M3 process information: ablation, redundancy, and sensor value

Why eight models?

The central 0.10 question is not whether adding sensors makes a model more complicated. It is whether a process channel provides additional measurement information for a defined target. Because RT, gaze and pupil may be correlated, their contributions are not assumed to add linearly.

M3 therefore defines the complete response-anchored lattice:

eyeprocess:::.ep10_m3_ablation_definitions()
#>          model    rt  gaze pupil                     channels
#> 1            R FALSE FALSE FALSE                     response
#> 2         R_RT  TRUE FALSE FALSE                response + RT
#> 3       R_GAZE FALSE  TRUE FALSE              response + gaze
#> 4      R_PUPIL FALSE FALSE  TRUE             response + pupil
#> 5    R_RT_GAZE  TRUE  TRUE FALSE         response + RT + gaze
#> 6   R_RT_PUPIL  TRUE FALSE  TRUE        response + RT + pupil
#> 7 R_GAZE_PUPIL FALSE  TRUE  TRUE      response + gaze + pupil
#> 8         FULL  TRUE  TRUE  TRUE response + RT + gaze + pupil

The eight models are response only; response + RT; response + gaze; response + pupil; each two-process-channel combination; and the full four-channel model.

Fit the ablation lattice

sim <- simulate_multimodal_m3(n_person = 100, n_item = 12, seed = 20260815)

ab <- multimodal_m3_ablation(
  sim,
  chains = 4,
  parallel_chains = 4,
  iter_warmup = 750,
  iter_sampling = 750,
  refresh = 0
)

info <- multimodal_m3_process_information(ab)
print(info)

multimodal_m3_process_information() uses response-target PSIS-LOO and posterior variance of person ability. It does not sum channel Fisher information under a correlated joint model.

Incremental pupil evidence

Pupil is compared with and without the channel in four contexts: response only, response + RT, response + gaze, and response + RT + gaze. The paired response-ELPD contrast is accompanied by a standard error and a descriptive evidence classification. The classification can return no_clear_incremental_pupil_information; this is an intended scientific outcome, not a failure of the package.

info$incremental_pupil
plot(info, type = "incremental_pupil")

Redundancy and complementarity

The non-additivity table contrasts the full model with the sum of single-channel additions and asks whether the incremental pupil gain is attenuated or amplified after RT and gaze are already included.

info$nonadditivity
plot(info, type = "redundancy")

These are model-conditional predictive contrasts. “Synergy” in this table means non-additivity on the response ELPD scale; it does not establish a causal interaction among psychological processes.

Sensor value and channel conflict

M3 adds two deliberately practical diagnostics. sensor_value reports pupil response-target gain per usable pupil observation and per analyst-supplied relative sensor cost. channel_conflict places predictive performance beside convergence/stability diagnostics. A sensor can improve an in-sample latent representation while worsening response prediction or computational geometry; M3 surfaces that conflict rather than hiding it behind one scalar rank.

info$sensor_value
info$channel_conflict
plot(info, type = "sensor_value")
plot(info, type = "conflict")

These diagnostics are not economic cost-effectiveness analyses and not causal estimates. Their role is to prevent “more modalities” from becoming an automatic conclusion of “more information.”