Frequentist Model-based Statistical Induction and the Replication Crisis

The prevailing view in the current replication crisis literature is that the non-replicability of published empirical studies (a) confirms their untrustworthiness, and (b) the primary source of that is the abuse of frequentist testing, in general, and the p-value in particular. The main objective of the paper is to challenge both of these claims and make a case that (a) non-replicability does not necessarily imply untrustworthiness and (b) the abuses of frequentist testing are only symptomatic of a much broader problem relating to the uninformed and recipe-like implementation of statistical modeling and inference that contributes significantly to untrustworthy evidence. It is argued that the crucial contributors to the untrustworthiness relate (directly or indirectly) to the inadequate understanding and implementation of the stipulations required for model-based statistical induction to give rise to trustworthy evidence. It is argued that these preconditions relate to securing reliable 'learning from data' about phenomena of interest and pertain to the nature, origin, and justification of genuine empirical knowledge, as opposed to beliefs, conjectures, and opinions.