Academic literature on the topic 'Debiased machine learning'

In this article, we introduce a package, ddml, for double/debiased machine learning in Stata. Estimators of causal parameters for five different econometric models are supported, allowing for flexible estimation of causal effects of endogenous variables in settings with unknown functional forms or many exogenous variables. ddml is compatible with many existing supervised machine learning programs in Stata. We recommend using double/debiased machine learning in combination with stacking estimation, which combines multiple machine learners into a final predictor. We provide Monte Carlo evidence to support our recommendation.

Chernozhukov et al. (2016) provide a generic double/de-biased machine learning (ML) approach for obtaining valid inferential statements about focal parameters, using Neyman-orthogonal scores and cross-fitting, in settings where nuisance parameters are estimated using ML methods. In this note, we illustrate the application of this method in the context of estimating average treatment effects and average treatment effects on the treated using observational data.

In this study, we investigate the estimation and inference on a low-dimensional causal parameter in the presence of high-dimensional controls in an instrumental variable quantile regression. Our proposed econometric procedure builds on the Neyman-type orthogonal moment conditions of a previous study (Chernozhukov et al. 2018) and is thus relatively insensitive to the estimation of the nuisance parameters. The Monte Carlo experiments show that the estimator copes well with high-dimensional controls. We also apply the procedure to empirically reinvestigate the quantile treatment effect of 401(k) participation on accumulated wealth.

Many causal and structural effects depend on regressions. Examples include policy effects, average derivatives, regression decompositions, average treatment effects, causal mediation, and parameters of economic structural models. The regressions may be high‐dimensional, making machine learning useful. Plugging machine learners into identifying equations can lead to poor inference due to bias from regularization and/or model selection. This paper gives automatic debiasing for linear and nonlinear functions of regressions. The debiasing is automatic in using Lasso and the function of interest without the full form of the bias correction. The debiasing can be applied to any regression learner, including neural nets, random forests, Lasso, boosting, and other high‐dimensional methods. In addition to providing the bias correction, we give standard errors that are robust to misspecification, convergence rates for the bias correction, and primitive conditions for asymptotic inference for estimators of a variety of estimators of structural and causal effects. The automatic debiased machine learning is used to estimate the average treatment effect on the treated for the NSW job training data and to estimate demand elasticities from Nielsen scanner data while allowing preferences to be correlated with prices and income.

Summary We propose double/debiased machine learning approaches to infer a parametric component of a logistic partially linear model. Our framework is based on a Neyman orthogonal score equation consisting of two nuisance models for the nonparametric component of the logistic model and conditional mean of the exposure with the control group. To estimate the nuisance models, we separately consider the use of high dimensional (HD) sparse regression and (nonparametric) machine learning (ML) methods. In the HD case, we derive certain moment equations to calibrate the first order bias of the nuisance models, which preserves the model double robustness property. In the ML case, we handle the nonlinearity of the logit link through a novel and easy-to-implement ‘full model refitting’ procedure. We evaluate our methods through simulation and apply them in assessing the effect of the emergency contraceptive pill on early gestation and new births based on a 2008 policy reform in Chile.

Summary This paper provides an orthogonal extension of the semiparametric difference-in-differences estimator proposed in earlier literature. The proposed estimator enjoys the so-called Neyman orthogonality (Chernozhukov et al., 2018), and thus it allows researchers to flexibly use a rich set of machine learning methods in the first-step estimation. It is particularly useful when researchers confront a high-dimensional data set in which the number of potential control variables is larger than the sample size and the conventional nonparametric estimation methods, such as kernel and sieve estimators, do not apply. I apply this orthogonal difference-in-differences estimator to evaluate the effect of tariff reduction on corruption. The empirical results show that tariff reduction decreases corruption in large magnitude.

Can machines outperform crowds in financial lending decisions? Using data from a crowd-lending platform, we show that, compared with portfolios created by crowds, a reasonably sophisticated machine can construct financial portfolios that provide better returns while controlling for risk. Further, we find that the machine-created portfolios benefit not only the lenders, but also the borrowers. Borrowers receive loans at a much lower interest rate as the machine can weed out the riskiest loans better than the crowds. We also find suggestive evidence of algorithmic bias in machine decisions. We find that, compared with women, men are more likely to receive loans by machine. We propose a general and effective “debiasing” method that can be applied to any prediction-focused machine learning (ML) applications. We show that the debiased ML algorithm, which suffers from lower prediction accuracy, still improves the crowd’s investment decisions in our context. Our results indicate that ML can help crowd-lending platforms better fulfill the promise of providing access to financial resources to otherwise underserved individuals and ensure fairness in the allocation of these resources.

Identifying causal effects from observational data is a pervasive challenge found throughout the empirical sciences. Very general methods have been developed to decide the identifiability of a causal quantity from a combination of observational data and causal knowledge about the underlying system. In practice, however, there are still challenges to estimating identifiable causal functionals from finite samples. Recently, a method known as double/debiased machine learning (DML) (Chernozhukov et al. 2018) has been proposed to learn parameters leveraging modern machine learning techniques, which is both robust to model misspecification and bias-reducing. Still, DML has only been used for causal estimation in settings when the back-door condition (also known as conditional ignorability) holds. In this paper, we develop a new, general class of estimators for any identifiable causal functionals that exhibit DML properties, which we name DML-ID. In particular, we introduce a complete identification algorithm that returns an influence function (IF) for any identifiable causal functional. We then construct the DML estimator based on the derived IF. We show that DML-ID estimators hold the key properties of debiasedness and doubly robustness. Simulation results corroborate with the theory.

This paper reveals a data bias issue that can profoundly hinder the performance of machine learning models in malicious URL detection. We describe how such bias can be diagnosed using interpretable machine learning techniques and further argue that such biases naturally exist in the real world security data for training a classification model. To counteract these challenges, we propose a debiased training strategy that can be applied to most deep-learning based models to alleviate the negative effects of the biased features. The solution is based on the technique of adversarial training to train deep neural networks learning invariant embedding from biased data. Through extensive experimentation, we substantiate that our innovative strategy fosters superior generalization capabilities across both CNN-based and RNN-based detection models. The findings presented in this work not only expose a latent issue in the field but also provide an actionable remedy, marking a significant step forward in the pursuit of more reliable and robust malicious URL detection.

La causalité est un concept fondamental en science et en philosophie. Dans un contexte où la collecte massive de données de grande complexité s’impose dans tous les domaines, les statistiques jouent un rôle crucial dans l'inférence des causes et des effets. Cette thèse explore des méthodes avancées d'inférence causale. Elle met l'accent sur l'apprentissage de politiques d’action (“politiques” dans la suite), les variables instrumentales (IV), et les approches de différences en différences (DiD).Les méthodes IV et DiD sont utilisées par les chercheurs en épidémiologie, médecine, biostatistique, économétrie et sciences sociales quantitatives. Elles reposent sur des hypothèses restrictives, telles que, d’une part, l'exigence que l'IV n’ait aucun effet direct sur le résultat autre qu’à travers le traitement et, d’autre part, l'hypothèse de tendances parallèles en DiD, qui peut être violée en présence de confusion non mesurée.Dans ce contexte, cette thèse propose une approche innovante de DiD instrumentalisée pour l'apprentissage de politiques. Cette combinaison permet de relâcher certaines des hypothèses clés des méthodes IV et DiD conventionnelles. Des résultats d'identification novateurs pour les politiques optimales en présence de confusion non mesurée sont établis, et une gamme d'estimateurs (de Wald ; par pondération inverse des probabilités ; semi-paramétriques efficaces et multiplement robustes) sont introduits. Des garanties théoriques multiplement robustes sont fournies, incluant le taux cubique de convergence pour les politiques paramétriques et une inférence statistique valide avec des algorithmes de machine learning (ML) flexibles pour l'estimation des paramètres de nuisance. Ces méthodes sont en outre étendues à la configuration de données de panel.La majorité des méthodes d'inférence causale dans la littérature dépendent fortement de trois hypothèses causales standard pour identifier les effets causaux et les politiques optimales. Bien que des progrès aient été réalisés pour relâcher les hypothèses de consistance et de non-confusion, les avancées pour traiter les violations de l'hypothèse de positivité sont restées limitées.Dans ce contexte, cette thèse présente un cadre novateur d'apprentissage des politiques qui ne repose pas sur l'hypothèse de positivité, se concentrant plutôt sur des politiques dynamiques et stochastiques pratiques pour des applications réelles. Des politiques de score de propension incrémentale, ajustant les scores de propension par des paramètres individualisés, sont proposées. Leur analyse ne met en jeu que les hypothèses de consistance et de non-confusion. Ce cadre améliore le concept d'effets d'intervention incrémentale, l'adaptant aux contextes de politique de traitement individualisée, et utilise la théorie semi-paramétrique pour développer des fonctions d'influence efficaces et des estimateurs ML dédiés. Des méthodes pour optimiser les politiques en maximisant la fonction de valeur sous des contraintes spécifiques sont également introduites.De plus, le régime de traitement individualisé optimal (ITR) appris d'une population source peut ne pas se généraliser bien à une population cible en raison des décalages de covariables. Un cadre d'apprentissage par transfert est proposé pour l'estimation de l'ITR dans des populations hétérogènes avec des données de survie censurées à droite, que l’on rencontre fréquemment dans les études cliniques. Un estimateur doublement robuste pour la fonction de valeur ciblée est proposé, qui accommode une large classe de fonctionnelles de distributions de survie. Pour une classe pré-spécifiée d'ITRs, un taux cubique de convergence pour le paramètre estimé indexant l'ITR optimal est établi. L'utilisation de procédures de cross-fitting (ajustement croisé) assure la consistance et la normalité asymptotique de l'estimateur de valeur optimal proposé, y compris lorsque l’on a recours à des méthodes ML flexibles pour estimer des paramètres de nuisance
Causality is a fundamental concept in science and philosophy, and with the increasing complexity of data collection and structure, statistics plays a pivotal role in inferring causes and effects. This thesis delves into advanced causal inference methods, with a focus on policy learning, instrumental variables (IV), and difference-in-differences (DiD) approaches.The IV and DiD methods are critical tools widely used by researchers in fields like epidemiology, medicine, biostatistics, econometrics, and quantitative social sciences. However, these methods often face challenges due to restrictive assumptions, such as the IV's requirement to have no direct effect on the outcome other than through the treatment, and the parallel trends assumption in DiD, which may be violated in the presence of unmeasured confounding.In that context, this thesis introduces an innovative instrumented DiD approach to policy learning, which combines these two natural experiments to relax some of the key assumptions of conventional IV and DiD methods. To the best of our knowledge, the thesis presents the first comprehensive study of policy learning under the DiD setting. The direct policy search approach is proposed to learn optimal policies, based on the conditional average treatment effect estimators using instrumented DiD. Novel identification results for optimal policies under unmeasured confounding are established. Moreover, a range of estimators, including a Wald estimator, inverse probability weighting (IPW) estimators, and semiparametric efficient and multiply robust estimators, are introduced. Theoretical guarantees for these multiply robust policy learning approaches are provided, including the cubic rate of convergence for parametric policies and valid statistical inference with flexible machine learning algorithms for nuisance parameter estimation. These methods are further extended to the panel data setup.The majority of causal inference methods in the literature heavily depend on three standard causal assumptions to identify causal effects and optimal policies. While there has been progress in relaxing the consistency and unconfoundedness assumptions, addressing the violations of the positivity assumption has seen limited advancements.In that context, this thesis presents a novel policy learning framework that does not rely on the positivity assumption, instead focusing on dynamic and stochastic policies that are practical for real-world applications. Incremental propensity score policies, which adjust propensity scores by individualized parameters, are proposed, requiring only the consistency and unconfoundedness assumptions. This approach enhances the concept of incremental intervention effects, adapting it to individualized treatment policy contexts, and employs semiparametric theory to develop efficient influence functions and debiased machine learning estimators. Methods to optimize policy by maximizing the value function under specific constraints are also introduced.Additionally, the optimal individualized treatment regime (ITR) learned from a source population may not generalize well to a target population due to covariate shifts. A transfer learning framework is proposed for ITR estimation in heterogeneous populations with right-censored survival data, which is common in clinical studies and motivated by medical applications. This framework characterizes the efficient influence function (EIF) and proposes a doubly robust estimator for the targeted value function, accommodating a broad class of survival distribution functionals. For a pre-specified class of ITRs, a cubic rate of convergence for the estimated parameter indexing the optimal ITR is established. The use of cross-fitting procedures ensures the consistency and asymptotic normality of the proposed optimal value estimator, even with flexible machine learning methods for nuisance parameter estimation

碩士
國立臺灣大學
經濟學研究所
107
The aim of this paper is to investigate estimation and inference on a low-dimensional causal parameter in the presence of high-dimensional controls in an instrumental variable quantile regression. The estimation and inference are based on the Neyman-type orthogonal moment conditions, that are relatively insensitive to the estimation of the nuisance parameters. The Monte Carlo experiments show that the econometric procedure performs well. We also apply the procedure to reinvestigate two empirical studies:the quantile treatment effect of 401(k) participation on accumulated wealth, and the distributional effect of job-training program participation on trainee earnings.

Machine learning learns patterns from data to improve the performance of the decision-making systems through computing, and gradually affects people’s lives. However, it shows that in current research machine learning algorithms may reinforce human discrimination, and exacerbate negative impacts on unprivileged groups. To mitigate potential unfairness in machine learning classifiers, we propose a fair classification approach by quantifying the difference in the prediction distribution with the idea of correlation alignment in transfer learning, which improves fairness efficiently by minimizing the second-order statistical distance of the prediction distribution. We evaluate the validity of our approach on four real-world datasets. It demonstrates that our approach significantly mitigates bias w.r.t demographic parity, equality of opportunity, and equalized odds across different groups in a classification setting, and achieves better trade-off between accuracy and fairness than previous work. In addition, our approach can further improve fairness and mitigate the fair conflict problem in debiased networks.

Knowledge graphs are important in human-centered AI because of their ability to reduce the need for large labelled machine-learning datasets, facilitate transfer learning, and generate explanations. However, knowledge-graph construction has evolved into a complex, semi-automatic process that increasingly relies on opaque deep-learning models and vast collections of heterogeneous data sources to scale. The knowledge-graph lifecycle is not transparent, accountability is limited, and there are no accounts of, or indeed methods to determine, how fair a knowledge graph is in the downstream applications that use it. Knowledge graphs are thus at odds with AI regulation, for instance the EU’s upcoming AI Act, and with ongoing efforts elsewhere in AI to audit and debias data and algorithms. This paper reports on work in progress towards designing explainable (XAI) knowledge-graph construction pipelines with human-in-the-loop and discusses research topics in this space. These were grounded in a systematic literature review, in which we studied tasks in knowledge-graph construction that are often automated, as well as common methods to explain how they work and their outcomes. We identified three directions for future research: (i) tasks in knowledge-graph construction where manual input remains essential and where there may be opportunities for AI assistance; (ii) integrating XAI methods into established knowledge-engineering practices to improve stakeholder experience; as well as (iii) evaluating how effective explanations genuinely are in making knowledge-graph construction more trustworthy.