Guest Post by Pablo Achard

This post is by Pablo Achard of the University of Geneva. It refers to  the Shanghai subject rankings. However, the problem of outliers in subject and regional rankings is one that affects all the well known rankings and will probably become more important over the next few years

How a single article is worth 60 places

We can’t repeat it enough: an indicator is bad when a small variation in the input is overly amplified in the output. This is the case when indicators are based on very few events.

I recently came through this issue (again) with Shanghai’s subject ranking of universities. The universities of Geneva and Lausanne (Switzerland) share the same School of Pharmacy and a huge share of published articles in this discipline are signed under the name of both institutions. But in the “Pharmacy and pharmaceutical sciences” ranking, one is ranked between the 101^st and 150^th position while the other is 40^th. Where does this difference come from?

Comparing the scores obtained under each category gives a clue

	Geneva	Lausanne	Weight in the final score
PUB	46	44.3	1
CNCI	63.2	65.6	1
IC	83.6	79.5	0.2
TOP	0	40.8	1
AWARD	0	0	1
Weighted sum	125.9	166.6

So the main difference between the two institutions is the score in “TOP”. Actually, the difference in the weighted sum (40.7) is almost equal to the value of this score (40.8). If Geneva and Lausanne had the same TOP score, they would be 40^th and 41^st. 

Surprisingly, a look at other institutions for that TOP indicator show only 5 different values : 0, 40.8, 57.7, 70.7 and 100. According to the methodology page of the ranking, “TOP is the number of papers published in Top Journals in an Academic Subject for an institution during the period of 2011-2015. Top Journals are identified through ShanghaiRanking’s Academic Excellence Survey […] The list of the top journals can be found here  […] Only papers of ‘Article’ type are considered.”

Looking deeper, there is just one journal in this list for Pharmacy: NATURE REVIEWS DRUG DISCOVERY. As its name indicates, this recognized journal mainly publishes ‘reviews’. A search on Web of Knowledge shows that in the period 2011-2015, only 63 ‘articles’ were published in this journal. That means a small variation in the input is overly amplified.

I searched for several institutions and rapidly found this rule: Harvard published 4 articles during these five years and got a score of 100 ; MIT published 3 articles and got a score of 70.7 ; 10 institutions published 2 articles and got a 57.7 and finally about 50 institutions published 1 article and got a 40.8.

I still don’t get why this score is so unlinear. But Lausanne published one single article in NATURE REVIEWS DRUG DISCOVERY and Geneva none (they published ‘reviews’ and ‘letters’ but no ‘articles’) and that small difference led to at least a 60 places gap between the two institutions.

This is of course just one example of what happens too often: rankers want to publish sub-rankings and end up with indicators where outliers can’t be absorbed into large distributions. One article, one prize or one  co-author in a large and productive collaboration all of the sudden makes very large differences in final scores and ranks.