JSM 2026 offered me a view of statistics centered on the connection among measurement, uncertainty, computation, and decisions.
When producing another model is the easy part of the meeting
An applied team can now produce several plausible models during the time once required to implement one. A standard library supplies reliable baselines. Automated tuning searches configurations. Foundation-model components widen the candidate set. AI-assisted coding reduces the time between a modeling idea and an executable comparison.
The resulting review meeting may still stall.
One candidate has the best discrimination, another the best calibration, and a third is easier to deploy. The outcome is defined through diagnosis codes whose relationship to disease varies with testing and access. The training population differs from the hospital that will use the model. Intervention capacity makes the threshold a resource decision. Subgroup performance changes depending on whether the team examines crude capture, risk-standardized behavior, or clinical benefit. Approximate computation has produced uncertainty estimates, although nobody has checked whether the approximation preserves the quantity that matters to the decision.
Producing another candidate does not resolve any of these questions. Lower implementation cost can actually make them more prominent because the organization has more models to compare, more outputs to govern, and more routes from a plausible score to a consequential action.
This was the boundary I went to JSM 2026 hoping to see: where statistical work concentrates in an AI/ML-heavy research environment. The answer I found is necessarily partial. I selected sessions according to my background in healthcare analytics, statistics, data science, and software. The itinerary favored problems I could understand and expected to reuse. It was not a representative sample of JSM, and another attendee could reasonably have found a different intellectual center.
Within that selected itinerary, the pattern was consistent enough to examine. As the technical synthesis in Part 2 showed, many of the talks began where a predictive-model comparison often ends. They asked what should be measured, which target should be optimized, what assumptions identify the comparison, what an approximation preserves, and what decision the result can support.
Easier construction moves effort into comparison and governance
Predictive modeling has not become universally easy. Specialized data, distribution shift, small samples, rare outcomes, latency constraints, and changing environments continue to make strong prediction difficult. A notebook that fits a model is far from a reliable production system.
The narrower change is still important. In many common workflows, the cost of producing a plausible candidate has fallen. A team can compare gradient boosting, generalized linear models, neural architectures, and foundation-model features without implementing each method from its mathematical definition. AI assistance can connect libraries, prepare evaluation code, and translate an idea into a runnable prototype. These capabilities expand access to model construction without establishing that the resulting model is appropriate.
More candidates increase the burden on evaluation. A team has to determine whether a small average-loss difference exceeds the uncertainty created by scoring models on the same observations, whether the least bad candidate is adequate, and whether a narrow gain from a complex model justifies its monitoring and explanation cost. The likelihood-as-data talk made statistical ties and poor absolute fit legitimate outcomes rather than inconveniences to be removed from a leaderboard.
Consider a hospital choosing among four readmission models. Automated tuning can return the best cross-validated score before the deployment group has agreed on whether the model will trigger a phone call, pharmacist review, or home visit. Those actions have different capacities and different costs of false positives. A small ranking advantage among models may disappear after the threshold is chosen for the actual service. The cheaper model search has created useful options, while the main design work now lies in specifying the intervention and evaluating the candidates under it.
This changes how effort should be planned. Model development cannot consume the entire analytical budget while validation, observation-process analysis, and deployment design are treated as final checks. As the cost of generating candidates declines, teams should expect more work after model fitting: independent evaluation, comparison under shared uncertainty, monitoring design, and explicit review of how outputs enter decisions.
Statistics defines the connection between data and claims
A predictive target can look obvious until the data-generating institution is examined. In EHR research, documented diagnosis is produced through disease, care access, testing, clinician judgment, and coding. A missing code combines true negatives with patients whose disease was never detected. Training on documentation can therefore produce an accurate model of the care process while being a poor model of underlying disease.
The underdiagnosis talk and its targeted-validation session treated this as a design problem. A small amount of gold-standard adjudication, allocated deliberately, can help separate disease risk from the probability of receiving a diagnosis. The useful records to validate depend on the final estimand. A study of hidden disease prevalence, a regression effect, and a subgroup contrast may require different Phase II samples.
The same reasoning appeared in geographic measurement. Straight-line distance to healthy food is inexpensive and can be badly misleading when highways, rivers, or street networks lengthen the feasible route. The proxy error can change estimated disparities and determine which communities cross a policy threshold. Deciding where to buy route-based measurements is part of the analysis because the measurement budget affects the precision and equity of the eventual claim.
Selective labels made the institutional point sharper. When a prior policy determines which outcomes are recorded, comparisons among new algorithms may be only partially identified; the resulting set shows what the existing evidence leaves unresolved.
These cases place target definition and observation-process analysis near the beginning of a project. Before optimizing a method, the team needs to state which scientific quantity or decision it wants to improve and how the available variables came to exist. Model documentation should record more than feature names and training ranges. It should describe which variables are direct measurements, which are institutional outputs, what remains unobserved, and what added assumptions connect the record to the intended claim.
Statistical reasoning does not own this work. Causal inference, machine learning, epidemiology, economics, measurement science, and domain practice all contribute. What stood out in the sessions I attended was how consistently statistical methods made these connections explicit and treated their uncertainty as part of the result.
Computation has to answer to the scientific quantity
Modern analytical systems are built from approximation. Mini-batches replace full-data gradients. Surrogates replace expensive simulators. Lower precision accelerates matrix operations. Compression reduces storage and movement. Basis truncation turns an impossible representation into a tractable one.
The relevant distinction is the target each approximation must preserve. Stochastic-gradient algorithms may estimate a population average adequately while distorting a small subgroup interval; climate emulators can reproduce average temperature and fail at the tails or dependence required for extreme-event research. Compression may preserve pointwise values while altering gradients or classifications, and a surrogate fitted across a broad parameter box can remain inaccurate near the posterior region used for calibration.
This is where statistics contributes an accountability rule for computation. Numerical error, Monte Carlo error, representation error, and model discrepancy should be connected to the quantity reported to scientists or decision-makers. “Fast” and “accurate” are incomplete descriptions until the workload and the protected summary are named.
The high-performance statistical computing session made that connection unusually concrete. Spherical representations created the structure needed for exascale climate emulation. Mixed precision and distributed runtimes created additional approximation and reproducibility questions. Scientific compression raised the possibility that stored bits contain predictable structure without contributing to a downstream analysis. Hardware and software choices had become part of the inferential path.
An applied team can turn this into reviewable practice. Define acceptance tolerances for the summaries that determine the scientific conclusion. Benchmark end-to-end analysis rather than isolated kernels. Record precision, libraries, random-number behavior, and hardware when they can affect results. Spend higher precision or additional computation where sensitivity analysis shows that the target needs it.
Approximation is unavoidable at scale. Accountability comes from stating what was preserved, what was allowed to change, and why that trade-off was adequate for the intended use.
Honest inference sometimes preserves several answers
Many reporting systems are designed to collapse uncertainty. Dashboards, model-selection pipelines, calibration reports, and fairness reviews are commonly expected to return a score, winner, parameter estimate, or parity judgment.
Several JSM sessions treated multiplicity of answers as substantive evidence. Selective labels can produce an identified set because unobserved outcomes do not justify a point comparison. Paired model losses can support a statistical tie. Computer-model calibration can yield several posterior modes when physically different parameter regimes produce similar observations. Crude and risk-standardized fairness metrics can disagree because one describes operational capture and the other isolates model behavior at a common risk distribution. Net benefit can change with prevalence, threshold, false-positive cost, and intervention capacity.
These outputs are harder to communicate, but forcing them into one number can remove the part of the analysis that a decision-maker most needs. Two tied models invite a choice based on simplicity, monitoring cost, or domain constraints. A multimodal calibration posterior identifies competing physical explanations and the additional measurement that might distinguish them. A fairness trade-off makes the value judgment visible rather than embedding it in an unexamined metric.
Uncertainty in this sense is broader than interval width. It can concern identification, model class, posterior topology, transport, or competing objectives. The analytical task is to represent the form of ambiguity that matters to the decision.
Teams need reporting formats that can carry these results. An identified region, practical-equivalence set, trade-off curve, or collection of mode-specific predictions may be more defensible than a rank. The accompanying discussion should state which additional data, assumption, or value judgment would narrow the choice. Decision-makers still have to act, but the analysis should not claim that the evidence made a choice it did not make.
Preserving several answers can also improve the next study. If two calibration modes imply different physical mechanisms, the difference suggests which measurement could separate them. If subgroup benefit depends on the threshold, the unresolved choice identifies the capacity or harm-benefit judgment that governance must supply. If selective labels leave a wide performance frontier, targeted outcome collection can be evaluated by how much it would narrow the relevant part of that frontier. Ambiguity becomes a guide to information collection instead of a disclaimer attached after the main result.
The boundary is organizational as well as mathematical
My previous conference experiences often began from different parts of an applied system. Healthcare meetings foreground clinical operations, payment, access, and policy consequences. Information-technology meetings foreground implementation, architecture, reliability, and scale. Those are tendencies rather than clean divisions; both kinds of conferences contain much more.
The statistical sessions I selected repeatedly examined the validity of the links among those concerns. They asked which records deserved expensive validation, what information could move across institutions, which feedback should be blocked between model modules, what approximation an engineering team could accept, and what uncertainty a governance group needed to see. Each question crossed an organizational boundary.
Working this way places statisticians in data collection before modeling begins. The role includes deciding which charts to adjudicate and which populations need stronger measurement, collaborating with engineers on numerical tolerances and reproducibility, and working with clinical or policy teams on thresholds and capacity. It also requires telling leadership when the evidence supports several alternatives instead of one.
This is a whole-system contribution, although it is not a solitary one. It requires domain experts who can define consequences, engineers who understand the implementation, and institutional knowledge about how data and decisions are produced. Statistical reasoning supplies methods for making the target, assumptions, uncertainty, and decision connection inspectable across those roles.
What I brought home from JSM
I attended JSM looking for a clean boundary between statistics and an AI/ML research environment. The selected sessions gave me something more useful: a division of attention rather than a division of tools.
Many methods were shared across fields. The distinctive concentration appeared in the questions attached to them: the quantity to estimate, the process that made a label observable, the population supporting the claim, the information worth collecting next, and the approximation introduced by computation. The same reasoning connected difficult posterior geometry to model diagnosis and predictive output to eventual benefit or harm.
Those questions will change how I review projects. I want the decision target stated earlier, before a convenient metric becomes entrenched. I want label-generating institutions described alongside data dictionaries. I want validation design treated as part of analysis and approximation judged by downstream consequences. When the evidence supports several explanations or policies, I want the report to preserve that structure instead of compressing it into one confident answer.
The itinerary remains an important limitation. I found this version of statistics partly because I chose sessions near the edge of my own competence and likely future work. Another path through JSM would have produced another map. Still, this selected path answered the practical question that brought me to the conference.
The boundary I saw did not depend on who could fit the most powerful model. It appeared in the work required to decide what the data support, which assumptions carry the result, what uncertainty survives computation, and what action the analysis can defend.