W3 R82: actual original and independent MCAR/MAR/MNAR source evidence ledger¶

Numerical source and plot provenance · 21 October 2026 · synthetic research only

Generating model and four paired observation processes¶

The R82 original prospective protocol freezes 32 independent new primitive generating blocks evaluated under 2 independent-unit configurations (72 separate curves or 36 participants each with three trials), 2 extreme-predictor support counts (2 and 16), and 2 unit-level variance laws (constant and predictor heteroskedastic). 256 true complete-source datasets each have 4 paired masks (COMPLETE/MCAR/MAR/MNAR): 1,024 mechanism evaluations, not 1,024 statistically independent replications.

The generating complete source is synthetic nonseparable Gaussian spatial/temporal noise with Student-t5 independent-unit radial shocks plus mean-zero rare asymmetric unit shocks. The observation laws are MCAR p=.12, MAR depending only on recorded predictor/time/trial, and MNAR also on the original subsequently missing response residual.

Figure 1. Analytic simulation probabilities, not estimated empirical gaze-dropout probabilities.

Main design: sample retention and refused grid locations¶

Figure 2. Primary adverse two-anchor predictor-heteroskedastic design, 32 independent primitive blocks/cell.

No grid position failing at least eight independent units, rank-two design and both ±4 independent-unit anchors receives an estimated slope. This is a prespecified conservative support rule, not an impossibility theorem for generic OLS. Every source under all missing-data laws preserved the stable native NaN rejection contract.

All eight design/support/variance cells: observed fraction and grid refusal¶

Each cell below has 32 independent original primitive source blocks. Values list MCAR / MAR / MNAR.

The difference between curves and participants reflects the explicitly prespecified choice to average only available trials per participant before regressing. Neither masked trial observations nor 1,024 mechanism rows are pseudo-replicated into 1,024 independent units.

Pointwise slope changes conditional on estimable positions¶

Figure 3. Original source-averaged absolute change in pointwise estimated slope between a missingness mask and that source's corresponding complete data, on masked-grid locations that passed the independent-unit/anchor support gate.

These selected-grid descriptive means are not MNAR causal effects or unbiased slope-bias estimates: both the set of estimable positions and observed retention differ across masking mechanisms. The original source archive additionally preserves per-source signed/absolute error against generating slope truth, observed native-unit counts, exact per-coordinate refusal and support matrices, whole-curve/whole-unit survival, and source/mask checksums.

Immutability and independent replication¶

R82-A independently reproduces the original, actually archived source, not just a synthetic fixture. It validates numerical agreement on this artificial data/observation generating mechanism. It does not establish that any real eye-tracker's loss follows these probabilities, that MAR is identified in field data, or that outcome-dependent missingness can be corrected from observed values alone.

Read scientific R82 article · Read methods companion · Previous fully observed R81 audit