To outsiders, statisticians inhabit a world of Spock-like logic governed by immutable laws that lead to unassailable conclusions. Statisticians, however, know this to be folly, and the folly can be traced back to an 18th century British clergyman and philosopher named Thomas Bayes (1701-1761). Bayes, whose eponymous theorem has simultaneously confused, enlightened, and divided scientists for centuries, was neither a trained statistician nor mathematician. He was a Presbyterian minister who was a Nonconformist-he did not use the Anglican Church's Book of Common Prayer as mandated by the 1660 Act of Uniformity. Bayes' place in statistics stems from his anonymously published defense of Newton's calculus and an unpublished essay presented to the Royal Society of London 2 years after his death (Bellhouse, 2004). That unpublished essay presented Bayes' theorem, a formula that calculates probability when relevant conditional probabilities are known. Today, many Bayesians advocate incorporating these relevant probabilities into statistical analyses as an alternative to traditional p value-based analyses.

Bayesians and non-Bayesians alike agree that the time-honored significance value of .05 is an arbitrary threshold, best discredited by Rosnow and Rosenthal's assertion that God loves a probability of .06 as much as .05 (Rosnow & Rosenthal, 1989); however, Bayesians take the argument further and assert that p values are inadequate for hypothesis testing because they do not consider effect sizes, clinical significance, predictive value, or the reproducibility of the findings. Instead, p values only express the probability of the obtained data if the null hypothesis were true. A Bayesian approach flips this notion on its head and expresses the probability of the hypothesis given the obtained data (Maxwell & Satake, 2017). Essentially, it tells us the probability that a real effect exists.

Ever since Sir Ronald Fisher proposed the p value threshold of .05 in 1925, mathematicians, scientists, and armchair statisticians have called for its abolition. And, these calls have become increasingly louder with the growth of technology-driven data science and its reliance on Bayesian inference to compute probabilities. So, why have we not relegated p values to the dustbin along with cold fusion, phrenology, and Martian canals? One reason is that Bayesian estimates do not provide conclusive answers to binary questions. One advantage of the frequentist (non-Bayesian) approach is that it tests a null hypothesis that can be rejected if the p value is below .05. Moreover, because frequentists divide the world into two possible states (the null and alternative hypotheses), they can further state that the alternative hypothesis represents the true state of the world. This simplifies decisions about whether a treatment is effective or not (Neyman & Pearson, 1933).

Bayesians, on the other hand, offer more opaque answers to questions that vex those charged with making high stakes data-driven decisions. This is for two reasons. The first is that they posit multiple states of the world and assign a probability or confidence statement to each. The second is that they continually revise their models for accuracy as new information becomes available. Although this makes Bayesian inference inherently more dynamic, it can make arriving at a single definitive conclusion more difficult.

Bayesian models abound in today's world. They underlie artificial intelligence and machine learning. They are also relevant to health care. The emergence of antibiotic-resistant organisms offers a lesson in Bayesian models at work. In response to the increase in antibiotic-resistant infections, the Centers for Disease Control and Prevention has accelerated initiatives to prevent food-borne and hospital-acquired infections, to promote responsible antibiotic stewardship, and to investigate nonantibiotic approaches that improve immune function, such as the use of probiotics.

So, is there a place for Bayesian estimates in nursing research? It depends on the research question. If the question is whether a new treatment is more effective than the usual treatment, then a frequentist approach is useful. If the question is whether a new treatment is effective over time, then a Bayesian approach is preferred. This is our current conclusion, but it can change with the addition of new information (or can it?).

Disclosures

The authors report no real or perceived vested interests related to this article that could be construed as a conflict of interest.

References