Bayesian and 3PL Process Diagnostics

Scope

This article adds two diagnostic layers that were present in the source research templates but should remain separate from substantive behavioral claims.

1. bayesian_process_diagnostics_dashboard() collects LOO, posterior convergence/effective-sample-size summaries, and optionally a Bayes factor for fitted brms process models.
2. fit_gaze_anchored_3pl_audit() fits a standard psychometric 3PL model and descriptively aligns its item lower-asymptote parameter with gaze, pupil, response-time, or accuracy summaries.

Neither function establishes a causal cognitive mechanism. In particular, a 3PL lower asymptote is an item parameter and is not a participant-level “guessing detector”.

Bayesian diagnostics

dash <- bayesian_process_diagnostics_dashboard(
  response_only = brms_response_model,
  response_plus_pupil = brms_pupil_model,
  compute_loo = TRUE,
  compute_bayes_factor = FALSE
)

bayesian_process_diagnostic_flags(dash)
plot(dash, type = "loo")
plot(dash, type = "rhat")

Bayes factors are deliberately opt-in because they require suitable model fitting settings and answer a different evidential question than predictive LOO comparison.

3PL response-process alignment

three_pl <- fit_gaze_anchored_3pl_audit(
  response_matrix = binary_response_matrix,
  process_data = binary_long,
  item = "item_id",
  process_features = c("ttff_ms", "dwell_ms", "pupil_peak", "rt_ms", "accuracy")
)

gaze_anchored_3pl_alignment(three_pl)
audit_3pl_process_signatures(three_pl)
plot(three_pl, type = "lower_asymptote")
plot(three_pl, type = "process_alignment", feature = "ttff_ms")

The resulting correlations and review flags are descriptive item-level diagnostics. They require independent substantive validation before any interpretation in terms of rapid responding, guessing, effort, engagement, or strategy.