User guide

The user guide explains the choices that sit between the quickstart and a defensible applied analysis. Read it by task. Use the Python API for signatures and the technical reference for derivations and evidence.

The public workflow has four visible stages:

CausalStudy -> identify(typed estimand) -> IdentifiedEffect -> estimate(method) -> CausalResult

The study design owns the column roles. The estimand owns the causal question. The method owns learning, targeting, inference, and runtime settings. That separation is what lets cleverly refuse an unsupported combination before it fits any nuisance model.

The examples here are explanatory. The behavioural guarantees live in the fast tests and registered validation studies, not in an example.

• Data and study design
  • Accepted table backends
  • Point treatment
  • Missing outcomes and missing treatments
  • Continuous treatment
  • Longitudinal treatment
  • Observation weights are not estimand weights
• Estimands and interventions
  • Arm-indexed questions
  • Known regimes
  • Learned treatment rules
  • Modified treatment policies
  • Incremental propensity-score interventions
  • Marginal structural models
  • Controlled direct effects
• Methods, learners, and cross-fitting
  • Method presets
  • Immutable configuration
  • Learner choices
  • Use GLM for the sequential baselines
  • Use GLM for the parametric g-formula
  • Two fold layers
  • Reuse an outer split
  • Targeting and bounds
• Results, inference, and assessment
  • Run the assessment battery
  • Read the support label
  • Result structure
  • Contrasts and simultaneous inference
  • Diagnostics
  • Sensitivity analysis
  • Persistence and replayability
• Longitudinal treatment
  • Declare temporal roles
  • Read a long panel
  • Declare a dynamic rule
  • Declare a stochastic policy
  • Survival outcomes
  • Repeated outcomes
  • Competing risks
  • A baseline treatment and a time-to-event outcome
  • A survival curve that never decreases
  • Fit the sequentially doubly robust estimator
  • Report the sequential baselines
  • Estimate structural blips
  • Simulate the parametric g-formula
  • Estimate an incremental intervention over time
  • Longitudinal MSM projections
  • Diagnostics
  • Persistence
• Capabilities and refusals
  • Shipped analysis families
  • Why a request can be refused
  • Explicitly unimplemented
  • Extending the package