Design-Aware Sequence Group Inference
In gp3sequences: Transparent Analysis of Ordered Categorical Sequences

knitr::opts_chunk$set(collapse = TRUE, comment = "#>")
library(gp3sequences)

Why declare the design?

Sequence rows are rarely independent. The comparison API requires the group column and independent unit, and can additionally record pairs or assignment clusters. It aggregates the selected metric before permutation or bootstrap resampling.

Synthetic randomized groups

paths <- replicate(20, sample(c("A", "B", "C"), 6L, replace = TRUE), simplify = FALSE)
data <- do.call(rbind, lapply(seq_along(paths), function(i) {
  data.frame(
    participant_id = paste0("p", i),
    sequence_id = paste0("s", i),
    sequence_order = seq_along(paths[[i]]),
    state = paths[[i]],
    group = if (i <= 10L) "control" else "treatment",
    stringsAsFactors = FALSE
  )
}))

Declare and test

design <- declare_sequence_comparison_design(
  group_col = "group",
  unit_col = "participant_id",
  design = "randomized"
)

result <- test_sequence_group_difference(
  data,
  design,
  metric = "state_prevalence",
  target_state = "A",
  n_permutations = 999L,
  seed = 10L
)
result$estimate

Supported metrics are deliberately limited to transparent quantities: sequence length, transition count, state prevalence, and declared subsequence presence.

Bootstrap interval

result <- bootstrap_sequence_group_difference(
  result,
  n_boot = 999L,
  level = 0.95,
  seed = 11L
)
summarise_sequence_group_inference(result)

plot_sequence_group_inference(result, type = "permutation")
plot_sequence_group_inference(result, type = "group_means")

Causal language

For design = "randomized", causal interpretation still depends on valid assignment, implementation, attrition handling, and an estimand consistent with the design. For design = "observational", the output explicitly describes an associational exchangeability-based contrast and does not license causal claims.