Model-readiness and sensitivity analysis

This article summarises model-readiness checks and sensitivity workflows.

Model-readiness principle

Before modelling, check whether the analysis table contains the expected participant, trial, condition, outcome, predictor, and covariate columns. Also inspect missingness, group sizes, influential observations, and scale compatibility.

Example workflow

model_data <- prepare_gazepoint_multimodal_data(
  face_windows = face_windows,
  gaze_data = gaze_summary,
  response_data = response_data,
  by = c('participant_id', 'trial_id'),
  predictors = c('AU12_r_mean', 'claim_dwell_ms'),
  outcome = 'rating'
)

fit <- fit_gazepoint_multimodal_response_model(
  model_data,
  outcome = 'rating',
  predictors = c('AU12_r_mean', 'claim_dwell_ms'),
  covariates = c('trial_order')
)

loo <- run_gazepoint_model_leave_one_out(fit)
nested <- compare_gazepoint_nested_models(fit_reduced, fit_full)
pred <- plot_gazepoint_model_predictions(fit)

Sensitivity examples

• leave-one-participant-out or leave-one-trial-out checks;
• nested model comparisons;
• robustness to covariates;
• robustness to exclusion thresholds;
• robustness to alternative AOI or baseline definitions.

Reporting note

Sensitivity analyses should be described as robustness checks. They should not be used to selectively report only favourable specifications.