Choosing a Sequence-Analysis Method • gp3sequences

Start with the research question

No single sequence method is universally preferable. Method choice should be driven by the declared structural question, the sequence representation, the intended comparison, and the assumptions that can be defended. Every workflow should begin with audit_sequence_data(), validate_sequence_data(), and, when needed, prepare_sequence_data().

method_guide <- data.frame(
  question = c(
    "Which exact contiguous patterns recur?",
    "What state is most supported at each aligned position?",
    "How do observed state, transition, or length summaries differ by group?",
    "How dissimilar are complete sequences?",
    "Can a declared distance reveal reproducible descriptive partitions?",
    "Which transitions and contexts organise the observed paths?",
    "Does recent state history improve next-state description?",
    "Can a compact latent categorical model summarise serial dependence?",
    "Is interoperability with a specialist package required?"
  ),
  method = c(
    "Contiguous motifs",
    "Aligned-position consensus",
    "Descriptive group comparison",
    "Sequence distance",
    "Clustering and stability",
    "Transition network",
    "Higher-order transition model",
    "Categorical HMM or mixture HMM",
    "Optional adapter"
  ),
  primary_functions = c(
    "extract_sequence_ngrams(); summarise_sequence_motifs()",
    "create_consensus_sequence(); summarise_consensus_agreement()",
    "compare_sequence_groups()",
    "compute_sequence_distance(); summarise_sequence_distance()",
    "cluster_sequences(); validate_sequence_clusters(); bootstrap_sequence_clusters()",
    "create_transition_network(); summarise_transition_centrality()",
    "fit_higher_order_transition_model(); predict_next_state()",
    "fit_sequence_hmm(); fit_sequence_hmm_mixture(); decode_sequence_states()",
    "as_traminer_sequences(); as_arules_sequences(); as_igraph_transition_network()"
  ),
  stringsAsFactors = FALSE
)

method_guide
#>                                                                  question
#> 1                                  Which exact contiguous patterns recur?
#> 2                  What state is most supported at each aligned position?
#> 3 How do observed state, transition, or length summaries differ by group?
#> 4                                  How dissimilar are complete sequences?
#> 5     Can a declared distance reveal reproducible descriptive partitions?
#> 6             Which transitions and contexts organise the observed paths?
#> 7               Does recent state history improve next-state description?
#> 8     Can a compact latent categorical model summarise serial dependence?
#> 9                 Is interoperability with a specialist package required?
#>                           method
#> 1              Contiguous motifs
#> 2     Aligned-position consensus
#> 3   Descriptive group comparison
#> 4              Sequence distance
#> 5       Clustering and stability
#> 6             Transition network
#> 7  Higher-order transition model
#> 8 Categorical HMM or mixture HMM
#> 9               Optional adapter
#>                                                                  primary_functions
#> 1                           extract_sequence_ngrams(); summarise_sequence_motifs()
#> 2                     create_consensus_sequence(); summarise_consensus_agreement()
#> 3                                                        compare_sequence_groups()
#> 4                       compute_sequence_distance(); summarise_sequence_distance()
#> 5 cluster_sequences(); validate_sequence_clusters(); bootstrap_sequence_clusters()
#> 6                   create_transition_network(); summarise_transition_centrality()
#> 7                        fit_higher_order_transition_model(); predict_next_state()
#> 8         fit_sequence_hmm(); fit_sequence_hmm_mixture(); decode_sequence_states()
#> 9   as_traminer_sequences(); as_arules_sequences(); as_igraph_transition_network()

Contiguous motifs

Use motifs when the unit of interest is an exact adjacent state window. Declare minimum and maximum length, overlap handling, prevalence denominators, filtering thresholds, and tie rules. Motifs do not capture non-contiguous subsequences unless an external specialist method is used.

Consensus sequences

Use aligned-position consensus when positions are meaningfully comparable across sequences. Declare the missing-position policy, weighting, state order, and deterministic tie method. A consensus is a modal structural summary, not a normative or ideal pathway.

Descriptive group comparison

Use compare_sequence_groups() when the objective is to compare observed state shares, transition shares, prevalence, or sequence lengths across declared groups. The function is deliberately descriptive and does not automatically perform significance tests or causal attribution.

Distances

compute_sequence_distance() supports four explicit families:

Levenshtein distance for unit-cost edits;
LCS distance for differences based on longest common subsequences;
configurable optimal matching for declared insertion/deletion and substitution costs;
transition-profile Euclidean distance for first-order transition distributions.

Normalisation and cost choices can materially change the geometry. Report them and avoid selecting a distance solely because it produces visually convenient clusters.

Clustering, representatives, ensembles, and stability

Clustering is a downstream description of a chosen distance matrix. Declare k, the clustering method, linkage, seed, and any optional-package settings. Use validation and resampling summaries to assess the supplied solution, not to claim that clusters are objectively true. Representatives minimise declared within-cluster distance; they are not automatically typical in a substantive sense.

Transition networks

Use first-order networks when nodes and directed edges are the relevant representation. Declare whether weights are counts, conditional proportions, or global shares; whether self-transitions are included; and whether grouping or smoothing is used. Centrality and communities are graph-structural outputs, not psychological attributes.

Higher-order transition models

Use higher-order models when recent context is part of the structural question. Declare the order, smoothing, context separator, and backoff policy. Inspect which context and order were actually used for every prediction, especially for unseen contexts.

Hidden and mixture sequence models

Use the native HMM helpers only for compact, time-homogeneous categorical workflows. Report initial values, seed, pseudocount, tolerance, convergence history, likelihood, AIC/BIC, and decoding method. Hidden-state labels and mixture components are exchangeable statistical constructs. Multiple seeded fits and specialist software are appropriate when the model is consequential or more complex.

Optional adapters

Adapters support TraMineR, arulesSequences, GrpString-style inputs, seqHMM, igraph, and common gp3tools-style column names without making those packages or formats mandatory. Use them when a specialist engine adds functionality that should not be duplicated inside gp3sequences.

Minimum decision checklist

Before analysis, record:

the sequence unit and ordering variable;
state definitions and any preprocessing policy;
whether positions are aligned and comparable;
the structural estimand or descriptive question;
method parameters, costs, thresholds, seeds, and normalisation;
how missing positions, ties, unseen contexts, and optional dependencies are handled;
sensitivity checks and failure cases;
the interpretation boundary separating structure from substantive claims.

Focused articles

The package website provides dedicated articles for contiguous motifs, consensus and group comparison, distances and clustering, transition networks, higher-order models, latent models, and optional adapters. The synthetic case study demonstrates how these layers can be combined without treating every method as mandatory.