CLARE：一种用于韧性估计的因果机器学习方法

11292 CLARE A Causal machine Learning Approachto Resilience Estimation Talip KilicMarco LettaPierluigi MontalbanoFederica Petruccelli Policy Research Working Paper 11292 Abstract that CLARE outperforms existing resilience metrics andalternative approaches to predict food insecurity out-of-sample—both in the future (dynamic forecasting) and inheld-out countries (cross-sectional prediction). The indexcan be decomposed to causally identify the relative impor-tance of resilience capacities that can insulate populationsfrom shocks. Thus, it can be operationalized in designing,targeting, and monitoring policies and investments thataim to strengthen resilience. CLARE’s deployment—paired This paper proposes a new resilience index, CLARE (Causalmachine Learning Approach to Resilience Estimation),which is rooted in an impact evaluation framework andcausal machine learning algorithms applied to longitudinalhousehold survey data. The indicator is model-agnos-tic, data-driven, scalable, and normatively anchored towellbeing thresholds, and can be either shock-specific ora general-purpose resilience metric. The paper providesan empirical demonstration of constructing the CLAREresilience index, leveraging more than 28,000 householdobservations from 19 nationally representative, longitu- The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about developmentissues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry thenames of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those Talip Kilic‡, Marco Letta§, Pierluigi Montalbano§, Federica Petruccelli JEL-Codes:C31; O12; O15. Keywords:resilience;causal machine learning;longitudinal household surveys,impact 1.Introduction In the era of polycrisis, identifying the subjects most in need of resilience-building interventions iscrucial. Essential requirements for effectively targeting and evaluating these interventions include thescalability of resilience indicators, their explicit anchoring to the myriad of shocks and stressors affectingwellbeing, and the alignment between the adopted resilience measures and the wellbeing outcomes theyare meant to reflect. In other words, resilience indicators should accurately predict wellbeing measuresout of sample. By ‘out-of-sample prediction’, we refer to the ability of a resilience indicator to correctlyidentify wellbeing outcomes when applied to previously unseen data. These can be either data from other We address these problems by leveraging Machine Learning (ML) algorithms to construct a novelindicator of household resilience. These algorithms excel at predicting out-of-sample and were designedto address prediction policy problems (Kleinberg et al., 2015; Athey & Imbens, 2019). However, we donot simply rely on predictive ML techniques, which have now become standard tools in the econometrictoolkit of empirical economists. Our approach incorporates recent innovations in causal ML (Wager & threshold under the shock.2The novelty lies in using causal ML to construct a data-driven weighting scheme for aggregating the resilience components, derived from the counterfactual evaluation of theunderlying causal relationships between wellbeing, shocks, and resilience drivers. As such, CLARE is basedon a linear aggregation method using weights estimated non-linearly,thereby balancing complexity and interpretability.In this way, CLARE addresses the challenging issue of crediblyaggregating different resilience subcomponents into a single composite indicator of resilience. Whilecomposite indicators offer a synthetic and comprehensive understanding of the phenomena under study, This is where our causal ML approach comes into play: we use it to derive importance weights for eachresilience subcomponent in intermediating the causal relationshipbetween the shock of interest and theoutcome to which resilience capacity is intended to be indexed.Consequently, the entire causal MLprocedure serves the main purpose of conducting a preliminary counterfactual estimation of therelationshipbetween wellbeing outcomes,covariate shocks,and intermediating variables.This Importantly, while in the illustrative application we show how to compute CLARE using a specifictechnique—causal forests (Athey et al., 2019; Wager & Athey, 2018)—CLARE is not tied to any specificmethodology. As long as the chosen approach estimates granular and heterogeneous treatment effects and enables the establishment of an objective hierarchy among the drivers of heterogeneity, any causal Despite being data-driven, CLARE is not a black box. First, the identification of the underlying causalrelationships is thoroughly rooted in the potential outcomes framework (Imbens & Rubin, 2015). Second,while CLARE remains agnostic regarding the relative importa