Lost in the Storm: The Academic Collaborations that Went Missing in Hurricane Isaac

By exploiting the cancellation of the 2012 American Political Science Association Annual Meeting, we investigate the role of conferences in facilitating academic collaboration. We assembled datasets comprising 17,468 academics and 86 million pairs of conference participants. In difference-in-differences analysis, we ﬁnd the conference cancellation led to a decrease in individuals’ likelihood of co-authoring an article with another attendant by 18 percent. Moreover, collaborations formed among attendants of (occurring) conferences are associated with more successful co-publications: an effect which is sharpest for teams that are new or non-collocated. These ﬁndings are novel and demonstrate the importance of conferences in scientiﬁc production.


Introduction
A phenomenon observed across all scientific disciplines -as noted for example by Wuchty   et al. (2007) -is the increasing prevalence of collaborative endeavour.An existing literature (Jones, 2009; Gans and Murray, 2014; Agrawal et al., 2016) has attributed this trend to the increasing challenges associated with pushing further outwards at the existing frontiers of knowledge, and of producing work generally at the standards required for success in an increasingly competitive academic environment.Co-authors bring, to a project, a wider pool of ideas and of specialist expertise, and scientific productivity therefore depends on co-authors becoming efficiently matched.And yet, there is also strong evidence, provided in Freeman and   Huang (2015), that some of the most productive scientific collaborations arise the least readily.
In this paper, we measure the extent to which academic conferences facilitate collaborations generally, and productive collaborations particularly.By exploiting a "natural experiment"the last-minute cancellation, due to "Hurricane Isaac", of the 2012 American Political Science Association (APSA) Annual Meeting -we are able to estimate the number and character of the collaborations that "went missing".From these estimates we draw inferences about the specific role of conferences in the formation of new scientific work, and also, more generally, about the role of network constraints in causing inefficient biases in co-author matchings.
The APSA meeting gathers around 3,000 presenters every year, and by the time of its cancellation in 2012 the conference program had been arranged and published.Our main hypothesis is that the cancellation decreased individuals' chances of collaborating with another conference participant.We run standard difference-in-difference regressions, examining the likelihood of collaboration among participants in the APSA conference in the 2009-2012 editions and, using as 'control' the chance of collaboration among participants attending a comparator conference (the Midwest Political Science Association Annual Meeting).To conduct this analysis, we assembled new datasets including 17,468 academics (attendants of the relevant meetings) and close to 86 million dyads of participants.This sample is representative of research active academics in the field of political science, accounting for 22% of published authors during the period.We matched these datasets to co-authored working papers and published articles to infer the occurrence of a collaboration.
We find that the 2012 APSA meeting cancellation led to an eighteen percent decrease in individuals' likelihood of co-authoring an output with another conference participant, and moreover that it was specifically the likelihood of collaborating with academic affiliated to a different institution that fell.In our regressions, we include controls for individual fixed effects and several covariates to control for individuals' time-varying productivity and propensity to collaborate.Our findings are robust to several econometric specifications and sample classifications.(Moreover, in the Appendix, we provide evidence that there were no systematic changes in attendants' characteristics across 'treatment' and 'control' conferences in the year of the cancellation.) We also find that collaborations forged between the attendants of occurring conferences lead to better publication outcomes.Of collaborations manifested as journal publications, those that were among academics scheduled to attend the cancelled meeting appeared, on average, in journals ranked 5 places lower than those that were among academics that attended a conference that actually took place.Pairs of collaborators that are not collocated or have not co-authored a paper together before, are the ones that benefit most from conferences, in terms of improving their ranking of publication.This premium may be driven by academics finding more suitable co-authors in academic meetings, or by already nascent collaborations benefitting from the faceto-face interaction afforded by the conference.In either of these cases, our results demonstrate that there are strikingly substantial, positive effects of relatively brief face-to-face meetings in the success of academic papers.
Chai (2014)  However, by using a natural experiment we offer here an identification strategy that more definitively avoids issues such as endogeneity of the acceptance status or decision to attend a conference.Outside of a typical conference setting, Boudreau et al. (2017) implemented a field experiment at Harvard Medical School to understand how search costs (within one institution) affect the formation of collaborations.Within a grant opportunity informational event, individuals were randomly assigned to small brainstorm sessions, and participants were subsequently 75% 1 Previous papers have however used experimental data to understand other forms of conference effect.Leon and  McQuillin (2016) used the same data and setting as the current paper to look at conference effects on the academic impact (as measured by citations) of presented papers.Blau et al. (2010) investigated the success of CeMENT -a mentoring workshop, arguably similar to a small conference, for female assistant professors -in increasing participants' publications and successful grant publications, based on randomized controlled trial data.more likely to write a grant application with an academic assigned to the same brainstorm room than with someone assigned to a different room.
Our results also contribute to a broader and growing literature on the determinants of the formation of academic collaborations.2A strand within this literature seeks to understand the role of communication costs and network constraints in determining patterns of academic collaborations. 3 Agrawal and Goldfarb (2008) and Ding et al. (2010) show that the introduction of Bitnet (an early version of the Internet) led to substantial increases in the rates of multi-institutional collaborations.Concomitantly, as shown in Kim et al. (2009), the research productivity effects of being placed in a top university diminished, as academics' dependency on colleagues, as possible co-authors, declined.Even so, it seems that opportunities for face-to-face interaction remain important: Boudreau et al. evidence this, and Catalini et al. (2016) show that decreasing air travel costs have also facilitated collaboration.A survey conducted by Freeman et al. (2015)   suggests that, still, most academic collaborations are among (presently or previously) collocated authors, and therefore that network constraints remain significant in affecting collaborations.Our results suggest that even a single conference relaxes these constraints with discernible effect.
We develop the remainder of the paper as follows.In section 2 we explain the data and describe the natural experiment.In section 3 we present the estimates of conference effects.We conclude, in section 4, with some interpretative discussion.

Background
The American Political Science Association (APSA) is a professional association of political science in the United States, and it publishes one of the preeminent journals in its field: The American Political Science Review.Its Annual Meeting is held immediately preceding Labor Day (in September) and gathers close to 3,000 presenters, from more than 700 institutions.
Participants present working papers in fifty two main themes encompassing a very broad spectrum of approaches and research topics across the field of political science.
The 2012 APSA meeting was due to take place in New Orleans and was scheduled to start on August 30.However, it was cancelled at less than 48 hours' notice due to the approach of "Hurricane Isaac".By the time of this cancellation, and indeed well before any genesis of tropical cyclone Isaac itself, the conference program was finalised and publicly available. 4The cancellation generated a group of participants that did not experience the network benefits of the meeting.
Our main hypothesis is that the individuals within this group (named in the 2012 APSA Meeting Programme) became less likely to form subsequent in-group collaborations than individuals in the groups that attended occurring conferences.
To quantify conference impacts, we conduct difference-in-differences regressions using data on conference participants from the 2009-2012 editions.This strategy is similar to the one we use in Leon and McQuillin (2016).We use, as a baseline group, collaborations formed among participants at a comparator conference: the Midwest Political Science Association (MPSA) Annual Meeting.The MPSA is also a professional association of political science scholars in the United States, and it publishes another leading journal within the discipline: The American Journal of Political Science.The APSA and the MPSA meetings are the largest conferences in the field of political science.They are similar in profile and almost identical in format.Each is a four-day event consisting of panels, posters, workshops, evening sessions, roundtables, and panel sessions including four presenting papers, discussants and a chair.Moreover, because the MPSA meeting (held in April) precedes the APSA meeting, the 2012 MPSA attendance would not have been impacted by the APSA cancellation.

Data Sources and Sample
We collected information on APSA and MPSA conference participants, from the meeting programmes available on the respective associations' websites. 5The programmes describe, for each session within the conference, the names and affiliations of the session chair(s) and discussants(s), and for each presenting paper within the session the names and affiliations of all of the authors.Also recorded is the theme with which the session is associated, and title (sub-theme) of the session. 6r presenting papers with more than one author, the conference programmes do not distinguish the presenting author, and therefore in our main analyses we will have included some "conference participants" who, as non-presenting co-authors, did not actually attend.Our main sample includes all individuals named in panel sessions, as an author, a chair or a discussant.
We also present the results for the subset of "sure participants": the 69 percent of conference participants who appear within the programme as a sole-author, chair or discussant. 7e assembled dataset comprises 39,586 conference-authors and 17,468 individuals.(They often attend many conferences.An individual who attended three conferences appears, in our dataset, as one individual -"Adam Adams" -and as three conference-authors: "APSA2009-Adam Adams", "MPSA2011-Adam Adams", and "APSA2012-Adam Adams".)In our analysis, we examine the data at two levels: (i) at the conference-author level in which the outcome is an indicator for whether the individual comes to collaborate with someone in the same conference, and (ii) and at the conference-author pair (dyad) level, where the outcome is whether the pair collaborates after the conference (and also whether the pair generates an output with specific characteristics).The number of observations is described in Table 1.For example, the 2009 APSA Meeting entailed a total of 4,007 participants (column 1) and therefore 4,007 x 4,006/2 = 8,026,021 potential collaboration dyads (column 2).However, we ignored collaborations formed among co-authors in a paper presented in the attending conference.This was to avoid the risk of misinterpreting a conference effect on the publication outcome of the presented paper as an impact on the formation of new work (the phenomenon of interest in 6 The APSA meeting has 52 main theme panels (that contain 90% of the articles) and 70 remaining themes that vary per year. 7Within the two conferences, panel sessions concentrate most of the presenting authors.70.8% of presenting papers are single-authored.By including individuals named in co-authored papers in our analysis, we may underestimate the conference effect, because we are likely to include authors who did not attend the conference.However, by excluding these papers we also underestimate the conference effect because we thereby exclude individuals that have a higher intrinsic likelihood to work in teams.this paper).8Hence, when analysing the data at the dyad level, we disregarded these pairs (14,252 dyads across eight conferences), and focused on the remaining number, shown in column 3.This exclusion comes at the expense of shutting down another possible channel of conference effect: on ongoing collaborations that might get reinforced, and turned into new research projects, because of feedback and suggestions offered during the conference.

T able1
To look for collaborations, we assembled a dataset of working papers and published papers in political science, using the Social Science Research Network (SSRN) and the Web of Science

Linking Datasets
We linked the SSRN and WoS data to conference participants using individuals' first and last name.A complication, in using this rule, is that some names are not unique among published authors or among conference participants, potentially leading to misattributions of collaborations (based on co-publications).There are, within our set of conference participants, 493 firstname/last-name combinations that appear with more than one associated middle initial across the set of all conference participants and WoS authors.(For example, we may have both a Jenny A Jones and a Jenny B Jones among the conference participants, or -more frequently -a Jenny Jones among the conference participants and both a Jenny A Jones and a Jenny B Jones among the WoS published authors.)We categorised these names as "ambiguous" and checked by hand all co-publications involving someone with an ambiguous name.

Measures of Collaboration and Explanatory Variables
Our main dependent variable and measure of collaboration is the existence of a co-authored paper after the conference, in a form of a SSRN working paper or a WoS published paper.A key decision was the time frame within which to look for such collaborations.Our main interest is in collaborative outputs that may have been in some sense generated by a conference, and we judged that papers appearing very soon after a conference (less than one year for a working paper, or less than approximately two and a half years for a published paper) were likely to reflect work that had been substantially completed by the time of the conference itself.We therefore consider, for each conference, working papers appearing in a two-year window commencing one year after the conference, and published papers appearing in the two complete calendar years commencing between two and three years after the conference.In Table A1 in the Appendix, we provide the precise starting and ending dates for these windows. 10 terms of explanatory variables, from the WoS data, we recover conference-author characteristics, as observed in a five year window prior to their attendance in the conference.
These are: the number of previous publications weighted by journal impact factor, the number of previous collaborators, and the number of previous collaborators attending the same conference.
From the conference programmes, we recovered each conference-author's affiliation and we associated geographic coordinates,11 an affiliation ranking, and affiliation "size".Affiliation rankings, coded 1 to 200 and "below top 200", were taken from Hix (2004).The "size" was based on the number of individuals (with a given affiliation) within the eight conferences (APSA2009-12 and MPSA 2009-12).For each conference-author pair we additionally calculated the spatial distance between affiliation coordinates (using the "geodist" command in Stata).

Summary Statistics
In Table 2, we present summary statistics describing the set of conference-authors.Most of the authors are affiliated to an institution in the US (81%), and they are roughly equally divided (30.9%, 30.3% and 38.5%) between institutions ranked, respectively, in the top 50, 51-200, and outside the top 200.Less than a third have co-authored an article (31.8%) and the average number of previous publications in the sample is 1.32.The APSA and MPSA meetings seem to host a large proportion of PhD students and early-career academics: the median participant has not published, or co-authored, a paper before the conference.However, the meetings also gather experienced authors.The academic in the 90th centile (95th centile) has published 4 papers (6 papers) in the five years preceding the conference, and has 3 (4) previous co-authors also attending the conference.In terms of outcomes, 15% of conference-authors subsequently co-author a paper with another participant in the conference. 12In Table A3 in the Appendix, we provide a picture of types of participants that come to collaborate.Conference collaborations are more likely to occur among academics that have closer research (are assigned to the same session and that have papers in the same theme), that work in the same institution and have collaborated before.

T able2
In the last two columns in Table 2, we present separate means for participants in the APSA (treatment) and MPSA (control) meetings.The MPSA meeting has a larger number of participants (and presenting papers) than the APSA: 4,700 vs 3,000 participants, on average.This difference reflects in participants' profiles: MPSA participants have fewer previous publications on average (1.03 vs 1.66) and fewer co-authors (1.34 vs 1.74).The diff-in-diff approach that we are using controls for systematic differences across conferences, such as different standards for article (and author) acceptance.In our main regressions we control for individual fixed effects, but it is also appropriate to check that the peer environments in neither the treated (APSA) nor control (MPSA) meeting series changed in any systematic way in 2012. 13We test for whether participants' pre-determined characteristics evolved in parallel over conference-years by running standard diff-in-diff regressions on these characteristics.We use, as dependent variables, several author characteristics predictive of collaborative behaviour (including previous collaborations and publication record).The results are reported in Table A2 in the Appendix.For most variables, we find no statistically significant effect and the p-values associated to the 2012 APSA coefficient are large.The parallel conference series trends are also noticeable visually in Figures 1A in the

Results
We present the analysis in two levels.We examine impacts at the conference-author level to better illustrate the magnitudes of the effects for individual academics.Then, we move the analysis to the dyad level to investigate the types of collaborations facilitated by a conference, and to examine the productivity of these collaboration-types.Lastly, we examine broader conference effects on academics' publication portfolios.

Effects of Conferences on Academic Collaborations
We begin by investigating how the 2012 APSA Meeting cancellation affected the likelihood of academic collaborations.Figures 1 and 2 illustrate this impact in terms of simple averages.Figure 1 shows the number of unique pairwise collaborations formed among conference participants.
While there is a visible upward trend in the number of collaborations among participants over the period 2009-11, there is then a marked fall among participants in the 2012 APSA Meeting.This suggests a conference effect.Based on a simple diff-in-diff calculation, 76 pairwise collaborations were lost due to the 2012 cancellation.In Figure 2, we show the effects at conference-author level (by aggregating pairwise collaborations).The graph shows the proportion of academics that subsequently co-author at least one article with another participant in the conference, and suggests that this percentage decreased by 1.6 percentage points as a consequence of the conference cancellation.

F igures1and2
Next, we investigate the effects of conferences on authors' chance of forming collaborations in a more controlled way.We estimate the following equation: where, i indexes for individual, s 2 fAP SA; M P SAg for conference series and t 2 {2009,2010,2011,2012} for conference year.An observation corresponds to a conference-author (a combination of individual, conference series and year).Collaboration ist is an indicator for whether the conference-author subsequently collaborated with another participant in the same meeting (the same conference series and year).The dummies [s = AP SA] and [t = T ] indicate respectively whether the conference-author is associated with an APSA meeting, and with a year T meeting.The terms i , X it , and ist denote, respectively, individual fixed effects, a vector for time-varying author characteristics (the number of previous publications weighted by journal impact factor, the number of previous co-authors, previous co-authors with attendants in the same conference and size of affiliated institution), and a random term.Robust standard errors are in brackets.
The results are shown in the first row of Table 3.The diff-in-diff estimates indicate that the 2012 APSA cancellation lead to a decrease in the likelihood of authors subsequently collaborating with others in the conference by 2.4 percentage points (around 18 percent).The magnitude and significance of this effect is robust to using time-varying controls (columns 3 and 6) and to restricting the sample to "sure-participants" (columns 4-6).
The next rows in Table 3 show three different classifications for the collaboration outcome.
First (rows 2-3), by collocation versus non-collocation: i.e. showing the conference effect on authors' likelihood of forming a subsequent collaboration with another meeting participant, (i) from their own institution, and then (ii) from a different institution.Second (rows 4-6), by a proxy for research-closeness: author's likelihood of forming a subsequent collaboration with another meeting participant, (i) from her own session(s) within the conference, (ii) from a session in the same theme, and then (iii) in a session from a different theme.And third (rows 7-8), by distinguishing between existing and new co-authors.
It is unclear whether the originally-observed effect is present among collocated co-authorships: the coefficients are no longer significantly different from zero (and indeed change their sign within the sure-participant sub-sample).But the effect is detected among inter-institutional coauthorships: it is specifically the likelihood of forming a collaboration with an author from a different institution that fell, by 22 percent.
The results in rows 5-6 suggest that conferences mainly facilitate collaborations between academics already working on closely related topics (indicated here by having a paper in the same theme): doubling the likelihood of such a collaboration.Curiously, however, the results in row 4 do not seem to indicate that conferences facilitate collaborations specifically between authors assigned to the same session. 14Finally, comparing rows 7 and 8, it seems to be that conferences facilitate collaborations between new, rather than existing, co-authors.
Altogether, therefore, the results in Table 3 suggest that conferences in some sense facilitate collaborations between participants at the conference, and that these are collaborations between individuals with closely related existing research interests, who are not collocated, and who have not previously collaborated.

T able3
For additional insights, we turn to analysis at the conference-author-pair (dyad) level.In particular, our analysis at this level -in Table 4 -is suggestive of a conference-effect both in terms of increasing the likelihood of collaboration and in terms of the quality of the co-authored output. 15ntrols used in the regression include, firstly, measures of "closeness": the geographical distance between the authors in the dyad, and then dummy indicators for whether the authors have the same affiliation, work in the same country, and whether they are previous co-authors.Further controls are proxies for productivity and propensity to collaborate.These are all measures at the author-level, but are included (at the dyad level) both as the average and as the absolute difference 14 We also did not detect effects of conferences increasing the chance of collaboration between within-session presenters-and-discussants, presenters-and-presenters, or presenters-and-chairs.(These results are not reported, but available under request.)These finding contrasts somewhat with the strong in-session effects reported in Boudreau  et al. (2017). 15In Table 4, we estimate the following equation: where, fijg indexes for a conference-author pair, s for conference series and t for conference year.An observation corresponds to a conference-author-dyad (a combination of author pair, conference series and year).Collaboration f[ij]stg is an indicator for whether the conference-author-dyad subsequently collaborated.The terms Z f[ij]tg and f[ij]stg denote, respectively, time-varying dyad characteristic controls, and a random term.
within the pair: the authors' total number of previous publications weighted by journal impact factor, the number of authors' previous co-authors, and size of own institution.
In Table 4, Panel A, Column 1 we show the conference effect on the likelihood of a collaboration among a pair of participants.In column 2, we focus on the chance of a co-publication only.In both cases, the coefficients are negative (i.e.suggesting that the cancellation led to a reduced likelihood of collaboration) but they are only (marginally, at the 10% level) statistically significant using the broader measure.In the remaining Panel A columns, we decompose the likelihood of a co-published paper into the chance of a co-publication in journals in the first, second, third and fourth quartiles, by impact factor.The results here are not significant.
In Table 4 Panel B we restrict the data to dyads that produced a co-publication (n=3,820).In the regressions, we use -as dependent variables -the journal ranking of the co-publication (varying from 1 to 149), the normalized journal impact factor and indicators for publication among journal impact factor quartiles.The estimates indicate that collaborations among 2012 APSA meeting participants led to co-publications placed, on average, in journals that are 4.97 points lower-ranked (column 1) and whose impact factor is 0.19 standard deviations lower (column 2).These impacts are statistically significant at the 1% level.The results for publication-quartile resemble the ones in Panel A, but gain statistical significance, suggesting in particular that co-publications among 2012 APSA meeting participants were shifted to lower ranked journals because of the conference cancellation.They were less likely to be placed in first-quartile journals (by 9 percentage points) and more likely to be published in second-(by 2.5 percentage points) , and third-quartile journals (by 6.7 percentage points).

T able4
A key question is then: why are collaborations formed within the cohorts of occurring conferences are associated with better publication outcomes?We conjecture two main possible explanations.One explanation is that conferences cause changes to patterns of co-authorship, in a direction which is productivity-enhancing.An alternative explanation is that the conference itself provides an important opportunity for co-authors to meet, discuss, and generally improve their work.The eventual outputs are better because of the face-to-face interaction between the co-authors that takes place within the conference.

Heterogeneous Effects of Conferences in Co-publication Outputs
In Table 5, we further explore the findings above.We ask which types of collaborations specifically benefitted from the conferences (or were most negatively impacted by the cancellation).Repeating the analysis of Table 4 Panel B Column 2, we again use as the dependent variable the normalised impact factor of the journal within which the co-publishing dyad's output appears. 16However, we now run separate regressions splitting the sample, (i) into "existing" versus "new" co-authors (rows 1-2), (ii) into pairs that are "collocated" versus "non-collocated" (rows 3-4), and (iii) into pairs whose authors are based in the same versus in a different country (rows 5-6).

T able5
We only detect statistically significant conference effects for collaborations formed between pairs that are non-collocated and that have not previously co-authored a paper (new co-authors).
These loci of benefit seem wholly consistent with the two explanations (for conference effects on publication outcomes) offered above.Co-authors that are collocated or that have previous experience collaborating together are unlikely to have depended on the conference as an opportunity for face-to face interaction to the same extent to as those that are non-collocated or newly collaborating.On the other hand, if a conference can help an author to find the right co-author then it is most likely to do so for authors looking for a new co-author, outside their own institution.17

Broader Conference Effects
Finally, we consider broader effects of conferences on academics' publications.In particular, it seems natural to consider whether collaborative outputs triggered by the conference represent an overall boost to academics' productivity in terms of published papers, or rather a displacement of other co-authors and projects.We estimate equation (1), using as the dependent variable various counts relating to the conference-author's published papers (within the same window as used for the analyses in Table 3).These include counts of published papers, co-published papers (in total and also disaggregated to papers with and without a conference co-author), and sole-authored papers.They also include counts of the conference-author's "new" co-authors (co-authors not appearing in any co-authored paper preceding the conference) and new co-authors from the conference.The results are shown in Table 6.
It should be noted that in this analysis it is natural to exclude from consideration authors that attended both the MPSA and APSA conferences in a given conference year.This is because several outcome measures will necessarily be the same for the MPSA conference-author as for the APSA conference-author. 18Though we report results for the full sample in column 1, we show results excluding authors that attended both conferences in column 2.
Few of the estimated impacts in Table 6 are statistically significant, and indeed it seems reasonable that the consequence of missing one conference will be scarcely discernible in this 'bigger picture'.However, the signs of coefficients are suggestive of a co-author substitution effect.Rows 3-4 show negative coefficients for the effect of the 2012 APSA cancellation on the numbers of published papers that are co-authored with a conference co-author (p-value=12.6%),and positive coefficients for the effects on the number of co-published papers with someone that is not in the conference programme (p-value=18%).The coefficients in rows 6-7 suggest that 2012 APSA conference-authors were also able, notwithstanding the conference cancellation, to find new co-authors, but that these became less likely to be from within the conference cohort.The significant result in row 7 corroborates the findings in Tables 3 and 4. It suggests that conferences are an important instrument for networking and that their effects are likely to survive in the long term, reflected in an academic's pool of collaborators.

T able6 4 Discussion
We have found that the cancellation of the 2012 APSA meeting reduced participants' likelihood of subsequently co-authoring a paper with another scheduled participant at the meeting.
It particularly reduced the likelihood of forming an inter-institutional collaboration.We have also observed that collaborations formed among participants in occurring conferences were associated with better publication outcomes than those formed among the participants of the conference that was cancelled.This may have been either because an occurring conference improves the matching of co-authors, or because it provides an invaluable opportunity for face-to-face interaction within early stages of a collaborative project.
There is no question that conferences feature prominently in academic and scientific life.Our findings give scientific corroboration to the previously untested (but commonly held) supposition that conferences are commensurably instrumental in the formation of scientific work.
To some extent, our findings also throw light on broader issues in a literature that more generally explores team formation and network effects in scientific production.Freeman and   Huang (2015) have suggested that the characteristics of collaborations that are most productive are not necessarily the characteristics that are most commonly observed. 19There could be two reasons for this.It could be that that the teams that commonly arise are relatively inefficient, and that working with distant co-authors improves the potential for a project; or it could be instead that the scientist's decision to work with different (new, less usual) co-authors is itself endogenous to the potential of the project.The fact that, in our results, an exogenously induced reduction in inter-institutional collaboration seems to have dented publication outcomes, could be viewed as support for the first of these two accounts.This then begs a further question: why do academics or scientists not reach out beyond their usual (for example, collocated) pool of co-authors more often?
The answer could lie in communication (or other) costs,20 or it could lie in network constraints that affect the pattern of co-author matching, and therefore reduce scientific productivity.A conference represents a (slight) relaxation of network constraints, and so the fact that academics respond by increasingly forming productive, inter-institutional and new collaborations suggests that (to some degree at least), the networks account lies behind existing inefficiencies. .

APSA Estimates
Notes: The unit of observation is at the pair (dyad) level.The outcome in column 1, Panel A refers to an indicator for whether the pair has produced a joint SSRN working paper or co-published a paper in the timeframe detailed in Table A1.The dependent variables in columns 3-6 are indicators for whether the pair has co-published a paper according to the impact factor journal quartile.All regressions include controls for: geographical distance between the authors in the dyad, same affiliation (dummy), same country (dummy), previous co-authors (dummy), number of previous publications weighted by journal impact factor (average and absolute difference), number of previous co-authors (  Notes: The sample specified in column 2 excludes conference-participants that attend both the APSA and the MPSA in the same year."Conference author" refers to an author attending the same conference.All regressions include author fixed effects, number of previous publications weighted by journal impact factor, the number of previous co-authors, and previous co-authors with attendants in the same conference and size of affiliated institution.Standard errors are in brackets.

APSA Estimates
also infers positive conference effects by comparing patterns of collaboration among attendees of the Gordon Research Conferences with patterns among a matched group of non-conference attendees.

(
WoS).The set of working papers comprises all papers posted in the SSRN Political Science Network from January 1996 to September 2015.The set of published papers comprises all articles published in the 155 WoS Political Science journals and in the top 20 WoS journals in Economics, Sociology, Law, History, and International Relations from 2004 to 2016. 9The list of journals is detailed in the Appendix.The sets include 113,895 working papers and 199,692 published papers respectively. Appendix.
average and absolute difference), and affiliation size (average and absolute difference).The number of observations in Panel A differs from 86,071,612 because data on controls are missing for 1,303,698 dyads.Standard errors are in brackets.***Significant at the 1 percent level.**Significant at the 5 percent level.*Significant at the 10 percent level.value for a test of [1]=[2] is 0.1131, for [3]=[4] is 0.2222, and for [5]=[6] is 0.6309.***Significant at the 1 percent level.**Significant at the 5 percent level.

Figure A2 -
Figure A2 -Conference-Authors that Petitioned Against the 2012 APSA Venue The main theme sections are Political Thought and Philosophy, Foundations of Political Theory, Normative Political Theory, Formal Political Theory, Political Psychology, Political Economy, Politics and History, Political Methodology, Teaching and Learning, Political Science Education, Comparative Politics, Comparative Politics of Developing Countries, The Politics of Communist and Former Communist Countries, Advanced Industrial Societies, European Politics and Society, International Political Economy, International Collaboration, International Security, International Security and Arms Control, Foreign Policy, Conflict Processes, Legislative Studies, Presidency Research, Public Administration, Public Policy, Law and Courts, Constitutional Law and Jurisprudence, Federalism and Intergovernmental Relations, State Politics and Policy, Urban Politics, Women and Politics Research, Race, Ethnicity, and Politics, Religion and Politics, Representation and Electoral Systems, Political Organizations and Parties, Elections and Voting Behavior, Public Opinion, Political Communication, Science, Technology, and Environmental Politics, Information Technology and Politics, Politics, Literature, and Film, New Political Science, International History and Politics, Comparative Democratization, Human Rights, Qualitative and Multi-method Research, Sexuality and Politics, Health Politics and Policy, Canadian Politics, Political Networks, Experimental Research.

Table 1 -Number of Observations
: The number of prospective collaborations (in column 3) is given by the number of conference-author dyads that have not collaborated in a paper presented in the conference. Note

Table 2 -Descriptives of Conference-Authors' Characteristics
Significant at the 1 percent level.**Significant at the 5 percent level.*Significant at the 10 percent level. ***

Table 3 -Effects of Conferences on the Formation of Academic Collaborations
Each entry in columns 2,3, 5 and 6 represents estimates for the 2012 APSA coefficient from a separate regression.Outcomes are indicator for whether the conference-author has come to collaborate with someone with the specific characteristics.All regressions include author fixed effects.Time varying controls include: total number of previous publications weighted by journal impact factor, number of previous co-authors, previous co-authors attending the same conference, and the size of own institution (as explained in Section 2.4).The definition of 'sure-participant' is explained in Section 2.2.Standard errors are in brackets. Notes:

Table 5 -Journal Impact Factor Outcomes of Collaborations Formed Among Conference Participants by Pair Characteristics
Each row in the table represents estimates for the 2012 APSA coefficient from a separate regression.Standard errors are in brackets.*Significant at the 5 percent level.*Significant at the 10 percent level. Notes:*