Beaumont bootstrap method.
Added new options to enable users to speed up the creation of replicate weights if they have the 'torch' package installed. For more details, see help('svrep-package-options', package = 'svrep').
Fix issue #32: stack_replicate_designs() would only accept designs with types known to the 'survey' package. Fixed to allow other design types such as the random-groups jackknife.
Added new helper function shuffle_replicates() to randomize the order of the columns of replicate weights. This is useful, for example, when replicates are created independently in each stratum and then combined.
Added new helper function subsample_replicates() to retain a random subset of replicates and accordingly increase the scale factor used for variance estimation.
Added new helper function add_inactive_replicates() which allows the user to add "inactive" replicates to a design object, so that the matrix of replicate weights has the desired number of columns.
The function as_fays_gen_rep_design() now has a default value of mse = TRUE. Setting mse = FALSE will produce a warning message, since Fay's generalized replication method can sometimes produce large underestimates of variance when mse = FALSE.
The function as_random_group_jackknife_design() now returns an object with class tbl_svy if the input was also an object with class tbl_svy.
Generalized replication based on the Yates-Grundy estimator is now much faster due to code refactoring.
Bug fixes:
Changes specifically for CRAN check:
Removed 12 unmarked UTF-8 strings causing a CRAN check note.
Removed the LaTeX 'cases' formatting from the documentation for as_random_group_jackknife_design(), since an old release on MacOS was throwing a LaTeX error when trying to build the manual. The formatting might be restored later when 'oldrel' on CRAN increases to 4.3.X.
Added support for Fay's generalized replication method, specifically the version proposed in Fay (1989): the key functions are as_fays_gen_rep_design() and make_fays_gen_rep_factors(), which are nearly identical to the generalized bootstrap functions as_gen_boot_design() and make_gen_boot_factors().
Added a new variance estimator, "Deville-Tille", useful for balanced sampling (including the cube method). Currently only works for single-stage designs.
as_gen_boot_design() and as_fays_gen_rep_design() have a new argument aux_var_names meant to be used for the "Deville-Tille" variance estimator. Similarly, make_gen_boot_factors() and make_fays_gen_rep_factors() have an argument named aux_vars.Added a function as_random_group_jackknife_design() to create random-group jackknife replicates.
The creation of generalized bootstrap replicates for designs with many observations but few degrees of freedom (e.g., stratified cluster samples) is now much faster and more efficient. This is based on using the 'Matrix' package--particularly its efficient representation of sparse matrices which arise for stratified designs--as well as using a compressed representation of designs that use cluster sampling.
Now using the 'Matrix' package to improve speed and memory usage for large quadratic forms. This is primarily helpful for making the generalized bootstrap computationally feasible for larger datasets.
Better documentation for the bootstrap methods covered by as_bootstrap_design().
The following functions now work for database-backed survey design objects (i.e., objects with the class DBIsvydesign):
as_data_frame_with_weights()as_gen_boot_design()as_bootstrap_design()redistribute_weights()calibrate_to_sample()calibrate_to_estimate()The function as_data_frame_with_weights() has gained an argument vars_to_keep which allows the user to indicate that they only want to keep a specific list of variables from the data. This can be useful, for example, if you only want to keep the weights and unique identifiers.
Minor updates and bug fixes:
The function as_bootstrap_design() now throws an informative error message when you supply an invalid value for the type argument.
Bug Fix: The "Deville-1" and "Deville-2" estimators threw errors for strata where one or more units were selected with certainty (i.e., had sampling probabilities of 1). This has now been fixed.
Bug Fix: The function as_gen_boot_design() could sometimes fail to detect that the input design is a PPS design, which caused it to give the user an unnecessary error message.
Added argument exact_vcov = TRUE to as_gen_boot_design() and make_gen_boot_factors(). This argument forces the generalized bootstrap variance-covariance estimates for totals to exactly match the target variance estimator. In other words, this eliminates bootstrap simulation error for variance estimates of totals. This is similar to how, for simple survey designs, the jackknife and BRR give variance estimates for totals that exactly match the Horvitz-Thompson estimates. Using exact_vcov requires that the number of replicates is strictly greater than the rank of the target variance estimator.
Added new variance estimators ("Deville 1" and "Deville 2") available to use for the generalized bootstrap, which are particularly useful for single-stage PPSWOR designs or for multistage designs with one or more stages of PPSWOR sampling. See updated documentation for as_gen_boot_design() and make_quad_form_matrix().
If the 'srvyr' package is loaded, then functions from 'svrep' that return survey design objects will always return a tbl_svy if their input was a tbl_svy. This makes it easier to use functions such as summarize() or mutate().
as_bootstrap_design() wouldn't create more than 50 replicates for the Rao-Wu, Preston, or Canty-Davison types.This release adds extensive new functionality for two-phase designs. The new vignette "Replication Methods for Two-phase Sampling" describes the new functionality as well as the underlying statistical methods.
The function as_gen_boot_design() can now create generalized bootstrap weights for two-phase survey design objects created with the 'survey' package's twophase() function. The user must specify a list of two variance estimators to use for each phase, e.g. list('Stratified Multistage SRS', 'Ultimate Cluster').
The function make_twophase_quad_form() can be used to create a quadratic form for a two-phase variance estimator, by combining quadratic forms from each phase.
The helper function get_nearest_psd_matrix() can be used to approximate a quadratic form matrix by the nearest positive semidefinite matrix. This can be particularly useful for two-phase designs, since the double expansion estimator commonly used in practice frequently does not have a variance estimator which is positive semidefinite.
The function as_gen_boot_design() has a new argument named psd_option, which controls what will happen if the target variance estimator has a quadratic form matrix which is not positive semi-definite. This can occasionally happen, particularly for two-phase designs. By default, the function will warn the user if the quadratic form is not positive semi-definite and then automatically approximate the matrix by the nearest positive semi-definite matrix.
Added a new function get_design_quad_form(), which determines the quadratic form matrix of a specified variance estimator, by parsing the information stored in a survey design object created using the 'survey' package.
Added a new function rescale_reps() which implements the rescaling of replicate adjustment factors to avoid negative replicate weights. This functionality already existed in as_gen_boot_design() and make_gen_boot_factors(), but now it is implemented with the help of this new function.
Added helper function is_psd_matrix() for checking whether a matrix is positive semi-definite, and added a helper function get_nearest_psd_matrix() for approximating a square matrix by the nearest positive semi-definite matrix.
Minor improvements to vignettes, particularly formatting.
Fix bug in #15, where bootstrap conversion of multistage survey design objects with as_bootstrap_design() would throw an error when user manually specified weights in svydesign().
Creation of Rao-Wu-Yue-Beaumont bootstrap replicate weights is now faster and takes less computer memory.
Typo fix in vignettes.
This release adds several functions for creating bootstrap and generalized bootstrap replicate weights. The new vignette "Bootstrap methods for surveys" provides guidance for choosing a bootstrap method and selecting the number of bootstrap replicates to use, along with statistical details and references.
Added function as_bootstrap_design() to convert a survey design object to a replicate design with replicate weights created using a bootstrap method. This is essentially a specialized version of as.svrepdesign() that supports additional bootstrap methods and has detailed documentation about which bootstrap methods can be used for different types of sampling designs.
Added function as_gen_boot_design() to convert a survey design object to a replicate design with replicate weights created using the generalized survey bootstrap. The user must supply the name of a target variance estimator (e.g., "Horvitz-Thompson" or "Ultimate Cluster") used to create the generalized bootstrap factors. See the new vignette for details.
Added functions to help choose the number of bootstrap replicates. The function estimate_boot_sim_cv() can be used to estimate the simulation error in a bootstrap estimate caused by using a finite number of bootstrap replicates. The new function estimate_boot_reps_for_target_cv() estimates the number of bootstrap replicates needed to reduce the simulation error to a target level.
Added function make_rwyb_bootstrap_weights(), which creates bootstrap replicate weights for a wide range of survey designs using the method of Rao-Wu-Yue-Beaumont (i.e., Beaumont's generalization of the Rao-Wu-Yue bootstrap method). This function can be used directly, or users can specify as_bootstrap_design(type = "Rao-Wu-Yue-Beaumont").
Added function make_gen_boot_factors() to create replicate adjustment factors using the generalized survey bootstrap. The key input to make_gen_boot_factors() is the matrix of the quadratic form used to represent a variance estimator. The new function make_quad_form_matrix() can be used to represent a chosen variance estimator as a quadratic form, given information about the sample design. This can be used for stratified multistage SRS designs (with or without replacement), systematic samples, and PPS samples, with or without replacement.
Minor Updates and Bug Fixes:
as_data_frame_with_weights(), ensure that the full-sample weight is named "FULL_SAMPLE_WGT" if the user does not specify something different.calibrate_to_estimate(), ensure that the output names the list of columns with perturbed control columns col_selection instead of perturbed_control_cols, so that the name matches the corresponding function argument, col_selection.Added helper function as_data_frame_with_weights() to convert a survey design object into a data frame with columns of weights (full-sample weights and, if applicable, replicate weights). This is useful for saving data and weights to a data file.
Added by argument to summarize_rep_weights() which allows the specification of one or more grouping variables to use for summaries (e.g. by = c('stratum', 'response_status') can be used to summarize by response status within each stratum).
Added a small vignette "Nonresponse Adjustments" to illustrate how to conduct nonresponse adjustments using redistribute_weights().
Minor Updates and Bug Fixes:
rho in calibrate_to_estimate().stack_replicate_designs() where designs created with as.svrepdesign(..., type = 'mrbbootstrap') or as.svrepdesign(..., type = 'subbootstrap') threw an error.Added functions calibrate_to_estimate() and calibrate_to_sample() for calibrating to estimated control totals with methods that account for the sampling variance of the control totals. For an overview of these functions, please see the new vignette "Calibrating to Estimated Control Totals".
The function calibrate_to_estimate() requires the user to supply a vector of control totals and its variance-covariance matrix. The function applies Fuller's proposed adjustments to the replicate weights, in which control totals are varied across replicates by perturbing the control totals using a spectral decomposition of the control totals' variance-covariance matrix.
The function calibrate_to_sample() requires the user to supply a replicate design for the primary survey of interest as well as a replicate design for the control survey used to estimate control totals for calibration. The function applies Opsomer & Erciulescu's method of varying the control totals across replicates of the primary survey by matching each primary survey replicate to a replicate from the control survey.
Added an example dataset, lou_vax_survey, which is a simulated survey measuring Covid-19 vaccination status and a handful of demographic variables, based on a simple random sample of 1,000 residents of Louisville, Kentucky with an approximately 50% response rate.
lou_pums_microdata provides person-level microdata from the American Community Survey (ACS) 2015-2019 public-use microdata sample (PUMS) data for Louisville, KY. The dataset lou_pums_microdata includes replicate weights to use for variance estimation and can be used to generate control totals for lou_vax_survey.