41. TSBI (Trek Separation Blocks Independence)
Command: tsbi
Kind: Independence test
The description below is Tetrad’s own, shown in the interface when you select this method. It is reproduced here so that the method is documented; it has not yet been edited for the manual, and it has no “when to use” guidance of the kind the hand-written pages carry.
A block-level conditional independence test based on trek separation, operating under linear latent factor measurement models. This class implements the Trek Separation Blocks Independence (TSBI) test, elaborating on the IndTestTrekSep class of Adam Brodie and Peter Spirtes. Setting Variables are organized into blocks, where each block corresponds to a set of observed indicator variables that measure a common latent factor. The independence test operates at the block level: X, Y, and the conditioning set Z = {Z1, …, Zk} are each a block (latent factor), and the test assesses whether the latent factor behind X is independent of the latent factor behind Y given the latent factors behind Z1, …, Zk. Test logic The test is based on the rank of the cross-covariance matrix between two constructed variable sets L and R: Each conditioning block Zi is split into two nearly equal halves ZiA and ZiB. The left side is formed as L = cols(X) ∪ cols(Z1A) ∪ … ∪ cols(ZkA). The right side is formed as R = cols(Y) ∪ cols(Z1B) ∪ … ∪ cols(ZkB). The rank of the cross-covariance submatrix Σ_{L,R} is estimated via RankTests.estimateWilksRank. Under a linear measurement model with k conditioning latent factors, trek separation implies that X ⊥ Y | Z1,…,Zk if and only if: rank(Σ_{L,R}) = sum of ranks(Zi) where the rank of each Zi block is specified externally via BlockSpec.ranks(). The test declares independence when the estimated rank equals this target sum. Splitting strategy: Each conditioning block Zi is split either deterministically or randomly: When randomizeSplits=false, the same fixed split (determined by splitSeed) is used for every query. When randomizeSplits=true (default), each trial uses a different seed derived from splitSeed + trialIndex, allowing the minimum rank over multiple trials to be taken. This reduces sensitivity to unlucky splits for odd-sized blocks. When a block has an odd number of indicators, the leftGetsSmallerHalfWhenOdd flag controls whether the left or right side receives the extra column. Multiple trials: When numTrials > 1 and randomizeSplits=true, the test runs the full rank estimation procedure numTrials times with different random splits and takes the minimum estimated rank. This provides a more conservative (less likely to falsely declare dependence) result when block sizes are small. Caching: Rank estimates are cached in a thread-safe LRU cache keyed by the left and right column sets together with all hyperparameters that affect the rank estimate (sample size, alpha, seed, randomize flag, and number of trials). The cache holds up to 400000 entries. P-values: This test does not produce p-values. All IndependenceResult objects returned carry NaN for both the p-value and the alpha-minus-p fields. The independence decision is made purely by comparing the estimated rank to the target rank.