Class BossMultipleImputationStudy
MiceLiteImputer) and conditional draws from an EM-fit
multivariate normal (MvnImputer). The prior contenders--listwise, testwise, EM covariance, and pairwise
covariance--are included for reference, along with mean imputation as the naive baseline and the complete data
as the ceiling.
Each multiple-imputation engine appears twice, because there are two defensible ways to pool m completed datasets for a graph search, and they answer different questions:
- Graph pooling: run BOSS on each completed dataset and keep the edges that appear in at least half
of the m graphs (the
ImputationSearchpath, which is what a user of the MULTIPLE_IMPUTATION policy gets). The pooled graph can carry conflicting orientations, so it need not be a legal PDAG and SHD can be undefined for it. - Covariance pooling: average the m completed-data covariance matrices and run BOSS once on the average, at the mean-pairwise effective sample size. This is the Rubin-style pooling of the sufficient statistic rather than of the answers, always yields a legal CPDAG, and is the natural head-to-head comparison with EM_COVARIANCE, from which it differs only by the imputation noise.
Run this from IntelliJ; it streams a line per failure as it goes and a summary table per mechanism. With the defaults below it is a few minutes, most of it in the EM fits inside the MVN imputer.
What to expect, stated before running
The simulated data are exactly multivariate normal, so the MVN imputer's model is correctly specified and
MICE with linear conditionals is as well; the two should land close to each other. Under MCAR every arm is
consistent and the comparison is of efficiency only; differences of kind need the MAR or self-masking
mechanism, which is one entry in MECHANISMS away. Under self-masking (MNAR) no arm is valid, and the
point of including it is to see how gracefully each degrades.
- Author:
- josephramsey
-
Constructor Summary
ConstructorsConstructorDescriptionDefault constructor for the BossMultipleImputationStudy class. -
Method Summary
Modifier and TypeMethodDescriptionstatic voidThe entry point: for each mechanism, simulates data, injects missingness, runs every arm on every rep, and prints a summary table of adjacency precision and recall, arrowhead precision and recall, structural Hamming distance where defined, and seconds per rep.
-
Constructor Details
-
BossMultipleImputationStudy
public BossMultipleImputationStudy()Default constructor for the BossMultipleImputationStudy class.
-
-
Method Details
-
main
The entry point: for each mechanism, simulates data, injects missingness, runs every arm on every rep, and prints a summary table of adjacency precision and recall, arrowhead precision and recall, structural Hamming distance where defined, and seconds per rep.- Parameters:
args- Command-line arguments; not used.- Throws:
Exception- If any error occurs during simulation or search.
-