Second Brain

Index: Dillon et al - SHAP and Causal Insights

5 min read

Dillon et al - SHAP and Causal Insights

Routing Summary

This folder covers the SHAP-documentation case study by Dillon, LaRiviere, Lundberg, Roth and Syrgkanis on why SHAP explanations of predictive models are not causal effects. It has 9 notes covering prediction vs causal tasks, a simulated subscriber-retention DGP with a known DAG, SHAP vs true-effect overlays, the independence condition, observed confounding with double ML (econML LinearDML), non-confounding redundancy (mediators), unobserved confounding, and IV/DiD/RD remedies. All 13 notebook cells are reproduced verbatim.

Concept Map

ConceptNoteTypeDepends OnKey Result
Prediction vs causal taskDillon et al - Prediction Tasks vs Causal TasksdefinitionPotential Outcomes Frameworkout-of-sample fit is silent on causal validity
Retention SCM and FixableDataFrame do-operatorDillon et al - Subscriber Retention Data-Generating Processdefinition / exampleDirected Acyclic GraphsAd spend coefficient 0; Discount effect positive; Bugs reported not a parent
marginal_effects true-effect overlayDillon et al - SHAP Plots vs True Causal EffectsalgorithmDGP noteSHAP flips Discount and Bugs reported and inflates Ad spend
Independence condition; redundancy dendrogramDillon et al - When SHAP Recovers Causal Effectstheorem / procedureConditional Independence AssumptionEconomy recovered; the dendrogram detects only observed redundancy
Observed confounder; double MLDillon et al - Observed Confounding and Double ML for Ad Spenddefinition / algorithmRegularization Bias and the Partially Linear ModelLinearDML: no Ad spend effect,
Non-confounding redundancy; control-or-drop ruleDillon et al - Non-Confounding Redundancy and the Control-or-Drop Ruledefinition / ruleCausal Estimandscontrolling for a mediator gives the direct effect only; “controlling … the safer option”
Unobserved confounding; IV/DiD/RDDillon et al - Unobserved Confounding and Quasi-Experimental Remediesdefinition / resultThe Experimental IdealDML gets the Discount sign wrong; randomization is required
Limitations and marketing checklistDillon et al - Summary, Limitations and Marketing Implicationssynthesisall of the abovea SHAP channel ranking is not an ROI ranking

Notes

  • Dillon et al - SHAP and Causal Insights - Overview — CONTAINS: research question, contribution list, key-findings table, Source Map (section → note, raw line ranges), figure and cell map, argument flow.
  • Dillon et al - Prediction Tasks vs Causal Tasks — CONTAINS: introduction thesis, surprising findings and business-partner explanations, Def. prediction task, Def. causal task, claim that one model can be right for one task and wrong for the other, example of one forecast with two consumers; complete quote of the second-team paragraph (raw line 187).
  • Dillon et al - Subscriber Retention Data-Generating Process — CONTAINS: full structural-equation table, logit renewal equation, 17-edge causal graph, training pipeline algorithm, FixableDataFrame do-operator definition, verbatim cells [1], [2], [5], code-quirk table, vault table of true interventional effects and correlations; complete quotes of the example opening (line 11) and the causal-graph framing (line 193).
  • Dillon et al - SHAP Plots vs True Causal Effects — CONTAINS: cells [3], [4], [6], Figs. 1, 2, 4; marginal_effects algorithm; Findings 1-3 (Bugs reported, Discount, Ad spend; line 290 quoted in full); Fig. 1 bar values against true effects (ranking close to inverted); Fig. 2 panel-by-panel table; eight-row Fig. 4 overlay table.
  • Dillon et al - When SHAP Recovers Causal Effects — CONTAINS: independence condition (Economy), Def. redundancy clustering, two-part causal-readiness procedure, cell [7], Fig. 5, experiment example, per-feature check table.
  • Dillon et al - Observed Confounding and Double ML for Ad Spend — CONTAINS: Def. observed confounder, regularization-redundancy mechanism, three-step DML algorithm with PLR formalization, Fig. 7 value reading (slope about -0.009, band covering 0), verbatim cell [8] and cell [9] (RegressionWrapper, double_ml, plot_effect), code walk-through, Fig. 7, result, all-confounders proviso, vault linear replication.
  • Dillon et al - Non-Confounding Redundancy and the Control-or-Drop Rule — CONTAINS: Def. non-confounding redundancy (downstream and upstream), credit-splitting result, adjustment-choice table for DML, control-or-drop rule with caveats, cells [10]-[11], Figs. 8-9, root-node explanation, vault replication, branded-search analogue; Fig. 9 value-by-value reading cross-checked against the +0.78 true total effect (about 73% recovered).
  • Dillon et al - Unobserved Confounding and Quasi-Experimental Remedies — CONTAINS: Def. unobserved confounding, DML-fails result for Discount, cells [12]-[13], Figs. 10-11 (Fig. 11 value reading: slope about -1.05 vs true +0.06), randomization claim, IV/DiD/RD table, vault replication, randomized discount-test design, advertising analogue.
  • Dillon et al - Summary, Limitations and Marketing Implications — CONTAINS: five summary points, consolidated decision table, seven limitations, weekly-media driver example, example of choosing among model fits by SHAP mass.

Sources

  • Dillon et al - Be Careful When Interpreting Predictive Models in Search of Causal Insights (SHAP docs) — SHAP documentation notebook by E. Dillon, J. LaRiviere, S. Lundberg, J. Roth and V. Syrgkanis (Microsoft).

See Also

Graph View

Backlinks