h t t p s : / / a r x i v . o r g / a b s / 2 1 0 8 . 1 3 5 1 8

Ingest: DoWhy: Addressing Challenges in Expressing and Validating Causal Assumptions

Type: kb/sources/types/ingest-report.md

Classification

A workshop/method paper describing a causal-inference framework and the challenge of expressing and partially validating assumptions. The genre recorded on the snapshot is correct. Author: Amit Sharma, Vasilis Syrgkanis, Cheng Zhang, and Emre Kiciman from Microsoft Research; strong practitioner-research authority for causal inference tooling, with the usual tool-builder interest in the framework's framing.

Summary

The paper argues that causal-effect estimation depends on assumptions about the data-generating process, and unlike predictive modeling there is no global validator for a causal estimate. DoWhy's response is to make assumptions explicit through causal graphs, use identification procedures such as graph-based criteria and do-calculus, estimate effects, and run refutation or validation tests for subsets of the assumptions. For this KB, the source is the boundary condition on causal reach assessment: causal formalism can assess reach only as far as the assumptions are declared and partially checkable.

Quotes

  • Source extract (verbatim): four steps: model, identify, estimate, and validate.
  • Source location: Introduction, “DoWhy: Expressing and validating assumptions”; short fragment required by the two-column PDF capture.
  • Source extract (verbatim): Before starting any causal analysis, DoWhy stipulates that
  • Source location: Section 3.1, first half of the user-graph requirement; two-column capture.
  • Source extract (verbatim): the user provide a causal graph over the observed variables.
  • Source location: Section 3.1, second half of the user-graph requirement; next captured line.
  • Source extract (verbatim): DoWhy uses graph-based criteria and do-calculus to find
  • Source location: Section 3.1, first half of the identification statement; two-column capture.
  • Source extract (verbatim): expressions that can identify the causal effect.
  • Source location: Section 3.1, second half of the identification statement; next captured line.
  • Source extract (verbatim): It is important to note here that causal assumptions cannot be fully verified. Rather, the intent is to validate some
  • Source location: Section 3.2, validation boundary; the right PDF column is contiguous across these two captured lines.

Connections Found

This source supports reach assessment and Formal symbolic systems assess reach only through causal and proof obligations by grounding the warning that causal discovery and causal inference are not assumption-free. It also fits Warranted autonomy is bounded by oracle reach as a domain-specific example: automation can run the causal pipeline, but the warrant stops where assumptions cannot be globally validated.

Extractable Value

  1. Assumptions are first-class artifacts -- A formal reach-assessment system must represent graph, confounding, mediation, instrument, and identification assumptions explicitly, not bury them in an estimator. [quick-win]
  2. There is no global causal validator -- This is the strongest caution against treating causal inference as a magic reach oracle. Validation is partial and assumption-specific. [quick-win]
  3. Do-calculus identifies effects, not graphs by itself -- Useful correction for the Gödel-machine speculation: adding do-calculus to axioms is not enough unless graph learning or graph assumptions are also present. [quick-win]
  4. Causal discovery and causal inference need integration -- The source names the gap between building the graph and estimating the effect, which maps to the design surface for a future formal symbolic reach-assessment system. [experiment]

Limitations (our opinion)

The source is partly a tool-framework argument, so it should not be read as independent proof that DoWhy's particular API solves causal validation. The paper is most valuable for its negative claim -- assumptions are unavoidable and only partly testable -- and for the workflow decomposition. It does not make causal assumptions true, and it does not supply the missing semantic judgment for natural-language claims.

Keep this source as the cautionary citation in Formal symbolic systems assess reach only through causal and proof obligations for the assumption boundary of causal reach assessment; no separate promotion is needed now.