bayesgmed_sens {BayesGmed}

Conduct sensitivity analysis for unmeasured confounder

Description

'bayesgmed_sens' is used to conduct sensitivity analysis for unmeasured confounders in mediation analysis.

Usage

bayesgmed_sens(
  outcome,
  mediator,
  treat,
  covariates = NULL,
  dist.y = "continuous",
  dist.m = "continuous",
  link.y = "identity",
  link.m = "identity",
  data,
  priors = NULL,
  ...
)

Arguments

outcome

a character string indicating the name of the outcome variable.

mediator

a character string indicating the name of the mediator variable.

treat

a character string indicating the name of the treatment variable. The treatment variable is considered binary and should be coded as 0 for control and 1 for treated.

covariates

a character vector indicating the name of the confounding variables.

dist.y

a character string indicating the family distribution of the outcome. E.g., dist.y = "bernoulli" will fit a logistic regression the outcome.

dist.m

a character string indicating the family distribution of the mediator E.g., dist.m = "bernoulli" will fit a logistic regression the mediator

link.y

a character string indicating the link function to be used for the outcome model.

link.m

a character string indicating the link function to be used for the mediator model.

data

A data.frame or a tibble object.

priors

A list of named values for the prior scale parameters. See details.

...

Other optional parameters passed to rstan::stan().

Details

Perform a sensitivity analysis for unmeasured confounding following the approach proposed by McCandless LC and Somers JM (2019). This is done by incorporating uncertainty about unmeasured confounding in the outcome and mediator model through a prior distribution. One can control the size of unmeasured confounding by varying the location and scale parameters of the prior distribution values for the bias parameter (i.e., gamma). See the Vignette for more details.

priors

Users may pass a list of named values for the priors argument. The values will be used to define the scale parameter of the respective prior distributions. This list may specify some or all of the following parameters: priors <- list( scale_m = 2.5diag(P_m) scale_y = 2.5diag(P_y), llocation_m = rep(0, P_m) location_y = rep(0, P_y), location_gamma = rep(0, 4), scale_gamma = 0.1*diag(4), scale_sd_y = 2.5, scale_sd_m = 2.5) where P_m is the number of regression parameters (including the intercept) in the mediator model and P_y is the number of regression parameters in the outcome model. Note that there are 4 bias parameters i.e., gamma).

Value

An object of 'S4' class 'stanfit', with all its available methods.

Author(s)

Belay Birlie Yimer [e-mail]

References

1. McCandless, L.C. and J.M. Somers, Bayesian sensitivity analysis for unmeasured confounding in causal mediation analysis. Statistical Methods in Medical Research, 2019. (28)(2): p. 515-531.

2. Comment, L., Coull, B. A., Zigler, C., and Valeri, L. (2019). Bayesian data fusion for unmeasured confounding. arXiv preprint arXiv:1902.10613.

Examples

## Run example using the example_data
data(example_data)
# priors including for unmeasured single confounder 
P <- 3 # number of covariates plus the intercept term 
priors <- list(scale_m = 2.5*diag(P+1), scale_y = 2.5*diag(P+2), location_m = rep(0, P+1), 
location_y = rep(0, P+2), location_gamma = rep(0,4), scale_gamma = 0.5*diag(4))
fit1 <- bayesgmed_sens(outcome = "Y", mediator =  "M", treat = "A", covariates = c("Z1", "Z2"), 
dist.y = "binary", dist.m = "binary", link.y = "logit", link.m = "logit", priors = priors,  
data = example_data)
bayesgmed_summary(fit1)