Examples & support

Binary action with observed covariates

Use Treatment A, Covariates C, and Outcome Y. Review C → A, C → Y, A → Y. Bind the CSV columns and declare their types. Identify checks the graph; APO calls fit_query_mean once. ATE calls it for the action and reference assignments, then subtracts the two returned means. If Outcomes contains Y1 and Y2, Identify submits the complete list once, APO fits it once, and ATE fits that same list for two regimes. Cards only display coordinates from the original SDK vector; they do not run separate outcome fits.

Read identification status, estimation status, exact method, options and diagnostics separately. The default score route is not the old three-estimator menu.

Front-door graph

Enter A, mediator M, outcome Y and any baseline covariates. Review all arrows and hidden common causes. Unobserved nodes have no CSV binding. The app submits this actual graph to the SDK; it does not assume the Front-door formula because of the template name.

Sequential actions

Use distinct names such as C1, A1, C2, A2 and Y. Review the time order and the complete fixed-action comparison. Submit the same rows and original graph for each regime. A failed graph-based fit is not replaced by a graph-free sequential call.

Ordered column bindings

A logical variable can bind ordered members such as X(x1, x2). Check both the graph metadata and Data columns. A treatment assignment must match the declared action dimension. Population support is declared explicitly; it is not inferred from the sample's unique values.

Unused ID or text columns need not enter the estimator. Repeated-person rows require appropriate study assumptions; a row identifier does not establish independent sampling.

Inspect and replay a result

Open Details to see the actual public API, SDK version, dependency versions, original request and returned diagnostics. Calculation code contains the adapter-generated public API replay. Download data preserves the observed rows used by the result.

The replay preserves both successful and unsupported requests. No model reference replaces a failed estimate. None remains None for an unavailable causal interval.

Current app boundaries

The default score route supports declared discrete actions. Continuous fixed-point interventions, arbitrary Conditioning on and counterfactual evidence are not implemented by this app's public SDK route. Unsupported cases return their actual reason.

Graph-free BD/SBD/FD/UCA and paired CATE/effect APIs have separate public package contracts. They are not automatic fallbacks for graph-based app requests.

IV monotonicity, bounds, Difference-in-Differences and Synthetic Control require separate assumptions and procedures; they are not active calculation options here.

Data sources and references

CSV, supported synthetic rows and compatible benchmark rows all require explicit graph bindings. Random Data model-reference comparisons remain N/A in this integration. A released benchmark reference can appear as separate source metadata after dataset ID, row count, factual-data hash, treatment/outcome bindings and requested regime match. Its definition and provenance are separate from the SDK result; LaLonde retains N/A with its source reason. Known simulation answers and benchmark effects are not estimation inputs or causal uncertainty intervals.