Figure 6

Evidence consistent with confirmation bias and a counter-effective role of cumulative advantage. We test whether the relative citation impact ${\tilde{z}}_i(n)$ decreases, increases, or is independent of n. While repeated publication in a highly competitive journals reflects the underlying quality of the researcher, it also indicates a strong role played by other factors such as author/institutional reputation and social ties with the journal editors and the referee base, and in the case of PNAS, membership in the US National Academy of Sciences. (A) Scientists with between 5 and 10 publications in the Nat./PNAS/Sci. arena. (B) Scientists with between 11 and 20 publications in the Nat./PNAS/Sci. arena. (C) Economists with between 4 and 9 publications in the top economics journal set arena. (D) Economists with between 10 and 20 publications in the top economics journal set arena. (A-D) For each cohort analyzed, the top panel shows a significant negative trend in $〈\tilde{z}(n)〉$ (black curve) with each successive publication. Linear regression of each $〈\tilde{z}(n)〉$ is shown by the dashed green line, with the best-fit slope and regression F-test p-value listed in each panel. In the lower half of each panel we show the empirical cumulative distribution $P(\le s_i)$, and list the number of trajectories analyzed and the mean value $〈s_i〉$ (indicated by the vertical solid blue line). For comparison, we also plot the $P(\le s_i)$ for the shuffled data (dashed black curve), with the mean shuffled value (vertical dashed gray line). We apply the Kolmogorov-Smirnov test between the empirical and shuffled distributions, and for each panel we list the p-values that confirm that the underlying $s_i$ values belong to different distributions. Only research profiles with $L\ge 5$ years were analyzed. In order to ensure that the relative citation impact $z_p$ of a given publication had sufficient time to stabilize within the journal set dataset, only publications published prior to 2002 for Nat./PNAS/Sci. (since the publication citation counts used were current as of census year 2009) and 2005 for Economics (since citation counts used were current as of census year 2012) were analyzed. In order to reduce censoring bias arising form careers that started before the beginning of each data sample, we only included trajectories with the first publication year $y_{i,0}^j\ge 1970$ for the Nat./PNAS/Sci. and $y_{i,0}^j\ge 1960$ for the economic sciences.