Skip to content

Reporting guide

A sequence-analysis report should make the data representation, method choices, validation, and interpretation boundary inspectable.

Minimum methods checklist

Component Report at minimum
Sequence unit what one sequence represents
Ordering order / time variable and tie handling
State definition state labels, coding rules, unknown-state policy
Missingness missing-state and missing-position handling
Repeated states preserve / collapse policy
Duration whether durations are analysed, weighted, or ignored
Motifs length range, overlap policy, prevalence denominator, filters
Distances distance family, normalisation, insertion/deletion/substitution costs
Clustering method, linkage, k, seed, validation and stability checks
Networks weight definition, normalisation, self-transition policy
HMMs state count, components/channels, seed, convergence, likelihood criteria
Group inference design declaration, statistic, assignment mechanism, permutations
Time-varying models target transition, smoothing / basis specification, prediction target
Software gp3sequencespy version and optional backends
Reproducibility seed(s), analysis audit, environment / lockfile where relevant

Example methods paragraph

Ordered categorical sequences were analysed with gp3sequencespy 0.1.1. Sequence order was defined explicitly by the event-order field. Missing states and duplicate positions were treated as errors and repeated states were preserved. Whole-sequence dissimilarity was quantified using normalized LCS distance. A two-cluster average-linkage hierarchical solution was summarised descriptively and evaluated with cluster validation and bootstrap stability. First-order transition networks used row-normalized outgoing transition proportions. All substantive interpretation was restricted to observed sequence structure.

Adapt that paragraph to the actual analysis. Do not copy settings that were not used.

Reporting motifs

State:

  • motif length range;
  • whether overlapping occurrences were allowed;
  • the sequence-prevalence denominator;
  • minimum occurrence / sequence / prevalence filters;
  • ranking and top-N rules.

Reporting distance and clustering

State:

  • distance method and normalization;
  • any non-default edit / substitution costs;
  • clustering algorithm and linkage;
  • k;
  • medoid / representative rule;
  • validation criteria;
  • bootstrap design, number of repetitions, and seed.

Reporting transition networks

State:

  • whether self-transitions were included;
  • whether edges are counts, global shares, or conditional outgoing shares;
  • any thresholding / smoothing;
  • how centrality and communities were computed.

Warning

Network centrality does not automatically represent attentional priority, importance, influence, or preference.

Reporting HMMs

State:

  • number of latent states;
  • mixture components or channels where relevant;
  • starting / seed strategy;
  • pseudocount or smoothing;
  • optimization tolerance and iteration cap;
  • convergence status;
  • likelihood, AIC/BIC where available;
  • decoding method;
  • sensitivity to alternative seeded fits.

Hidden-state labels are exchangeable statistical labels.

Reporting randomization inference

State:

  • declared design;
  • grouping / assignment variable;
  • exchangeability or randomization unit;
  • statistic;
  • permutation count;
  • seed;
  • effect estimate separately from the randomization p-value.

A p-value does not repair an invalid assignment mechanism.

Reporting software and provenance

import gp3sequencespy as g

print(g.__version__)
print(g.sequence_capabilities())

Also retain the analysis environment or lockfile when the work is intended to be exactly reproducible.

Citation

Cite the software version used and archive DOI:

DOI: 10.5281/zenodo.22166449

See the repository CITATION.cff for machine-readable citation metadata.

Interpretation language

Prefer:

  • “sequence structure differed between groups”;
  • “the observed transition share was higher”;
  • “the clustering solution separated these trajectories under the declared distance”;
  • “the latent model assigned observations to two statistical states”.

Avoid unsupported language such as:

  • “participants paid more attention”;
  • “the cluster represents a cognitive strategy”;
  • “centrality proves importance”;
  • “the HMM identified emotional states”;
  • “the group difference is causal” without a valid design.