Skip to content

Interpreting AOI Assignment Stability

Assignment stability compares each perturbed branch with the nominal AOI assignment. It is a measurement-robustness diagnostic, not an estimate of whether one geometry is scientifically “true.”

Core quantities

The framework reports proportions unchanged, newly assigned, lost, and reassigned; AOI-to-AOI reassignment matrices; AOI-level stability; and optional participant- and trial-level stability.

Fixation versus sample level

Declare the row type explicitly with observation_level="fixation" or "sample".

Both levels use the same geometry and assignment engine. Recomputed features always include universal observation_count and first_observation fields. Fixation input additionally populates fixation_count and first_fixation; sample input populates sample_count and leaves fixation-specific fields missing. Sample rows are therefore never relabeled as fixations.

Zero is not missing

When a complete AOI universe is supplied by the sensitivity runner, participant/trial × AOI cells are retained even when nothing lands in an AOI.

  • 0 means the group had valid assignment opportunities but none were assigned to that AOI.
  • Missing count/inspection means the group had no valid assignment opportunity, for example because all coordinates were missing.
  • If assigned observations have missing duration values, dwell remains missing rather than being partially summed.
  • First-observation timing is never inferred from row order.

These rules keep denominators auditable across perturbation branches.

Missing coordinates and overlap

Missing x/y values remain missing and are excluded from comparable-assignment denominators. They are never converted to zero or outside-AOI observations.

Multiple simultaneous memberships default to __ambiguous__. This prevents boundary overlap from being silently converted into a unique AOI.

Assignment frequency

estimate_fixation_assignment_probability() reports the empirical frequency with which an observation maps to each AOI across the declared perturbation set. Despite the historical function name, this is not a posterior probability of true membership.

Low stability signals sensitivity to geometry; it does not identify which boundary is substantively correct.

Continue to Propagating AOI Uncertainty Into Statistical Models or use the decision clinic to diagnose a stability pattern.