Skip to contents

Introduction

Graphical approaches for multiple comparison procedures (MCPs) are a general framework to control the familywise error rate strongly as a pre-specified significance level 0<α<10<\alpha<1. This approach includes many commonly used MCPs as special cases and is transparent in visualizing MCPs for better communications. graphicalMCP is designed to design and analyze graphical MCPs in a flexible, informative and efficient way.

Basic usage

Initial graph

The base object in graphicalMCP is an initial_graph, which consists of a vector of hypothesis weights and a matrix of transition weights (Bretz et al. 2009). This object can be created via graph_create(). In the graphical representation, hypotheses are denoted as nodes (or vertices) associated with hypothesis weights. A directed edge from a hypothesis to another indicates the direction of propagation of the hypothesis weight from the origin hypothesis to the end hypothesis. The edge is weighted by a transition weight indicating the proportion of propagation.

library(graphicalMCP)
# A graph of two primary hypotheses (H1 and H2) and two secondary hypotheses (H3 and H4)
hypotheses <- c(0.5, 0.5, 0, 0)
transitions <- rbind(
  c(0, 0, 1, 0),
  c(0, 0, 0, 1),
  c(0, 1, 0, 0),
  c(1, 0, 0, 0)
)
hyp_names <- c("H1", "H2", "H3", "H4")
example_graph <- graph_create(hypotheses, transitions, hyp_names)
example_graph
#> Initial graph
#> 
#> --- Hypothesis weights ---
#> H1: 0.5
#> H2: 0.5
#> H3: 0.0
#> H4: 0.0
#> 
#> --- Transition weights ---
#>     H1 H2 H3 H4
#>  H1  0  0  1  0
#>  H2  0  0  0  1
#>  H3  0  1  0  0
#>  H4  1  0  0  0
plot(example_graph, vertex.size = 60)

Update graph

When a hypothesis is removed from the graph, hypothesis and transition weights of remaining hypotheses should be updated according to Algorithm 1 in Bretz et al. (2011). For example, assume that hypotheses H1, H2 and H4 are removed from the graph. The updated graph after removing three hypotheses is below.

updated_example <- graph_update(
  example_graph,
  delete = c(TRUE, TRUE, FALSE, TRUE)
)

updated_example
#> Initial and final graphs -------------------------------------------------------
#> 
#> Initial graph
#> 
#> --- Hypothesis weights ---
#> H1: 0.5
#> H2: 0.5
#> H3: 0.0
#> H4: 0.0
#> 
#> --- Transition weights ---
#>     H1 H2 H3 H4
#>  H1  0  0  1  0
#>  H2  0  0  0  1
#>  H3  0  1  0  0
#>  H4  1  0  0  0
#> 
#> Updated graph after deleting hypotheses 1, 2, 4
#> 
#> --- Hypothesis weights ---
#> H1: NA
#> H2: NA
#> H3:  1
#> H4: NA
#> 
#> --- Transition weights ---
#>     H1 H2 H3 H4
#>  H1 NA NA NA NA
#>  H2 NA NA NA NA
#>  H3 NA NA  0 NA
#>  H4 NA NA NA NA
plot(updated_example, vertex.size = 60)

Perform graphical MCPs

Given the set of p-values of all hypotheses, graphical MCPs can be performed using graph_test_shortcut() to determine which hypotheses can be rejected at the significance level alpha. Assume p-values are 0.01, 0.005, 0.03, and 0.01 for hypotheses H1-H4. With a one-sided significance level alpha = 0.025, rejected hypotheses are H1, H2, and H4. More details about the shortcut testing can be found in vignette("shortcut-testing").

test_results <- graph_test_shortcut(
  example_graph,
  p = c(0.01, 0.005, 0.03, 0.01),
  alpha = 0.025
)
test_results$outputs$rejected
#>    H1    H2    H3    H4 
#>  TRUE  TRUE FALSE  TRUE

A similar testing procedure can be performed using the closure principle. This will allow more tests for intersection hypotheses, e.g., Simes, parametric and a mixture of them. If the test type is Bonferroni, the resulting closed procedure is equivalent to the shortcut procedure above. Additional details about closed testing can be found in vignette("closed-testing").

test_results_closed <- graph_test_closure(
  example_graph,
  p = c(0.01, 0.005, 0.03, 0.01),
  alpha = 0.025,
  test_types = "bonferroni",
  test_groups = list(1:4)
)
test_results_closed$outputs$rejected
#>    H1    H2    H3    H4 
#>  TRUE  TRUE FALSE  TRUE

Power simulations

With multiplicity adjustment, such as graphical MCPs, the “power” to reject each hypothesis will be affected, compared to its marginal power. The latter is the power to rejected a hypothesis at the significance level alpha without multiplicity adjustment. The marginal power is usually obtained from other pieces of statistical software. graph_calculate_power() performs power simulations to assess the power after adjusting for the graphical MCP (Bretz, Maurer, and Hommel 2011). Assume that the marginal power to reject H1-H4 is 90%, 90%, 80%, and 80% and all test statistics are independent of each other. The local power after the multiplicity adjustment is 87.7%, 87.7%, 67.2%, and 67.2% respectively for H1-H4. Additional details about power simulations can be found in vignette("shortcut-testing") and vignette("closed-testing").

set.seed(1234)
power_results <- graph_calculate_power(
  example_graph,
  sim_n = 1e6,
  power_marginal = c(0.9, 0.9, 0.8, 0.8)
)
power_results$power$power_local
#>       H1       H2       H3       H4 
#> 0.875760 0.876295 0.670470 0.671431

References

Bretz, Frank, Willi Maurer, Werner Branath, and Martin Posch. 2009. “A Graphical Approach to Sequentially Rejective Multiple Test Procedures.” Statistics in Medicine 53 (4): 586–604. https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.3495.
Bretz, Frank, Willi Maurer, and Gerhard Hommel. 2011. “Test and Power Considerations for Multiple Endpoint Analyses Using Sequentially Rejective Graphical Procedures.” Statistics in Medicine 30 (13): 1489–1501. https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.3988.
Bretz, Frank, Martin Posch, Ekkehard Glimm, Florian Klinglmueller, Willi Maurer, and Kornelius Rohmeyer. 2011. “Graphical Approaches for Multiple Comparison Procedures Using Weighted Bonferroni, Simes, or Parametric Tests.” Biometrical Journal 53 (6): 894–913. https://onlinelibrary.wiley.com/doi/10.1002/bimj.201000239.