Bareinboim Pearl 2016 - Causal Inference and Data Fusion - Overview

Summary

The paper unifies three classical biases under one structural-causal-model language and one inference method (do-calculus on a graph): confounding bias (tasks 1 and 2), sample selection bias (task 3) and transportability / cross-population bias (task 4). Each “design”, defined by the triple (population, regime, sampling method) of an available dataset, is a different syntactic form of data, and the strategy is always to transform the target query into an expression estimable from the datasets in hand. The authors claim complete graphical and algorithmic solutions for policy evaluation, transportability and data fusion in nonparametric DAG models.

Overview

Question. Given several heterogeneous datasets (different populations, observational vs experimental regimes, different sampling mechanisms), when and how can a causal effect $P(y\mid do(x))$ in a target population be computed?

Thesis. “Not all data are created equal”: the design of each dataset must be encoded mathematically, and different designs support different inferences. Big data does not replace causal assumptions; it magnifies the need to reason about how data were acquired.

Contributions.

1. A four-task taxonomy of generalization problems (Fig. 1).
2. Statement of the two principles of causal inference (structural counterfactuals, structural independences) from which the rest follows.
3. Review of identification results: backdoor, front-door, z-identification (auxiliary experiments) with the three do-calculus rules.
4. The selection-backdoor criterion (Definition 4, Theorem 1) for recovering causal effects from sampling-biased data, and the clinical-trial generalization pattern (Eq. 14).
5. Transport formulas (Eqs. 15-17), the complete decision procedure for transportability, and an assurance of completeness for multi-source data fusion (Summary Results 1-3).

Main Content

Source Map

Paper section	PDF pages	Content	Covered in
Abstract, Introduction: Causal Inference and Big Data, Fig. 1	1-2	tasks 1-4, designs, outline	this note, Bareinboim 2016 - SCM Principles and Confounding Bias (Fig. 1 table)
The Structural Causal Model; Two Principles	2-3	$⟨U,V,F⟩$, Eq. 1 counterfactual, d-separation (Def. 1), Fig. 2A	Bareinboim 2016 - SCM Principles and Confounding Bias
Policy Evaluation and the Problem of Confounding	3-5	Eqs. 2-4, Def. 2, rules (Eqs. 5-7), backdoor (Def. 3, Eqs. 8-9), front-door (Eq. 10)	Bareinboim 2016 - SCM Principles and Confounding Bias
Identification Through Auxiliary Experiments	5-6	z-identification, Fig. 3, Eqs. 11-12, Summary Result 1	Bareinboim 2016 - SCM Principles and Confounding Bias
Sample Selection Bias	6-7	Fig. 4, Eq. 13, Def. 4 S-backdoor, Theorem 1	Bareinboim 2016 - Sample Selection Bias and Recoverability
The Generalizability of Clinical Trials	7	Fig. 4F, Eq. 14, S-ignorability, Summary Result 2	Bareinboim 2016 - Sample Selection Bias and Recoverability
Transportability and the Problem of Data Fusion	7-8	selection diagrams, Fig. 5, Eqs. 15-17, multi-population fusion, Summary Result 3	Bareinboim 2016 - Transportability and Data Fusion
Conclusion	8	external validity “solved”, future work	Bareinboim 2016 - Transportability and Data Fusion
References 1-55	8	bibliography	cited inline

Key results at a glance

Problem	Data available	Result
Confounding (task 1)	$P(v)$	backdoor adjustment ${\sum }_{z}P(y\mid x,z)P(z)$; front-door if no backdoor set
Auxiliary experiment (task 2)	$P(v)$, $P(v\mid do(z))$	z-identifiable iff $X$ intercepts all directed $Z\to Y$ paths and $P(y\mid do(x))$ identifiable in $G_{\overline{Z}}$
Selection bias (task 3)	$P(v\mid S=1)$ (+ population-level $P(z)$)	S-backdoor admissible $Z$: $P(y\mid do(x))={\sum }_{z}P(y\mid x,z,S=1)P(z)$
Transportability (task 4)	$P(v\mid do(x))$ in $\pi$, observational in ${\pi }^{\ast }$	transport formula via selection diagram; e.g. ${\sum }_{z}P(y\mid do(x),z)P^{\ast }(z)$

Examples

Three recoverability/transport designs in one table (details in the topic notes):

Design	Query	Estimand
Fig. 4B: $X\to S$, no confounding	$P(y\mid do(x))$	$P(y\mid x,S=1)$
Fig. 4F clinical trial, post-treatment $Z$	$P(y\mid do(x))$	${\sum }_{z}P(y\mid do(x),z,S=1)P(z\mid x)$ (Eq. 14)
Fig. 5C biomarker	$P^{\ast }(y\mid do(x))$	${\sum }_{z}P(y\mid do(x),z)P^{\ast }(z\mid x)$ (Eq. 17)

Connections

• Direct successor and detailed derivations: Pearl Bareinboim 2014 - External Validity and Transportability - Overview.
• Estimator-level counterpart (what to compute once identification is established): Degtiar Rose 2023 - Generalizability and Transportability Review - Overview, which cites this paper as the identifiability-focused review.
• Underlying identification machinery: DAGs and Causal Identification, Do-Calculus in Summary Causal DAGs, Directed Acyclic Graphs.
• Marketing: a measurement stack typically has an observational panel (task 1), geo experiments (tasks 2/4) and platform-biased user data (task 3). The paper’s contribution for a practitioner is a checklist for which combinations of these answer a given query. Concrete mapping: Bareinboim 2016 - Transportability and Data Fusion and Pearl 2014 - Marketing Application Transporting Geo and Lift Results.

See Also

• Potential Outcomes Framework, Propensity Score Matching - Overview, Instrumental Variables
• Sibling sources in this cluster: Dahabreh 2020 - Extending Inferences from a Randomized Trial - Overview (identification and doubly robust estimation for trial-to-target inference) and Allcott 2015 - Site Selection Bias in Program Evaluation - Overview (site selection bias, marketing implications).