How to build root-cause attribution sketch for multi- location KPI diagnosis

The 4-skill bundle on the root-cause-attribution agent

One agent. Four coordinated skills. The Decompose + Counterfact + Test + Audit bundle runs above the warehouse + causal-inference + Bayesian + MMM + behavior-analytics substrate and writes one canonical per-KPI per-location per-tactic attribution record.

Compliance overlay

Five anchors run per-KPI per-location per-tactic before any attribution distributes to downstream decision systems. The first anchor is operationally distinctive: Shapley decomposition + counterfactual decomposition + multi-touch attribution + BSTS + CausalImpact + MMM + replication- crisis statistical discipline converge on every attribution result.

FAQ

What is root-cause attribution sketch for multi-location KPI diagnosis — and what is the Shapley-times-counterfactual-times-BSTS-times-CausalImpact-times-MMM-times-multi-touch-attribution-times-replication-crisis-discipline problem distinctive to this skill?

A multi-location retail operator with 50-300 stores ships per-KPI per-location per-tactic root-cause attribution diagnosis when KPIs swing (revenue + traffic + conversion + AOV + CAC + LTV + retention + churn + NPS). Standard BI tools (Looker + Tableau + Power BI + Mode + Hex + Sigma + Lightdash + Metabase + Apache Superset) show the KPI moved but cannot decompose root-cause across overlapping tactics + per-location overlapping causes + correlated marketing channels. The four-skill bundle on the root-cause-attribution agent — Decompose, Counterfact, Test, Audit — sits above the warehouse + causal-inference + Bayesian + MMM + behavior-analytics substrate (Snowflake + BigQuery + Databricks + dbt + Airflow + SHAP + Shapley + LIME + DoWhy + EconML + causalml + PyMC + Stan + CausalImpact + BSTS + Robyn + LightweightMMM + PyMC-Marketing + GA4 + Adobe Analytics + Mixpanel + Amplitude). The operationally distinctive anchor: per-KPI per-location per-tactic root-cause attribution algorithm (Shapley value decomposition + Owen Shapley + Shapley regression + Lloyd Shapley 1953 cooperative game theory + monotonicity + symmetry + efficiency + null-player axioms + Aumann-Shapley extension + SHAP integration + tree-SHAP + kernel-SHAP + deep-SHAP) + counterfactual decomposition (do-calculus + Pearl back-door + front-door criterion + g-formula + targeted maximum likelihood estimation (TMLE) + double-robust estimation + augmented inverse-probability-weighting (AIPW) + matching + propensity-score weighting) + multi-touch attribution (Markov chain + Shapley + position-based + linear + time-decay + first-touch + last-touch + data-driven) + Bayesian structural time-series (BSTS) + Google CausalImpact + StateSpace models + ARIMA + VAR + Prophet causal + Average Treatment Effect (ATE) + Conditional Average Treatment Effect (CATE) + Individual Treatment Effect (ITE) + Local Average Treatment Effect (LATE) + Marketing Mix Modeling (MMM) + Bayesian MMM + per-perturbation A/B test + per-cohort placebo test + Granger causality + reverse-causality test + replication-crisis statistical discipline (Ioannidis 2005 + Amrhein Greenland McShane 2019 + Benjamini-Hochberg FDR + Bonferroni + E-value + Rosenbaum sensitivity (Γ) + Cornfield inequality + falsification test + negative-control outcome).

Why do Looker + Tableau + Power BI + Mode + GA4 + Adobe Analytics + Mixpanel + Amplitude + Robyn + LightweightMMM break at multi-location-overlapping-tactics-overlapping-causes scale?

Each BI + analytics + MMM vendor ships per-account flat attribution primitive at single-touchpoint level. None coordinates per-KPI per-location per-tactic Shapley decomposition + counterfactual decomposition + multi-touch attribution + BSTS + CausalImpact + MMM + per-perturbation A/B + per-cohort placebo + Granger causality + reverse-causality simultaneously. None handles per-location overlapping causes (per-location promo + per-location SEO + per-location SEM + per-location social + per-location email + per-location SMS + per-location loyalty + per-location BD + per-location operations + per-location staffing + per-location weather + per-location competitive) at decomposition layer. None gates against replication-crisis statistical discipline (Bonferroni + BH FDR + E-value + Rosenbaum sensitivity + negative-control outcome + per-cohort placebo). None enforces FDD Item 19 financial performance representations when attribution shared with franchisees + FINRA Rule 2210 when public-company attribution + SOX 302/404/906 + FASB ASC 280 segment reporting. None writes a per-KPI per-location per-tactic WORM attribution audit trail. The four-skill bundle Decompose + Counterfact + Test + Audit sits above the warehouse + causal-inference + Bayesian + MMM + behavior-analytics substrate — it does not replace it.

How does Decompose + Counterfact work?

Decompose runs per-KPI per-location per-tactic Shapley value decomposition: Lloyd Shapley 1953 cooperative game theory + monotonicity + symmetry + efficiency + null-player axioms + Aumann-Shapley extension + SHAP integration with gradient-boosting (tree-SHAP + kernel-SHAP + deep-SHAP). Per-KPI per-location per-tactic Shapley regression (gradient-boosting + XGBoost + LightGBM + CatBoost feature attribution + Owen Shapley + LIME + InterpretML + Anchors). Per-KPI Marketing Mix Modeling (MMM) decomposition (Robyn + LightweightMMM + PyMC-Marketing + Bayesian MMM + Lifesight + Recast + Northbeam + Triple Whale + Rockerbox). Counterfact runs per-KPI counterfactual decomposition: Pearl do-calculus + back-door criterion + front-door criterion + g-formula + targeted maximum likelihood estimation (TMLE) + double-robust estimation + augmented inverse-probability-weighting (AIPW) + matching + propensity-score weighting + propensity-score stratification (via DoWhy + EconML + causalml). Per-KPI Bayesian structural time-series (BSTS) + Google CausalImpact + StateSpace models + ARIMA + VAR + Prophet causal counterfactual (what KPI would have done without tactic). Per-KPI causal-inference estimates: Average Treatment Effect (ATE) + Conditional Average Treatment Effect (CATE) + Individual Treatment Effect (ITE) + Local Average Treatment Effect (LATE).

What does Test + Audit do?

Test runs per-KPI per-location per-tactic per-perturbation A/B test + per-cohort placebo test + Granger causality + reverse-causality test. Per-test severity classification: P0 spurious-attribution risk (E-value < 1.5 + Rosenbaum sensitivity Γ < 1.5) + P1 multiple-comparison failure 72-hour (Bonferroni + Benjamini-Hochberg FDR fail) + P2 counterfactual-overlap violation 7-day + P3 Granger-causality reverse-causal 30-day + P4 docs-only. Per-KPI per-location per-tactic test result + per-anchor risk classification. Gate runs 5 anchors per-KPI per-location per-tactic before any attribution distributes to downstream decision systems. (1) Per-KPI per-location per-tactic Shapley decomposition + counterfactual decomposition + multi-touch attribution + BSTS + CausalImpact + MMM + per-perturbation A/B + per-cohort placebo + Granger causality + reverse-causality + replication-crisis discipline (Ioannidis + Amrhein-Greenland-McShane + Bonferroni + BH FDR + E-value + Rosenbaum Γ + Cornfield + falsification + negative-control). (2) FTC Section 5 + Pfizer 1972 + CFPB UDAAP + Lanham + USPTO + Robinson-Patman + FDD Item 19 financial performance representations when attribution shared with franchisees + 15-state franchise + per-state attorney advertising. (3) HIPAA + state mini-HIPAA + FINRA Rule 2210 + Rule 3110 + SEC Regulation FD + per-state professional licensing. (4) EU AI Act Article 50 transparency when AI-generated attribution + Article 13/14/15 + Annex III when AI-ML attribution drives capital/inventory decisions + Article 6/27 FRIA + DSA + DMA + GDPR Article 6/7/22/28/30 (Article 22 automated decision-making consent) + LGPD + DPDP + PIPEDA + Quebec Law 25 + CCPA + CPRA + 18-state. (5) WCAG 2.2 AA + ARIA + EAA + ADA Title III + Section 508 + SOX 302/404/906 + COSO + Exchange Act 13(b)(2) + FASB ASC 280 segment reporting + SEC Reg S-K. Audit writes a per-KPI per-location per-tactic WORM attribution record: KPI snapshot + Shapley decomposition + counterfactual decomposition + multi-touch attribution + BSTS + CausalImpact + MMM + per-anchor gate-pass + AI-ML provenance + EU AI Act FRIA. Retention: 7-year FTC + 7-year IRS + 7-year HIPAA + 7-year state bar + 6-year SEC + 3-year FINRA + 7-year SOX + GDPR Article 30 + EU AI Act Article 12 + SOC 2 CC7/CC8.

What does this skill connect to on the root-cause-attribution agent and across the swarm?

On the root-cause-attribution agent: per-KPI per-location per-tactic attribution + per-location peer-cohort benchmarking + per-location per-cohort anomaly detection. Across the swarm: per-location AI-calibrated forecasting (#600 same per-location-cohort substrate + same replication-crisis discipline) + per-location per-cohort two-sigma anomaly detection (same per-location-cohort substrate) + integration-drift-monitor agent (#562 + #569 + #570) + governance-decision-router five-destination routing + tiered pre-filter deterministic gates + per-state-overlay-composer (#599 UPSTREAM canonical for FDD Item 19 + per-state attorney advertising + FINRA per-state) + real-time change-event emission (#603 UPSTREAM canonical for per-KPI event stream). Commercial-pillar parent: /pre-emptive-churn-and-cohort-relative-trends.

What does the 6-workstream pre-engagement-baseline reporting cycle look like for this skill?

Every two weeks during the Tier 3 Fractional CMO with AI Swarm engagement, six workstreams report against the pre-engagement baseline. Workstream 1: per-portfolio per-KPI per-location per-tactic attribution coverage — KPIs monitored + locations covered + tactics decomposed. Workstream 2: Decompose per-KPI Shapley + MMM flow — Shapley value + Owen Shapley + Shapley regression + SHAP + Robyn + LightweightMMM + PyMC-Marketing decomposition absorbed. Workstream 3: Counterfact per-KPI counterfactual flow — do-calculus + back-door + front-door + TMLE + AIPW + BSTS + CausalImpact counterfactual + ATE + CATE + ITE + LATE absorbed. Workstream 4: Test per-KPI per-perturbation flow — per-perturbation A/B + per-cohort placebo + Granger causality + reverse-causality test result + replication-crisis discipline check. Workstream 5: Regulatory-defense audit coverage — Shapley + counterfactual + BSTS + CausalImpact + MMM + replication-crisis discipline + EU AI Act Article 50 + FDD Item 19 + FINRA 2210 + SOX + FASB ASC 280. Workstream 6: FBC feedback-loop pattern-learning — per-KPI realized-vs-predicted attribution + per-tactic decomposition retrospective + per-counterfactual robustness retrospective.

Related reading

• Parent commercial pillar: pre-emptive churn and cohort- relative trends
• Sibling build-pillar: per-location AI-calibrated forecasting (#600 same per-location-cohort substrate + same replication-crisis discipline)
• Sibling build-pillar: real-time change-event emission (#603 UPSTREAM canonical for per-KPI event stream)
• Sibling build-pillar: per-state overlay configuration (#599 UPSTREAM canonical for FDD Item 19 + per-state attorney advertising + FINRA per-state)
• Fractional CMO with AI Swarm
• AI Readiness Assessment