Last week I gave the Suzanne Scotchmer Memorial Lecture in Toulouse. I chose as my subject, my paper with Fiona Murray on “Markets for Scientific Attribution” because it was part of a research line that was one of Suzanne‘s last discussions before she passed away a few years ago. The topic of that research line is the division of credit in scientific collaborations and this paper was an attempt to endogenise how the “market” parsed individual contributions when there were collaborations. This was a topic I had been interested in for many years beginning with a multi-authored Journal of Political Economy paper on alphabetical ordering norms in economics; “First Author Conditions.”
One the problems with our current paper was that, while we motivated it using a setting familiar in natural scientists (that of a scientist-owned lab employing postdocs or graduate students), apart from lose discussions of the mistreatment of underlings in these settings, it was admittedly difficult to motivate why we would care about the processes of scientific credit in applications. Specifically, were they operating in ways that really led to major distortions? This is hard to disentangle in a theoretical paper that showed only qualitative effects.
But as Diane Coyle noted, the day after my lecture, the issue of whether scientific credit in collaborations came roaring onto the New York Times with this article by Justin Wolfers. He reported on research conducted by Harvard PhD student Heather Sarsons that presented “new evidence suggests that the underrepresentation of women reflects a systemic bias in that marketplace: a failure to give women full credit for collaborative work done with men.” As reported by Wolfers:
While women in the field publish as much as men, they are twice as likely to perish. And this higher rate for women being denied tenure persists even after accounting for differences in tenure rates across universities, the different subfields of economics that women work in, the quality of their publications and other influences that may have changed over time.
But Ms. Sarsons discovered one group of female economists who enjoyed the same career success as men: those who work alone. Specifically, she says that “women who solo author everything have roughly the same chance of receiving tenure as a man.” So any gender differences must be because of the differential treatment of men and women who work collaboratively.
The result was summarised by this figure from the paper.
Suffice it to say, Wolfers was right that this result would generate discussion. It has been the talk of economics for the past week although it hasn’t received much attention in the blogs. In particular, the paper and certainly the NYT article suggested that in tenure review meetings, women were receiving no credit for collaborations with men which was not true in reverse. This was a strong result that suggested at least an implicit bias and also a strong one that I believe it is fair to say that many who had participated in those meetings were hitherto unaware of. Moreover, this was somewhat unique to economics and did not arise in sociology. Suffice it to say, if it was going on in economics, then it was important that awareness was raised.
Given my long-standing interest in precisely the issue underlying these findings, I thought I would dive deeply into Sarsons’ paper. What I found, however, was not quite the same set of collusions a reading of the New York Times piece would give. That is not to say there aren’t issues in economics and elsewhere on gender and scientific credit, it is just that the mechanism is likely different from what was implied by the paper.
The Sarsons paper has two parts: theoretical and empirical and, interestingly, the theoretical predictions do not end up quite matching the empirical results. The theoretical model involves individuals receiving signals about the ‘quality’ of potential collaborators, choosing whether to collaborate or go it alone and then being promoted on the basis of signals from their research output. [Interestingly, the language of the paper is gendered with the employer who is responsible for promoting being a ‘he.’ My personal preference in presenting theory is to keep things gender neutral because of research elsewhere about the biases such non-neutrality can generate.] All of the action occurs in the thresholds that the employer infers from decisions of whether you collaborate and which type of person you collaborate with. It is assumed that type matters because everyone assesses that, on average, a woman is less likely to be of high quality than a man. These assumptions are tricky to make but are often used in these sorts of models because it is hard to generate different outcomes with at least someone believing there are differences. In this paper, everyone is similarly biased which is a modelling strength. But similarly everyone wants the individual treatment to rise above the statistical assumptions and so intentions are not biased against one gender.
The statistical bias harms women. Unless they can overcome it, even for sole authored papers, they are less likely to be treated as well as men. Time is a friend here as the more papers you have the less likely the statistical bias will translate to the individual. With regard to collaborations, the statistical bias is potentially an opportunity for women. After all, if people assume women of high ability in short supply, then if women collaborate with others that is because those others have likely seen something the “market” has not and these are the, supposedly rare, high ability women. By contrast, no similar signal can be sent by the men. So if you work with a women, you are more likely to be promoted than if you work with a man. Nonetheless, working alone is still best as it sends a cleaner signal to the employer.
There is a bias, however, in that high ability men and women, will prefer to work with men rather than women. That is, each collaboration involves two sides and this is the side that drives a problem. Because of this, women will be involved in fewer collaborations but when they do, they will be of higher ability than man. Employers realise all of this, of course, but if workers do, then high ability women are going to end up choosing to work alone at a higher rate than anyone else.
The Empirical Results
The theoretical model is of some interest but is really not the main take-away in the paper. The punchline comes in the empirics, so let me turn to that.
To understand any empirical result, it is important to have a solid grasp of the ‘data generating process.’ That is, any data that an economist has used has come from somewhere. You need to think about that carefully as you draw inferences from it and other variables it correlates with.
The paper is written, as is Wolfers’ article, that the independent variable of interest is whether an academic economist received tenure or not; that is, the core promotion decision in an academic’s life and our key indicator of whether you have made it. Tenure is something that can be observed but careers are messier than that. People may move, get extended clocks, take parental leave, have spousal location issues and idiosyncratic preferences. All of these impact on whether you actually take part in a tenure decision.
Interestingly, Sarsons sidesteps all of that by not actually coding tenure decisions themselves. Instead she collects the CVs of economists at the top-30 PhD granting universities in the US between 1975 and 2014. From that it is easy to see who was granted tenure but it is harder to see whether someone went up for tenure and was denied it. Thus, near as I can tell, what the paper does is take the length of time for each academic since they received their PhD and then sort them into three bins. The first bin is whether they received tenure at the institution they received their first job. The second bin is whether they moved to a similarly ranked university after 6-8 years. The final bin is the rest. They are the people Sarsons concludes were denied tenure. Comparing the first two bins to the last reveals a difference between men and women with 77% of men and only 52% of women being in this first two bins. Sarsons aims at trying to explain this gender gap.
This is a very different variable of interest than the outcomes of tenure decision meetings. It is likely correlated with that but all of the things that impact on career mobility also will impact on this variable. For instance, if it was the case that women were more likely to move and sacrifice institution-status than men, that could generate the observed gender gap. I am not saying that would excuse the gap — just the opposite — but it does alter where we think the problem is arising. For instance, when people move, it could be because of differences in negotaiting power with Deans (something for which there is existing work on gender differences) or it could be because of spousal issues and over that time period it is uncontroversial to suggest that women were more likely to have a spousal co-location issue than men. That means women were statistically more likely to have moved for reasons unrelated to their ability.
Which makes it all the more intruiging that the gender gap as it pertained to this variable appeared to be accounted for by differences between men and women in the impact of papers co-authored with women. That result suggests that comparing people who stayed at equivalent institutions versus those who did not, the probability of maintaining rank was not correlated with whether you were a women who produced an additional co-authored paper but was correlated with all other additional papers produced. It is because we can presume that people have limited capacity to produce papers that it can then be argued that, for women, reallocating a paper from a collaboration to a sole authored paper is correlated with their likelihood of maintaining rank after 6-8 years (as per the above graph).
One mechanism that could explain this is the mechanism described by the paper and Wolfers that in tenure meetings, the median voter is completely discounting women’s collaborations. Why they should do this is hard to work out. After all, they are being presented with a portfolio that includes usually a variety of collaborations and sole authored work. It could be that absence of a sole authored signal is very costly to women compared with men but that is not what the result is. It is that collaborations, independent of other measures of quality, do not help women. It is as if they are struck from their CV.
Another possibility is related to the seminal work of Paul Milgrom and Sharon Oster on the “Invisibility Hypothesis” (a work that is surprisingly not cited by Sarsons especially since it is the classic on the issue of teams and gender discrimination). That paper shows how employers may hide talented employees who are in a discriminated against class because they can retain them at lower wages despite their high productivity. The idea is that you want to discourage those individuals from sending external signals. How best to do that than encourage them to collaborate more! In this situation, those individuals who go against the trend will actually have an advantage — specifically, there will be a relative bump to clear signals for them. I have to admit, this would have been my starting place for thinking through the mechanisms here.
Sarsons dismisses some aspects of strategy but she does so on the basis of a survey of academics conducted today. Suffice it to say, I am not sure whether that is enough to give me confidence regarding behaviours over the last four decades. Moreover, elsewhere, academics are very strategic with regard to credit. For instance, this paper by Daniel Garcia and Joshua Sherman does an impressive job of parsing the data on authorship choices in relation to alphabetic norms and demonstrates strategic choices being made by those blessed with high alphabetic names and those cursed with low ones.
Finally, if we lump in all of the other gender biases to do with spousal situations, I suspect we can also account for the gender gap in the maintenance of equivalent rank that Sarsons identified. In particular, the very fact that women are underrepresented may play a statistical role here. The paper does not consider any of these but it would surprise me if they did not play a role. This is especially the case considering the 40 years under study and the changes in society over that time. Sarsons does include tenure year fixed effects but does not tell us whether there is anything interesting there.
This research has brought an interesting set of potential facts to light on the relationship between the gender composition of collaborations in economics and career trajectories of men and women. While it may be that there is a credit bias going on, it could also reflect the myriad of other biases that have impacted on women’s academic careers. Teasing out the mechanism in a convincing manner is important because if it is about the allocation of credit, we need to know it and understand it so that it can be corrected. Otherwise, we are in danger of second guessing those decisions along variables that may not be relevant.