The research team, led by Nu Hoang and Thin Nguyen of Deakin University’s Applied Artificial Intelligence Institute, together with collaborators from the university’s Global Centre for Preventive Health and Nutrition, set out to solve a problem that has long undermined evaluations of community interventions. Traditional statistical approaches, such as linear regression, demand careful selection of covariates to ensure valid causal inference. Choosing the wrong variables, either omitting a genuine confounder or adjusting for an inappropriate one, can bias estimates of a program’s effect. Even with the right covariates, misspecifying the functional form of the model can still produce misleading conclusions. The team’s answer was not a new algorithm but a principled, end-to-end workflow that chains together causal discovery, formal identification through do-calculus, double machine learning, and interpretable subgroup discovery.

The first stage of the pipeline is causal discovery, the task of inferring cause-and-effect relationships directly from data rather than assuming them from prior theory. The researchers employed a hybrid causal discovery model designed for mixed-type data, meaning it can handle both continuous measurements, such as body mass index, and categorical variables, such as sex. The model works in two phases: a randomized conditional independence test builds the skeleton of the causal graph, and a cross-validation-based scoring function then orients and refines the edges. Because community intervention data are inherently longitudinal, the team adapted the model with temporal constraints, learning a graph over baseline variables first and then forcing edges to flow forward in time. This two-step design proved its worth in validation: it achieved significantly higher log-likelihood on held-out data than a one-step alternative, with a p-value below 0.0001.

The resulting causal graph is a map of how children’s characteristics and behaviors interconnect. Baseline age stood out as the most influential upstream variable, with the highest mean out-degree of 13.4 across bootstrap resamples, linking to physical activity, sedentary behavior, active transport, fruit intake, sweet beverage consumption, takeaway food consumption, and the outcome itself: the change in body mass index z-score over time. Baseline zBMI and sex were the next most recurrent upstream variables. Critically, the confounding pathways from sex, baseline age, and baseline zBMI to program participation appeared with a bootstrap frequency of 1.0, meaning they were perfectly stable across repeated resampling of the data. This stability gave the researchers confidence that the adjustment set identified from the graph could reliably strip away spurious correlations.

With the causal structure established, the team turned to double machine learning, a technique from econometrics that allows flexible machine learning models to be used in causal estimation without sacrificing statistical consistency. The method models both the outcome and the treatment assignment as functions of the covariates, then exploits two tricks to avoid the biases that normally plague machine learning estimators. Cross-fitting splits the data into partitions so that models are never evaluated on the data used to train them, guarding against overfitting bias. Orthogonalization, rooted in the classical Frisch-Waugh-Lovell theorem, regresses out nuisance functions and estimates the causal parameter from residuals, delivering a root-N consistent estimator free of regularization bias. Here, LightGBM models served as the nuisance learners, with hyperparameters tuned by fivefold cross-validated grid search.

The headline result of the estimation stage is a picture of profound heterogeneity. The average treatment effect across all children was a modest reduction of 0.02 in zBMI, with a 95 percent confidence interval spanning from minus 0.09 to plus 0.05, a range that crosses zero. But the individualized estimates told a far richer story, ranging from minus 0.25 to plus 0.15. Roughly 18.7 percent of children had confidence intervals entirely below zero, indicating a genuine benefit, while 8.7 percent had intervals entirely above zero, suggesting the program may have been counterproductive for them. The remaining 72.6 percent had intervals that included zero, reflecting the substantial uncertainty inherent in individual-level causal estimates. The researchers are careful to stress that these individualized figures describe broad patterns of heterogeneity rather than decision-grade predictions for any single child.

To probe what drives this variation, the team first examined linear correlations between the estimated effects and baseline variables. Age showed the strongest association, with a Pearson coefficient of 0.59, while no other baseline variable correlated significantly. Because linear correlation can miss nonlinear structure, the researchers then trained a shallow decision tree to classify children into positive-effect, negative-effect, and no-effect groups based on nine baseline characteristics. The tree partitioned the data using just two variables: age and weekly takeaway food consumption. Children under 8.9 years old formed the most robustly benefited subgroup, with a dominant-class probability of 0.93 and a mean individual treatment effect of minus 0.096, more than four times the pooled average. Children aged 8.9 to 13.6 who ate takeaway food less than once per week also benefited, with a mean effect of minus 0.038. In contrast, adolescents above 15.2 years showed a clearly negative response, with a mean effect of plus 0.018.

The robustness of these findings was tested from multiple angles. Bootstrap validation of the decision tree across 100 resampled datasets confirmed that age was selected as a splitting variable in every single run, while takeaway food consumption appeared in 40 percent of runs and no other variable appeared at all. Refutation tests bolstered the causal estimate itself: a placebo treatment test, which replaces the real treatment with random values, yielded an effect of minus 0.0038 that was statistically indistinguishable from zero, exactly as expected if the original estimate reflects a genuine causal relationship. Subset validation produced an effect of minus 0.016, not significantly different from the original. An E-value sensitivity analysis indicated that an unmeasured confounder would need to be associated with both treatment and outcome by a risk ratio of at least 1.23 to explain away the average effect, though the subgroup effect for younger children is substantially larger and more resilient.

The authors are candid about the limitations of their work. Communities self-selected into the intervention programs, so unmeasured factors such as local political support or socioeconomic resources could confound the results, potentially overstating benefits. Spillover effects, in which children in comparison communities are indirectly exposed to intervention activities, could bias estimates toward the null. The pooled average effect is small and its confidence interval crosses zero, so the researchers emphasize that the study’s main contribution lies in demonstrating how heterogeneity can be identified and characterized, not in claiming a large overall benefit. The individualized estimates are explicitly framed as exploratory and potentially model-dependent, useful for revealing broad subgroup structure rather than for clinical decision-making at the level of a single child.

Even with those caveats, the implications are considerable. If community-based obesity prevention is genuinely most effective for younger children and least effective for older adolescents, then policymakers have a concrete, actionable signal: interventions may need age-sensitive redesign, with different strategies for teenagers than for primary-school children. The interaction analysis adds nuance, finding that the combination of age and takeaway food consumption drives heterogeneity beyond age alone, and that physically active children who eat fewer takeaway meals benefit more. Beyond obesity, the framework is explicitly generalizable. The authors argue that the same workflow, causal discovery to build the graph, do-calculus to identify the adjustment set, double machine learning to estimate effects, and tree-based subgroup analysis to interpret them, could be applied to any complex community intervention where average effects conceal the individuals the program actually helps. In an era when public health budgets are finite and one-size-fits-all programs increasingly look inadequate, that may be the study’s most viral-worthy message: the average is a lie, and the tools to see past it now exist.

Subject of Research: Causal machine learning analysis of heterogeneous treatment effects in community-based childhood obesity prevention interventions

Article Title: Causal machine learning for understanding heterogeneous effects of childhood obesity prevention

Article References: Hoang, N., Nguyen, T., Duong, B., Nichols, M., Brown, V., Backholer, K., Allender, S., & Nguyen, T. (2026). Causal machine learning for understanding heterogeneous effects of childhood obesity prevention. International Journal of Data Science and Analytics, 22(1), Article 326. https://doi.org/10.1007/s41060-026-01156-z

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01156-z

Keywords: causal machine learning, causal discovery, causal inference, childhood obesity, community-based interventions, heterogeneous treatment effects, double machine learning, BMI z-score, public health, age heterogeneity, subgroup analysis, preventive health

A key contribution of the review is a conceptual distinction the authors argue has been chronically blurred in the literature: the difference between AI-as-intervention and AI-for-trial-operations. When an algorithm is itself the treatment being tested—say, a diagnostic tool randomised against standard care—the relevant standard is prospective clinical evaluation, with explicit estimands, protocol-level oversight of model behaviour, and reporting under extensions such as SPIRIT-AI and CONSORT-AI. When AI instead supports trial logistics—feasibility modelling, prescreening, monitoring, or adjudication—the central questions become operational: Does it change workflow? Is it transportable across sites? Is it safe under real-world constraints? The two categories overlap, but they carry different burdens of proof, and conflating them has allowed retrospectively validated models to be promoted on evidence that would never suffice for a drug or device.

The review organises AI’s potential around three intersecting pillars: mechanism, measurement, and efficiency. Mechanism refers to network-medicine approaches that map disease pathways and link multi-omics data, phenotypes, and exposures to define rational patient stratification and biologically grounded endpoints. Measurement centres on digital health technologies—wearables and sensors that capture passive gait speed, respiratory patterns, arrhythmic burden, or step-level fatigue—offering dense longitudinal data on how patients actually function between clinic visits. Efficiency, the most commercially advanced pillar, involves large language models and natural language processing systems that convert free-text eligibility criteria into computable logic, match patients to trials, and extract structured evidence from unstructured records. The authors stress that these pillars are complementary rather than interchangeable, and that a system strong in one dimension cannot substitute for weakness in another.

The recruitment domain offers the most concrete evidence of benefit. Tools such as Criteria2Query and its successors translate eligibility criteria into database queries, while newer systems like TrialMatchAI use retrieval-augmented generation to process structured and unstructured patient data, retrieve candidate trials through hybrid search, and apply chain-of-thought reasoning at the level of individual criteria. In real-world evaluation, TrialMatchAI placed 92 percent of oncology patients somewhere within its top twenty trial recommendations, with expert assessment confirming accuracy above 90 percent. Even more striking, a randomised evaluation of a retrieval-augmented GPT-4 prescreening system for heart failure trials found that AI-assisted eligibility review matched the accuracy of experienced coordinators, reduced false-negative screening errors, and more than doubled prescreening throughput while maintaining transparent, citation-linked reasoning. Crucially, however, those gains materialised only when the AI was coupled to clinician review, site-specific workflow integration, and clear error-mitigation steps.

The authors repeatedly caution that operational AI cannot repeal biology or logistics. Machine learning models can now forecast trial accrual, duration, and the probability of early termination using protocol features, disease epidemiology, and historical site performance, and uncertainty-aware deep learning systems can generate interval-based enrolment estimates rather than single numbers. Digital twins—patient-specific counterfactual simulations—can even test protocol variations and enrichment strategies before a site is activated. Yet if a study depends on slow event accrual, long follow-up, or minimum safety exposure, no algorithm can eliminate the required observation time. A scoping review of 142 studies cited in the analysis found that most feasibility models remain retrospective and confined to individual risk domains, meaning that prospective decision-impact—the demonstration that a model changes what trial teams actually do, early enough to matter—has yet to be shown for most applications.

Endpoint assessment presents a subtler challenge. Digitally derived measures can reduce participant burden and increase measurement density, but the review warns that technological novelty often outpaces endpoint validity. An ambiguous clinical concept does not become more scientific when measured by a sensor; instead, uncertainty simply shifts from coordinators and endpoint committees to the algorithm. The remedy is rigorous endpoint specification: predefining the clinical construct, the derivation from raw data, the assessment threshold and window, and the handling of missing or poor-quality data. Natural language processing pipelines for adjudicating heart failure hospitalisations in cardiovascular trials have already reduced manual review while preserving agreement with clinician committees, and emerging LLM-enabled adjudication may scale this further—provided such systems are validated against adjudicated reference standards, tested for calibration and subgroup performance, and kept subordinate to predefined human review pathways.

In the inferential realm, machine learning methods for estimating heterogeneous treatment effects are spreading rapidly. A scoping review of 32 randomised trials found that causal-forest and related algorithms are now the most common tools for this purpose, and a causal-tree analysis of the DECAAF II trial illustrated how such methods can generate structured, hypothesis-generating insights about age-based differential benefit from fibrosis-guided ablation even when the overall intention-to-treat result was null. But the review insists these analyses be anchored to explicit ICH E9(R1) estimands, with pre-specification in statistical analysis plans, multiplicity control, overlap diagnostics, calibration checks, and Data and Safety Monitoring Board oversight. The same discipline applies to external control arms, where hidden confounding and unclear comparability can silently invalidate an entire comparison. These are not optional regulatory formalities, the authors argue; they are the mechanisms by which trials protect inference from operational complexity.

Governance gaps loom largest for continuously learning and agentic systems. The protocol should state whether a model is locked, periodically recalibrated, or adaptively updated, with defined update cadences, approval pathways, shadow-mode evaluation, drift thresholds, rollback triggers, and version freezes around interim analyses and database lock. Agentic AI—systems that plan, retrieve, and revise outputs across multiple steps—remains, in the authors’ assessment, an emerging architectural pattern rather than a mature standard of practice. Because failure can arise in intermediate steps such as retrieval, normalisation, ranking, or tool invocation, generic benchmark performance is insufficient for trial-critical use. Until stronger evidence exists, the review recommends confining such systems to bounded decision support in shadow mode: drafting eligibility logic, summarising protocol deviations, or prioritising safety review, never autonomously determining enrolment or altering endpoint definitions. Algorithmic bias compounds these concerns, since a prescreening tool with lower sensitivity in underrepresented populations produces not merely uneven performance but inequitable access to trials and a less generalisable evidence base.

The authors close with a practical agenda: prospective, randomised operational trials comparing AI-assisted and standard recruitment, site selection, adjudication, and monitoring; evaluation metrics that capture decision consequences—time saved, screen-failure burden, missed eligibility—rather than discrimination statistics alone; fairness monitoring embedded in the protocol itself, with denominator tracking at every stage from screening to outcome ascertainment; and data infrastructure capturing device versions, sampling cadence, and missingness patterns alongside raw streams. Privacy-preserving techniques such as federated learning earn a pointed caveat: they do not preserve privacy by default, since model updates can leak information, and they solve none of the deeper problems of inconsistent definitions, site heterogeneity, and subgroup imbalance. The bottom line is measured but firm. AI may accelerate learning and lighten workloads today, but trustworthy trials still depend on careful design, explicit human accountability, and—for better or worse—time for outcomes to emerge.

Subject of Research: The state of evidence, gaps, and governance for artificial intelligence across the clinical-trial lifecycle

Article Title: Artificial intelligence in clinical trials—state of the evidence, gaps, and next steps

Article References: Armoundas, A. A., Tarabanis, C., & Loscalzo, J. (2026). Artificial intelligence in clinical trials—state of the evidence, gaps, and next steps. eClinicalMedicine, 100, Article 104196. https://doi.org/10.1016/j.eclinm.2026.104196

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104196

Keywords: artificial intelligence, clinical trials, large language models, eligibility screening, digital health endpoints, event adjudication, heterogeneous treatment effects, algorithmic bias, regulatory governance, federated learning, agentic AI, pharmacovigilance

The authors trace the intellectual lineage of experimentation back to Ronald Fisher’s The Design of Experiments and Jerzy Neyman’s work on agricultural trials, noting that the same statistical machinery now powers decisions in economics, political science, public health, education and digital product development. Randomized evaluations involving millions of observations have transformed social science, generating causal evidence at a pace and scale once unimaginable. Yet as companies and institutions run thousands of experiments each year, often across decentralized teams with competing incentives, the assumptions that made textbook methods reliable no longer hold automatically. The Perspective argues that the research community must re-engage with these practical frictions, drawing on sustained collaboration between academics and the practitioners who operate experimentation platforms at scale.

The first challenge concerns organizational incentives and experimental governance. In large companies, experiments are not conducted by neutral statisticians but by teams whose careers depend on the results. Researchers have begun modeling this as a principal-agent problem, where the people commissioning a test may have strategic reasons to select metrics, stopping rules or reporting practices that favor a preferred outcome. Recent work on principal-agent hypothesis testing and screening for experiments formalizes how such misaligned incentives can distort what gets run and what gets reported. The authors also draw on Friedrich Hayek’s insight about the use of knowledge in society, noting that decentralized organizations generate enormous experimental activity, but coordinating that activity requires governance structures that few firms have built carefully. Democratizing experimentation, they suggest, can accelerate learning, but only if paired with guardrails that prevent local optimization from damaging global objectives.

Privacy, fairness and ethics form the second area of concern. Online experiments involve human subjects who rarely know they are being studied, a situation that sits uneasily with the ethical principles articulated in the Belmont Report. The authors review recent advances in differential privacy, including federated experiment designs that allow companies to run randomized controlled trials while limiting how much any individual’s data can leak into published results. Synthetic data generation offers another path, with new methods producing privacy-conscious datasets suitable for causal effect estimation. Fairness raises its own difficulties: treatments that improve average outcomes may harm specific subgroups, and protected attributes are often unobserved, complicating any assessment of disparate impact. New bounding techniques can estimate the fraction of a population negatively affected by a treatment even when individual-level harm cannot be directly measured, giving practitioners a way to quantify risk rather than ignore it.

The third challenge is estimating long-term impact. Most experiments run for days or weeks, but the decisions they inform concern effects that unfold over months or years. The literature on surrogate end points, from clinical trials in medicine to proxy metrics in technology companies, shows how easily short-term indicators can mislead. A famous cautionary example comes from cardiac medicine, where drugs that suppressed arrhythmias ultimately increased mortality, a lesson the authors invoke to highlight the danger of optimizing the wrong outcome. Newer approaches include the surrogate index, which combines short-term proxies to estimate long-term treatment effects, and methods that pool information across many weak experiments to learn the covariance of treatment effects. Evaluations at Netflix, drawing on hundreds of A/B tests, suggest these tools can improve decision-making, but imperfect surrogates remain an unsolved problem, and sensitivity analysis techniques for unmeasured confounding are increasingly seen as essential.

Fourth, the Perspective examines time-adaptive experimental studies, in which the design of the experiment itself changes as data accumulates. Classical sequential testing, pioneered by Abraham Wald, addressed the problem of peeking at results, but modern platforms monitor experiments continuously by default. Always-valid inference methods and confidence sequences now allow researchers to check results at any time without inflating false positive rates, and these techniques are being deployed in enterprise A/B testing platforms. Beyond monitoring, adaptive assignment schemes such as multi-armed bandits reallocate traffic toward better-performing treatments, trading statistical rigor for user benefit. Phased release strategies using batched bandits balance risk and reward during rollouts, and switchback experiment designs allow causal inference in settings like marketplaces where treatments must alternate over time. Inference after adaptive experiments remains technically demanding, but recent work on demystifying such inference and on conformal methods for distribution shifts is narrowing the gap.

Heterogeneous treatment effects constitute the fifth area. An average treatment effect can conceal enormous variation: a feature that helps most users may actively harm a vulnerable minority, and a policy that works in one region may fail in another. Machine learning methods, including metalearners, causal random forests and recursive partitioning, now allow researchers to estimate how effects differ across individuals, while calibration techniques help discover stable, interpretable subgroups. The challenge is statistical as much as computational: hunting for subgroups multiplies hypothesis tests and invites false discoveries, requiring multiple testing corrections and careful validation. New approaches also bridge prediction and causal targeting, distinguishing the question of who will benefit from the question of who is likely to respond, a distinction that matters enormously when experiments inform personalized product decisions. Interpretable personalized experimentation is emerging as a practical goal, letting non-experts understand which groups an experiment affects and why.

The final and perhaps most timely challenge comes from generative artificial intelligence. Large language models are being proposed both as tools within experiments and as simulated participants, raising fundamental questions about what a digital-era experiment should be. Research on simulated economic agents, sometimes described as Homo Silicus, and on using language models to replicate human subject studies suggests these systems can mimic certain human response patterns, but their biases, training data provenance and tendencies toward sycophancy remain poorly understood. Warnings about the illusions of understanding that AI can create in scientific research loom large. The authors also note that AI agents themselves are becoming subjects of experimentation, with large-scale autonomous negotiation competitions already underway. Whether generative AI will serve as a substitute for human experiments, an augmentation of them, or a new experimental subject entirely is one of the field’s most consequential open questions.

Underlying all six areas is a shared theme: the infrastructure of experimentation has outpaced its statistical and ethical scaffolding. P-hacking, publication bias and the winner’s curse in estimating effects across many experiments are old problems given new urgency by sheer volume. Post-selection inference, always-valid confidence sequences and empirical Bayes methods offer partial remedies, but the Perspective is explicit that many open problems remain unsolved and that progress will require methodological innovation grounded in real operational constraints. The authors, who include researchers from Columbia Business School, Harvard Business School, Stanford University, Cornell University and MIT alongside their industry collaborators, describe their goal as surfacing practical challenges that merit greater attention from the research community.

The stakes extend well beyond technology companies. As governments experiment with digital public services, health systems test behavioral interventions and educators evaluate online learning at scale, the methods developed for platform experimentation are migrating into domains with far less tolerance for error. The authors hope the Perspective will act as a research agenda, encouraging statisticians, economists, computer scientists and social scientists to work directly with the practitioners running experiments on millions of people. In the digital era, they argue, the future of large-scale experimentation depends less on any single technical breakthrough than on building institutions, methods and norms capable of keeping rigorous causal inference aligned with human welfare at unprecedented scale.

Subject of Research: Methodological and organizational challenges facing large-scale randomized experiments in the digital era

Article Title: The future of large-scale experiments and their challenges in the digital era

Article References: Holtz, D., Bojinov, I., Johari, R., Kallus, N., Lal, A., Anand, S., Carlson, K., Cohn, B., Cunningham, T., Deng, A., Dimakopoulou, M., Gandhi, A., Kostyuk, V., Kumar, M., Loh, S.-M., Machmouchi, W., Mao, J., McQueen, J., Meakin, J., … Tingley, M. (2026). The future of large-scale experiments and their challenges in the digital era. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02582-6

Image Credits: AI Generated

DOI: 10.1038/s41562-026-02582-6

Keywords: randomized experiments, A/B testing, causal inference, experimental design, differential privacy, fairness, surrogate metrics, multi-armed bandits, anytime-valid inference, heterogeneous treatment effects, generative AI, large language models