Specifying and Fitting the M4 Reference Model
Source:vignettes/m4-reference-model-fitting-0-11.Rmd
m4-reference-model-fitting-0-11.RmdThe reference model uses a marginalized finite-state Markov process over explicitly ordered participant/session sequences. K=1 is a formal null and K=2 is the conservative default.
sim <- simulate_multimodal_m4(n_person = 30, n_item = 8, seed = 20260820)
audit <- audit_multimodal_m4_identifiability(sim, include_posterior = FALSE)
audit
#> <eye_multimodal_m4_identifiability>
#> overall: REVIEW
#> domain criterion status severity value
#> sequence sequence_count PASS none 30
#> sequence minimum_sequence_length PASS none 8
#> sequence transition_count PASS none 210
#> data response_observed PASS none 240
#> data rt_observed PASS none 240
#> data gaze_observed PASS none 240
#> data pupil_observed PASS none 240
#> data maximum_process_missing_fraction PASS none 0
#> model_complexity trait_conditioned_markov REVIEW moderate 2
#> threshold
#> >=2 preferred
#> >=2
#> >=20
#> >0
#> >0
#> >0
#> >0
#> <=0.50 preferred
#> 0 transition traits for unconditional Markov
#> message
#> Number of independent ordered sequences.
#> Shortest sequence relative to declared fitting requirement.
#> Available within-sequence transitions.
#> Observed response measurements.
#> Observed rt measurements.
#> Observed gaze measurements.
#> Observed pupil measurements.
#> Largest process-channel missing fraction.
#> Trait-conditioned Markov dynamics add person-level process slopes whose posterior identification cannot be established from structural counts alone.
#> recommendation
#> Use multiple persons/sequences for population-level state inference.
#> Do not silently discard short sequences; revise design or explicit threshold.
#> Transition parameters may be prior-dominated when few transitions are observed.
#> M4 cannot identify the requested response contribution without observed data.
#> M4 cannot identify the requested rt contribution without observed data.
#> M4 cannot identify the requested gaze contribution without observed data.
#> M4 cannot identify the requested pupil contribution without observed data.
#> Study missingness sensitivity; M4 reference fitting assumes ignorable missingness.
#> Treat this specification as gated: require satisfactory posterior R-hat/ESS, stable state occupancy across chains, and separated state emissions before interpretation.A deliberate CmdStanR fit is required for posterior inference; documentation does not compile or sample automatically.
fit <- fit_multimodal_m4(sim, spec = multimodal_m4_spec())
summary(fit)
fitted(fit, type = "state")