Frequentist vs Bayesian statistics - a non-statisticians view

People who by training end up dealing with probabilities (“statisticians”) roughly fall into one of two camps. One is either a frequentist or a Bayesian. To a scientist, who needs to use probabilities to make sense of the real world, this division seems sometimes baffling. I will argue that science mostly deals with Bayesian questions. Yet the dominance of frequentist ideas in statistics points many scientists in the wrong statistical direction.

💡 Research Summary

The paper presents a non‑statistician’s perspective on the two dominant schools of statistical thought—frequentist and Bayesian—and argues that scientific inquiry is fundamentally Bayesian in nature. It begins by outlining the philosophical underpinnings of each approach. Frequentist statistics define probability as the long‑run relative frequency of an event in repeated trials and treat model parameters as fixed but unknown quantities. In contrast, Bayesian statistics interpret probability as a degree of belief, allowing parameters themselves to be random variables with prior distributions that encode existing knowledge.

The author contends that scientists routinely ask questions that align with the Bayesian framework: “Given what we already know, how plausible is a new hypothesis after observing the data?” This process is naturally expressed by Bayes’ theorem, which updates prior beliefs with the likelihood of the observed data to produce a posterior probability. The paper illustrates this with examples from medical research (evaluating a new treatment’s efficacy) and astronomy (assessing the existence of a newly detected object), showing how prior information—previous studies, theoretical constraints, expert opinion—plays a crucial role in shaping inference.

In contrast, the frequentist paradigm relies on p‑values and confidence intervals to test a null hypothesis. The author points out that a p‑value measures only the probability of obtaining data as extreme as observed under the null, not the probability that the hypothesis itself is true. This leads to common misinterpretations, such as equating a small p‑value with strong evidence for the alternative hypothesis. The paper highlights how such misunderstandings contribute to the reproducibility crisis, especially when researchers engage in “p‑hacking” or selective reporting to achieve statistical significance.

A structural analysis follows, showing that frequentist methods dominate curricula, textbooks, and journal guidelines because they are mathematically simpler and historically entrenched. Consequently, many scientists are trained to apply null‑hypothesis significance testing even in situations with limited data, high measurement noise, or costly experiments—contexts where Bayesian methods can be more informative. The author cites the rise of computational tools (Markov chain Monte Carlo, variational inference, probabilistic programming) that have dramatically lowered the practical barriers to Bayesian analysis, making it feasible for routine use.

The paper enumerates four key advantages of the Bayesian approach: (1) explicit incorporation of prior knowledge, enabling cumulative science; (2) meaningful inference from small samples, which is vital in rare‑event or high‑cost studies; (3) results expressed as intuitive probabilities (e.g., “there is an 85 % chance the treatment is effective”), facilitating decision‑making for policymakers and clinicians; and (4) flexibility to model complex hierarchical structures and uncertainty at multiple levels.

To shift the scientific culture, the author proposes concrete actions: integrate Bayesian fundamentals and hands‑on computing labs into undergraduate and graduate programs; encourage journals to accept and highlight Bayesian analyses alongside frequentist ones; and require transparent reporting of prior specifications and sensitivity analyses. By doing so, the community can reduce the over‑reliance on p‑values, improve the interpretability of findings, and enhance the reproducibility of research.

In conclusion, the paper asserts that while frequentist tools remain useful for certain exploratory tasks, the core scientific question—how plausible is a hypothesis given all available evidence—is best addressed by Bayesian statistics. Scientists should therefore develop fluency in both paradigms, select the method that aligns with their research goals, and recognize that the Bayesian framework offers a more direct, coherent, and decision‑relevant pathway for turning data into knowledge.