13. Generalized Information Criterion (GIC) Scores
13.1. Summary
GIC (Generalized Information Criterion) Scores form a family of criteria that generalize AIC and BIC by allowing more flexible penalty terms. In Tetrad, GIC scores provide a tunable framework for balancing fit and complexity beyond the standard BIC penalty.
13.2. When to use
You want more flexibility than BIC or EBIC in penalizing model complexity.
You are exploring alternative trade-offs between fit and sparsity.
You are running score-based or hybrid algorithms and wish to compare different penalty regimes.
13.3. Model class
Follows the same model classes as the underlying likelihood (SEM, discrete BN, etc.), but with a generalized penalty structure.
13.4. Score form (conceptual)
A generic GIC score can be written as:
GIC = 2 * logL − c(N, p, k)
where c(N, p, k) is a user-defined or theory-motivated penalty function
depending on sample size N, dimension p, and parameter count k.
13.5. Parameters
Parameter (camelCase) |
Description |
|---|---|
|
Choice of generalized information criterion rule for SEM. Selects which GIC formula is applied (for example, a Zhang–Shen–type rule versus more BIC-like variants). In the GUI, this is exposed as a drop-down of named options; the underlying code uses an integer or enum to represent the choice. |
|
Double ≥ 0.0. Penalty discount (or weight) specific to the Zhang–Shen–style GIC rule. Larger values impose a stronger complexity penalty under that rule and tend to yield sparser graphs; smaller values allow denser graphs. Has an effect only when a Zhang–Shen–type GIC rule is selected. |
|
Boolean. If |
|
Double. Handles singular or nearly singular covariance matrices. If |
|
Double > 0, or |
13.6. Strengths
Highly flexible; can replicate AIC, BIC, EBIC, and other criteria as special cases.
Useful for sensitivity analysis and method comparison.
13.7. Limitations
Theoretical guarantees depend on the specific penalty chosen.
Parameter tuning can be nontrivial.
13.8. References
Kim, Y., Kwon, S., & Choi, H. (2012). Consistent model selection criteria on high dimensions. Journal of Machine Learning Research, 13(1), 1037–1057.