Non-dichotomous inference using bootstrapped evidence

The problems with classical frequentist statistics have recently received much attention, yet the enthusiasm of researchers to adopt alternatives like Bayesian inference remains modest. Here I present the bootstrapped evidence test, an objective resampling procedure that takes the precision with which both the experimental and null hypothesis can be estimated into account. Simulations and reanalysis of actual experimental data demonstrate that this test minimizes false positives while maintaining sensitivity. It is equally applicable to a wide range of situations and thus minimizes problems arising from analytical flexibility. Critically, it does not dichotomize the results based on an arbitrary significance level but instead quantifies how well the data support either the alternative or the null hypothesis. It is thus particularly useful in situations with considerable uncertainty about the expected effect size. Because it is non-parametric, it is also robust to severe violations of assumptions made by classical statistics.