From raw gaze to a defensible analysis¶
A gaze file becoming a dataframe is the start of an analysis, not evidence that the measurement is understood.
Preserve source identity¶
Keep the source export, tracker or acquisition identity, participant/session/trial identifiers, timestamp representation, coordinate basis, geometry, nominal sampling rate, and observed timestamp cadence explicit.
A parser that can read a file establishes compatibility. It does not establish device validity, calibration quality, or scientific suitability.
Canonicalise without erasing provenance¶
A canonical schema should make downstream analysis easier while retaining enough information to reconstruct what was transformed. Unit conversion, coordinate conversion, duplicate handling, timestamp repair, and sample-rate assumptions should therefore be explicit.
Treat QC as evidence¶
Missingness, discontinuity, extreme velocity, geometry violations, and sampling irregularity can be measured. Those diagnostics should remain attached to the data rather than becoming hidden deletion rules.
Separate review from exclusion¶
A QC rule can identify observations requiring review. The decision to retain or exclude belongs in a separate ledger with the reason, decision level, denominator impact, and whether the rule was prespecified.
Derive events and AOIs after the measurement contract is clear¶
Fixations, saccades, dynamic AOIs, semantic assignments, and scanpaths depend on sampling, geometry, detector definitions, stimulus support, and review status.
Preserve the inferential unit¶
Samples, fixations, trials, stimuli, AOIs, and participants are not interchangeable independent observations. Before statistical modelling, preserve the grouping structure required by the design.
Report the evidence class you actually have¶
Reproducible code, a successful parser, a synthetic benchmark, held-out dataset evidence, native-device validation, and construct validity answer different questions.
A defensible analysis makes visible what was observed, transformed, reviewed, modelled, and still uncertain.