Put the causal question before the model

Goal & fit

Produce a reproducible causal-analysis notebook and an assumptions memo, with effect estimates conditional on a defensible design. This is not a recipe for converting predictive accuracy into causality.

Suitable when

Experimental data or observational data with a defensible identification strategy, enough relevant variation and a researcher able to assess estimator assumptions. Heterogeneous-effect estimation should answer a real question, not merely produce a colorful subgroup chart.

Not suitable when

A dataset with no credible treatment-assignment story, uncontrolled post-treatment adjustment, severe lack of overlap or insufficient information about important confounders. A more flexible estimator does not repair these problems.

Inputs to gather

• A precise treatment, outcome, target population, time horizon and causal estimand.
• A documented dataset with units, timestamps, missingness and treatment-assignment information; use only data you are permitted to process.
• A causal design or diagram, subject-matter expertise, and an analysis environment where code and data access can be reviewed.

Suggested workflow

The sequence below is editorial guidance. Numbered references identify sources for tool capabilities, not evidence that this entire workflow was tested.

1. Write the estimand in words

   Specify the intervention and comparison, who the effect is about and over what period. Distinguish an average effect from a conditional or subgroup effect. Document whether the question concerns assignment or receipt of treatment; they need not have the same estimand.

2. Make identification assumptions explicit

   Draw the timing and causal relationships that justify adjustment. State the design-specific assumptions, including confounding, consistency, interference and overlap where relevant. Have a domain reviewer challenge them before selecting an estimator. If identification is not credible, report an association or redesign the study.

3. Audit the data without outcome-driven fishing

   Check units, treatment coding, missingness, duplicate observations and timing. Separate pre-treatment covariates from mediators and other post-treatment measures. Inspect support for the comparison and define exclusions transparently. Preserve raw data and record transformations in reviewable code.

4. Choose a method that matches the design

   EconML describes individualized causal-response estimators [1]; CausalML describes conditional average treatment effect methods [2]. Compare the selected estimator's assumptions with the design rather than treating those libraries as interchangeable black boxes. Plan appropriate splitting or cross-fitting, clustered sampling and uncertainty estimation for the chosen method.

5. Challenge the estimate

   Compare with a simple, design-appropriate baseline and inspect overlap and nuisance-model diagnostics. Use sensitivity analyses or falsification checks that address plausible threats, not a ritual list of tests. Check behavior on suitable simulated or known-effect data before trusting custom implementation. Passing a check cannot prove the absence of unmeasured confounding.

6. Report uncertainty and reproduce the result

   Restart the analysis in a clean, approved environment and reproduce tables from the saved data version and code. Report the target population, estimand, interval method, exclusions and sensitivity results. Label exploratory subgroup results and multiplicity concerns; do not turn uncertain individual-effect estimates into automatic high-stakes decisions.

Alternative tools & routes

Choose by project needs and governance, not by an assumed quality ranking. Current costs, licences and service terms need checking before adoption.

Expected output

An assumptions memo, data dictionary, reviewable analysis notebook, versioned environment description and effect table with uncertainty and sensitivity results. Include a simple design-appropriate statistical baseline; AI assistance is optional, not a condition of credible causal analysis.

Validation checklist

Use these as acceptance checks in your project record. This page does not save completion state.

• Treatment, outcome, target population and time horizon match the stated estimand.
• Adjustment variables precede treatment where required, and identification assumptions have a substantive justification.
• Overlap and missing-data decisions are documented, including any change to the target population.
• The uncertainty method respects the estimator and sampling design; subgroup exploration is disclosed.
• A clean rerun reproduces the reported table, and sensitivity findings are reported even when inconvenient.

Privacy & research integrity

• Keep restricted microdata in the approved environment. Do not send identifiable rows, secrets or confidential outputs to a coding assistant without explicit authorization.
• Generated analysis code must be read and tested. Do not execute downloaded agent skills or grant an assistant unrestricted filesystem or network access as a shortcut to reproducibility.
• Current package compatibility and operational availability were not tested here. Neither primary overview establishes an empirical benefit for this workflow, and no causal estimate has been independently reproduced by the directory.

Linked primary sources

Based on stored primary-source extracts accessed on 15 October 2026. They were re-read for these guides, not freshly fetched or independently reproduced. Features and terms may change.

1. EconML — official overview

   Project documentation of causal-response estimation; not evidence that a particular dataset identifies a causal effect.

   Source accessed: 2026-10-15

2. CausalML — About

   Project documentation of uplift and CATE methods; no analysis was executed for this guide.

   Source accessed: 2026-10-15