Class WitteReplicationStudy
Their claims under test, labeled as in our review:
- W1: test-wise deletion and multiple imputation both clearly beat list-wise deletion and single mean/mode imputation (which produces too-dense graphs);
- W2: for Gaussian data under MCAR and MAR, MI beats test-wise deletion, with the gap invisible at n = 100 and largest at n = 5000;
- W3: under their MNAR design (key-variable deletion, built to satisfy the admissible separator condition), test-wise deletion stays valid while MI does not;
- W4: for mixed data, test-wise deletion beats MI in all their scenarios;
- W5: their warning against predictive mean matching (increased type I error, results not shown). Our MICE arms use PMM, so depressed MICE precision under Gaussian MCAR at large n would support them, and clean precision there would not.
Design: two phases (Gaussian p = 12, 17 edges, their ECOLI shape; mixed p = 8, 9 edges, half discrete with 3 categories, their HEALTHCARE shape); three mechanisms at their 18% overall rate (MCAR; their ampute-style MAR with exactly one group member missing per row, chosen by the other members' values, topped up by one MCAR column; their MNAR with the key variable's largest rows deleting the key and four subordinates); n in {100, 1000, 5000}.
Arms per phase: ORACLE BOSS on the complete data; LISTWISE PC (complete rows only); MEAN/MODE-IMPUTE PC;
TESTWISE PC (Fisher-Z or DG-LRT with the testwise spec); MICE-PMM m = 5 with test-level pooling inside PC,
which is their actual procedure (Rubin-pooled Fisher z in the Gaussian phase; Meng-Rubin D3 pooling of the
DG likelihood ratio via PooledDgLrtTest in the mixed phase); MICE-PMM m = 5 + IMaGES BOSS (SEM-BIC
or DG components); and, Gaussian only, MVN-EM covariance BOSS. Including both MI arms separates "MI versus
deletion" from "PC versus BOSS", which their design confounds.
Parallelism: repetitions run concurrently (RandomUtil is thread-local, and each rep seeds its own thread's instance from the condition and rep index, so results are reproducible regardless of scheduling), and the m MICE chains within an arm run concurrently (the imputer takes an explicit seed and uses no shared state across instances). Arms within a rep run sequentially on the rep's thread.
Run from IntelliJ; one line per arm per rep streams as it finishes, and a table prints per condition. Nothing outside this file is touched.
- Author:
- josephramsey
-
Method Summary
-
Method Details
-
main
-