Class NabiBoundaryStudy

java.lang.Object
edu.cmu.tetrad.study.missingness.NabiBoundaryStudy
public final class NabiBoundaryStudy extends Object
Companion study to NabiIdentificationProbe: does the Nabi et al. (2020) full-law identification boundary (no self-censoring, no colluders) predict where practical causal discovery methods degrade?

To be explicit about what this is NOT: none of the arms below implements the paper's odds-ratio identifying functional, and full-law identification is neither necessary nor sufficient for any of them -- test-wise deletion is governed by admissible-separator-type conditions (Tu et al. 2019; Witte et al. 2021), imputation by MAR-style recoverability, and the paper itself notes the target law can be identified when the full law is not. So this study measures whether the theoretical boundary happens to coincide with the practical-degradation boundary, which is an empirical question the literature leaves open.

Linear Gaussian SEM, p = 10, 20 edges, 4 variables incomplete at a 30% rate each (NaN-marked missingness, unlike the zero-coded dropout study). Three mechanisms, one per side of their boundary:

  • ICIN_MNAR (identified side): each incomplete variable's indicator depends on the TRUE values of two OTHER incomplete variables -- genuinely MNAR, since the drivers are themselves sometimes missing -- with no self-censoring edges, and no edges between indicators, hence no colluders. By their Theorem 1 the full law is identified;
  • SELF_CENSOR (not identified): each indicator depends on its own variable's value (logistic self-masking, the NaN-marked cousin of the dropout study's mechanism);
  • COLLUDER (not identified): incomplete variables are paired; within a pair (a, b), R_b depends on the value of a third variable, and R_a depends on the VALUE of X_b and on R_b, forming the colluder X_b -> R_a <- R_b.

Arms: ORACLE BOSS on the complete data; LISTWISE BOSS; TESTWISE PC (Fisher-Z, testwise spec); MICE-PMM m = 5 + IMaGES BOSS (chains in parallel); MVN-EM covariance BOSS; MEAN-IMPUTE BOSS. All continuous.

Predictions on record: SELF_CENSOR hurts every arm (both boundaries agree there). The interesting cells are ICIN_MNAR -- identified full law, yet no arm uses the identifying functional, so degradation here would show the boundary is not sufficient for practical methods -- and COLLUDER, where degradation beyond ICIN_MNAR's would be consistent with the boundary mattering, though not proof that it does.

Repetitions run concurrently (RandomUtil is thread-local; each rep seeds its own thread). Run from IntelliJ; one line per arm per rep streams as it finishes. Nothing outside this file is touched.

Author:
josephramsey

  • Method Details

    • main

      public static void main(String[] args) throws Exception
      Entry point: one table per mechanism, reps in parallel.
      Parameters:
      args - Unused.
      Throws:
      Exception - If a condition fails irrecoverably.