Talks & courses · Talk

Beyond p < 0.05

João Muzzi, Ph.D · Graduate seminars in Parasitology, Pathology and Microbiology – Federal University of Paraná (UFPR), Brazil · October 2026 · 50 minutes, discussion format · Delivered in Portuguese

The p-value is the most widely used metric for deciding whether a result "works", and the 0.05 threshold is often treated as a natural boundary. This talk shows that the p-value reflects the distribution of the data and the choices of the analyst, and can therefore mislead when read on its own.

Open the interactive presentation (Portuguese)

The presentation runs R simulations on a Quantum server and may take a few seconds to load. Use the arrow keys to navigate and the + and − keys to change the font size.

Estimates from 29 teams that analysed the same dataset on red cards in soccer — 29 teams analysed the same data and reached 29 different estimates. Data from Silberzahn et al. (2018), *Adv Methods Pract Psychol Sci* 1(3):337–356; figure adapted from "Science Isn't Broken" (FiveThirtyEight, 2015).

What we discuss

What the p-value is, and the most common misinterpretations.
Where 0.05 comes from, the evidence scale, and thresholds suited to each context, from exploratory studies to genomics.
The p-value as a function of sample size: irrelevant differences become "significant" with a large n, and large effects go unnoticed with a small n.
Type I and II errors and statistical power: why p > 0.05 in an underpowered study is inconclusive, not negative.
Multiple comparisons, from the dead salmon in the MRI scanner to spurious correlations.
Meta-analysis: when weak studies add up and when the synthesis misleads.
Analytic flexibility: same data, different results.
The American Statistical Association statement (2016) and good practices for reporting results.

Reproducibility

Every simulation in the presentation can be reproduced: each slide includes the corresponding R code.