Beyond the Average

A doctoral student presents a finding in a research seminar. The coefficient on the focal variable is significant at the one percent level. The sign is in the predicted direction. The standard errors are clustered appropriately. The robustness checks are thorough. A senior faculty member in the back asks whether the finding holds for the firms that matter, and the student does not immediately understand the question. The faculty member clarifies. The coefficient describes a firm at the conditional mean of the outcome distribution. Most of the firms in the sample are near the conditional mean by definition. The question is whether the same coefficient describes the firms in the top decile of the outcome distribution, whose behavior is the reason the field cares about this variable in the first place. The student answers honestly that the analysis was not designed to address that question.

This exchange is common. It is also diagnostic. The question the student was asked and the question the student answered are not the same question, and the gap between them is not a gap in the student's competence or the quality of the data. The gap is architectural. The method the student used expresses a particular form of causation, and that form of causation is not the form the question requires.

The four architectures

This paper argues that four architectures of causation are relevant to business research, that the dominant methods express only one of them, and that the mismatch is invisible because the method of choice is the default of the field. The four architectures are single-metric causation, in which one measurement determines the outcome; averaging causation, in which many factors contribute additively and compensate for each other; weakest-link or necessary-alignment causation, in which every critical dimension must clear a floor and a deficit on any one constrains the outcome regardless of strength elsewhere; and top-versus-bottom contrast, in which the mechanisms that distinguish transformative performers from failing performers are not the same mechanisms that describe the central tendency of the sample.

The averaging architecture is built into ordinary least squares regression, into analysis of variance and its multivariate extension, and into the linear-index form of the Cox proportional hazards model. It is the default architecture of the field. It is also invisible, because the methods that express it are the methods we are trained to reach for first. A graduating doctoral candidate has typically seen several hundred regression tables and perhaps a dozen fsQCA tables. She designs around the tools she has practiced with. This is rational behavior. It is also the mechanism by which the dominant architecture reproduces itself.

The provocation

In a well-studied research area, where the averaging architecture has already been applied in dozens or hundreds of studies, the reanalysis of the same phenomenon under a non-averaging architecture is by construction a theoretical contribution. It is not a replication. It is the demonstration that the received findings rest on an architectural commitment the theory itself does not require and may not support. When the averaging architecture obscures a floor, a configuration, or a top-versus-bottom divergence, the reanalysis surfaces a finding that could not have been seen through the original method, and the new finding is in direct dialogue with the prior literature.

The paper develops this argument through four illustrations: the healthcare organization that performs in the top decile on clinical quality but fails financially because its revenue cycle is broken; the 737 MAX certification process in which excellent averages obscured a weakest-link dependency on a single sensor; the Equifax breach in which a mature security program averaged to strong but failed catastrophically on one unpatched server; and the architecture of extraordinary post-IPO value creation in technology companies, where Innovation Alignment Theory specifies a conjunctive rather than compensatory architecture for the conditions that produce transformative outcomes.

The methodological toolkit

The paper maps research questions to the methods that express each architectural form. Necessary Condition Analysis and geometric-mean composites express the weakest-link architecture. Fuzzy-set Qualitative Comparative Analysis expresses the configurational architecture. Quantile regression estimates how the effect of an explanatory variable varies across the conditional distribution. Stochastic frontier analysis and Data Envelopment Analysis locate the production frontier rather than the conditional mean. Survival analysis with time-varying covariates models the arrival of events rather than their average. Bayesian hierarchical models carry heterogeneity forward rather than averaging it away. The top-versus-bottom decile contrast, advanced here as a fourth architecture and developed across more than 200 hospital engagements by the present author, isolates the processes that distinguish transformative performance from failing performance.

The infrastructure

The paper closes with a claim about the infrastructure that makes methodological pluralism practical. Until recently, the time required to specify, estimate, and compare a research question under six or eight architectural forms on the same data was prohibitive for a single doctoral project. The Karpathy Loop and the auto-research patterns that have emerged around it change this calculus. A bounded specification, a single testable metric, a fixed time budget, and a complete audit trail make it practical to test a theory under the architecture it claims, against the architecture the dominant literature has assumed, on identical data, overnight. The discipline of pre-registered boundaries, single testable metrics, complete audit trails, and full reversibility is what separates legitimate architectural pluralism from the specification search that is currently understood as a methodological failure.

The invitation

The invitation of this paper is directed to doctoral peers identifying the novelty of their dissertation contribution and to the faculty who advise them. Take a research question the field has answered under averaging assumptions. State in plain terms the architectural form the underlying theory actually claims. Test the phenomenon under that architecture alongside the averaging architecture the field has assumed. In many cases, the architecture the theory claims is not the architecture the dominant methods express. When that mismatch is present, the path to theoretical novelty runs directly through it. The data have often been available for decades. The architecture through which the data are examined is what has lagged.

Companion to The High Stakes Decision

The High Stakes Decision is the executive-facing version of this argument. It speaks to CEOs and C-suite leaders about why the AI infrastructure most organizations buy cannot serve the decisions that most define them. Beyond the Average is the methodologist-facing version, written for doctoral peers and faculty rethinking the default analytical toolkit. The two papers share an architectural premise. The audiences and the registers differ.