Publication Bias Adjustment — publication

Fits publication-bias models using estimates from a fitted mars object. Currently implements the beta selection model described by Citkowicz and Vevea (2017), where selection weights are modeled as a beta kernel over p-values. Also supports a beta-binomial selection model over equal-width p-value bins.

Usage

publication_bias(
  mars_object,
  method = c("beta_selection", "beta_binomial", "logistic_selection"),
  p_value_tail = c("two.sided", "right", "left"),
  n_bins = 10,
  p_value_min = NULL,
  maxit = 2000,
  integration_points = 121,
  integration_limit = 8,
  funnel_plot = FALSE,
  funnel_args = list(),
  control = list(),
  ...
)

Arguments

mars_object: A fitted object returned by mars.
method: Publication-bias method. One of "beta_selection" or "beta_binomial" or "logistic_selection".
p_value_tail: Which p-value tail to use for the selection function. One of "two.sided", "right", or "left".
n_bins: Number of p-value bins used when method = "beta_binomial". Ignored for "beta_selection". For "beta_binomial", p-values are assigned to equal-width bins \([j / n\_bins, (j + 1) / n\_bins)\) for \(j = 0, \ldots, n\_bins - 1\), with \(p = 1\) assigned to the final bin.
p_value_min: Lower/upper censoring point for p-values used inside the selection likelihood. When NULL, defaults to 1e-5 for method = "beta_selection" and 1e-12 otherwise. This setting is most relevant for "beta_selection"; for "beta_binomial", it typically only affects exact boundary values such as \(p = 0\) or \(p = 1\).
maxit: Maximum number of optimization iterations passed to optim.
integration_points: Number of grid points used for numeric integration of the normalization term in the selection likelihood.
integration_limit: Absolute z-limit for integration grid.
funnel_plot: Logical; if TRUE, draws a funnel plot from the fitted publication-bias object.
funnel_args: Optional named list passed to funnel_plot when funnel_plot = TRUE.
control: Optional named list of control values passed to optim.
...: Unused. Optimization controls should be supplied through control; optimization uses "BFGS" internally.

Value

A list of class pub_bias containing adjusted coefficients, selection parameters, optimization diagnostics, and an adjusted-vs-unadjusted coefficient summary table.

References

Citkowicz, M., & Vevea, J. L. (2017). A parsimonious weight function for modeling publication bias. Psychological Methods, 22(1), 28-41.

Examples

# \donttest{
fit <- mars(
  data = teacher_expectancy,
  studyID = "study",
  effectID = NULL,
  sample_size = NULL,
  formula = yi ~ 1,
  variance = "vi",
  varcov_type = "univariate",
  structure = "univariate"
)
bias_fit <- publication_bias(fit, method = "beta_selection")
summary(bias_fit)
#> Publication Bias Model (MARS)
#> Method: beta_selection 
#> Structure: univariate 
#> Effects: 19  | Studies: 19 
#> P-value tail: two.sided 
#> Convergence: 0 - Converged 
#> Objective: 3.113 
#> 
#> Selection Parameters:
#>  parameter estimate
#>      alpha   0.9907
#>       beta   1.0564
#>       tau2   0.0186
#> 
#> Selection Weight Profile (p-value grid):
#>           model p_value weight
#>  beta_selection  0.0100 1.0430
#>  beta_selection  0.1189 1.0127
#>  beta_selection  0.2278 0.9991
#>  beta_selection  0.3367 0.9870
#>  beta_selection  0.4456 0.9745
#>  beta_selection  0.5544 0.9606
#>  beta_selection  0.6633 0.9440
#>  beta_selection  0.7722 0.9221
#>  beta_selection  0.8811 0.8878
#>  beta_selection  0.9900 0.7712
#> 
#> Bias-Adjusted Coefficients:
#>         term estimate std_error z_value p_value
#>  (Intercept)   0.0855    0.0253  3.3753   7e-04
#> 
#> Adjusted vs Unadjusted Summary:
#>         term unadjusted_estimate unadjusted_se unadjusted_ci_lower
#>  (Intercept)              0.0837        0.0516             -0.0175
#>  unadjusted_ci_upper estimate std_error adjusted_ci_lower adjusted_ci_upper
#>               0.1849   0.0855    0.0253            0.0358            0.1351
#>  estimate_diff percent_change direction_changed ci_overlap
#>         0.0018         2.1322             FALSE       TRUE
#> 
#> Small-Study Effects Check (Egger-type):
#> Model: SND ~ precision 
#> Method: Egger-type small-study-effects regression with study-cluster robust SE (CR1) 
#> Intercept: 0.9059 SE: 0.8531 t: 1.0619 df: 18 p: 0.3023 
#> 
#> SE note:
#> SE rule: reported SEs use the larger of the Hessian-based and OPG-based values when both are finite, otherwise the finite one. 
# }