Causal Machine Research Articles

Application of Artificial Intelligence to Analyze Data from Randomized Controlled Trials: An Example from DECAAF-II.

Causal machine learning (ML) provides an efficient way of identifying heterogeneous treatment effect groups from hundreds of possible combinations, especially for randomized trial data. The aim of this paper is to illustrate the potential of applying causal ML on the DECAAF II trial data. We proposed a causal ML model to predict the treatment response heterogeneity. We applied causal tree learning to the DECAAF II trial data as an example of real applications, identifying subgroups that may be superior when subject to one of the treatments over the other through an easily interpretable process. For each subgroup identified, the characteristics were summarized, and the relationship between treatment arms and risk for atrial tachyarrhythmia (aTA) recurrence among subjects was assessed. Causal tree learning demonstrated that, among all the pre-ablation predictors, dividing subgroups according to age, with a cutoff of 58 years, provides the most heterogeneous subgroups in response to fibrosis-guided ablation in addition to PVI compared to PVI alone. The difference in the risk of aTA recurrence between two treatments was non-significant in older patients (HR= 1.06 95% CI (0.77 - 1.47); P=0.72). However, among the younger patients, the risk of aTA recurrence was significantly lower in the fibrosis-guided ablation group compared to PVI-only (HR= 0.50 95% CI (0.28 - 0.90); P=0.02). Applying causal ML on RCT datasets helped us identify groups of patients that profited from the treatment of interest in an efficient and unbiased manner.

What-if: A causal machine learning approach to control-oriented modelling for building thermal dynamics

Operational optimization of buildings can improve efficiency, and reduce costs and emissions. This optimization typically relies on a model of the building thermal dynamics, which is used by a predictive controller to obtain optimal schedules for heating, ventilation and air conditioning. This model has historically been constructed by domain experts using physical properties of the building. However, this approach scales poorly as more and more residential and commercial buildings need to be modelled and controlled. As a consequence, researchers and practitioners have turned to data-driven models, trained only on observational data. However, this alternative is no panacea: such models often fail to generalize to truly unseen conditions due to their inability to learn cause–effect relationships - i.e. they are not causal, rather they only learn correlational associations between input and target variables. In this paper, we demonstrate this problem using classical machine learning models trained on data from two different use cases (a simulated RC building and nine real-world Dutch buildings). Our results show that, unlike commonly used data-driven methods, causal machine learning (CML) algorithms trained on debiased data can produce accurate models necessary for control-oriented applications which outperform baseline models by over 40%, besides learning the correct causal associations which we verify using a custom testing environment as well as SHAP feature analysis. These results emphasize the need to move beyond simplistic data-driven methods if control-oriented applications are to be realized in a feasible manner.

Adversarial Missingness Attacks on Causal Structure Learning

Causality-informed machine learning has been proposed as an avenue for achieving many of the goals of modern machine learning, from ensuring generalization under domain shifts to attaining fairness, robustness, and interpretability. A key component of causal machine learning is the inference of causal structures from observational data; in practice, this data may be incompletely observed. Prior work has demonstrated that adversarial perturbations of completely observed training data may be used to force the learning of inaccurate structural causal models (SCMs). However, when the data can be audited for correctness (e.g., it is cryptographically signed by its source), this adversarial mechanism is invalidated. This work introduces a novel attack methodology wherein the adversary deceptively omits a portion of the true training data to bias the learned causal structures in a desired manner (under strong signed sample input validation, this behavior seems to be the only strategy available to the adversary). Under this model, theoretically sound attack mechanisms are derived for the case of arbitrary SCMs, and a sample-efficient learning-based heuristic is given. Experimental validation of these approaches on real and synthetic datasets, across a range of SCMs from the family of additive noise models (linear Gaussian, linear non-Gaussian, and non-linear Gaussian), demonstrates the effectiveness of adversarial missingness attacks at deceiving popular causal structure learning algorithms.

Using GPT-4 to guide causal machine learning

Using GPT-4 to guide causal machine learning

Energy optimizing evaporating crystallizers through machine learning

This study presents a machine learning-based approach to optimize energy use in batch-operated evaporating crystallizers within the sugar industry. Leveraging causal machine learning with LightGBM, the study introduces a method to balance high- and low-energy steam usage by dynamically adjusting setpoints in the crystallization process, thus minimizing energy consumption while maintaining optimal batch cycle times. The solution, implemented in collaboration with Nordic Sugar A/S Nakskov, demonstrated a 39.8% increase in lower-energy steam utilization, leading to significant energy savings, including an approximate reduction of 3000 L of diesel consumption per crystallizer per campaign. These findings support a pathway toward sustainable energy practices in industrial crystallization.

Medication Prescription Policy for US Veterans With Metastatic Castration-Resistant Prostate Cancer: Causal Machine Learning Approach.

Prostate cancer is the second leading cause of death among American men. If detected and treated at an early stage, prostate cancer is often curable. However, an advanced stage such as metastatic castration-resistant prostate cancer (mCRPC) has a high risk of mortality. Multiple treatment options exist, the most common included docetaxel, abiraterone, and enzalutamide. Docetaxel is a cytotoxic chemotherapy, whereas abiraterone and enzalutamide are androgen receptor pathway inhibitors (ARPI). ARPIs are preferred over docetaxel due to lower toxicity. No study has used machine learning with patients' demographics, test results, and comorbidities to identify heterogeneous treatment rules that might improve the survival duration of patients with mCRPC. This study aimed to measure patient-level heterogeneity in the association of medication prescribed with overall survival duration (in the form of follow-up days) and arrive at a set of medication prescription rules using patient demographics, test results, and comorbidities. We excluded patients with mCRPC who were on docetaxel, cabaxitaxel, mitoxantrone, and sipuleucel-T either before or after the prescription of an ARPI. We included only the African American and white populations. In total, 2886 identified veterans treated for mCRPC who were prescribed either abiraterone or enzalutamide as the first line of treatment from 2014 to 2017, with follow-up until 2020, were analyzed. We used causal survival forests for analysis. The unit level of analysis was the patient. The primary outcome of this study was follow-up days indicating survival duration while on the first-line medication. After estimating the treatment effect, a prescription policy tree was constructed. For 2886 veterans, enzalutamide is associated with an average of 59.94 (95% CI 35.60-84.28) more days of survival than abiraterone. The increase in overall survival duration for the 2 drugs varied across patient demographics, test results, and comorbidities. Two data-driven subgroups of patients were identified by ranking them on their augmented inverse-propensity weighted (AIPW) scores. The average AIPW scores for the 2 subgroups were 19.36 (95% CI -16.93 to 55.65) and 100.68 (95% CI 62.46-138.89). Based on visualization and t test, the AIPW score for low and high subgroups was significant (P=.003), thereby supporting heterogeneity. The analysis resulted in a set of prescription rules for the 2 ARPIs based on a few covariates available to the physicians at the time of prescription. This study of 2886 veterans showed evidence of heterogeneity and that survival days may be improved for certain patients with mCRPC based on the medication prescribed. Findings suggest that prescription rules based on the patient characteristics, laboratory test results, and comorbidities available to the physician at the time of prescription could improve survival by providing personalized treatment decisions.

Interpretable causal machine learning optimization tool for improving efficiency of internal carbon source-biological denitrification

Interpretable causal machine learning (ICML) was used to predict the performance of denitrification and clarify the relationships between influencing factors and denitrification. Multiple models were examined, and XG-Boost model provided the best prediction (R2 = 0.8743). Based on the ICML framework, hydraulic retention time (HRT), mixture chemical oxygen demand/total nitrogen (COD/TN = C/N), mixture COD concentration, and pretreatment technology were identified as important features affecting the denitrification performance. Further, tapping point and partial dependence analyses provided the range of key factors that precisely regulate denitrification. In the application analysis, HRT (6–10.5 h), mixture C/N (6–12), and mixture COD concentration (300–600 mg L−1) were the appropriate operating ranges, achieving TN removal of approximately 73 %–77 %. The effluent TN and COD concentrations met the discharge standards for wastewater in China (class 1A) and EU. These findings provide support for regulating excess sludge as internal carbon source to promote denitrification.

Do non-farmers pay more for land than farmers?

Abstract The increase in farmland prices in many parts of the world over the past decade has sparked discussions about whether non-farmers pay higher prices for farmland. This study uses a causal machine learning approach with a rich data set of land transactions in Germany to quantify the potential price premium paid by non-farmers. By applying the causal forest method, we uncover the heterogeneity of price premiums and reveal moderating effects of covariates. We find that the average positive price premium by non-farmers decreases with parcel size and distance to a highway exit.

Policy Learning for Many Outcomes of Interest: Combining Optimal Policy Trees with Multi-objective Bayesian Optimisation

AbstractMethods for learning optimal policies use causal machine learning models to create human-interpretable rules for making choices around the allocation of different policy interventions. However, in realistic policy-making contexts, decision-makers often care about trade-offs between outcomes, not just single-mindedly maximising utility for one outcome. This paper proposes an approach termed Multi-Objective Policy Learning (MOPoL) which combines optimal decision trees for policy learning with a multi-objective Bayesian optimisation approach to explore the trade-off between multiple outcomes. It does this by building a Pareto frontier of non-dominated models for different hyperparameter settings which govern outcome weighting. The method is applied to a real-world case-study of pricing targetting subsididies for anti-malarial medication in Kenya.

Investment in intangible assets and economic complexity

We study the nexus between a country's economic complexity and its investment level in intangible assets. Our data spans 27 countries, all sector classifications and 8 intangible categories, which allows us to consider over 188 indicators per country. Our approach offers a complementary more policy-oriented perspective, for economic complexity, compared to relatedness. Results underscore the significance of high knowledge intensity intangibles, such as research and development, in explaining economic complexity. Policy recommendations advocate for prioritizing R&D in manufacturing, alongside fostering complementary activities like employee training, design, and branding in the same sector. In the first part of our results, we introduce an intangible-complexity score for policymakers, enabling the assessment of a country's relative performance in ensuring complexity through investments in various forms of intangible capital, on a country basis. Our second set of estimates, which control for a vast set of potential confounders (almost 800), offer a precise measure of the average impact of intangible investments on complexity, and allow us to argue for the empirical superiority of recent advances in causal machine learning when modeling complexity.

The heterogeneous response of real estate prices during the Covid-19 pandemic

Abstract We estimate the transmission of the pandemic shock in 2020 to the residential and commercial real estate market by causal machine learning, using granular data for Germany. We exploit differences in the incidence of Covid infections and short-time work at the municipal level for the identification of epidemiological and economic effects of the pandemic. We find that (i) a larger incidence of Covid infections temporarily reduced rents for retail real estate; (ii) a larger incidence of short-time work temporarily reduced rents of office real estate; (iii) the pandemic increased prices, particularly in the top price segment of commercial real estate.

Exploring the mechanism of path-creating strategy for latecomers: a combined approach of econometrics and causal machine learning

In this paper, the traditional econometrics method and causal machine-learning method are combined to study the mechanism of a path-creating strategy of latecomers to influence latecomers’ catch-up performance. A total of 283 high-tech manufacturing enterprises listed on the Shanghai and Shenzhen stock exchanges from 2007 to 2019 were selected for the study. OLS linear regression model verifies that path-creating has a positive impact on latecomers’ technological catch-up performance, technological capability plays an intermediary role between path-creating and technology catch-up performance, technological innovation appropriability positively moderates the effect of path-creating on technological capability, and technological innovation cumulativeness negatively moderates the effect of path-creating on technological catch-up but positively moderate the effect of technological capability on catch-up performance. Through machine learning, on the one hand, a conclusion basically consistent with the linear regression model is obtained, but on the other hand, a more heterogeneous situation is presented. Through analyses of the individual treatment effect of a path-creating strategy of latecomers, the Shapley value graph shows the complex influence of different features on the treatment effect of the enterprise using the path-creating strategy. Through the decision tree, some more complex patterns are found. In addition, the decision tree model based on causal analysis can also assist enterprises in making strategic decisions.

Development and Validation of a Policy Tree Approach for Optimizing Intravenous Fluids in Critically Ill Patients with Sepsis and Acute Kidney Injury.

Intravenous fluids are mainstay of management of acute kidney injury (AKI) after sepsis but can cause fluid overload. Recent literature shows that restrictive fluid strategy may be beneficial in some patients with AKI, however, identifying these patients is challenging. We aimed to develop and validate a machine learning algorithm to identify patients who would benefit from a restrictive fluid strategy. We included patients with sepsis who developed AKI within 48 hours of ICU admission and defined restrictive fluid strategy as receiving <500mL fluids within 24 hours after AKI. Our primary outcome was early AKI reversal within 48 hours of AKI onset, and secondary outcomes included sustained AKI reversal and major adverse kidney events (MAKE) at discharge. We used a causal forest, a machine learning algorithm to estimate individual treatment effects and policy tree algorithm to identify patients who would benefit by restrictive fluid strategy. We developed the algorithm in MIMIC-IV and validated it in eICU database. Among 2,091 patients in the external validation cohort, policy tree recommended restrictive fluids for 88.2%. Among these, patients who received restrictive fluids demonstrated significantly higher rate of early AKI reversal (48.2% vs 39.6%, p<0.001), sustained AKI reversal (36.7% vs 27.4%, p<0.001) and lower rates of MAKE by discharge (29.3% vs 35.1%, p=0.019). These results were consistent in adjusted analysis. Policy tree based on causal machine learning can identify septic patients with AKI who benefit from a restrictive fluid strategy. This approach needs to be validated in prospective trials.

Neyman meets causal machine learning: Experimental evaluation of individualized treatment rules

Abstract A century ago, Neyman showed how to evaluate the efficacy of treatment using a randomized experiment under a minimal set of assumptions. This classical repeated sampling framework serves as a basis of routine experimental analyses conducted by today’s scientists across disciplines. In this article, we demonstrate that Neyman’s methodology can also be used to experimentally evaluate the efficacy of individualized treatment rules (ITRs), which are derived by modern causal machine learning (ML) algorithms. In particular, we show how to account for additional uncertainty resulting from a training process based on cross-fitting. The primary advantage of Neyman’s approach is that it can be applied to any ITR regardless of the properties of ML algorithms that are used to derive the ITR. We also show, somewhat surprisingly, that for certain metrics, it is more efficient to conduct this ex-post experimental evaluation of an ITR than to conduct an ex-ante experimental evaluation that randomly assigns some units to the ITR. Our analysis demonstrates that Neyman’s repeated sampling framework is as relevant for causal inference today as it has been since its inception.

Double marginalization in the pricing of complements: The case of US freight railroads

AbstractMonopolists selling complementary products charge a higher price in a static equilibrium than a single (multiproduct) monopolist would, reducing both the industry profits and consumer surplus. Firms could instead reach a Pareto improvement by lowering prices to the single‐monopolist level. We analyze pricing data of railroad coal shipping in the United States. We compare a coal producer that needs to ship from A to C, with the route passing through B, in two cases: (1) the same railroad owning AB and BC and (2) different railroads owning AB and BC. We do not find that the price in case (2) is higher than the price in case (1), suggesting that the complementary monopolist pricing inefficiency is absent in this market. Our findings are robust to propensity score blocking, causal machine learning algorithms, and difference‐in‐differences analysis. Our results have implications for vertical mergers, tragedy of the anticommons, mergers of firms selling complements, elimination of double marginalization, and royalty stacking and patent thickets.

Glucagon-like Peptide 1 Receptor Agonists and Asthma Exacerbations: Which Patients Benefit Most?

Rationale: Although recent evidence suggested that glucagon-like peptide 1 receptor agonists (GLP1RAs) might reduce the risk of asthma exacerbations, it remains unclear which subpopulations might derive the most benefit from GLP1RA treatment. Objectives: To identify characteristics of patients with asthma that predict who might benefit the most from GLP1RA treatment using real-world data. Methods: We implemented an active-comparator, new-user design analysis using commercially ensured patients 18-65 years of age from MarketScan data for 2007-2019 and identified two cohorts: GLP1RAs versus thiazolidinediones and GLP1RAs versus sulfonylureas. The outcome was acute exacerbation of asthma (hospital admission or emergency department visit for asthma) within 180 days after initiation. We applied iterative causal forest, a novel causal machine learning subgrouping algorithm, to assess heterogeneous treatment effects. In identified subgroups, we predicted propensity score, conducted propensity score trimming, and then estimated adjusted risk differences for the effect of GLP1RAs relative to comparators on asthma exacerbation using inverse probability treatment weighting in the propensity score-trimmed subpopulation. Results: Among 10,989 patients initiating GLP1RAs or thiazolidinediones and 17,088 patients initiating GLP1RAs versus sulfonylurea, GLP1RA initiators had fewer exacerbations, with adjusted risk differences of -0.5% (95% confidence interval [CI], -1.1% to 0.1%) and -1.6% (95% CI, -2.2% to -1.1%), respectively. In the GLP1RA versus sulfonylurea cohort, in which we observed a beneficial effect, our iterative causal forest analysis identified five subgroups with different treatment effects, defined by the number of emergency department visits, the number of prescriptions for short-acting β2-agonists, the number of prescriptions for inhaled steroids and long-acting β-agonists (either combination therapy or concurrent use), and age ≥ 50 years. Among these, patients with two or more emergency department visits during the 12-month baseline period had the largest absolute exacerbation risk reduction, with a decrease of 2.8% for GLP1RAs (95% CI, -4.8% to -0.9%). Conclusions: GLP1RAs demonstrated a beneficial effect on reducing asthma exacerbation relative to sulfonylureas. Patients with asthma with two or more emergency department visits (a proxy for disease severity) benefit most from GLP1RAs. Emergency department visit frequency, the number of maintenance and reliever inhalers, and age might help individualize prediction of the short-term benefit of GLP1RAs on asthma exacerbation.

Recommending Career Transitions to Job Seekers Using Earnings Estimates, Skills Similarity, and Occupational Demand

This paper describes a career recommendation algorithm that uses government administrative data to help job seekers discover new careers that similar job seekers have successfully switched to in the past. Algorithm development was motivated by workers and contractors who were displaced by the COVID-19 economic crisis and by workers in declining industries seeking new careers in growing ones. Traditional job boards available through state government websites list all available jobs but do little to remove uncertainty associated with moving to a new industry or occupation. Our recommendation algorithm can lower this uncertainty. It uses causal machine learning techniques and administrative data on the universe of individual-level employment histories and earnings to identify career transitions that have resulted in increased earnings and employment for previous job seekers. We combine these estimates with measures of skill similarity across occupations, derived from natural-language processing of millions of full-text job descriptions, and with occupational demand, as measured by nightly job posting volume. The algorithm parses applicant resumes and returns recommended careers that use similar skills, have available jobs, and are estimated to lead to higher earnings and employment. We have implemented our algorithm in production workforce development systems in five U.S. states.

Causal Machine Learning in Marketing

This study reviews three primary purposes of causal machine learning (CML) in marketing, merging impact evaluation of marketing interventions with machine learning algorithms for learning statistical patterns from data. Firstly, CML enables more credible impact evaluation by considering important control variables that simultaneously influence the intervention and business outcomes (such as sales) in a data-driven manner. Secondly, it facilitates the data-driven detection of customer segments for which a marketing intervention is particularly effective or ineffective, a process known as effect moderation or heterogeneity analysis. Thirdly, closely related to the second point, it allows for optimal customer segmentation into groups that should and should not be targeted by the intervention to maximize overall effectiveness. The discussion is grounded in recent empirical applications, all of which aim to enhance decision support in marketing by leveraging data-driven evaluation and optimization of interventions across different customer groups.

The effect of plough agriculture on gender roles: A machine learning approach

SummaryThis paper undertakes a replication in a wide sense of a recent study that examines the relationship between historical plough agriculture and current gender roles. We revisit the main research question with recently developed causal machine learning methods, which allow researchers to model the relationship of covariates with the treatment and the outcomes in a more flexible way, while also including interactions and nonlinearities that were not considered in the original analysis. Our results suggest an even larger negative effect of the historical plough adoption on female labor force participation than what the original analysis found. The paper highlights the benefits of using causal machine learning methods in applied empirical economics.

Returns to solar panels in the housing market: A meta learner approach

This paper aims to estimate the returns to solar panels in the UK residential housing market. Our analysis applies a causal machine learning approach to Zoopla property data containing about 5 million observations. Drawing on meta-learner algorithms, we provide strong evidence documenting that solar panels are directly capitalized into sale prices. Our results point to a selling price premium above 6% (range between 6.1% to 7.1% depending on the meta-learner) associated with solar panels. Considering that the average selling price is £230,536 in our sample, this corresponds to an additional £14,062 to £16,368 selling price premium for houses with solar panels. Our results are robust to traditional hedonic pricing models and matching techniques, with the lowest estimates at 3.5% using the latter. Despite the declining trend, the additional analyses demonstrate that the positive premium associated with solar panels persists over the years.