Statistical Errors in Co-authorship Networks

The study investigates the quality of psychology research in collaborative context. We use co-authorship as a proxy for collaboration and error rate in reporting the results of null hypothesis statistical testing as a proxy for research quality. We downloaded 24 547 full text psychology articles published in PLoS journals in 2003-mid.2019 and successfully checked 6 842 of them for statistical test reporting consistency using R package statcheck. 36.3% of the checked papers contained at least one inconsistency, and 10.2%, at least one gross inconsistency. The results were sligtly lower than those of the original study involving statcheck, in which the respective rates were 49.6% and 12.9%. We used these statistical test reporting consistency data from to calculate author error rate - the number of inconsistent p-values to total number of p-values reported - for 12 106 authors. We mapped the error rate as node attribute with co-authorship data obtained from Web of Science for PLoS and other seven psychology journals featured in Nuijten's study and transformed it to network format with WoS2Pajek software, and then extracted locally important collaborations using the node island procedure, which were then closely analysed. We performed in-depth analysis on a set of 18 locally important co-authorship networks of size 14 to 31 nodes. We described the obtained data at a level of single nodes, four-node regular graphs, and networks. We described the associations between node attribute - each author's error rate - with node centrality measures and structural positions of nodes, which we summarise using graphlet analysis. We grouped similar networks together using visual inspection, network and non-network measures and explored the associations between network characteristics - their structure, transitivity, error rate assorativity, main mode characteristics, and error rate distribution patterns. The results suggest presence of assortative mixing of author error rate as node attribute in locally important collaborations and certain associations between the structural positions of authors and their error rates.