Skip to contents

Control Parameters for Inference

Source: R/control_inference.R
control_inf.Rd

control_inf constructs a list with all necessary control parameters for statistical inference.

Usage

control_inf(
  var_method = c("analytic", "bootstrap"),
  rep_type = c("subbootstrap", "auto", "JK1", "JKn", "BRR", "bootstrap", "mrbbootstrap",
    "Fay"),
  vars_selection = FALSE,
  vars_combine = FALSE,
  bias_correction = FALSE,
  num_boot = 100,
  alpha = 0.05,
  cores = 1,
  keep_boot = TRUE,
  nn_exact_se = FALSE
)

Arguments

var_method

the variance method (default "analytic"). Note that the doubly robust analytic variance is derived under the logistic propensity model (Chen, Li & Wu 2020, Theorem 2); for the "probit" and "cloglog" selection links it is a conservative approximation (it tends to over-estimate the standard error and can be numerically unstable when fitted propensities approach 1), so var_method = "bootstrap" is recommended for doubly robust inference with those links.

rep_type

the replication type for weights in the bootstrap method for variance estimation passed to survey::as.svrepdesign(). Default is "subbootstrap".

vars_selection

logical scalar (default FALSE); if TRUE, then the variables selection model is used.

vars_combine

logical scalar indicating whether variables should be combined after variable selection for doubly robust estimators (default FALSE)

bias_correction

logical scalar (default FALSE); if TRUE, the doubly-robust mean is estimated by jointly solving the Yang-Kim-Song (2020) eq. (9) system in (theta, beta). When vars_selection = FALSE this gives the low-dimensional Kim & Haziza (2014) joint estimator; when vars_selection = TRUE and vars_combine = TRUE it gives the high-dimensional two-step Yang-Kim-Song estimator. A svydesign argument is required (joint estimation needs individual-level probability sample data).

num_boot

the number of iteration for bootstrap algorithms.

alpha

significance level (default 0.05).

cores

the number of cores in parallel computing (default 1).

keep_boot

a logical value indicating whether statistics from bootstrap should be kept (default TRUE)

nn_exact_se

a logical value indicating whether to compute the exact standard error estimate for nn or pmm estimator. The variance estimator for estimation based on nn or pmm can be decomposed into three parts, with the third computed using covariance between imputed values for units in the probability sample using predictive matches from the non-probability sample. In most situations this term is negligible and is very computationally expensive so by default it is set to FALSE, but the recommended option is to set this value to TRUE before submitting the final results.

Value

A list with selected parameters.

See also

nonprob() – for fitting procedure with non-probability samples.