Sequentially rejective graphical multiple comparison procedures based on Bonferroni tests

Code
library(gt)
library(graphicalMCP)

Motivating example

Consider a confirmatory clinical trial comparing a test treatment (treatment) against the control treatment (control) for a disease. There are two doses of treatment: the low dose and the high dose. There are two endpoints included in the multiplicity adjustment strategy, which are the primary endpoint (PE) and the secondary endpoint (SE). In total, there are four null hypotheses: \(H_1\) and \(H_3\) are the primary and the secondary hypotheses respectively for the low dose versus control; \(H_2\) and \(H_4\) are the primary and the secondary hypotheses respectively for the high dose versus control.

There are clinical considerations which constrain the structure of multiple comparison procedures, and which can be flexibly incorporated using graphical approaches. First, the low and the high doses are considered equally important, which means that rejecting the primary hypothesis for either dose versus control leads to a successful trial. Regarding secondary hypotheses, each one is tested only if its corresponding primary hypothesis has been rejected. This means that \(H_3\) is tested only after \(H_1\) has been rejected; \(H_4\) is tested only after \(H_2\) has been rejected.

In addition, there are some statistical considerations to complete the graph. The primary hypotheses \(H_1\) and \(H_2\) will have an equal hypothesis weight of 0.5. The secondary hypotheses will have a hypothesis weight of 0. When a primary hypothesis has been rejected, its weight will be propagated along two outgoing edges: One to the other primary hypothesis and one to its descendant secondary hypothesis. The two edges will have an equal transition weight of 0.5. When both the primary and the secondary hypotheses have been rejected for a dose-control comparison, their hypothesis weights will be propagated to the primary hypothesis for the other dose-control comparison. With these specifications, we can create the following graph.

Create the graph

Code
hypotheses <- c(0.5, 0.5, 0, 0)
transitions <- rbind(
  c(0, 0.5, 0.5, 0),
  c(0.5, 0, 0, 0.5),
  c(0, 1, 0, 0),
  c(1, 0, 0, 0)
)
hyp_names <- c("H1", "H2", "H3", "H4")

g <- graph_create(hypotheses, transitions, hyp_names)

plot(g, vertex.size = 60)

Perform the graphical multiple comparison procedure

Adjusted p-values and rejections

Given a set of p-values for \(H_1, \ldots, H_4\), the graphical multiple comparison procedure can be performed to control the familywise error rate (FWER) at the significance level alpha. The graph_test_shortcut function is agnostic to one-sided or two-sided tests. For one-sided p-values, alpha is often set to 0.025 (default); for two-sided p-values, alpha is often set to 0.05. We consider one-sided tests here. A hypothesis is rejected if its adjusted p-value is less than or equal to alpha. After running the procedure, hypotheses \(H_1\), \(H_2\), and \(H_4\) are rejected with their adjusted p-value calculated.

Code
p_values <- c(0.018, 0.01, 0.105, 0.006)
test_results <- graph_test_shortcut(g, p = p_values, alpha = 0.025)

test_results$outputs$adjusted_p # Adjusted p-values
#>    H1    H2    H3    H4 
#> 0.024 0.020 0.105 0.024
test_results$outputs$rejected # Rejections
#>    H1    H2    H3    H4 
#>  TRUE  TRUE FALSE  TRUE

Obtain final and intermediate graphs after rejections

The final graph is the graph after removing rejected hypotheses \(H_1\), \(H_2\), and \(H_4\). It can be obtained via test_results$outputs$graph. Rejected hypotheses get a hypothesis weight of NA and a transition weight of NA. This is based on the printing method of print.updated_graph. We can also obtain the non-NA graph by calling hypothesis weights and transition weights separately via test_results$outputs$graph$hypotheses and test_results$outputs$graph$transitions. Note in this case, rejected hypotheses get a 0 hypothesis weight and a 0 transition weight. This is mainly for internal calculation and updating of graphs.

If we are also interested in intermediate graphs - for example, the graph after \(H_2\) and \(H_1\) are rejected - we can specify verbose = TRUE in graph_test_shortcut. Note that intermediate graphs depend on the order of rejections, i.e., the sequence of hypotheses being rejected. The default order is defined by the increasing adjusted p-values, followed by the earlier hypothesis numbering in the case of ties. In this example, the default order of rejection is \(H_2\rightarrow H_1\rightarrow H_4\). To obtain intermediate graphs based on this order of rejection, one can specify test_results_verbose$details$results. For example, the graph after \(H_2\) and \(H_1\) being rejected is given by test_results_verbose$details$results[[3]].

Code
test_results$outputs$graph # Final graph after H1, H2 and H4 rejected (as NA's)
#> Updated graph
#> 
#> --- 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
test_results$outputs$graph$hypotheses # Hypothesis weights of the final graph
#> H1 H2 H3 H4 
#>  0  0  1  0
test_results$outputs$graph$transitions # Transition weights of the final graph
#>    H1 H2 H3 H4
#> H1  0  0  0  0
#> H2  0  0  0  0
#> H3  0  0  0  0
#> H4  0  0  0  0

test_results_verbose <- graph_test_shortcut(
  g,
  p = p_values,
  alpha = 0.025,
  verbose = TRUE
)

# Intermediate graph after H1 and H2 rejected
test_results_verbose$details$results[[3]]
#> Updated graph
#> 
#> --- Hypothesis weights ---
#> H1:  NA
#> H2:  NA
#> H3: 0.5
#> H4: 0.5
#> 
#> --- Transition weights ---
#>     H1 H2 H3 H4
#>  H1 NA NA NA NA
#>  H2 NA NA NA NA
#>  H3 NA NA  0  1
#>  H4 NA NA  1  0

Obtain possible orders of rejections

The order of rejections may not be unique and not all orders are valid. For this example, the rejected hypotheses are \(H_1\), \(H_2\) and \(H_4\). The default order of rejections is \(H_2 \rightarrow H_1 \rightarrow H_4\). Another valid order of rejections is \(H_2 \rightarrow H_4 \rightarrow H_1\). However, the first rejected hypothesis can not be \(H_1\) or \(H_4\). To obtain all possible rejection orders, one can use the function graph_rejection_orderings. Then intermediate and final graphs can be obtained by using the function graph_update with a particular order of rejections.

Code
# Obtain all valid orders of rejections
orders <- graph_rejection_orderings(test_results)$valid_orderings
orders
#> [[1]]
#> H2 H1 H4 
#>  2  1  4 
#> 
#> [[2]]
#> H2 H4 H1 
#>  2  4  1

# Intermediate graphs following the order of H2 and H4
graph_update(g, delete = orders[[2]])$intermediate_graphs[[3]]
#> Updated graph
#> 
#> --- Hypothesis weights ---
#> H1:  1
#> H2: NA
#> H3:  0
#> H4: NA
#> 
#> --- Transition weights ---
#>     H1 H2 H3 H4
#>  H1  0 NA  1 NA
#>  H2 NA NA NA NA
#>  H3  1 NA  0 NA
#>  H4 NA NA NA NA

Obtain adjusted significance levels

An equivalent way to obtain rejections is by adjusting significance levels. A hypothesis is rejected if its p-value is less than or equal to its adjusted significance level. The adjusted significance levels are calculated in the same order as adjusted p-values: \(H_2 \rightarrow H_1 \rightarrow H_4\), and there are four steps of checking for rejections. First, \(H_2\) is rejected at an adjusted significance level of 0.5 * alpha. Second, \(H_1\) is rejected at an adjusted significance level of 0.75 * alpha, after \(H_2\) is rejected. Third, \(H_4\) is rejected at an adjusted significance level of 0.5 * alpha, after \(H_1\) and \(H_2\) are rejected. Fourth and finally, \(H_3\) cannot be rejected at an adjusted significance level of alpha, after \(H_1\), \(H_2\) and \(H_4\) are rejected. These results can be obtained by specifying test_values = TRUE.

Code
test_results_test_values <- graph_test_shortcut(
  g,
  p = p_values,
  alpha = 0.025,
  test_values = TRUE
)

test_results_test_values$test_values$results
#>   Step Hypothesis     p Weight Alpha Inequality_holds
#> 1    1         H2 0.010   0.50 0.025             TRUE
#> 2    2         H1 0.018   0.75 0.025             TRUE
#> 3    3         H4 0.006   0.50 0.025             TRUE
#> 4    4         H3 0.105   1.00 0.025            FALSE

Power simulation

Given the above graph, we are interested in the “power” of the trial. For a single null hypothesis, the power is the probability of a true positive - that is, rejecting the null hypothesis at the significance level alpha when the alternative hypothesis is true. For multiple null hypotheses, there could be multiple versions of “power”. For example, the power to reject at least one hypothesis vs the power to reject all hypotheses, given the alternative hypotheses are true. With the graphical multiple comparison procedures, it is also important to understand the power to reject each hypothesis, given the multiplicity adjustment. Sometimes, we may want to customize definitions of power to define success. Thus power calculation is an important aspect of trial design.

Input: Marginal power for primary hypotheses

Assume that the primary endpoint is about the occurrence of an unfavorable clinical event. To evaluate the treatment effect, the proportion of patients with this event is calculated, and a lower proportion is preferred. Assume that the proportions are 0.181 for the low and the high doses, and 0.3 for control. Using the equal randomization among the three treatment groups, the clinical trial team chooses a total sample size of 600 with 200 per group. This leads to a marginal power of 80% for \(H_1\) and \(H_2\), respectively, using the two-sample test for difference in proportions with unpooled variance each at one-sided significance level 0.025. In this calculation, we use the marginal power to combine the information from the treatment effect, any nuisance parameter, and sample sizes for each hypothesis. Note that the significance level used for the marginal power calculation must be the same as alpha, which is used as the significance level for the FWER control. In addition, the marginal power has a one-to-one relationship with the noncentrality parameter, which is illustrated below.

Code
alpha <- 0.025
prop <- c(0.3, 0.181, 0.181)
sample_size <- rep(200, 3)

unpooled_variance <-
  prop[-1] * (1 - prop[-1]) / sample_size[-1] +
  prop[1] * (1 - prop[1]) / sample_size[1]

noncentrality_parameter_primary <-
  -(prop[-1] - prop[1]) / sqrt(unpooled_variance)

power_marginal_primary <- pnorm(
  qnorm(alpha, lower.tail = FALSE),
  mean = noncentrality_parameter_primary,
  sd = 1,
  lower.tail = FALSE
)

names(power_marginal_primary) <- c("H1", "H2")
power_marginal_primary
#>        H1        H2 
#> 0.8028315 0.8028315

Input: Marginal power for secondary hypotheses

Assume that the secondary endpoint is about the change in total medication score from baseline, which is a continuous outcome. To evaluate the treatment effect, the mean change is calculated, and greater reduction is preferred. Assume that the mean change from baseline is the reduction of 7.5 and 8.25, respectively for the low and the high doses, and 5 for control. Further assume a known common standard deviation of 10. Given the sample size of 200 per treatment group, the marginal power is 71% and 90% for \(H_3\) and \(H_4\), respectively, using the two-sample \(z\)-test for the difference in means each at the one-sided significance level 0.025.

Code
mean_change <- c(5, 7.5, 8.25)
sd <- rep(10, 3)
variance <- sd[-1]^2 / sample_size[-1] + sd[1]^2 / sample_size[1]

noncentrality_parameter_secondary <-
  (mean_change[-1] - mean_change[1]) / sqrt(variance)

power_marginal_secondary <- pnorm(
  qnorm(alpha, lower.tail = FALSE),
  mean = noncentrality_parameter_secondary,
  sd = 1,
  lower.tail = FALSE
)

names(power_marginal_secondary) <- c("H3", "H4")
power_marginal_secondary
#>        H3        H4 
#> 0.7054139 0.9014809

Input: Correlation structure to simulate test statistics

In addition to the marginal power, we also need to make assumptions about the joint distribution of test statistics. In this example, we assume that they follow a multivariate normal distribution with means defined by the noncentrality parameters above and the correlation matrix \(R\). To obtain the correlations, it is helpful to understand that there are two types of correlations in this example. The correlation between two dose-control comparisons for the same endpoint and the correlation between endpoints. The former correlation can be calculated as a function of sample size. For example, the correlation between test statistics for \(H_1\) and \(H_2\) is \(\rho_{12}=\left(\frac{n_1}{n_1+n_0}\right)^{1/2}\left(\frac{n_2}{n_3+n_0}\right)^{1/2}\). Under the equal randomization, this correlation is 0.5. The correlation between test statistics for \(H_3\) and \(H_4\) is the same as the above. On the other hand, the correlation between endpoints for the same dose-control comparison is often estimated based on prior knowledge or from previous trials. Without the information, we assume it to be \(\rho_{13}=\rho_{24}=0.5\). In practice, one could set this correlation as a parameter and try multiple values to assess the sensitivity of this assumption. Regarding the correlation between test statistics for \(H_1\) and \(H_4\) and for \(H_2\) and \(H_3\), they are even more difficult to estimate. Here we use a simple product rule, which means that this correlation is a product of correlations of the two previously assumed correlations. For example, \(\rho_{14}=\rho_{12}*\rho_{24}\) and \(\rho_{23}=\rho_{12}*\rho_{13}\). In practice, one may make further assumptions instead of using the product rule.

Code
corr <- matrix(0, nrow = 4, ncol = 4)

corr[1, 2] <-
  corr[3, 4] <-
  sqrt(
    sample_size[2] / sum(sample_size[1:2]) *
      sample_size[3] / sum(sample_size[c(1, 3)])
  )

rho <- 0.5
corr[1, 3] <- corr[2, 4] <- rho
corr[1, 4] <- corr[2, 3] <- corr[1, 2] * rho
corr <- corr + t(corr)
diag(corr) <- 1
colnames(corr) <- hyp_names
rownames(corr) <- hyp_names
corr
#>      H1   H2   H3   H4
#> H1 1.00 0.50 0.50 0.25
#> H2 0.50 1.00 0.25 0.50
#> H3 0.50 0.25 1.00 0.50
#> H4 0.25 0.50 0.50 1.00

User-defined success criteria

As mentioned earlier, there are multiple versions of “power” when there are multiple hypotheses. Commonly used “power” definitions include:

These are the default outputs from the graph_calculate_power function. In addition, a user can customize success criteria to define other versions of “power”.

Code
success_fns <- list(
  # Probability to reject H1
  H1 = function(x) x[1],

  # Expected number of rejections
  `Expected no. of rejections` = function(x) x[1] + x[2] + x[3] + x[4],

  # Probability to reject at least one hypothesis
  `AtLeast1` = function(x) x[1] | x[2] | x[3] | x[4],

  # Probability to reject all hypotheses
  `All` = function(x) x[1] & x[2] & x[3] & x[4],

  # Probability to reject both H1 and H2
  `H1andH2` = function(x) x[1] & x[2],

  # Probability to reject both hypotheses for the low dose or the high dose
  `(H1andH3)or(H2andH4)` = function(x) (x[1] & x[3]) | (x[2] & x[4])
)

Output: Simulate power

Given the above inputs, we can estimate “power” via simulation for the graphical multiple comparison procedure at one-sided significance level alpha = 0.025 using sim_n = 1e5 simulations and the random seed 1234. The local power is 0.758, 0.765, 0.689, and 0.570, respectively for \(H_1, \ldots, H_4\). Note that the local power is lower than the marginal power because the former is adjusted for multiplicity. The power to reject at least one hypothesis is 0.856 and the power to reject all hypotheses is 0.512. The expected number of rejections is 2.782. For the last two user-defined success criteria, the probability to reject both \(H_1\) and \(H_2\) is 0.667, and the probability to reject at least one pair of \((H_1\) and \(H_3)\) and \((H_2\) and \(H_4)\) is 0.747.

Code
set.seed(1234)
power_output <- graph_calculate_power(
  g,
  alpha = 0.025,
  sim_corr = corr,
  sim_n = 1e5,
  power_marginal = c(power_marginal_primary, power_marginal_secondary),
  sim_success = success_fns
)

power_output$power
#> $power_local
#>      H1      H2      H3      H4 
#> 0.76396 0.75887 0.56767 0.69133 
#> 
#> $rejection_expected
#> [1] 2.78183
#> 
#> $power_at_least_1
#> [1] 0.85557
#> 
#> $power_all
#> [1] 0.51205
#> 
#> $power_success
#>                         H1 Expected no. of rejections 
#>                    0.76396                    2.78183 
#>                   AtLeast1                        All 
#>                    0.85557                    0.51205 
#>                    H1andH2       (H1andH3)or(H2andH4) 
#>                    0.66726                    0.74695

To see the detailed outputs of all simulated p-values and rejection decisions for all hypotheses, specify verbose = TRUE. This will produce a lot of outputs. To allow flexible printing functions, a user can change the following:

Code
set.seed(1234)
power_verbose_output <- graph_calculate_power(
  g,
  alpha = 0.025,
  sim_corr = corr,
  sim_n = 1e5,
  power_marginal = c(power_marginal_primary, power_marginal_secondary),
  sim_success = success_fns,
  verbose = TRUE
)

head(power_verbose_output$details$p_sim, 10)
#>                 H1           H2           H3           H4
#>  [1,] 0.0308204265 0.0120653993 0.0041185823 9.324338e-02
#>  [2,] 0.0007933716 0.0006499046 0.0245177515 2.965604e-03
#>  [3,] 0.0302991819 0.0595395828 0.0543082956 2.625834e-02
#>  [4,] 0.0097433244 0.0033185711 0.0007417213 4.024688e-04
#>  [5,] 0.0197134942 0.0086161835 0.0164182765 2.418325e-07
#>  [6,] 0.0031206572 0.0067023099 0.0137441457 2.751703e-04
#>  [7,] 0.0302208038 0.1423757994 0.0060382838 2.117403e-02
#>  [8,] 0.0024975725 0.0294025573 0.0004142729 2.207786e-03
#>  [9,] 0.0618994292 0.0387257108 0.3166125781 5.699791e-02
#> [10,] 0.3677921053 0.1895975134 0.0702264885 1.189651e-02

print(power_verbose_output, indent = 4, precision = 6)
#> 
#> Test parameters ($inputs) ------------------------------------------------------
#>     Initial graph
#> 
#>     --- Hypothesis weights ---
#>     H1: 0.5
#>     H2: 0.5
#>     H3: 0.0
#>     H4: 0.0
#> 
#>     --- Transition weights ---
#>         H1  H2  H3  H4
#>     H1 0.0 0.5 0.5 0.0
#>     H2 0.5 0.0 0.0 0.5
#>     H3 0.0 1.0 0.0 0.0
#>     H4 1.0 0.0 0.0 0.0
#> 
#>     Alpha = 0.025
#> 
#>     Test types
#>     bonferroni: (H1, H2, H3, H4)
#> 
#> Simulation parameters ($inputs) ------------------------------------------------
#>     Testing 100,000 simulations with multivariate normal params:
#> 
#>                           H1       H2       H3       H4
#>     Marginal power: 0.802831 0.802831 0.705414 0.901481
#> 
#>     Correlation:      H1   H2   H3   H4
#>                  H1 1.00 0.50 0.50 0.25
#>                  H2 0.50 1.00 0.25 0.50
#>                  H3 0.50 0.25 1.00 0.50
#>                  H4 0.25 0.50 0.50 1.00
#> 
#> Power calculation ($power) -----------------------------------------------------
#>                                      H1      H2      H3      H4
#>                    Local power: 0.76396 0.75887 0.56767 0.69133
#> 
#>     Expected no. of rejections: 2.78183
#>      Power to reject 1 or more: 0.85557
#>            Power to reject all: 0.51205
#> 
#>                Success measure   Power
#>                             H1 0.76396
#>     Expected no. of rejections 2.78183
#>                       AtLeast1 0.85557
#>                            All 0.51205
#>                        H1andH2 0.66726
#>           (H1andH3)or(H2andH4) 0.74695
#> 
#> Simulation details ($details) --------------------------------------------------
#>         p_sim_H1    p_sim_H2    p_sim_H3    p_sim_H4 rej_H1 rej_H2 rej_H3 rej_H4
#>      3.08204e-02 1.20654e-02 4.11858e-03 9.32434e-02  FALSE   TRUE  FALSE  FALSE
#>      7.93372e-04 6.49905e-04 2.45178e-02 2.96560e-03   TRUE   TRUE   TRUE   TRUE
#>      3.02992e-02 5.95396e-02 5.43083e-02 2.62583e-02  FALSE  FALSE  FALSE  FALSE
#>      9.74332e-03 3.31857e-03 7.41721e-04 4.02469e-04   TRUE   TRUE   TRUE   TRUE
#>      1.97135e-02 8.61618e-03 1.64183e-02 2.41833e-07   TRUE   TRUE   TRUE   TRUE
#>      3.12066e-03 6.70231e-03 1.37441e-02 2.75170e-04   TRUE   TRUE   TRUE   TRUE
#>      3.02208e-02 1.42376e-01 6.03828e-03 2.11740e-02  FALSE  FALSE  FALSE  FALSE
#>      2.49757e-03 2.94026e-02 4.14273e-04 2.20779e-03   TRUE  FALSE   TRUE  FALSE
#>      6.18994e-02 3.87257e-02 3.16613e-01 5.69979e-02  FALSE  FALSE  FALSE  FALSE
#>      3.67792e-01 1.89598e-01 7.02265e-02 1.18965e-02  FALSE  FALSE  FALSE  FALSE
#>     ... (Use `print(x, rows = <nn>)` for more)

Reference

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.