Multimodal Process IRT: Responses, Time, Gaze, and Missingness

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

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

Measurement channels, not feature dumping

The 0.7 architecture treats response-process observations as explicit measurement channels. A process variable is not automatically useful merely because it predicts an outcome. It should have a declared role, latent target, family, provenance, and validation programme.

spec <- irt_model_spec(
  id = "accuracy_time_gaze",
  latent = c("ability", "speed", "engagement"),
  channels = list(
    response = irt_response_channel("2pl"),
    rt       = irt_rt_channel("lognormal"),
    gaze     = irt_count_channel("negative_binomial")
  ),
  status = "experimental"
)
spec

Other channels include nominal choices, survival/event time, compositional AOI measurements, process sequences, functional trajectories, and bounded continuous process measures.

irt_continuous_channel("censored_normal", value = "evidence_dwell_proportion")
irt_sequence_channel("scanpath", family = "hmm")

Registry

list_irt_models()

Models can be registered and later promoted only after their validation evidence passes an explicit gate.

register_irt_model(spec)
validate_irt_model("accuracy_time_gaze", validation_data)
promote_irt_model("accuracy_time_gaze", evidence = evidence_object)

Response + response time + gaze

fit_joint_gaze_rt_irt() supports two roles:

• engine = "reference" gives a transparent crossed-effects decomposition for development and validation;
• engine = "brms" builds a multivariate Bayesian model with shared grouping identifiers, which is the preferred route when a full Bayesian joint model is scientifically required.

fit <- fit_joint_gaze_rt_irt(
  data = trials,
  response = "correct",
  rt = "rt_ms",
  gaze = "fixation_count",
  person = "person_id",
  item = "item_id",
  gaze_family = "negative_binomial",
  engine = "brms"
)
plot(fit)

The function does not claim that a convenient reference engine is identical to the published three-way Bayesian model. That distinction is kept in the fit metadata.

Graded responses

The same idea extends to ordinal/graded outcomes:

fit_joint_graded_rt_process_irt(
  data = trials,
  response = "rating",
  rt = "rt_ms",
  process = "fixation_count",
  person = "person_id",
  item = "item_id",
  engine = "brms"
)

This is experimental until parameter recovery and external validation are completed.

Nominal distractors + option gaze

Binary correct/incorrect scoring discards which alternative was selected. A nominal process model can retain both the selected option and visual consideration of each option.

fit <- fit_nominal_gaze_irt(
  data = option_trials,
  response_option = "choice",
  option_gaze = c("dwell_A", "dwell_B", "dwell_C", "dwell_D"),
  item = "item_id",
  person = "person_id"
)

option_process_information(fit)
distractor_process_map(fit)
audit_distractor_attention(fit)
plot(fit)

Interpretation should stay process-based: an option attracted or retained more visual processing. This does not establish why.

Missingness as a process

missing <- classify_item_missingness(
  trials,
  response = "response",
  reached = "reached",
  inspected = "inspected_response_region",
  started = "started_response"
)

audit <- fit_omission_survival_irt(
  data = missing,
  response = "correct",
  response_time = "rt",
  omission_time = "elapsed",
  reached = "reached",
  person = "person_id",
  item = "item_id"
)
plot(audit)

The classification separates not reached, reached but not inspected, inspected omission, and started-but-unanswered cases instead of converting them all to NA.

Device and algorithm facets

facet_fit <- fit_manyfacet_process_irt(
  data = trials,
  response = "correct",
  process = "fixation_count",
  person = "person_id",
  item = "item_id",
  device = "device",
  session = "session",
  algorithm = "fixation_algorithm"
)

facet_effects(facet_fit)
audit_process_measurement_invariance(facet_fit)
plot(facet_fit)

A complementary generalizability_process_study() decomposes variance before a full measurement model is attempted.

Bounded gaze measures

AOI proportions and similar process quantities often have real mass at 0 and 1. The conditional censored-normal calibration helper respects those bounds rather than silently applying ordinary Gaussian regression.

cn <- fit_censored_normal_process_irt(
  response_matrix = aoi_proportion_matrix,
  theta = calibration_theta,
  lower = 0,
  upper = 1
)
predict(cn, theta = seq(-2, 2, length.out = 9))

This is conditional calibration given supplied theta; it is not labelled as the full marginal EM estimator from the 2026 CNRM paper.

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