At JSM 2026, the main constraint was deciding where limited attention could produce learning that would survive the conference.
The schedule was larger than my capacity to learn
Before the first full day of the Joint Statistical Meetings (JSM) this year, I had the conference program open in front of me and several reasonable choices occupying the same time block.
One session sat close to work I already knew. I could expect to follow the vocabulary and recognize the main problems, although much of the conceptual structure might be familiar. Another moved into methods I had encountered only at an introductory level. A third looked intellectually attractive while assuming more theory than a short conference talk could possibly rebuild. A less conspicuous option seemed likely to affect how I would approach a healthcare or software project within the next year.
All four could be called relevant. That did little to settle the choice.
JSM was my first statistics-focused conference. At healthcare conferences, the clinical or policy problem usually provides an immediate filter: I can ask whether a session speaks to a population, payment rule, care process, or decision I work with. IT/CS conferences offer a different filter through systems, tools, and implementation. A meeting as broad as JSM put methods, applications, theory, computing, and professional communities into one program (note: therefore joint makes much sense!). Hundreds of sessions were plausible, many overlapped, and their titles were imperfect signals of what I would actually be ready to learn in the room.
Comprehensive coverage was impossible. Switching rooms after every individual talk would also impose a cost. Each move meant reconstructing vocabulary, assumptions, and the speaker’s position in a new technical conversation. The schedule therefore presented a resource-allocation problem before I reached any of the formal decision-theory talks.
The program could tell me who was speaking, how the session had been assembled, and what its abstracts promised. It could not tell me how much background a presenter would assume, whether four related talks would reinforce one another, or whether the most useful point would appear in a discussant’s synthesis. A title sometimes described an application I knew while hiding an unfamiliar estimand. Another advertised an advanced method whose motivating problem was immediately relevant. Selection had to be made from these partial signals.
There was also a difference between exposure and learning. I could enter a distant session to discover its vocabulary and leave with a reading list. That is a legitimate use of conference time, but it is different from following an argument closely enough to carry it into future work. I wanted the schedule to contain both kinds of encounters without confusing one for the other.
I needed a criterion stronger than “this sounds interesting.”
Relevance and learning value can separate
The safest conference schedule stays near current work. Familiar subjects reduce setup cost, make questions easier to formulate, and allow a listener to notice refinements that a newcomer would miss. That can be valuable. It can also produce a day of confirming a map I already have.
The clinical fairness session was a useful test. True-positive-rate gaps, calibration, and subgroup evaluation were already connected to my work. If the session had stopped at those familiar metrics, its marginal value would have been limited. The talks instead moved toward risk-standardized metrics, subgroup net benefit, resource constraints, and data-driven discovery of differential performance. Those additions changed the question from whether a metric differed to what produced the difference and what decision consequence followed.
This was a more demanding form of relevance than matching a keyword in the program. The subject had to contain an unresolved edge for me. In fairness, that edge was the separation among population composition, model behavior, clinical utility, and eventual health outcomes. In another familiar area it might come from a new data source, an unusual failure mode, or a speaker whose application exposes assumptions that routine examples leave hidden.
That is a good reason to attend a familiar topic: the application, unresolved question, or analytical distinction extends the structure one already knows. Topic labels alone cannot reveal that value. I began asking what a session might change in how I frame, design, diagnose, or implement a project.
Maximum unfamiliarity creates the opposite problem. A 15- or 25-minute conference talk compresses years of background into a few slides. The speaker cannot reconstruct every definition, notation choice, and foundational result. When most of my attention is spent identifying the objects in an equation, little remains for understanding why the contribution matters.
Recognizing a few terms is not the same as having the scaffold needed to learn the frontier. Some sessions would have been more valuable after a textbook chapter or foundational paper. The conference could show me that such a field existed and why I might return to it, but it could not substitute for the preparation required to follow its current argument.
That observation made it easier to decline certain sessions without treating the choice as a judgment of their quality. The issue was timing. A highly theoretical contribution could be the most important item in the program and still offer less learning to me that morning than a session whose main objects I could already recognize. I could record the distant topic for later reading and spend the live hour where questions, comparisons, and speaker emphasis would add something that a paper alone might not.
This left a middle region. I wanted enough familiarity to recognize the problem, enough novelty to revise my understanding, and enough likely use that the idea would be retrieved after I returned home.
The useful zone sits near the edge of competence
A light formalization helped me describe the selection rule. I would like to model that, for a session $s$, its learning value may be illustrated by:
$$ V(s)=C(s)\times N(s)\times T(s), $$where $C(s)$ is comprehensibility given my current knowledge, $N(s)$ is novelty, and $T(s)$ is transfer potential. Note that this is not a scoring model, but an expression to reflect that a session can lose value when any one element is close to zero.
Near-zero comprehensibility leaves me aware of a field without teaching me much of its present contribution. Little novelty makes a talk comfortable to follow while adding little. Weak transfer potential allows an interesting idea to disappear from working memory because no future problem calls it back.
Several sessions occupied the productive region for different reasons. Targeted validation connected directly to healthcare data quality, yet it treated the acquisition of accurate measurements as part of statistical design rather than as a cleanup step before modeling. High-performance statistical computing connected to software and numerical work, while extending the problem into architecture-aware inference, mixed precision, and scientific compression. Bayesian calibration of computer models began with uncertainty and simulation, then pushed toward modular information flow, expensive design, and posterior geometry.
I could recognize the applied stakes in each case. I could not predict the full answer before the talks began. That combination kept the material both legible and capable of moving my understanding.
Transfer potential also turned out to be broader than immediate implementation. Some methods were candidates to use in a project. Others gave me a distinction I expected to reuse: documented diagnosis versus disease, numerical precision versus scientific fidelity, a difficult sampler versus a genuinely ambiguous model. A durable distinction can be worth carrying home even when its method is never installed or coded.
A well-built session can teach more than isolated talks
My initial rules focused on individual topics: avoid material that required missing foundations, avoid spending most of the program on concepts I already knew, and prefer work likely to change future practice. During JSM, the construction of the session itself became a fourth criterion.
A coherent session lowers the cost of each successive talk. The first speaker establishes vocabulary and the shape of the problem. Later speakers can vary the estimand, assumptions, application, or computational strategy without forcing the audience to restart. The repetition is productive because it preserves the question while changing what counts as an adequate answer.
The fairness session built from crude performance toward risk adjustment, clinical benefit, policy constraints, and subgroup discovery. Targeted validation applied one limited budget for gold-standard information to underdiagnosed EHR outcomes, geographic access, cancer-screening tests, and several downstream models. Calibration connected simulator design and modular Bayes to sequential computation and the possibility that poor mixing reflected scientific nonidentifiability.
An isolated talk from any of these sessions could have supplied a method. Staying for the conversation supplied a boundary. I could see which premises the speakers shared, where their goals differed, and why no one metric or algorithm resolved the whole problem.
This changed how I interpreted the program. A session was more than a container for several titles. It could be a compact literature review designed by people who knew where the disagreements were. When that coherence was visible in the abstract, it deserved weight in the schedule.
What I would change at the next conference
The selection rule worked best when I already had a modest scaffold. Next time, I would build more of that scaffold deliberately. One foundational paper before a technically adjacent session could free attention that would otherwise be spent decoding notation. It would also make it easier to recognize which part of a talk was established background and which part was the contribution.
I would leave more open space in the schedule. An abstract cannot reveal how useful a session will be once its shared vocabulary and disagreements become visible. A packed itinerary makes it difficult to stay with a conversation that turns out to be better than expected or to change direction after discovering that a topic requires more preparation.
I would still select at least one coherent session per day, then record one transferable idea from each talk. That note should answer what the idea might change rather than merely preserve a method name. I also found it useful to distinguish material worth implementing from material worth remembering. The second category is larger and often more durable.
For example, I may never implement an exascale climate emulator. I expect to reuse the question it raised: which downstream scientific quantities must an approximation preserve? Likewise, I may not need a particular computer-model calibration algorithm, while the distinction between poor computation and genuine posterior ambiguity can change how I diagnose a difficult Bayesian model. The method name helps me retrieve the source; the transferable question is what gives the note a future use.
Finally, I would review the notes at the end of the day, while the relationship among talks remained available. Conference notes decay quickly into disconnected titles and formulas. A short synthesis can preserve why two talks disagreed, which assumption connected them, and where a current project might retrieve the idea.
These practices do not maximize the number of sessions attended. They make limited attention more likely to produce a coherent map of adjacent fields.
The map I brought home from JSM was shaped by this method. My itinerary was a deliberately selected path through the conference, informed by my own background and expected use. It was never a neutral sample of JSM.
That limitation also made the next question possible. Once I chose sessions near the edge of my competence, what technical problems kept appearing across decision theory, validation, fairness, Bayesian modeling, and scientific computing?