Experimental Process-IRT Methods and Evidence Gates

In eyeprocess: Harmonize Eye-Tracking, Pupillometry, Biometrics, and Psychometric Process Data

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

Why these functions are gated

Some methods are scientifically attractive but too new, too estimator-specific, or too computationally demanding to present as production estimators without a validated implementation. In those cases eyeprocess provides one of three things:

• a transparent reference model;
• an adapter to an established external package;
• an explicit external-engine gate that refuses to fake the estimator.

Process-state HMM + IRT

hmm <- fit_process_hmm_irt(
  data = events,
  sequence_id = "person_item",
  process_features = c("stem_dwell", "option_dwell", "transition_rate"),
  response = "correct",
  person = "person_id",
  item = "item_id",
  n_states = 3
)
process_state_occupancy(hmm)
process_state_transition_summary(hmm)
plot(hmm)

The internal HMM is an interpretable two-stage reference engine. State labels are statistical summaries and should not be named as unobserved mental states without independent validation.

Cognitive diagnosis and latent process classes

cdm <- fit_cognitive_diagnosis_process(
  response_matrix = response_matrix,
  q_matrix = q_matrix,
  process_data = process_data,
  process_features = process_features
)

mix <- fit_latent_class_process_irt(
  data,
  process_features = c("fixation_count", "rt", "transition_entropy"),
  response = "correct",
  person = "person_id",
  item = "item_id",
  n_classes = 3
)

Latent-space IRT

fit_latent_space_irt() is an adapter to LSMjml, avoiding a home-grown approximation when a current R implementation exists.

ls <- fit_latent_space_irt(response_matrix, dimensions = 2)
map <- process_residual_map(ls)
plot(ls)

validate_latent_space_process_similarity(
  ls,
  process_matrix = scanpath_feature_matrix
)

This allows a new validation question: do person-item residual proximities agree with independently measured process similarity?

Process-adjusted DIF

surrogate <- process_dif_nuisance_surrogate(
  data,
  process_features = c("rt", "fixation_count", "stem_revisits")
)

audit_process_adjusted_dif(
  data,
  response = "correct",
  ability = "theta",
  group = "group",
  item = "item_id",
  process_features = c("rt", "fixation_count", "stem_revisits"),
  person = "participant_id"
)

Process adjustment should be reported transparently: which nuisance surrogate was used, whether conclusions changed, and whether the process channel itself may be group-dependent.

Sequence representations

ngrams <- process_ngram_features(sequences, n = 2:4)
emb <- process_sequence_embedding(sequences, dimensions = 8)
fit_response_process_embedding_irt(
  data,
  sequences = sequences,
  response = "correct",
  person = "person_id",
  item = "item_id"
)

Flexible item-response curves

gp <- fit_gpirt(response_matrix, engine = "spline_reference")
compare_parametric_nonparametric_irf(gp)
audit_irf_shape(gp)
plot(gp)

The spline reference is a shape audit, not a Gaussian-process posterior. Exact GPIRT remains behind external_engine until a validated engine is chosen.

fit_dynamic_gpirt(data, external_engine = my_validated_dynamic_gpirt)
fit_continuous_time_irt(data, external_engine = my_validated_ct_irt)
fit_flow_mirt(response_matrix, external_engine = my_validated_flow_mirt)
fit_variational_irt(response_matrix, external_engine = my_validated_vi_engine)

A missing engine produces a deliberate error instead of silently substituting a different model.

Linking and person fit

link <- equate_irt_scales(reference_parameters, new_parameters,
                          method = "stocking_lord")
plot(link)

pf <- process_person_fit(
  joint_fit,
  data = trials,
  person = "person_id"
)
plot(pf)

Person-fit output describes model-process inconsistency. It must not be relabelled as cheating, deception, disengagement, or pathology without separate evidence.

Process-aware adaptive testing

info <- process_item_information(theta, a, b,
                                 process_information = process_information,
                                 rt_information = rt_information,
                                 weights = c(response = 1, rt = .25, process = .25))
expected_process_information(info)
select_next_item_process(theta, item_bank)
simulate_process_cat(item_bank, true_theta = 0, n_items = 10)

The adaptive functions are research utilities. They should not be deployed in a high-stakes adaptive assessment until item-selection bias, exposure, fairness, measurement invariance, and stopping rules have been separately validated.

Promotion rule

Experimental methods should remain experimental until they pass the same simulation, calibration, misspecification, preprocessing, and external-validation contract as the simpler models. Novelty is not evidence.

eyeprocess documentation built on Sept. 28, 2026, 5:08 p.m.