You are here

The Subjectivity of Scientists and the Bayesian Approach

S. James Press and Judith M. Tanur
Publisher: 
Dover Publications
Publication Date: 
2016
Number of Pages: 
274
Format: 
Paperback
Price: 
19.95
ISBN: 
9780486802848
Category: 
Monograph
[Reviewed by
Tom Schulte
, on
09/14/2016
]

In his book Clinical vs. Statistical Prediction: A Theoretical Analysis and a Review of the Evidence (University of Minnesota, 1954), psychoanalyst Paul Meehl gave evidence that statistical models almost always yield better predictions and diagnoses than trained professionals. The authors here examine the cases of selected pioneers in science and how bias-driven subjectivity played a significant role in directing their advances. That role of subjectivity is the main thrust of this book. The formal application of Bayesian inference in deriving the posterior probability based on prior probability and a likelihood function derived from a statistical model for the observed data is not examined as deeply as the title may suggest. This is really a history of science via chapter-length biographies for the layman with a Bayesian introduction for the nonmathematician as almost an appendix.

Scientists, especially the visionary ones profiled here, often combine something between informed judgment and zealous belief with experimental data to validate theories. At times, they stray into such questionable areas as the outright fabrication of data or merely data manipulation to better fits the model. The authors class these cases as “extreme subjectivity” (Kepler, Mendel, Millikan, etc.). Kepler probably fudged his data on the motion of Mars to bolster support for his model. Mendel, although he rose above the expectations of a B-team substitute teacher to become the father of genetic theory, reported only data that conformed to his theory. Millikan while measuring the charge on the electron omitted publishing a significant number of measurements that disagreed “with the results of other observations.”

The subject of why scientists should publish even failed experiments has circulated from inside scientific circles out to popular discourse being published in Time (“Why Scientists Should Celebrate Failed Experiments”, Aug. 28, 2014). Extreme subjectivity crops up outside even the dedicated chapter, as in Newtown “doctoring his figures”. While this can strike the reader as something between quaint and inappropriate, the case of Louis Pasteur is more of an ethical dilemma. Pasteur had not even completed the animal trials before using the “outcomes” as a basis for what was, basically, hopeful experimentation on humans.

From the thought experiments of Galileo to Einstein, Einstein himself observed

There is no empirical method without speculative concepts … no speculative thinking whose concepts do not reveal…

By and large, the profiled scientists wedded intelligence and experience to vision and conviction, often in the face of entrenched beliefs and inadequate measurement technology, to transcend limitations insurmountable to others. This history is fascinating and enlightening, despite making no strong case for how the questions of personality, religious belief (e.g., Faraday and Sandemanianism, Chassidism for Freud), and backlash-inducing paradigm shattering could have gone differently in a scenario involving Bayesian techniques.

Bayesian inference provides a formal way to combine prior information, e.g. personal belief, with observed data. The authors make the point generally and specifically that this could have played a role in the cases cited. When William Harvey sought to prove the systemic circulation of blood in the human body, he could have “formed the posterior odds ratio for the probability of the pulsation model for blood compared with the probability of the circulation model.” As when Lavoisier dispelled phlogiston, the dispute “might have been resolved more readily had the posterior probability calculations been presented…” And, the significant achievements of Marie Curie could have been even greater had “A Bayesian model comparison of alternative theories … directed her to such insights.” However, the authors do not provide any example of such an application to data from the given exemplars where it was most needed. This disassociates the nearly two hundred pages of science history with the less than ten pages touching on Bayesian inference and strikes me as a missed opportunity.


Tom Schulte is a software architect at Plex Systems in Troy, Michigan.

The table of contents is not available.