Causal inference has had a remarkably productive decade. The literature in 2026 is not the literature I would have described in 2016, and practitioners today have tools that would have been considered research-frontier seven years ago. The maturity of the field creates a particular problem. It is harder to tell where the remaining intellectual leverage is. The easy wins have been made. What remains requires judgment about which open problems matter and which are artifacts of the methods-development tradition.
I want to make a specific, opinionated call about three research frontiers where the marginal research dollar produces the most value over the next three years. I will name the problems, explain why they are underdeveloped, identify the researchers whose work is closest, and sketch what a serious program on each would look like. Other calls are defensible. These are mine.
The three frontiers: heterogeneous treatment effect estimation under panel-data constraints, causal evaluation frameworks for LLM-driven decision agents, and observational-experimental data fusion for elasticity estimation at scale. Each is a frontier I would staff if I were leading a causal-ML research group.
Frontier 1: Panel data with genuinely heterogeneous effects
The modern causal-ML literature has two pieces that have not been stitched together. On one side, the heterogeneous-treatment-effect literature has matured (causal forests, DML, X-learners, R-learners). On the other, the panel-data DiD literature has revolutionized what we know about staggered-adoption bias (Callaway-Sant’Anna, Sun-Abraham, de Chaisemartin-D’Haultfœuille, Goodman-Bacon). The stitching is incomplete.
In industry, panel data is almost always unbalanced, staggered, and genuinely heterogeneous. Firms observe the same customers or merchants or markets repeatedly over time. Treatment is rolled out across cohorts. Treatment effects vary across units. The standard tools fail in specific, diagnosable ways.
A causal forest fit to panel data without respecting the within-unit fixed effects gives wrong-sign CATE estimates when the variation across units is larger than the variation within them. This is not a theoretical possibility. It is a failure mode I have watched in production.
A DiD estimator on a balanced-panel subset with heterogeneous effects gives a weighted average of treatment effects where the weights depend on cohort sizes, not on whatever the decision-maker cares about. A two-way-fixed-effects regression on a staggered panel with heterogeneous effects is, by now famously, biased in both sign and magnitude in the worst case.
What the field has not done is develop estimators that handle the three things simultaneously. Heterogeneity, meaning varying treatment effects across units and over time. Panel structure, meaning within-unit fixed effects absorbing unit-level variation. Staggered adoption, meaning different cohorts treated at different times. Each sub-problem has been solved in isolation. The joint problem is where industrial practice lives and where the methodology is weakest.
The closest current work. Callaway-Sant’Anna’s group-time ATT estimator handles staggered adoption implicitly but assumes homogeneity within cohort. Athey-Imbens 2018 on panel-data matching handles fixed effects but not heterogeneity. Wager-Athey causal forests handle heterogeneity without the panel-structure adjustment. A unified estimator, a causal forest with within-unit fixed effects residualized on staggered-adoption panels, would close the gap. The estimator would need to residualize within-unit fixed effects before the causal-forest machinery runs, respect cohort-specific adoption times in the group-time-ATT framework, and deliver pointwise CATE estimates with valid confidence intervals.
A serious research program on this problem would proceed as follows. Formalize the estimator. Prove -consistency and asymptotic normality. Validate on simulation studies with explicit panel-heterogeneity-staggered DGPs. Apply to canonical public datasets (Lalonde, Dehejia-Wahba, NSW) with synthetic panel extensions. Open-source the implementation.
I expect this program to produce two or three papers over 24 months and an open-source library that becomes the default tool for the intersection of the two subfields. The researchers best positioned to lead the work are the Athey-Wager-Tibshirani group at Stanford, the Callaway-Sant’Anna collaboration, and the Chernozhukov group at MIT. Much of the work is also reachable from industrial applied-research groups. The research is more likely to be practically useful if the team has regularly seen the real messiness of industrial panels.
Frontier 2: Causal evaluation of LLM-driven decision agents
This frontier does not yet exist in the published literature, which is why it is a frontier. Production systems are increasingly driven by LLM-based agents that recommend policies, prices, contracts, and operational decisions. The agents are trained and evaluated on benchmarks that test their task completion. Almost no research measures their causal impact on the downstream outcomes they are supposed to affect.
The gap is consequential. If an agent recommends a pricing change, we want to know whether the pricing change caused the subsequent revenue shift. If the agent has been operating for six months, we need a causal methodology for evaluating it that handles the time-varying treatment assignment the agent itself controls.
This is exactly the kind of problem that causal-inference theory was built for. The theory, however, assumes the treatment is either randomized or exogenously assigned. An LLM agent that makes its own recommendations based on the data it has seen is neither. It is making endogenous decisions that affect the distribution of subsequent data. The standard identification assumptions fail.
A rigorous research program on agent causal evaluation would need to do four things. Formalize the agent as a treatment-assignment policy and characterize its identifying restrictions. Propose estimators that handle the policy-driven endogeneity, possibly via instrumental-variables approaches using randomization in the agent’s internal reasoning (random temperature perturbations, for instance). Handle the time-varying selection problem, where the agent’s recommendations this week affect the state the agent observes next week. Develop sensitivity analyses for the inevitable residual unmeasured confounding.
The closest current work is scattered across adjacent literatures. The reinforcement-learning causal-inference literature (Ernst et al. 2005 fitted Q-iteration, Precup 2000 importance-weighted off-policy evaluation, Athey-Wager 2021 policy learning) addresses part of the picture but focuses on single-episode decisions. The sequentially-randomized-experiments literature (Robins et al. 1999, Hernán-Robins 2020 Chapter 19) handles time-varying treatments but was developed for clinical trials. The gap between those literatures and the problem of evaluating a shipped LLM agent is substantial.
The substantive reason this frontier matters: without rigorous causal evaluation, agent-driven decision systems are trusted or not trusted on the basis of intuition. That is not a sustainable governance approach for systems making substantial decisions. Firms that deploy agents responsibly will need the causal methodology to justify the deployment. Regulators who audit such systems will need it to verify the justification. The research that develops the methodology will be disproportionately cited by both practitioners and regulators.
I would argue this is the single most important unsolved methodology problem in applied causal inference for 2026-2028. If I were recruiting into a causal-ML research group, this is the problem I would use to screen candidates. The right candidate should be able to articulate the problem in fifteen minutes and sketch an identification strategy in the remaining time.
Frontier 3: Observational-experimental data fusion
The third frontier is the most mature of the three but still under-exploited. It sits at the intersection of randomized-experiment analysis and observational-data analysis. The question: given both randomized-experimental data (small sample, clean identification) and observational panel data (large sample, confounded), how do you combine them to produce an estimator with better properties than either alone?
In industrial settings, this is the practical question. A firm runs A/B tests on a subset of its users, producing clean ATE estimates but with limited sample and often restricted to specific subgroups. The firm also has panel data on all users, with much more data but subject to selection bias and unmeasured confounding. The ATE from the experiment is identified but noisy. The ATE from the panel is biased but precise. The right answer is a fusion.
The statistical literature has a scattered collection of partial answers. Athey-Imbens 2021 on combining experiments and observational data provides one approach. The meta-learner literature (S-, T-, X-, R-learners from Künzel et al. 2019) implicitly addresses the question for heterogeneous treatment effects. Bayesian hierarchical models (Chipman et al. BART, Hahn et al. BCF) offer another route by pooling information across data sources via shared priors. The targeted maximum likelihood estimation literature (van der Laan, Rose) handles the multi-source case under unified notation.
What is missing is a clear, opinionated practical framework for applied researchers. The existing work is scattered across traditions that do not read each other’s papers. A practitioner who wants to combine an experiment and observational data does not have a single canonical reference to consult. The research opportunity is to produce that reference: a paper, plus software, plus examples that a working data scientist can apply on Monday morning.
The technical content is substantial but not unprecedented. Specify the data sources formally (experiment plus panel) and their identifying assumptions. Propose a fused estimator, perhaps a weighted combination of the experimental ATE and the panel CATE-at-experimental-subgroup, or a more sophisticated shrinkage estimator. Characterize the efficiency gain as a function of sample sizes and effect heterogeneity. Provide diagnostics for when the fusion is dominated by one source or the other. Ship a reference implementation.
This is the research program I have planned as my Flagship 5 paper, targeted for submission in late 2026. It is the most applied of the three frontiers in the sense that the methodological novelty is modest but the practical utility is large. I estimate fifteen hours per week over six months would produce a publishable paper plus software at the level of quality the applied-ML community would adopt.
Why these three, not others
Other candidates I considered and did not name.
Sensitivity analysis for unmeasured confounding. Important, but the infrastructure (Rosenbaum -bounds, E-values, negative controls) is in acceptable shape. Incremental research here has diminishing returns.
High-dimensional instrumental variables. The literature (Belloni-Chernozhukov 2014, among others) is mature. The applied settings that benefit are narrow.
Transportability across populations. Important conceptually, but a narrow applied market. The research belongs in epidemiology and public health more than industrial ML.
Causal discovery from observational data. Intellectually interesting, but the published discovery algorithms rarely work well in industrial settings with mixed data types, high dimensionality, and strong correlations.
Synthetic control improvements. The Abadie-Diamond-Hainmueller framework plus Arkhangelsky et al.’s synthetic DiD have the territory well-covered for most applied needs. Incremental improvements are research-valid but not transformative.
The three frontiers I named share specific properties that make them high-leverage right now.
They are underdeveloped relative to the demand. Panel-data heterogeneity is a problem faced by every industrial analytics team; serious methodology for it is rare. Agent causal evaluation does not yet exist as a literature. Experimental-observational fusion has scattered pieces that nobody has synthesized.
They have clear paths to publication. Each research direction maps to natural target venues (NeurIPS, ICML, AEA, EC, Econometrica, JASA) and has comparison baselines that the research community can evaluate against.
They are suited to applied-research groups. None of them requires a twenty-billion-dollar compute budget. They reward careful methodological work, clean empirical examples, and shipped tooling.
They produce methodology researchers will cite. A researcher known for the panel-data-heterogeneity solution, or the agent-causal-evaluation framework, or the experimental-observational fusion paper, will be recognized by a specific intersection of the applied-causal and applied-ML communities. That recognition is the professional good that research directors trade in.
The research group I would lead
If I were recruiting into a causal-ML research group focused on these three frontiers, the team would be about eight scientists split roughly three-three-two across the frontiers. The group would operate on five habits.
Publish. Three papers per year across the three frontiers, in venues that applied-ML communities respect.
Ship code. One substantial open-source tool per year, maintained at production quality. Tools that other practitioners adopt are the durable credential.
Consult. Regular engagement with two or three industrial partners who are running exactly the problems the research is solving. The consulting produces data for the research and credibility for the team.
Host interns. One PhD intern per researcher per year. Interns do substantive research and often produce the first paper in a new direction. Their presence also recruits future faculty-track researchers.
Run a reading group. Monthly convening on causal-inference and agent-research papers, cross-listed for adjacent groups. The reading group builds intellectual community and surfaces research opportunities.
Researchers I would recruit have a strong economics or biostatistics background plus demonstrated applied-ML fluency. They have shipped at least one production system. They can write clearly in non-technical prose as well as technical LaTeX. Pure methodology researchers who have not interacted with production data are a worse fit than applied researchers who are willing to learn methodology.
Conclusion
The three frontiers (panel-data heterogeneity, agent causal evaluation, observational-experimental fusion) represent where I think the marginal research investment in causal ML pays off over the next three years. They are underdeveloped relative to need, accessible to a small research group, and high-leverage when done well.
A research director in 2026 should have a specific opinion about where to invest. Mine is these three. I am prepared to defend the choice, execute on it, and course-correct as the research produces evidence. The institution that wants to own these three intellectual territories will have a distinctive research program that is not a copy of any frontier lab’s LLM-scaling effort. That distinctiveness is what a serious research group should be selling.
Comments and critique welcome. I am particularly interested in hearing the case for research directions I explicitly set aside.