Big science doesn’t seem to carry its marginal weight

A new interesting paper in PLOS One. Here is the abstract:

Agencies that fund scientific research must choose: is it more effective to give large grants to a few elite researchers, or small grants to many researchers? Large grants would be more effective only if scientific impact increases as an accelerating function of grant size. Here, we examine the scientific impact of individual university-based researchers in three disciplines funded by the Natural Sciences and Engineering Research Council of Canada (NSERC). We considered four indices of scientific impact: numbers of articles published, numbers of citations to those articles, the most cited article, and the number of highly cited articles, each measured over a four-year period. We related these to the amount of NSERC funding received. Impact is positively, but only weakly, related to funding. Researchers who received additional funds from a second federal granting council, the Canadian Institutes for Health Research, were not more productive than those who received only NSERC funding. Impact was generally a decelerating function of funding. Impact per dollar was therefore lower for large grant-holders. This is inconsistent with the hypothesis that larger grants lead to larger discoveries. Further, the impact of researchers who received increases in funding did not predictably increase. We conclude that scientific impact (as reflected by publications) is only weakly limited by funding. We suggest that funding strategies that target diversity, rather than “excellence”, are likely to prove to be more productive.

From: Fortin J-M, Currie DJ (2013) Big Science vs. Little Science: How Scientific Impact Scales with Funding. PLoS ONE 8(6): e65263. doi:10.1371/journal.pone.0065263

In other words, the marginal dollar is better spent on additional smaller projects rather than on larger projects. Of course, this is a correlation only as this paper doesn’t identify marginal dollars but it is still different. The counter-bias goes the other way so the correlation may be under-stated rather than over-stated.