bayeslist: Fitting Bayesian sensitive item models

Description

The main function for estimating Bayesian sensitive item models. The function returns a bayeslist object that can be further investigated using standard functions such as summary, plot, print, predict, and coef. The model can be passed using a formula as in lm(). Convergence diagnotics can be performed using either print(object, "mcmc") or plot(object, "trace").

Usage

bayeslist(
  formula,
  data,
  treat,
  J,
  type = "outcome",
  nsim = 1000,
  burnin = NULL,
  thin = 1,
  CIsize = 0.95,
  nchain = 1,
  seeds = 12345,
  vb = FALSE,
  only_vb = FALSE,
  prior = NULL,
  direct_item = NULL,
  direct_item_misreport = NULL,
  double_list = NULL,
  double_list_treat = NULL,
  aux_info = NULL,
  aux_g = NULL,
  aux_h = NULL,
  BL_a = NULL,
  BL_b = NULL,
  conjugate_distance = FALSE,
  conjugate_k = NULL,
  predictvar = NULL,
  predictvar_type = "binary",
  parallel = TRUE,
  robust = FALSE
)

Value

A bayeslist object. An object of class bayeslist contains the following elements

Call: The matched call.
formula: Symbolic representation of the model.
type: Model type
nsim: Number of iterations.
Burnin: Number of burnin iterations.
thin: Thinning.
seeds: Random seeds for reproducibility. The default is 12345.
CIsize: Size of the posterior confidence interval.
data: Data used.
X: Independent variables.
Y: Dependent variables.
xnames: Names of the independent variables.
stanfit: Output from stan.
sampledf: Posterior samples.
summaryout: Summary of the stan-fit object.
npars: Number of control variables.
only_vb: Whether only viariational approximation is used.
prior: Informative prior types.
direct_item: Direct item.
double_list: The second list.
aux_info: Auxiliary information.
ulbs: Upper and lower bounds based on the specified confidence interval.
means: Mean estimates.
treat: Treatment.
outcome: Outcome to be predicted.
direct: Direct item for the misreport model.
robust: Robust indicator.

Arguments

formula: An object of class "formula" (or one that can be coerced to that class): a symbolic description of the model to be fitted.
data: A data frame containing the variables in the model.
treat: Variable name of the treatment.
J: Number of control items.
type: Type of the model. Options include "outcome", "predict", "misreport", for the sensitive item outcome model, predictor model and misreport model, respectively.
nsim: The number of iterations.
burnin: The number of burnin iterations.
thin: Thinning parameter.
CIsize: The size of posterior confidence interval.
nchain: The number of parallel chains.
seeds: Random seeds to replicate the results.
vb: Logic. If TRUE, variational approximation will be used to supply initial values. The default is FALSE.
only_vb: Logic. If TRUE, only variational approximation will be calculated. The default is FALSE.
prior: Prior types. Options include "auxiliary", "double_list", "direct_item", and "BL" for beta-logistic prior. If NULL, no informative priors will be used.
direct_item: Variable name of the direct item.
direct_item_misreport: Variable name of the direct item for the misreporting model.
double_list: Variable name of the second list.
double_list_treat: Treatment variable of the second list.
aux_info: Auxiliary information for the informative priors. list(G,h,g), where: G (number of subgroups), h (auxiliary information for each subgroup), and g (subgroup indicator). If is.NULL, the following two parameters need to be specified when estimating the model with prior = "auxiliary".
aux_g: Auxiliary information for the informative priors: name of the variable indicating the group of each observation.
aux_h: Auxiliary information for the informative priors: name of the variable containing information of prevalence for each group
BL_a: The first shape hyperparameter for the beta-logistic prior, indicating the prior number of affirmative answers to the sensitive item.
BL_b: The second shape hyperparameter for the beta-logistic prior, indicating the prior number of non-affirmative answers to the sensitive item.
conjugate_distance: Logic. Indicating whether conjugate distance prior should be used. The default is FALSE.
conjugate_k: Degrees of freedom to be scaled by conjugate distance prior. The default is NULL.
predictvar: Variable name of the outcome to be predicted.
predictvar_type: The type of the outcome variable to be predicted. Options include "linear" and "binary". The default is "binary".
parallel: Logic. Indicating whether to do paralell computing. The default is TRUE.
robust: Logic. Indicating whether to impose robust constraints on the intercept-only model. The default is FALSE.

References

Lu, X. and Traunmüller, R. (2021). Improving Studies of Sensitive Topics Using Prior Evidence: A Unified Bayesian Framework for List Experiments, SSRN, tools:::Rd_expr_doi("10.2139/ssrn.3871089").

Examples

Run this code

# Estimate sensitive item outcome model using Sri Lanka data on male sexual violence
# Load Sri Lanka list experiment data
data(srilanka)

# Model 1: intercept-only outcome model without prior information:
mod1 <- bayeslist(sexaussault ~ 1, data = srilanka, treat = "treatment", J = 3,
type = "outcome", nsim = 200, thin = 1, CIsize = 0.95, nchain = 1,
seeds = 342321, prior = NULL, parallel = TRUE)
summary(mod1) # summary of estimates
predict(mod1) # predicted prevalence for each observation
plot(mod1,"trace") # trace plot
plot(mod1,"coef") # coefficient plot
plot(mod1, only_prev = TRUE) # prevalence plot

# \donttest{
# Model 2: multivariate outcome model without prior information:
mod2 <- bayeslist(sexaussault ~ age + edu, data = srilanka, treat = "treatment", J = 3,
type = "outcome", nsim = 200, thin = 1, CIsize = 0.95, nchain = 1,
seeds = 342321, prior = NULL, parallel = TRUE)
summary(mod2) # summary of estimates
predict(mod2) # predicted prevalence for each observation
plot(mod2,"trace") # trace plot
plot(mod2,"coef") # coefficient plot
plot(mod2) # prevalence + coefficient plot

# Model 3: intercept-only outcome model with prior information from medicolegal reports, i.e.,
# with a prior beta-logistic distribution BL(38, 146).
a <- 38; b <-146
mod3 <- bayeslist(sexaussault ~ 1, data = srilanka, treat = "treatment", J = 3,
type = "outcome", nsim = 200, thin = 1, CIsize = 0.95, nchain = 1,
seeds = 342321, prior = "BL", BL_a = a, BL_b = b,, parallel = TRUE)
summary(mod3)
predict(mod3)
plot(mod3,"trace")
plot(mod3,"coef")
plot(mod3, only_prev = TRUE)

# Model 4: multivariate outcome model with prior information from a direct item.
# Load London list experiment data
data(london)
mod4 <- bayeslist(listCount ~ agegrp + gender + social_grade + qual,data = london, J = 4,
treat = "listTreat", seeds = 4597, nsim = 200, nchain = 1,
prior = "direct_item", direct_item = "baselineTurnout")
summary(mod4)
predict(mod4)
plot(mod4,"trace")
plot(mod4,"coef")
plot(mod4)
# }

Run the code above in your browser using DataLab