Capabilities¶
The full inventory of what Antecedent implements, area by area. The README carries the highlights; this page is the reference list. For what is deliberately not implemented, see Comparison.
Graphs¶
Supported graph classes:
- DAG;
- ADMG;
- CPDAG;
- PAG;
- temporal DAG;
- temporal CPDAG;
- temporal PAG.
Graph operations:
- d-separation;
- m-separation;
- districts;
- latent projection;
- Markov-equivalence completions;
- definite-status separation;
- temporal unfolding;
- intervention overlays.
Static and temporal graphs have separate semantics. A static graph is not interpreted as temporal by default.
Graph interchange is available through NetworkX, DOT, JSON, GML, and versioned CBOR artifacts.
Discovery¶
Static¶
- PC
- FCI
- RFCI
- GES
- DirectLiNGAM
- NOTEARS
Temporal and multi-context¶
- PCMCI
- PCMCI+
- LPCMCI
- J-PCMCI+
- regime-specific RPCMCI workflows
Bayesian structure learning¶
- exact DAG posterior;
- order MCMC;
- structure MCMC;
- CI-screened graph posterior;
- DBN posterior.
Posterior graph samples can be propagated into downstream effect analyses.
Conditional independence tests¶
- partial correlation;
- weighted and robust partial correlation;
- regression CI;
- k-nearest-neighbour CI;
- mixed k-nearest-neighbour CI;
- symbolic conditional mutual information;
- GPDC;
- G²;
- oracle tests;
- Bayesian CI tests.
Multiplicity corrections include BH, BY, Bonferroni, and Holm.
Discovery stability tools include block bootstrap, lag and threshold sensitivity, orientation stability, environment holdout, synthetic-null checks, and permutation or phase-randomized surrogates.
Identification¶
Antecedent reports whether a query is:
- nonparametrically identified;
- partially identified;
- graph-dependent;
- not identified.
Implemented identification strategies:
- backdoor adjustment;
- efficient backdoor adjustment;
- front-door identification;
- instrumental variables;
- sharp regression discontinuity;
- ID and IDC for DAGs and ADMGs;
- hedge certificates;
- nonparametric path-specific identification;
- generalized adjustment for partial graphs;
- unfolded temporal backdoor;
- temporal mediation.
AutoIdentifier reports applicable strategies. It does not silently choose an
estimator.
For PAGs, Antecedent uses identification envelopes or explicit graph completions. Full PAG-native ID and IDC are outside the supported scope.
Estimation¶
Frequentist¶
- linear and generalized-linear outcome regression;
- g-computation;
- inverse probability weighting;
- propensity matching;
- covariate-distance matching;
- stratification;
- AIPW;
- front-door two-stage estimation;
- Wald estimation;
- 2SLS;
- sharp local-linear regression discontinuity;
- linear conditional effect models;
- temporal adjustment;
- temporal mediation;
- functional plug-in estimation.
Bayesian¶
- Bayesian g-computation;
- temporal Bayesian g-computation;
- conjugate Gaussian models;
- Laplace GLM approximation;
- HMC GLMs;
- graph-by-effect posterior envelopes;
- same-design prior transfer;
- effect-level and mapped prior transfer;
- prior catalogs and compatibility filtering;
- power-prior mixtures;
- conflict-sensitive prior weighting;
- transport policies across compatible designs.
Unidentified graph-posterior mass is retained rather than silently renormalized away.
Interventions and counterfactuals¶
Antecedent includes a structural causal model layer.
Supported mechanisms:
- linear-Gaussian models;
- constant mechanisms;
- discrete mechanisms;
- hierarchical linear and generalized-linear models;
- Minnesota BVAR;
- linear Gaussian state-space models;
- Gaussian-process mechanisms.
Supported interventions:
- hard interventions;
- soft interventions;
- stochastic interventions;
- sequenced interventions;
- temporal policies;
- dynamic policies;
- mechanism overrides.
Do-sampling methods include weighting, KDE, and MCMC.
Counterfactual support:
- abduction–action–prediction;
- nested counterfactuals;
- temporal trajectories;
- unit-level counterfactual analysis.
Attribution and diagnostics¶
Antecedent can analyze:
- anomalous outcomes;
- distribution shifts;
- structural changes;
- mechanism changes;
- change points;
- unit-level change;
- path contributions;
- arrow strength;
- feature relevance;
- root-cause rankings.
Implemented techniques:
- likelihood-ratio tests;
- mean-difference tests;
- classifier-based tests;
- MMD;
- Gaussian KL divergence;
- CUSUM-style scans;
- Shapley attribution;
- coalition caching.
Validation and sensitivity¶
Estimate validation:
- placebo refuters;
- random common-cause refuters;
- unobserved common-cause refuters;
- bootstrap refuters;
- data-subset refuters;
- dummy-outcome refuters;
- overlap diagnostics;
- E-values;
- graph refutation.
Sensitivity methods:
- linear sensitivity;
- partial-linear sensitivity;
- nonparametric sensitivity;
- Reisz sensitivity.
Bayesian validation:
- prior predictive checks;
- prior sensitivity;
- MCMC diagnostics;
- simulation-based calibration hooks.
Resampling support:
- IID bootstrap;
- Bayesian bootstrap;
- moving-block bootstrap;
- circular-block bootstrap;
- column permutation;
- phase-randomized surrogates.
Experimental design¶
Antecedent can rank candidate actions such as:
- measuring a variable;
- intervening on a variable;
- observing an environment;
- changing a sampling plan.
Ranking criteria:
- expected information gain;
- probability of identification;
- expected effect-interval width;
- decision utility.
The design layer supports batched Monte Carlo evaluation, common random numbers, and early stopping.
Incremental state¶
CausalState supports stateful and online workflows.
Available components:
- explicit invalidation;
- incremental OLS;
- streaming covariance;
- particle-filter state-space models;
- local score caches;
- rolling mechanism diagnostics;
- configurable cache budgets;
- prepared analyses;
- progressive and cancellable execution;
- adaptive resampling.
Invalidation does not automatically rerun an analysis.
Data support¶
Antecedent supports:
- tabular data;
- time series;
- panel data;
- multi-environment data;
- event data converted into temporal frames.
Python interfaces support NumPy, pandas, and Arrow CDI. Rust uses TableView.
Artifacts¶
Versioned artifacts:
- graphs;
- graph posteriors;
- model bundles;
- analysis traces;
- causal state.
Artifacts use schema-versioned CBOR containers with optional Zstandard-compressed sections, selective reads, and memory-mapped access.