Feller-Umezaki Correspondence Index?

Context and Origin

The correspondence arose from a specific medical crisis. By 1972, serum therapy had proven clinically effective for certain bacterial infections, particularly endocarditis and meningitis, but no standardized statistical framework existed for measuring its efficacy. Traditional statistical methods assumed rapid cure or rapid failure—the condition of most nineteenth-century treatments. Serum therapy worked differently: patients improved slowly and incompletely, with survival extending over months or years, sometimes without apparent cure. Paul Kaplan had developed the Kaplan-Meier estimator in 1958 for industrial reliability testing, a method that could measure outcomes when the event of interest was rare or took years to occur. Umezaki recognized that this approach could reshape how serum therapy trials reported results, but the mathematical translation from industrial reliability to medical treatment required both microbiological intuition and mathematical precision.

Umezaki and Feller met at a 1973 conference on statistical methods in clinical medicine held in Princeton. Feller, approaching the end of his career at Princeton University, had been thinking about how classical survival analysis could be extended to account for the particular characteristics of passive immunotherapy: varying serum quality, patient genetic variation, competing causes of death. Umezaki brought detailed trial data from the Infectious Disease Research Centre in Cambridge, where she had been running comparative trials of serum therapy against azo drugs in cases of bacteraemia.

Contents

The earliest letters, from autumn 1973, are largely exploratory. Feller writes about the problem of censoring—the situation where a patient leaves a trial, dies of an unrelated cause, or simply disappears from medical records—a difficulty magnified by serum therapy's long time course. Umezaki responds with descriptions of actual trial dropout patterns and infection mortality rates from her Cambridge cohort. By early 1974, the correspondence becomes technical, with Feller sending pages of mathematical proofs and Umezaki responding with critiques grounded in laboratory observation.

It is at the level of the animal serum itself where your mathematics fails. The quality of the serum varies between animals, between seasons, between production runs. You cannot model this as random variation in a population parameter. The variation is structural.

-- Letter from Dorothy Umezaki to William Feller, March 1974

The index notes that this critique produced a decisive shift in Feller's approach. The subsequent letters, through 1975-1976, develop what became the foundational method: a Kaplan-Meier framework explicitly stratified for serum source and batch, allowing the estimator to account for biological variation as a measurable covariate rather than as uncontrolled noise.

By 1978, the correspondence had moved toward publication strategy. The two exchanged drafts of a paper that became "Survival Analysis in Serum Therapy: Statistical Methods for Passive Immunotherapy Trials," published in 1980 in the Journal of the American Statistical Association. The later letters, through Feller's death in 1986, focus on applications: other researchers writing to ask how the method should be adapted for tuberculosis, for diphtheria antitoxin, for the emerging problem of resistant bacterial strains.

Archival Organization

Access and Use