Skip to main content

Making Scientific Inferences More Objective

Periodic Reporting for period 3 - Objectivity (Making Scientific Inferences More Objective)

Reporting period: 2017-11-01 to 2019-04-30

The project engages in improving the objectivity of scientific inferences. This is important because science undergoes a trust crisis at the moment. Replications of famous experiments yield much weaker effects than the original experiments. Questionable research practices in statistical inference are widespread in the scientific community, although they bias the results of research papers. Pressure to publish "significant" findings and to suppress results that are hard to interpret leads to biased pictures of reality. In a time where powerful politicians publicly express their dismissal of science, it is more important than ever to strengthen the reliability of scientific research.

The project rethinks the notion of objectivity in science and uses its insights in order to develop more objective inference tools in three important domains of scientific reasoning: statistical inference, causal inference (i.e. cause-effect relations), and explanatory inference (i.e. inferring the best explanation of an observed phenomenon). It contributes to a more reliable, robust and objective science that can be trusted by policy-makers and the general public alike.
We argue for a new understanding of scientific objectivity, especially in the context of statistical inference. While previous research focused on objectivity as value freedom and correspondence to facts, we point out that important aspects of being objective involve questions such as transparency, balancing different perspectives, and robustness to model violations. In general, we point out that the social dimension of scientific research has been underestimated in discussing the question of objectivity. This perspective leads to a re-assessment of the objectivity of frequently used statistical procedures. For example, we conclude, seemingly paradoxically, that statistical inference based on subjective probabilities (degrees of belief) can be objective, too.

We also axiomatize different probabilistic measures of causal strength and give recommendations for their use in scientific practice. Moreover, we examine different determinants of judgments of explanatory power, such as probability, causality and generalizability. Currently, we are conducting simulations of meta-analysis where different statistical paradigms are pitched against each other. We compare their performances under ideal circumstances, and in a more realistic setting where data and interpretation are biased in various ways.
The contributions of the project tackle novel research questions in an unprecedented way, combining mathematical modeling, conceptual analysis, empirical data and computational techniques. It is expected that our research leads to a more nuanced assessment of scientific inference techniques, particularly in statistics, and ultimately to more reliable scientific inferences and conclusions. This has the potential to address the reproducibility crisis in science, to increase public trust in scientific inferences, and to integrate philosophical and scientific perspectives on scientific reasoning.