Description
Making research Evidence-Based ruins error control from the sampling spaceIn 2009, a ‘radical’ idea was proposed to reduce research waste in biomedical sciences: informing new research by past results. This recommendation has been reiterated in every research waste paper since, and embraced by the movement to do Evidence-Based Research. Surprisingly, this proposal is actually radical, when we care about (frequentist) error control and coverage of intervals, that guarantee properties averaged over the sample space. The reason is that we cannot define the sample space of two studies – let’s say two randomized clinical trials – if the design of the second study is informed by the first. If depending on the findings in the first, the population definition (e.g. inclusion criteria) is decided, the outcome measure might be defined in a new way, or a second trial is unethical because the first showed treatment harm. This talk will discuss accumulation bias: if the existence of future studies depend on the results of earlier ones in the same meta-analysis, or its timing might be changed. It will also discuss the need for new foundations of statistics: if the sample space is so flexible that it cannot even be described by standard probability (measure) theory. And it will discuss so-called anytime-valid statistics based on e-values and foundations of probability and statistics based on betting, that can resolve both.
| Period | 1 Mar 2024 |
|---|---|
| Event title | Perspectives on Scientific Error |
| Event type | Workshop |
| Location | Eindhoven, NetherlandsShow on map |
| Degree of Recognition | International |