The Reliability of Replications: A Study in Computational Reproductions

Le résumé fourni par la source

This paper reports findings from a crowdsourced replication. Eighty-five independent teams attempted a computational replication of results reported in an original study of policy preferences and immigration by fitting the same statistical models to the same data. The replication involved an experimental condition. Random assignment put participating teams into either the transparent group that received the original study and code, or the opaque group receiving only a methods section, rough results description and no code. The transparent group mostly verified the numerical results of the original study with the same sign and p-value threshold (95.7%), while the opaque group had less success (89.3%). Exact numerical reproductions to the second decimal place were far less common (76.9% and 48.1%), and the number of teams who verified at least 95% of all effects in all models they ran was 79.5% and 65.2% respectively. Therefore, the reliability we quantify depends on how reliability is defined, but most definitions suggest it would take a minimum of three independent replications to achieve reliability. Qualitative investigation of the teams’ workflows reveals many causes of error including mistakes and procedural variations. Although minor error across researchers is not surprising, we show this occurs where it is least expected in the case of computational reproduction. Even when we curate the results to boost ecological validity, the error remains large enough to undermine reliability between researchers to some extent. The presence of inter-researcher variability may explain some of the current “reliability crisis” in the social sciences because it may be undetected in all forms of research involving data analysis. The obvious implication of our study is more transparency. Broader implications are that researcher variability adds an additional meta-source of error that may not derive from conscious measurement or modeling decisions, and that replications cannot alone resolve this type of uncertainty.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.