Causal Decision Agent

Overview Schema Source Related Servers Discussions Admin

dml_model_averaging

Read-only

Averages double-ML causal estimates from multiple nuisance models, returning risk-weighted treatment effect with adjusted SEs. Use to combine learners for robust unconfounded inference.

Instructions

Model-averaging DML (PLR) per Ahrens et al. (2025, JAE). Fits DML-PLR under multiple candidate nuisance learners and reports a risk-weighted (or equal/single-best) average of their theta estimates with a covariance-adjusted SE. Assumptions: Unconfoundedness (IRM/PLR) or instrument validity (IIVM/PLIV) given the covariates; Overlap / positivity; Neyman-orthogonal score plus cross-fitting; nuisance estimators converge fast enough (o(n^-1/4)). Pre-conditions: Covariates, treatment, and outcome with enough data for cross-fitted machine-learning nuisances; For instrumented variants (PLIV / IIVM): an instrument as well. Failure modes: Propensity scores near 0/1 -- overlap failure inflates variance and bias -> Trim extreme scores, restrict the estimand to the overlap region, or report sensitivity (dml_sensitivity). Alternatives: sp.tmle, sp.auto_cate, sp.causal_forest. Typical minimum N: 500.

Input Schema

TableJSON Schema

Name	Required	Description	Default
`y`	Yes	Outcome column
`seed`	No	Random seed for reproducible stochastic steps.
`alpha`	No	Significance level for confidence intervals and tests.
`treat`	Yes	Treatment column
`detail`	No	Payload depth: 'minimal' (~150 tokens) for sub-step calls where only the point estimate is needed; 'standard' (~1K tokens) for diagnostics + coefficient table; 'agent' (~2K tokens, default) adds violations / next_steps / suggested_functions so the LLM can plan its next call without another round-trip.	agent
`n_folds`	No	Number of cross-fitting or cross-validation folds.
`as_handle`	No	If true, cache the fitted result on the server and return result_id + result_uri alongside the JSON payload so a subsequent tools/call can chain without re-running.
`data_path`	Yes	Absolute path or URL to a data file. Supported: .csv / .tsv / .txt (delimited), .parquet / .pq, .feather / .arrow, .xlsx / .xls, .dta (Stata), .json / .jsonl. Schemes: file://, s3://, gs://, https://.
`result_id`	No	Optional handle to a previously-fitted result (returned by an earlier call when as_handle=true). Tools that operate on a fitted object accept this in place of re-supplying data_path + columns.
`candidates`	No	List of (ml_g, ml_m, label) sklearn triples; defaults to Lasso/Ridge/RF/GBM
`covariates`	Yes	Covariate columns X
`weight_rule`	No	Weighting of candidate estimators	short_stacking
`data_columns`	No	Optional column projection. Parquet/Feather/Stata loaders honour this for fast partial reads.
`data_sample_n`	No	Optional uniform random subsample size (seed=0, deterministic) — useful on huge panels.

Output Schema

TableJSON Schema

Name	Required	Description	Default
No arguments

Schema Changelog

Changes observed during successful MCP inspections.

First observedv0.1.0Sep 15, 2026

TDQS

A4.3/5.0

Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Since annotations already declare readOnlyHint=true, the safety profile is covered; the description adds valuable behavioral detail such as cross-fitting, Neyman-orthogonal scores, nuisance convergence rates, and the overlap-failure mode with mitigations. It does not mention that as_handle=true can cache a fitted result server-side, but that is communicated in the schema rather than hidden.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is dense yet well organized: method first, then assumptions, preconditions, failure modes, alternatives, and min N. Every section earns its place and no filler or repetition of schema content appears.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given 14 parameters, a full schema, and an output schema, the description covers the key contextual gaps: statistical assumptions, data requirements, failure modes, mitigations, alternatives, and sample-size guidance. Nothing essential for an agent to decide whether to invoke this tool is missing.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 100%, so the baseline is 3. The description reinforces the meaning of candidate nuisance learners and weight_rule options, but it does not add significant parameter-level guidance beyond what the schema already documents for data_path, n_folds, detail, or seed.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description opens with a specific method name and citation, then states the exact action: 'Fits DML-PLR under multiple candidate nuisance learners and reports a risk-weighted (or equal/single-best) average of their theta estimates with a covariance-adjusted SE.' This clearly distinguishes it from plain DML, sensitivity tools, and causal-forest alternatives.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description lists assumptions, preconditions, failure modes, a minimum sample size, and even names alternatives such as 'sp.tmle, sp.auto_cate, sp.causal_forest' and 'dml_sensitivity' for overlap problems. However, it never explicitly states when to choose this tool over those alternatives, so the decision rule is implied rather than spelled out.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Deploy Server