General Transportability of Soft Interventions: Completeness Results

Juan Correa
Juan Correa

NIPS 2020, 2020.

Cited by: 0|Bibtex|Views6
EI
Other Links: dblp.uni-trier.de|academic.microsoft.com
Weibo:
We hope that this series of results related to soft interventions, and knowledge of its relationship with atomic interventions, can help data scientists to apply causal inference in broader and more realistic scenarios

Abstract:

The challenge of generalizing causal knowledge across different environments is pervasive in scientific explorations, including in AI, ML, and Data Science. Experiments are usually performed in one environment (e.g., in a lab, on Earth) with the intent, almost invariably, of being used elsewhere (e.g., outside the lab, on Mars), where the...More

Code:

Data:

0
Introduction
  • Generalizing causal knowledge across disparate domains is at the heart of many inferences across the empirical sciences as well as AI [26, 33, 29].
  • The authors design an efficient algorithm to determine the existence of an estimand for the effect of a non-atomic intervention as a function of the available distributions.
  • The authors prove that the σ-calculus is necessary and sufficient for the task of transportability when both the input and the output distributions involve soft interventions.
Highlights
  • Generalizing causal knowledge across disparate domains is at the heart of many inferences across the empirical sciences as well as AI [26, 33, 29]
  • We studied the problem of transporting effects of soft interventions from knowledge encoded in the form of a selection diagram and a combination of observational and experimental data from multiple, different domains
  • We showed how the problem can be solved by transporting the effect of an atomic intervention from the same input (Thm. 1)
  • We conclude that σ-calculus together with basic probability axioms are complete for the soft transportability task (Cor. 2)
  • We described a complete graphical condition to determine the transportability of any transportability instance (Cor. 3)
  • We hope that this series of results related to soft interventions, and knowledge of its relationship with atomic interventions, can help data scientists to apply causal inference in broader and more realistic scenarios
Results
  • Conditional and stochastic interventions allow the intervened variable to change as a deterministic function or a conditional probability distribution of a set of observable parents.
  • Given a causal diagram Gi = V, E and domain discrepancies ∆i, let S = {Sv | ∃ni=1V ∈ ∆i} be called selection variables.
  • The effect intervention σX on a set of outcome variables Y, conditional on W, P ∗(y|w; σX), in a target environment π∗, is said to be transportable from G∆, Z , if it is uniquely computable from the set of distributions Z for every assignment (y, w) and every set of models {Mi}πi∈Π inducing G∆ and Z.
  • Given the tightness of the reduction provided by Thm. 1, one may surmise that it is possible to blindly use existing transportability algorithms (e.g., GTR [22]) to solve for soft interventions.
  • Both input and output of the transportability task refer to probability distributions within different domains and for different interventions.
  • Let Y, X ⊆ V be any two sets of variables, and let σX=σX∗ be an atomic, conditional or stochastic intervention.
  • Input: G∆ selection diagrams over variables V for domains Π; Y, W ⊆ V disjoint subsets of variables; an intervention σX∗ defined over a set X ⊆ V; and available distribution specification Z.
  • Building on the observations and results the authors have so far, the authors design the algorithm σ-TR (Alg. 1) that takes as input the variables defining a query, the specification of σX, a set of available distributions (Z), and the selection diagrams.
  • Σ-TR uses the subroutine IDENTIFY from [36] that applies Lemma 1 systematically to obtain a C-factor Q[A] from Q[B], where A ⊆ B, and the subroutine ‘REPLACE’ to determine the factors of intervened variables according to the particular type of intervention.
Conclusion
  • Given query P ∗(y; σX=σX∗ ), selection diagram G∆, and the distribution specified by Z, let A be defined as in Thm. 2.
  • The authors studied the problem of transporting effects of soft interventions from knowledge encoded in the form of a selection diagram and a combination of observational and experimental data from multiple, different domains.
  • The authors hope that this series of results related to soft interventions, and knowledge of its relationship with atomic interventions, can help data scientists to apply causal inference in broader and more realistic scenarios
Summary
  • Generalizing causal knowledge across disparate domains is at the heart of many inferences across the empirical sciences as well as AI [26, 33, 29].
  • The authors design an efficient algorithm to determine the existence of an estimand for the effect of a non-atomic intervention as a function of the available distributions.
  • The authors prove that the σ-calculus is necessary and sufficient for the task of transportability when both the input and the output distributions involve soft interventions.
  • Conditional and stochastic interventions allow the intervened variable to change as a deterministic function or a conditional probability distribution of a set of observable parents.
  • Given a causal diagram Gi = V, E and domain discrepancies ∆i, let S = {Sv | ∃ni=1V ∈ ∆i} be called selection variables.
  • The effect intervention σX on a set of outcome variables Y, conditional on W, P ∗(y|w; σX), in a target environment π∗, is said to be transportable from G∆, Z , if it is uniquely computable from the set of distributions Z for every assignment (y, w) and every set of models {Mi}πi∈Π inducing G∆ and Z.
  • Given the tightness of the reduction provided by Thm. 1, one may surmise that it is possible to blindly use existing transportability algorithms (e.g., GTR [22]) to solve for soft interventions.
  • Both input and output of the transportability task refer to probability distributions within different domains and for different interventions.
  • Let Y, X ⊆ V be any two sets of variables, and let σX=σX∗ be an atomic, conditional or stochastic intervention.
  • Input: G∆ selection diagrams over variables V for domains Π; Y, W ⊆ V disjoint subsets of variables; an intervention σX∗ defined over a set X ⊆ V; and available distribution specification Z.
  • Building on the observations and results the authors have so far, the authors design the algorithm σ-TR (Alg. 1) that takes as input the variables defining a query, the specification of σX, a set of available distributions (Z), and the selection diagrams.
  • Σ-TR uses the subroutine IDENTIFY from [36] that applies Lemma 1 systematically to obtain a C-factor Q[A] from Q[B], where A ⊆ B, and the subroutine ‘REPLACE’ to determine the factors of intervened variables according to the particular type of intervention.
  • Given query P ∗(y; σX=σX∗ ), selection diagram G∆, and the distribution specified by Z, let A be defined as in Thm. 2.
  • The authors studied the problem of transporting effects of soft interventions from knowledge encoded in the form of a selection diagram and a combination of observational and experimental data from multiple, different domains.
  • The authors hope that this series of results related to soft interventions, and knowledge of its relationship with atomic interventions, can help data scientists to apply causal inference in broader and more realistic scenarios
Tables
  • Table1: Summary of the types of interventions considered. Each row contains the type of intervention, its representation using the regime indicator and the way the corresponding replacement function
Download tables as Excel
Funding
  • Acknowledgments and Disclosure of Funding This research was supported by grants from IBM Research, Adobe Research, NSF IIS-1704352, and IIS-1750807 (CAREER)
Reference
  • A. V. Banerjee, S. Cole, E. Duflo, and L. Linden. Remedying Education: Evidence from Two Randomized Experiments in India*. The Quarterly Journal of Economics, 122(3):1235–1264, 2007.
    Google ScholarLocate open access versionFindings
  • E. Bareinboim and J. Pearl. Transportability of causal effects: Completeness Results. In Proceedings of the 26th AAAI Conference on Artificial Intelligence, CA, 201Department of Computer Science, University of California, Los Angeles.
    Google ScholarLocate open access versionFindings
  • E. Bareinboim and J. Pearl. Causal Inference by Surrogate Experiments: z-Identifiability. In N. d. F. Murphy and Kevin, editors, Proceedings of the Twenty-Eighth Conference on Uncertainty in Artificial Intelligence, pages 113–120. AUAI Press, 2012.
    Google ScholarLocate open access versionFindings
  • E. Bareinboim and J. Pearl. A general algorithm for deciding transportability of experimental results. Journal of Causal Inference, 1(1):107–134, 2013.
    Google ScholarLocate open access versionFindings
  • E. Bareinboim and J. Pearl. Transportability from Multiple Environments with Limited Experiments: Completeness Results. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 280–288. Curran Associates, Inc., 2014.
    Google ScholarLocate open access versionFindings
  • M. Bertrand, D. Karlan, S. Mullainathan, E. Shafir, and J. Zinman. What’s Advertising Content Worth? Evidence from a Consumer Credit Marketing Field Experiment*. The Quarterly Journal of Economics, 125(1):263–306, 2 2010.
    Google ScholarLocate open access versionFindings
  • N. Cartwright. Hunting Causes and Using Them: {A}pproaches in Philosophy and Economics. Cambridge University Press, New York, NY, 2007.
    Google ScholarFindings
  • J. D. Correa and E. Bareinboim. From Statistical Transportability to Estimating the Effects of Stochastic Interventions. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 1661–1667. International Joint Conferences on Artificial Intelligence Organization, 2019.
    Google ScholarLocate open access versionFindings
  • J. D. Correa and E. Bareinboim. A Calculus For Stochastic Interventions: Causal Effect Identification and Surrogate Experiments. In Proceedings of the 34th AAAI Conference on Artificial Intelligence. AAAI Press, 2020.
    Google ScholarLocate open access versionFindings
  • A. P. Dawid. Influence diagrams for causal modelling and inference. International Statistical Review, 70:161–189, 2002.
    Google ScholarLocate open access versionFindings
  • A. P. Dawid, V. Didelez, and others. Identifying the consequences of dynamic treatment strategies: A decision-theoretic overview. Statistics Surveys, 4:184–231, 2010.
    Google ScholarLocate open access versionFindings
  • E. Duflo, R. Glennerster, and M. Kremer. Using Randomization in Development Economics Research: A Toolkit. In T. P. Schultz and J. A. Strauss, editors, Handbook of Development Economics, volume 4, chapter 61, pages 3895–3962.
    Google ScholarLocate open access versionFindings
  • F. Eberhardt and R. Scheines. Interventions and causal inference. Philosophy of science, 74(5): 981–995, 2007.
    Google ScholarLocate open access versionFindings
  • J. J. Heckman. The Scientific Model of Causality. Sociological Methodology, 35:1–97, 2005.
    Google ScholarLocate open access versionFindings
  • Y. Huang and M. Valtorta. Identifiability in Causal Bayesian Networks: A Sound and Complete Algorithm. In Proceedings of the Twenty-First National Conference on Artificial Intelligence (AAAI 2006), pages 1149–1156. AAAI Press, Menlo Park, CA, 2006.
    Google ScholarLocate open access versionFindings
  • Y. Huang and M. Valtorta. On the completeness of an identifiability algorithm for semiMarkovian models. Annals of Mathematics and Artificial Intelligence, 54(4):363–408, 2008.
    Google ScholarLocate open access versionFindings
  • D. Karlan and J. Zinman. Expanding credit access: Using randomized supply decisions to estimate the impacts. The Review of Financial Studies, 23(1):433–464, 2010.
    Google ScholarLocate open access versionFindings
  • K. Korb, L. Hope, A. Nicholson, and K. Axnick. Varieties of Causal Intervention. In Lecture Notes in Artificial Intelligence (Subseries of Lecture Notes in Computer Science), volume 3157, pages 322–331, 2004.
    Google ScholarLocate open access versionFindings
  • S. Lee and V. Honavar. Causal Transportability of Experiments on Controllable Subsets of Variables: z-Transportability. In A. Nicholson and P. Smyth, editors, Proceedings of the TwentyNinth Conference on Uncertainty in Artificial Intelligence (UAI), pages 361–370. AUAI Press, 2013.
    Google ScholarLocate open access versionFindings
  • S. Lee and V. Honavar. m-Transportability: Transportability of a Causal Effect from Multiple Environments. In M. desJardins and M. Littman, editors, Proceedings of the Twenty-Seventh National Conference on Artificial Intelligence, pages 583–590, Menlo Park, CA, 2013. AAAI Press.
    Google ScholarLocate open access versionFindings
  • S. Lee, J. D. Correa, and E. Bareinboim. General Identifiability with Arbitrary Surrogate Experiments. In Proceedings of the Thirty-Fifth Conference Annual Conference on Uncertainty in Artificial Intelligence, Corvallis, OR, 2019. AUAI Press, in press.
    Google ScholarLocate open access versionFindings
  • S. Lee, J. Correa, and E. Bareinboim. Generalized Transportability: Synthesis of Experiments from Heterogeneous Domains. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, Menlo Park, CA, 2020. AAAI Press.
    Google ScholarLocate open access versionFindings
  • J. K. Lunceford, M. Davidian, and A. A. Tsiatis. Estimation of Survival Distributions of Treatment Policies in Two-Stage Randomization Designs in Clinical Trials. Biometrics, 58(1): 48–57, 2002.
    Google ScholarLocate open access versionFindings
  • S. A. Murphy. An experimental design for the development of adaptive treatment strategies. Statistics in Medicine, 24(10):1455–1481, 2005.
    Google ScholarLocate open access versionFindings
  • J. Pearl. Causal diagrams for empirical research. Biometrika, 82(4):669–688, 1995.
    Google ScholarLocate open access versionFindings
  • J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, New York, NY, USA, 2nd edition, 2000.
    Google ScholarFindings
  • J. Pearl. Review of N. Cartwright ‘Hunting Causes and Using Them’. Economics and Philosophy, 26:69–77, 2010.
    Google ScholarLocate open access versionFindings
  • J. Pearl and E. Bareinboim. Transportability of Causal and Statistical Relations: A Formal Approach. In Proceedings of the Twenty-Fifth Conference on Artificial Intelligence (AAAI-11), pages 247–254, Menlo Park, CA, 8 2011.
    Google ScholarLocate open access versionFindings
  • J. Pearl and D. Mackenzie. The Book of Why. Basic Books, New York, 2018.
    Google ScholarFindings
  • J. Pearl and J. M. Robins. Probabilistic evaluation of sequential plans from causal models with hidden variables. In P. Besnard and S. Hanks, editors, Uncertainty in Artificial Intelligence 11, pages 444–453. Morgan Kaufmann, San Francisco, 1995.
    Google ScholarLocate open access versionFindings
  • I. Shpitser and J. Pearl. Identification of Joint Interventional Distributions in Recursive semiMarkovian Causal Models. In Proceedings of the Twenty-First AAAI Conference on Artificial Intelligence, volume 2, pages 1219–1226, 2006.
    Google ScholarLocate open access versionFindings
  • I. Shpitser and E. Sherman. Identification of Personalized Effects Associated With Causal Pathways. In Proceedings of the 34th Conference on Uncertainty in Artificial Intelligence, pages 530–539, 2018.
    Google ScholarLocate open access versionFindings
  • P. Spirtes, C. N. Glymour, and R. Scheines. Causation, Prediction, and Search. MIT Press, 2nd edition, 2001.
    Google ScholarFindings
  • J. Tian. Identifying Conditional Causal Effects. In Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence, UAI ’04, pages 561–568, Arlington, Virginia, United States, 2004. AUAI Press.
    Google ScholarLocate open access versionFindings
  • J. Tian. Identifying Dynamic Sequential Plans. In In Proceedings of the Twenty-Fourth Conference Annual Conference on Uncertainty in Artificial Intelligence (UAI-08), page 554–561, Corvallis, Oregon, 2008. AUAI Press.
    Google ScholarLocate open access versionFindings
  • J. Tian and J. Pearl. A General Identification Condition for Causal Effects. In Proceedings of the Eighteenth National Conference on Artificial Intelligence (AAAI 2002), pages 567–573, Menlo Park, CA, 2002. AAAI Press/The MIT Press.
    Google ScholarLocate open access versionFindings
  • J. Tian and J. Pearl. On the Testable Implications of Causal Models with Hidden Variables. Proceedings of the Eighteenth Conference on Uncertainty in Artificial Intelligence (UAI-02), pages 519–527, 2002.
    Google ScholarLocate open access versionFindings
  • J. Woodward. Making Things Happen. Oxford University Press, New York, NY, 2003.
    Google ScholarFindings
Full Text
Your rating :
0

 

Tags
Comments