## wpiea2020144-print-pdf

## Source details

**Canonical URL:** [wpiea2020144-print-pdf](https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020144-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2020/english/wpiea2020144-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2020/english/wpiea2020144-print-pdf.pdf.json)

---

### Data
- Universe: IMF Working Papers published between 1990 and 2017.
- Total working papers collected: 6,152.
- Distinct authors in the universe: 3,918.
- Excluded working papers with no authors listed: 44 papers (all from the 1990s).
- PDF files not available for: 822 publications, primarily prior to 1997.
- For each working paper (where available) compiled: list of authors; number of pages; author emails; Journal of Economic Literature (JEL) codes (JEL codes available starting in 1997).
- Employee-level dataset coverage: years 1999 to 2017; annual data on job title, year of birth (in five-year brackets), tenure, department, citizenship (by region), gender.
- Job-title to grade-level mapping:
  - support-level staff (research assistants): grades A01–A10,
  - economists: grades A11–A15,
  - managerial staff: grades B01–B05.
- Authors’ age at publication: calculated based on the first year of the five-year birth bracket.
- Non-IMF authors coverage: Working Papers published between 1997 and 2017; main affiliation compiled from emails and manual online searches.
- External affiliations categories: universities; other international organizations; central banks; government agencies; non-profit organizations; private companies.
- Citation source: Google Scholar; citations cumulative as of January 2019; citation counts annualized by dividing by the number of years since publication.
- Definition of active author set S_t: authors from earliest year of publication t_0 to latest year t_2; pair active from t_{i,j,0} = max{t_i0, t_j0} to t_{i,j,2} = min{t_i2, t_j2}.
- First-collaboration indicator fc_{i,j,t}: defined when pair active; equals 1 in year of first joint publication.
- Subsequent-collaboration indicator sc_{i,j,t}: defined during active period after first publication; equals 1 in year(s) of publications other than the first.

### Stylized Facts — Publications and Authors
- Total authors and papers (1990–2017):
  - 3,918 authors published 6,152 IMF Working Papers.
- Publications per year:
  - Increased steadily up to 2006 and stabilized at around 250 to 300 thereafter.
- Authors per year:
  - Increased faster than publications, reaching almost 600 in 2017.
- Collaboration patterns:
  - In 2017 three fourths of the publications were conducted in collaboration.
  - Over careers most authors worked with just one to three co-authors; some collaborated with over 100 distinct co-authors.
- Author activity and productivity:
  - In any given year, about 80 percent of the authors wrote just one paper.
  - Around one fourth of publications was produced by first-time authors.
  - Majority of authors wrote only one paper over their IMF publishing career.
  - For authors with more than one paper, median number of papers was four; very few authors produced more than 20 papers.
  - Number of years an author remained active—excluding one-off papers—was six on average, with half active for five years or less.

### Stylized Facts — Author Composition, Departments, Gender, Region
- Position by share of publishing authors (1999–2017):
  - Economists (grades A11–A15) account for over 80 percent of all publishing authors.
  - Support staff (grades A6–A10) rose to about ten percent since 1999.
  - Managerial-level positions (grade B1–B5) declined from nearly 20 percent to ten percent.
- Publication propensity as percent of the workforce (2017):
  - About 20 percent of economists published at least one paper.
  - Seven percent of support staff published at least one paper.
  - 15 percent of managers (focusing on managers with an economic background) published at least one paper.
- Departmental composition:
  - Around 40 percent of publishing authors work in area departments (AFR, APD, EUR, MCD, WHD); 60 percent mostly from functional departments (RES, SPR, STA, FAD, MCM, ICD).
  - RES stands out with the highest publication rate measured as publishing economists as percent of economists employed by the department.
  - Since 2009 among area departments: decline in EUR and increases in WHD and APD in the share of economists who publish.
- Gender:
  - Share of female authors gradually increased to about 30 percent in recent years.
  - In 2017 about 20 percent of both male and female economists published at least one paper (no systematic gender differences once workforce shares are taken into account).
- Region of citizenship (2017):
  - Slightly less than half of publishing authors were from Europe and North America, with their share decreasing over time; transition countries and East Asia largely compensated for the decline.
  - Publication rates as percent of IMF economists by region broadly tracked each other, with slightly higher rates for transition countries and East Asia in recent years.

### Stylized Facts — Author Pairs, Pages, Citations, Topics
- Active pairs:
  - Number of active pairs peaked at over 400,000 in 2011–2013.
  - On average just over 0.1 percent of active pairs actually published together.
- Duration and productivity of pairs:
  - Great majority of author pairs active for only one year.
  - Most author pairs that published released only a single paper.
- Pages:
  - Number of pages ranged, on average, between 30 and 40 over the years.
  - Median pages lower than mean, indicating right-skewed distribution.
- Citations (cumulative as of January 2019, annualized):
  - Average number of citations hovered between five and ten.
  - Median number of citations ranged between zero and five.
  - Distribution right-skewed with substantial dispersion.
- Relationship with number of authors:
  - Publications with six or more authors are about 15 pages longer than single-authored publications (median values).
  - Papers with three to four authors tend to get the highest number of citations.
- Research topics (JEL letters):
  - Majority classified under E and F; E and F account for about half of total publications.
  - Adding G and H covers about 75 percent of all publications.
  - In recent years about 80 percent of publications report more than one JEL letter; 30 percent report more than two.
  - Emerging upticks: climate change (JEL Q54) since 2012; gender (JEL J16) since 2015 (from a very low base).
- Gender and departmental topic patterns:
  - Male and female authors publish within the same JEL codes in broadly equal proportions; female authors marginally fewer under E and somewhat more across most other JEL codes.
  - Departments align with mandates: FAD publishes more under H; MCM more under G.

### Author-Level Network — Summary Statistics (1990–2017)
- Total nodes (authors): 3,918
- Total edges (co-authorship): 9,027
- Average degree: 4.608
- Size of giant component: 3,206
- Share of giant component (percent): 81.86
- Average distance in giant component: 4.961
- Average clustering coefficient: 0.614
- Subperiod (1990–1999): nodes 923; edges 1,257; average degree 2.724; giant component 594; share 64.48; average distance 5.132; clustering 0.443.
- Subperiod (2000–2009): nodes 1,891; edges 4,081; average degree 4.316; giant component 1,598; share 84.58; average distance 4.878; clustering 0.504.
- Subperiod (2010–2017): nodes 2,192; edges 6,058; average degree 5.527; giant component 1,916; share 87.44; average distance 4.640; clustering 0.613.
- Comparative reference (EconLit 1990–1999):
  - Total nodes 81,217; total edges 67,897; average degree 1.672; giant component 33,027; share 40.7; average distance 9.470; clustering 0.157.
- Structural interpretation:
  - Giant component covers 3,206 nodes (~82 percent).
  - Average shortest path length about five (exactly 4.961).
  - High clustering (exactly 0.614) and evidence of interlinked “star” authors (negative relation between degree and local clustering coefficient).

### Affiliation Composition and Core Network
- IMF staff unique authors: 2,125 nodes.
- Other affiliations (unique authors):
  - Academics: 964
  - Central bankers: 258
  - Other international organizations: 109
  - Government agencies: 72
  - Private companies: 52
  - Non-profits: 35
- Core author network (1990–2017) definition: authors with at least ten publications and at least ten different co-authors (minimum degree ten).
  - Core network size: 233 authors.
  - Most prolific and best-connected authors predominantly IMF staff; non-IMF core mainly academics visiting or employed at the IMF.

### Institution-Level Network (1997–2017)
- Distinct institutions IMF staff collaborated with: 639
  - Universities: 380 institutions
  - Central banks: 77 institutions
- Core institution network (institutions with at least ten publications):
  - Single most important institutional collaboration: World Bank — distinct publications 98; distinct authors 83.
  - Academic collaborators mainly U.S. and U.K. universities; no academic institutions from emerging market and developing economies in the core.
  - Core central banks include Federal Reserve Board and European Central Bank.
  - OG Research is only private company in the core (more than ten publication involvements but just three authors).
  - OECD and Inter-American Development Bank contributed to about 20 publications each in the core.
  - Government agencies and non-profits not represented in the core institutional network.
- Departmental collaboration (1999–2017):
  - RES most prominent in publication co-authorship: academics contributed 1,308 publications vs. RES staff 1,077 publications.
  - Departments with involvements in more than 600 publications: RES, MCM, FAD, EUR.
  - RES collaborates mostly with EUR, WHD, APD, SPR; RES’s collaborations with external universities or central banks exceed collaborations with other IMF departments.
  - Several departments (ICD, MCM, FAD, EUR, SPR) collaborate more with universities than with other IMF departments.
  - Within-department collaborations more common than cross-department collaborations.

### Evolution of the Network
- Growth and integration:
  - Number of authors: "just over 900" in 1990–1999; "almost 2,200" in 2010–2017.
  - Number of co-authorship links: "nearly quintupled" between 1990–1999 and 2010–2017.
  - Average degree increased from "2.7" to "5.5".
  - Giant component coverage: "64 percent" in the 1990s; "almost 85 percent" in the 2000s; "more than 87 percent" in the 2010s.
  - Average distance within the giant component declined from "5.1" to "4.6".
  - Average clustering coefficient rose considerably over time.
- Distributional changes:
  - Degree distribution right-skewed in all periods; mode between "one and two".
  - Clustering coefficient distribution heavily right-skewed; mode close to "zero".
  - Distance between unconnected authors: mode moved from "five" in 1990s to "four" in 2010s; minimum value "two".
  - Number of shortest paths: mode moved from "zero" to "one" by the 2010s.
- External collaboration shifts (2000s → 2010s):
  - Slight increase in diversity away from academics toward more central bankers and non-profits (think tanks).
- Evolution of core:
  - Core defined as authors with at least "ten" publications in each ten-year period; set of most prolific authors evolved with few appearing in the core for more than one period.
- Cluster structure among 2010–2017 core:
  - Departmental clustering important; clusters align with JEL areas (e.g., RES with C and E; SPR cluster; FAD with H; RES-WHD with F and G).

### Empirical Strategy for Collaboration Determinants
- Estimation: logit specification on sample of active pairs:
  - Pr(y_{i,j,t} = 1) = Φ(α + β N_{i,j,t} + γ C_{i,j,t−1} + δ HR_{i,j,t} + ε_{i,j,t})
  - Dependent variable y_{i,j,t} alternately: dummy for first collaborations or for subsequent collaborations.
- Regressor groups:
  - Network variables N_{i,j,t}: proximity (inverse shortest distance in G_t excluding direct co-authorship), number of shortest paths (excluding direct link).
  - Pair characteristics C_{i,j,t}: lagged average productivity and lagged absolute difference in productivity (productivity q_{i,t} = points accumulated between t−9 and t where points = citations / (number of authors + 1)); lagged average propensity to collaborate and lagged absolute difference; research overlap (cosine similarity of JEL first-letter distributions).
  - Employment and demographic HR_{i,j,t}: same department dummies, same gender dummy, same region of citizenship dummy.
- Network window: graph G_t uses collaborations between t−9 and t (ten-year window); proximity minimum distance = "two"; maximum proximity = "0.5"; unconnected → proximity = "zero".
- Sample sizes noted:
  - Initial collaborations: "6,040,424" observations for "2,310,620" pairs.
  - Subsequent collaborations: "14,893" observations for "3,325" pairs.

### Results — Determinants of Starting a New Collaboration (Table 2 highlights)
- Network distance is a key determinant of first collaborations:
  - Proximity odds ratio: 6.309*** (standard error 0.141).
  - Ln number of shortest paths odds ratio: 3.826*** (0.149).
- Pair characteristics (lagged):
  - Lag avg. productivity:
    - Column (1): 1.003 (0.002)
    - Column (2): 1.006*** (0.002)
    - Column (3)/(4): 1.007*** (0.002)
    - A one standard deviation increase in average productivity associated with a 9 to 10 percent higher probability of starting a new collaboration (text).
  - Lag abs. diff. in productivity:
    - Column (1): 0.996** (0.002)
    - Column (2): 0.997** (0.001)
    - Column (3)/(4): 0.998 (0.001) — effect not robust when all determinants included.
  - Lag avg. propensity to collaborate:
    - Column (1): 0.821*** (0.006)
    - Column (2): 0.819*** (0.006)
    - Column (3): 0.795*** (0.007)
    - Interpretation: authors with many connections are less likely to start new collaborations.
  - Lag abs. diff. in propensity to collaborate:
    - Column (1): 1.070*** (0.005)
    - Column (2): 1.061*** (0.005)
    - Column (3): 1.082*** (0.006)
    - Interpretation: pairs with different propensities to collaborate are more likely to start a collaboration.
- Research overlap:
  - Lag of research overlap: Column (3) 1.030 (0.075); Column (4) 0.951 (0.083) — research overlap insignificant once employment/demographic variables included.
- Employment and demographic effects (Table 2, column (4)):
  - Same department: 5.096*** (0.238).
  - Same region of citizenship: 1.278*** (0.067).
  - Same gender: 1.007 (0.047) — not significant.
- Table 2 sample and fit statistics:
  - Observations: 6,040,424 (col (1)), 3,729,804 (col (2)), 2,374,994 (col (3)), 1,921,742 (col (4)).
  - Pairs: 2,310,620 (col (1)), 875,197 (col (2)), 682,589 (col (3)), 526,841 (col (4)).
  - Pseudo R2: 0.361 (col (1)), 0.431 (col (2)), 0.478 (col (3)), 0.511 (col (4)).

### Results — Determinants of Continuing to Collaborate (Table 3 highlights)
- Network variables matter but less than for first collaborations:
  - Proximity (examples):
    - Column (1): 1.011 (0.022)
    - Column (2): 1.054** (0.026)
    - Column (3): 1.040 (0.025)
    - Column (4): 1.094** (0.041)
    - Example interpretation: reducing distance from three to two almost doubles chances of continuing (text).
  - Ln number of shortest paths:
    - Column (1): 1.507*** (0.077)
    - Column (2): 2.282*** (0.141)
    - Column (3): 2.217*** (0.138)
    - Column (4): 2.508*** (0.196)
    - Doubling number of shortest paths increases probability of continuing by about 50 percent (text).
- Pair characteristics:
  - Lag avg. productivity:
    - Column (1): 1.014*** (0.002)
    - Column (2): 1.016*** (0.002)
    - Column (3): 1.012*** (0.002)
  - Lag abs. diff. in productivity:
    - Column (1): 0.992*** (0.002)
    - Column (2): 0.992*** (0.002)
    - Column (3): 0.994*** (0.002)
    - Larger productivity differences reduce likelihood of continuation.
  - Lag avg. propensity to collaborate:
    - Column (1): 0.900*** (0.011)
    - Column (2): 0.919*** (0.011)
    - Column (3): 0.926*** (0.011)
    - Larger number of established connections reduces probability of continuing.
  - Lag abs. diff. in propensity to collaborate:
    - Column (1): 1.067*** (0.007)
    - Column (2): 1.056*** (0.007)
    - Column (3): 1.054*** (0.008)
    - Larger differences in propensity increase probability of continuing.
- Research overlap and employment/demographic:
  - Lag of research overlap:
    - Column (1): 0.573*** (0.063)
    - Column (3): 0.898 (0.109)
    - Insignificant when employment/demographic included (column (4)).
  - Same department (column (4)): 3.270*** (0.253).
  - Same gender (column (4)): 1.242*** (0.104).
  - Same region of citizenship (column (4)): 1.223** (0.098).
- Table 3 sample and fit statistics:
  - Observations: 14,893 (col (1)), 14,589 (col (2)), 14,207 (col (3)), 12,645 (col (4)).
  - Pairs: 3,325 (col (1)), 3,289 (col (2)), 3,244 (col (3)), 2,865 (col (4)).
  - Pseudo R2: 0.011 (col (1)), 0.033 (col (2)), 0.036 (col (3)), 0.090 (col (4)).

### Model Performance and Robustness
- ROC (richest specifications, column (4)):
  - AUROC for initial collaborations: 0.94.
  - AUROC for subsequent collaborations: 0.74.
- Robustness checks (selected):
  - Excluding non-IMF employees (columns (1) initial; (2) subsequent):
    - Initial collaborations proximity: 13.454*** (1.023).
    - Subsequent collaborations proximity: 1.113** (0.057).
  - Excluding pairs with a common coauthor dropped 17 initial pairs and 8 subsequent pairs; results close to baseline.
  - Redefining network over [t−4, t] largely confirms baseline; lag avg. productivity turns insignificant for initial collaborations in that specification.
- Robustness examples (Table 4 selected entries):
  - Proximity examples:
    - Col (1): 13.454*** (1.023)
    - Col (2): 1.113** (0.057)
    - Col (3): 14.247*** (1.099)
    - Col (4): 1.067* (0.037)
    - Col (5): 7.036*** (0.338)
    - Col (6): 1.291*** (0.041)
  - Ln number of shortest paths examples:
    - Col (1): 15.214*** (1.443)
    - Col (2): 2.818*** (0.273)
    - Col (5): 8.629*** (0.660)
    - Col (6): 2.369*** (0.186)
  - Observations/pairs examples for robustness:
    - Observations: 1,449,272 (col (1)); 9,962 (col (2)); 1,921,631 (col (3)); 12,346 (col (4)); 2,039,812 (col (5)); 13,026 (col (6)).
    - Pairs: 421,217 (col (1)); 2,495 (col (2)); 526,824 (col (3)); 2,857 (col (4)); 559,031 (col (5)); 2,901 (col (6)).

### Summary Conclusions (selected)
- Co-authorship network (1990–2017) features:
  - Increasing number of co-authors per paper over time.
  - Large majority of authors published only one paper; very few wrote more than 20.
  - Increasing diversity: women and non-U.S. and non-European authors increased their shares.
  - Primary topics: macroeconomics, international economics, financial economics, public economics; recent increases in climate change and gender topics.
- Network composition and structure:
  - IMF staff dominate in number of authors and papers; followed by academics, central bankers, staff from other international organizations.
  - Recurrent institutional links: World Bank, U.S. and U.K. universities, other European universities, advanced-economy central banks.
  - Within-IMF department collaborations more common than cross-department or external collaborations.
  - Network exhibits "small world" properties: many authors with few direct co-authors but short indirect chains and significant clustering.
- Main empirical findings on co-authorship formation:
  - Proximity in existing co-authorship network strongly increases likelihood of collaboration, especially for first-time collaborations.
  - Higher average productivity in a pair increases likelihood of starting and continuing collaborations.
  - Authors with established networks are less likely to seek new collaborations, but pairs with differing network sizes are more likely to start and maintain collaborations.
  - Same-department employment and same-region citizenship increase both starting and maintaining collaborations; broader research-area overlap does not consistently increase collaboration.

*Source: IMF Working Papers dataset description and analysis (1990–2017).*

### 2.1    Data  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 

### 2.1    Data

### Dataset coverage and publication-level variables
- Universe: IMF Working Papers published between 1990 and 2017.
- Total working papers collected: 6,152.
- Distinct authors in the universe: 3,918.
- Exclusions and missing files:
  - Excluded working papers with no authors listed: 44 papers (all from the 1990s).
  - PDF files not available for: 822 publications, primarily prior to 1997.
- For each working paper (where available) compiled:
  - list of authors,
  - number of pages,
  - author emails,
  - Journal of Economic Literature (JEL) codes (JEL codes available starting in 1997).

### IMF author employment and demographic data
- Employee-level dataset coverage: years 1999 to 2017.
- Employee-level information includes annual data on:
  - job title,
  - year of birth (in five-year brackets),
  - tenure,
  - department,
  - citizenship (by region),
  - gender.
- Name matching: fuzzy matching between working-paper authors and employee-level data plus extensive manual checks to merge records (to address spelling differences, middle names, maiden/married names, nicknames, and alternative Westernized names).
- Job-title to grade-level mapping used to distinguish:
  - support-level staff (research assistants): grades A01–A10,
  - economists: grades A11–A15,
  - managerial staff: grades B01–B05.
- Authors’ age at publication: calculated based on the first year of the five-year birth bracket.
- Time-invariant author characteristics were extended back prior to 1999.

### Non-IMF authors and institutional classification
- Coverage for non-IMF authors: Working Papers published between 1997 and 2017.
- Main affiliation compiled from author emails (where available) and manual online searches (CVs, LinkedIn, other websites).
- External affiliations organized into six mutually exclusive categories:
  - universities,
  - other international organizations,
  - central banks,
  - government agencies,
  - non-profit organizations (including think tanks and independent research institutes),
  - private companies (including consultancy firms and financial institutions).
- IMF visiting and resident scholars were classified as “non-IMF” authors and their home universities were looked up.

### Citation data
- Citation source: Google Scholar.
- Citations collected are cumulative as of January 2019.
- To account for differing publication dates, citation counts were annualized by dividing by the number of years since publication.
- Rationale: Google Scholar provides broader coverage (books, dissertations, grey literature, non-English publications) compared with other citation sources.

### Definitions and construction of author-pair activity variables
- Definition of active author set in year t: S_t includes each author from their earliest year of publication (t_0) to their latest year of publication (t_2).
- A pair of authors {i, j} ∈ S_t is considered active from t_{i,j,0} = max{t_i0, t_j0} until t_{i,j,2} = min{t_i2, t_j2}.
- First-collaboration indicator: fc_{i,j,t}
  - Defined only when the author pair is active.
  - fc_{i,j,t} = 1 in the year authors i and j collaborate for the first time; 0 for years before that.
- Subsequent-collaboration indicator: sc_{i,j,t}
  - Defined only during the period in which the pair is active and after the first publication.
  - sc_{i,j,t} = 1 in the year or years the pair publishes any paper other than the first; 0 otherwise.
- Visualization: Figure 1 (Lifetime of a Pair of Authors) illustrates the timing and lifetime of first and subsequent collaborations for pairs that co-author more than one paper.

*Source: IMF Working Papers dataset description (1990–2017) and associated employee-level data and methods.*

### 2.2  Stylized Facts

### 2.2  Stylized Facts

### Publications
- Between 1990 and 2017, 3,918 authors published a total of 6,152 IMF Working Papers.
- Number of publications per year:
  - Increased steadily up to 2006 and stabilized at around 250 to 300 thereafter.
- Number of authors:
  - Increased faster than publications, reaching almost 600 in 2017.
- Collaboration patterns:
  - Authors per working paper increased over time; in 2017 three fourths of the publications were conducted in collaboration.
  - Over their entire IMF publishing career most authors worked with just one to three co-authors, though some researchers collaborated with over 100 distinct co-authors.
- Author activity and productivity:
  - In any given year, about 80 percent of the authors wrote just one paper.
  - Around one fourth of publications was produced by first-time authors.
  - The large majority of authors wrote only one paper over their publishing career at the IMF.
  - For authors with more than one paper, the median number of papers was four, with very few authors producing more than 20 papers.
  - The number of years an author remained active—excluding one-off papers—was six on average, with half of the authors being active for five years or less.

### Authors
- Analysis covers 1999 to 2017 (employee-level data availability).
- Position by share of publishing authors:
  - Economists (grades A11–A15) account for over 80 percent of all publishing authors.
  - Support staff (grades A6–A10) rose to about ten percent since 1999.
  - Managerial-level positions (grade B1–B5) declined from nearly 20 percent to ten percent.
- Publication propensity as percent of the workforce (2017):
  - About 20 percent of economists published at least one paper.
  - Seven percent of support staff published at least one paper.
  - 15 percent of managers (focusing on managers with an economic background) published at least one paper.
- Departmental composition:
  - Around 40 percent of publishing authors work in area departments (AFR, APD, EUR, MCD, WHD).
  - The other 60 percent mostly come from functional departments (RES, SPR, STA, FAD, MCM, ICD).
  - In recent years, the largest share of authors was from RES, followed closely by WHD.
  - RES stands out with the highest publication rate when measured as publishing economists as percent of economists employed by the department.
  - Among area departments, since 2009 there has been a decline in EUR and increases in WHD and APD in the share of economists who publish.
- Gender and publication rates:
  - The share of female authors gradually increased over time and reached about 30 percent in recent years.
  - Once workforce shares are taken into account, there are no systematic gender differences in the percent of economists who publish.
  - In 2017 about 20 percent of both male and female economists published at least one paper.
- Region of citizenship:
  - In 2017 slightly less than half of the publishing authors were from Europe and North America, but their share has been decreasing over time.
  - Transition countries and East Asia largely compensated for this decline.
  - Publication rates expressed as percent of IMF economists by region have broadly tracked each other, with slightly higher rates for transition countries and East Asia in recent years.

### Author Pairs
- Active pairs:
  - Number of active pairs peaked at over 400,000 in 2011–2013.
  - Reference: IMF has around 3,000 employees; if all were active researchers there would have been around 4.5 million active pairs (3,000×2,999/2).
  - Since 2014, the number of active pairs has declined.
  - On average just over 0.1 percent of active pairs actually published together.
- Duration and productivity of pairs:
  - The great majority of author pairs were active for only one year.
  - Most author pairs that published released only a single paper.
- Pair attributes and collaboration propensity (observed patterns with caveat about denominators):
  - Women appear to collaborate relatively more with authors of the same sex than men do.
  - Economists collaborate more among themselves.
  - Authors who share the same region of citizenship or are employed in the same department have higher propensities to collaborate.
- Tenure and pair composition:
  - The vast majority of active pairs with at least one publication worked at the IMF for less than ten years.
  - Junior-senior collaborations are relatively common, whereas senior-senior collaborations are less frequent.

### Publication Length and Citations
- Pages:
  - Number of pages ranged, on average, between 30 and 40 over the years.
  - Median number of pages tends to be lower than the mean, indicating a right-skewed distribution driven by some exceptionally long publications.
- Citations (cumulative citations as of January 2019, annualized by years since publication):
  - The average number of citations hovered between five and ten.
  - The median number of citations ranged between zero and five.
  - Distribution of citations is right-skewed with substantial dispersion around the median.
- Relationship with number of authors:
  - Number of pages correlates with number of authors.
  - Publications with six or more authors are about 15 pages longer than single-authored publications, based on median values.
  - Papers with three to four authors tend to get the highest number of citations.
- Differences by gender and department:
  - No discernible difference in number of pages for male and female authors.
  - Papers by authors in RES and in MCM tend to be slightly longer.
  - Number of citations for male and female authors is similar.
  - Papers written by authors working in RES tend to have noticeably higher citation counts.

### Research Topics
- JEL classification (first-letter codes):
  - Majority of publications classified under E (Macroeconomics and Monetary Economics) and F (International Economics); these two codes account for about half of total publications.
  - Adding G (Financial Economics) and H (Public Economics) covers about 75 percent of all publications.
- Multi-topic and cross-listing patterns:
  - In recent years about 80 percent of total publications report more than one JEL letter.
  - 30 percent reports more than two JEL letters.
  - Two letters is the most commonly used configuration.
  - Most authors use two JEL letters over their active publishing years at the IMF, though many authors have used a wider range of JEL codes.
  - Heat map of bivariate JEL combinations shows E, F, G, and H often used in combination; recurrent combinations also include C (Mathematical and Quantitative Methods) and O (Economic Development).
- Emerging topics:
  - Uptick in papers on climate change (JEL code Q54) since 2012.
  - Uptick in papers on gender (JEL code J16) since 2015, albeit from a very low base.
- Gender and departmental differences in topics:
  - Male and female authors publish within the same JEL codes in broadly equal proportions.
  - Female authors publish marginally fewer papers under code E and somewhat more across most other JEL codes.
  - Departments show heterogeneity consistent with mandates: for example, researchers in FAD publish more under H (Public Economics), while researchers in MCM publish more under G (Financial Economics).

*Source: IMF Working Paper chapter excerpt — 2.2 Stylized Facts*

### 3.1  Author-Level Network

### 3.1 Author-Level Network

### Network description and visualization
- Nodes represent individual authors of IMF Working Papers published between 1990 and 2017; undirected edges represent co-authorship.
- Node size is proportional to the number of publications an author wrote over the entire period (including single-authored papers).
- Edge width is proportional to the number of repeated collaborations between two authors (including multiple publications within the same year).
- A publication with three authors implies three collaborations and thus three edges; a publication with four authors results in six edges, and so on (number of collaborations given by n(n−1)/2, where n is the number of authors).
- Node colors denote the mode of an author’s affiliation category over the years in which she/he published (categories: IMF; universities; other international organizations; central banks; government agencies; non-profit organizations; private organizations; N/A).

### Network summary statistics (IMF Working Papers, 1990–2017) and comparison
- Total nodes (authors): 3,918
- Total edges (co-authorship): 9,027
- Average degree: 4.608
- Size of giant component: 3,206
- Share of giant component (percent): 81.86
- Average distance in giant component: 4.961
- Average clustering coefficient: 0.614

- Comparative reference (EconLit articles, 1990–1999, from Goyal et al. (2006), shown in Table 1):
  - Total nodes (authors): 81,217
  - Total edges (co-authorship): 67,897
  - Average degree: 1.672
  - Size of giant component: 33,027
  - Share of giant component (percent): 40.7
  - Average distance in giant component: 9.470
  - Average clustering coefficient: 0.157

- Subperiod statistics for IMF Working Papers (from Table 1):
  - 1990–1999: Total nodes 923; Total edges 1,257; Average degree 2.724; Size of giant component 594; Share of giant component (percent) 64.48; Average distance 5.132; Average clustering coefficient 0.443.
  - 2000–2009: Total nodes 1,891; Total edges 4,081; Average degree 4.316; Size of giant component 1,598; Share of giant component (percent) 84.58; Average distance 4.878; Average clustering coefficient 0.504.
  - 2010–2017: Total nodes 2,192; Total edges 6,058; Average degree 5.527; Size of giant component 1,916; Share of giant component (percent) 87.44; Average distance 4.640; Average clustering coefficient 0.613.

### Structural interpretation and "small world" properties
- Most authors are connected to each other; the network has relatively few isolated authors or author groups.
- The giant component covers 3,206 nodes, or almost 82 percent of all authors in the entire network.
- Average shortest path length in the giant component is about five (exactly 4.961).
- Average clustering coefficient is more than 0.6 (exactly 0.614), indicating high overlap in co-authorship.
- The IMF Working Papers network is considerably more integrated than the EconLit network (higher average degree, larger share of giant component, higher clustering).
- The network exhibits clear “small world” properties: many authors have few direct co-authors but the network is highly integrated through short chains of co-authors, and there is a high extent of overlap in co-authorship relative to a random network.
- Evidence suggests an architecture of interlinked “star” authors (researchers who collaborate with many others who are not direct co-authors themselves), supported by the observed negative relation between authors’ degree and local clustering coefficient.

### Degree vs. clustering coefficient (empirical pattern)
- Many researchers with only two or three co-authors have at the median around 80 percent (100 percent) of their co-authors collaborating with each other.
- Researchers with over 20 co-authors seldom have clustering coefficients in excess of 0.2.
- Figure 8 plots individual authors’ degree against their clustering coefficient for authors with minimum degree of two (both axes logarithmic); size of dots is proportional to total publications by the author.

### Affiliation composition of authors
- IMF staff:
  - Unique authors: 2,125 nodes
  - IMF staff account for the largest number of publications by those authors (largest node sizes in Figure 7).
- Other affiliations (unique authors, nodes):
  - Academics (including resident and visiting scholars at the IMF): 964
  - Central bankers: 258
  - Staff from other international organizations: 109
  - Government agencies: 72
  - Private companies: 52
  - Non-profits: 35
- Many authors may have switched affiliations over time; node color denotes the mode of an author’s affiliation category.

### Core author network (1990–2017)
- Core sample definition: authors who contributed to at least ten publications and collaborated with at least ten different co-authors (minimum degree of ten).
- Core network size: 233 authors.
- Within the core network, the most prolific and best-connected authors are predominantly IMF staff; the few non-IMF authors in the core are mainly academics who were visiting or employed at the IMF.
- Most nodes in the core figure are connected by less than ten edges because collaborations with “non-core” authors are counted but not shown.

*Source: IMF Working Papers co-authorship data, 1990–2017 (figures and Table 1).*

### 3.2 Institution-Level Network

### Institution-level aggregation and visualization
- Institution-level network collapses individual co-authorship relations by institution for 1997–2017.
- Node size is proportional to the number of publications to which at least one author affiliated with the institution contributed (contributions by multiple authors of the same institution to a single publication are counted only once).
- Edge width is proportional to the number of publication collaborations between two institutions (collaborations by the same pair of institutions within a single publication are counted only once).
- Node colors follow institutional categories: IMF; other international organizations; universities; central banks; government agencies; non-profit organizations; private organizations.

### Breadth of institutional collaborations
- Number of distinct institutions (excluding self-employed consultants) that IMF staff collaborated with: 639
  - Universities: 380 institutions
  - Central banks: 77 institutions
- Very few papers are written solely by authors with a non-IMF affiliation (i.e., nodes not connected to the central IMF node).

### Core institution network (institutions with at least ten publications, 1997–2017)
- Single most important institutional collaboration: World Bank
  - Distinct publications: 98
  - Distinct authors: 83
- Academic collaborators are mostly universities in the United States and the United Kingdom; a limited number of other universities in Europe, Australia, Canada, and Israel are present in the core.
- No academic institutions from emerging market and developing economies are observed among the core.
- Central banks in the core include the Federal Reserve Board, the European Central Bank, and a handful of mostly advanced-economy national central banks.
- OG Research is the only private company in the core (more than ten publication involvements but just three authors).
- Organisation for Economic Co-operation and Development and the Inter-American Development Bank contributed to about 20 publications each in the core.
- Government agencies and non-profits are not represented in the core institutional network.

### IMF departments and cross-institution collaboration (1999–2017)
- Disaggregation of IMF authors by department (collapsing non-IMF affiliations into six categories) shows:
  - RES is the most prominent IMF department in terms of publication (co-)authorship.
    - Overall publications: academics contributed 1,308 publications vs. RES staff 1,077 publications.
  - Departments with involvements in more than 600 publications: RES, MCM, FAD, EUR.
- RES has the most collaborations with universities, central banks, and international organizations.
- Within the IMF, RES collaborates mostly with EUR, WHD, APD, and SPR, but RES’s collaborations with external universities or central banks exceed its collaborations with other IMF departments.
- Several IMF departments other than RES—especially ICD, MCM, FAD, EUR, and SPR—collaborate more with universities than with other IMF departments.
- Collaboration across area departments is relatively limited.

### Heat map of collaborations (1999–2017)
- Figure 13 quantifies the number of publications by pair of departments or institutional categories (collaborations by the same pair within a single publication are counted only once).
- Numbers on the diagonal refer to collaborations between authors of the same department or institutional category (i.e., within-department collaborations).
- The heat map demonstrates that research collaborations within the same department are more common than co-authorship across different departments.
- 186 authors affiliated with the IMF for which no information on the department was available are excluded from department-level figures.

*Source: IMF Working Papers co-authorship data, institution- and department-level aggregations, 1997–2017 (figures and heat map).*

### 3.3  Evolution of the Network

### 3.3  Evolution of the Network

### Network growth and integration over time
- Number of authors:
  - "just over 900" during 1990–1999.
  - "almost 2,200" during 2010–2017.
- Number of co-authorship links:
  - "nearly quintupled" between 1990–1999 and 2010–2017.
- Average degree:
  - increased from "2.7" to "5.5".
- Giant component coverage:
  - "64 percent" of authors in the 1990s.
  - "almost 85 percent" in the 2000s.
  - "more than 87 percent" in the 2010s.
- Average distance within the giant component:
  - declined from "5.1" to "4.6" between the 1990s and 2010s.
- Average clustering coefficient:
  - rose considerably over time; part of the increase attributable to a higher share of multi-authored publications in recent years.

### Changes in distributions of key network metrics (three sub-periods: 1990–1999, 2000–2009, 2010–2017)
- Degree distribution (Panel 15a):
  - Right-skewed in all periods.
  - 1990s: most authors had between "zero and two" co-authors; mode = "one".
  - 2000s and 2010s: right tail became fatter while mode remained between "one and two", indicating that already well-connected authors increased collaborations.
- Clustering coefficient distribution (Panel 15b):
  - Heavily right-skewed; mode remained close to "zero" across periods.
  - Right tail became fatter over time.
  - Distribution presents a bump at the end of the right tail, reflecting authors with most co-authors connected to each other (partly due to single publications with two or more co-authors).
- Distance between unconnected authors (Panel 15c):
  - Minimum value is "two" for authors sharing a common co-author.
  - Mode moved from "five" in the 1990s to "four" in the 2010s; distribution tightened around the mode, indicating unconnected authors became closer over time.
- Number of shortest paths between authors (Panel 15d):
  - Unconnected authors have value "zero".
  - Mode of the density moved from "zero" to "one" by the 2010s.
  - Right tail became fatter, indicating authors now have more equally-distant paths to reach others.

### Affiliation composition and external collaboration patterns
- IMF authorship:
  - IMF staff represented "just over 60 percent" of all nodes in both the 2000s and the 2010s (and most likely the 1990s).
- External collaboration shifts from the 2000s to the 2010s:
  - Slight increase in diversity of external collaborators away from academics (academics remained the most important external category) toward more central bankers and researchers affiliated with non-profits (mostly think tanks).

### Evolution of the core (most prolific authors)
- Core network defined as authors with at least "ten" publications in each ten-year period (1990–1999, 2000–2009, 2010–2017).
- The set of most prolific authors evolved over time; only a few authors appear in the core for more than one period.

### Cluster structure among core authors (2010–2017)
- Clustering performed on the 2010–2017 core using the modularity algorithm of Blondel et al. (2008).
- Main insight: working or having worked in the same IMF department is an important driver of first-time and subsequent research collaborations.
- Departmental clusters tend to correlate with research areas (JEL codes), for example:
  - A sky blue cluster: authors who worked in RES and wrote primarily on macro-monetary economics with a mathematical component (JEL codes C and E).
  - A green cluster: RES and SPR economists.
  - A pink cluster: FAD economists concentrating on public economics (JEL code H).
  - A light blue cluster (bottom-right): mostly RES and WHD staff working at the intersection of international and financial economics (JEL codes F and G).

### Empirical approach to determinants of collaborations (transition to Section 4)
- Objective: study determinants of first and subsequent collaborations between pairs of authors.
- Estimation: logit specification on the sample of active pairs:
  - Pr(y_{i,j,t} = 1) = Φ(α + β N_{i,j,t} + γ C_{i,j,t−1} + δ HR_{i,j,t} + ε_{i,j,t})
  - Φ is the logit function; dependent variable y_{i,j,t} is alternatively the dummy for first collaborations or for subsequent collaborations.
- Regessor groups:
  - Network variables N_{i,j,t}.
  - Pairs’ characteristics C_{i,j,t} (research ability, propensity to collaborate, research overlap).
  - Author-level IMF employment and demographics HR_{i,j,t}.
- Sample size notes:
  - For initial collaborations: "6,040,424" observations for "2,310,620" pairs.
  - For subsequent collaborations: "14,893" observations for "3,325" pairs.
- Network construction:
  - Graph G_t uses collaborations between t−9 and t (ten-year window).
  - Proximity defined as inverse of shortest distance d_{i,j,t} in G_t, computed excluding direct co-authorship between i and j; minimum distance = "two"; maximum proximity = "0.5"; no link → proximity = "zero".
  - Number of shortest paths between i and j in G_t computed excluding direct link; unconnected authors → value = "zero".
- Pair characteristics:
  - Research productivity score q_{i,t} computed as points accumulated between t−9 and t, where:
    - Points = citations / (number of authors + 1)
    - The +1 in the denominator avoids very high scores for papers with few co-authors.
  - Pair-level measures: lagged average productivity and lagged absolute difference in productivity.
  - Propensity to collaborate proxied by number of co-authors each author had over the past ten years; pair measures include lagged average propensity and lagged absolute difference in propensities.
  - Research overlap measured via cosine similarity of JEL first-letter distributions x_{i[t−9,t],f}; lagged overlap included in regressions; for subsequent collaborations, overlap excludes papers previously co-authored by the pair.
- Additional controls:
  - Dummy variables for whether pair worked in the same IMF department or did so in the past year.
  - Dummy for same gender.
  - Dummy for same region of citizenship.

*Source: IMF Working Papers co-authorship analysis, section 3.3 "Evolution of the Network", 1990–2017.*

### 4.2  Results

### 4.2 Results

### Determinants of starting a new collaboration
- Network distance is a key determinant of first collaborations.
  - Proximity (Table 2, column (1)) odds ratio: 6.309*** (standard error 0.141).
  - Example: for a pair with proximity 0 that becomes proximity 0.5, odds of starting a new collaboration are about 30 times larger (textual illustration).
  - Ln number of shortest paths (Table 2, column (1)) odds ratio: 3.826*** (0.149); doubling the number of shortest paths increases the probability of initiating a collaboration by a factor of nearly four (text).
- Including pairs’ characteristics (lagged) reduces sample to author pairs active at least two years and increases distance effects.
  - Lag avg. productivity:
    - Column (1): 1.003 (0.002)
    - Column (2): 1.006*** (0.002)
    - Column (3)/(4): 1.007*** (0.002)
    - A one standard deviation increase in average productivity is associated with a 9 to 10 percent higher probability of starting a new collaboration (text).
  - Lag abs. diff. in productivity:
    - Column (1): 0.996** (0.002)
    - Column (2): 0.997** (0.001)
    - Column (3)/(4): 0.998 (0.001) — effect not robust when all determinants included (text).
  - Lag avg. propensity to collaborate:
    - Column (1): 0.821*** (0.006)
    - Column (2): 0.819*** (0.006)
    - Column (3): 0.795*** (0.007)
    - Interpretation: authors with many connections are less likely to start new collaborations (text).
  - Lag abs. diff. in propensity to collaborate:
    - Column (1): 1.070*** (0.005)
    - Column (2): 1.061*** (0.005)
    - Column (3): 1.082*** (0.006)
    - Interpretation: pairs with different propensities to collaborate are more likely to start a collaboration (text).
- Research overlap:
  - Lag of research overlap: Column (3) 1.030 (0.075); Column (4) 0.951 (0.083) — research overlap turns insignificant once employment and demographic variables are included (text).
- Employment and demographic effects (Table 2, column (4)):
  - Same department: 5.096*** (0.238) — odds of starting a new collaboration increase by a factor of more than five (text).
  - Same region of citizenship: 1.278*** (0.067) — increases odds by a factor of about 1.3 (text).
  - Same gender: 1.007 (0.047) — not significant for starting collaborations (text).

- Table 2 key sample and fit statistics (reported):
  - Observations: 6,040,424 (col (1)), 3,729,804 (col (2)), 2,374,994 (col (3)), 1,921,742 (col (4))
  - Pairs: 2,310,620 (col (1)), 875,197 (col (2)), 682,589 (col (3)), 526,841 (col (4))
  - Pseudo R2: 0.361 (col (1)), 0.431 (col (2)), 0.478 (col (3)), 0.511 (col (4))

### Determinants of continuing to collaborate (subsequent collaborations)
- Network variables matter but less than for first collaborations.
  - Proximity (Table 3):
    - Column (1): 1.011 (0.022)
    - Column (2): 1.054** (0.026)
    - Column (3): 1.040 (0.025)
    - Column (4): 1.094** (0.041)
    - Example: reducing distance from three to two almost doubles chances of continuing (text) — compared to factor of 11 for initial collaborations.
  - Ln number of shortest paths:
    - Column (1): 1.507*** (0.077)
    - Column (2): 2.282*** (0.141)
    - Column (3): 2.217*** (0.138)
    - Column (4): 2.508*** (0.196)
    - Doubling number of shortest paths increases probability of continuing by about 50 percent (text).
- Author pair characteristics:
  - Lag avg. productivity:
    - Column (1): 1.014*** (0.002)
    - Column (2): 1.016*** (0.002)
    - Column (3): 1.012*** (0.002)
    - Interpretation: higher average productivity increases likelihood of continuing collaboration (text).
  - Lag abs. diff. in productivity:
    - Column (1): 0.992*** (0.002)
    - Column (2): 0.992*** (0.002)
    - Column (3): 0.994*** (0.002)
    - Larger productivity differences reduce likelihood of continuation (text).
  - Lag avg. propensity to collaborate:
    - Column (1): 0.900*** (0.011)
    - Column (2): 0.919*** (0.011)
    - Column (3): 0.926*** (0.011)
    - Larger number of established connections reduces probability of continuing (text).
  - Lag abs. diff. in propensity to collaborate:
    - Column (1): 1.067*** (0.007)
    - Column (2): 1.056*** (0.007)
    - Column (3): 1.054*** (0.008)
    - Larger differences in propensity increase probability of continuing (text).
- Research overlap and employment/demographic variables:
  - Lag of research overlap:
    - Column (1): 0.573*** (0.063)
    - Column (3): 0.898 (0.109)
    - Research overlap negative and significant in column (3) but insignificant when employment/demographic included (column (4)) (text).
  - Same department (Table 3, column (4)): 3.270*** (0.253) — increases odds by more than a factor three (text).
  - Same gender (column (4)): 1.242*** (0.104) — being same gender increases likelihood of continuing collaboration by 24 percent (text).
  - Same region of citizenship (column (4)): 1.223** (0.098) — increases odds by 22 percent (text).

- Table 3 key sample and fit statistics (reported):
  - Observations: 14,893 (col (1)), 14,589 (col (2)), 14,207 (col (3)), 12,645 (col (4))
  - Pairs: 3,325 (col (1)), 3,289 (col (2)), 3,244 (col (3)), 2,865 (col (4))
  - Pseudo R2: 0.011 (col (1)), 0.033 (col (2)), 0.036 (col (3)), 0.090 (col (4))

### Model performance (ROC)
- ROC analysis for richest specifications (column (4) of Table 2 for first collaborations; column (4) of Table 3 for subsequent collaborations):
  - AUROC for initial collaborations: 0.94 (text).
  - AUROC for subsequent collaborations: 0.74 (text).
  - Interpretation: model assigns a higher predicted value to a randomly selected collaboration than to a random non-collaboration with probability 94 percent (initial) and 74 percent (subsequent) (text).
  - Conclusion: logit models have satisfactory performance separating collaborating from non-collaborating pairs; weaker performance for subsequent collaborations may relate to reduced sample size (text).

### Robustness checks
- Robustness modifications applied to richest specifications and results (Table 4):
  - Excluding non-IMF employees (columns (1) initial; (2) subsequent):
    - Initial collaborations proximity: 13.454*** (1.023)
    - Subsequent collaborations proximity: 1.113** (0.057)
    - Results similar to baseline (text).
  - Excluding collaborations of pairs with a common coauthor (to neutralize "middleman/woman effect") (columns (3) initial; (4) subsequent):
    - Exclusion led to dropping 17 author pairs for initial and 8 author pairs for subsequent collaborations (text).
    - Results remain close to baseline (text).
  - Redefining network over [t−4, t] instead of [t−9, t] (columns (5) initial; (6) subsequent):
    - Allows larger sample; results largely confirm baseline. Only lag avg. productivity turns insignificant for initial collaborations (text).
- Table 4 selected entries (examples):
  - Proximity:
    - Col (1): 13.454*** (1.023)
    - Col (2): 1.113** (0.057)
    - Col (3): 14.247*** (1.099)
    - Col (4): 1.067* (0.037)
    - Col (5): 7.036*** (0.338)
    - Col (6): 1.291*** (0.041)
  - Ln number of shortest paths (examples):
    - Col (1): 15.214*** (1.443)
    - Col (2): 2.818*** (0.273)
    - Col (5): 8.629*** (0.660)
    - Col (6): 2.369*** (0.186)
  - Observations and pairs for robustness:
    - Observations: 1,449,272 (col (1)); 9,962 (col (2)); 1,921,631 (col (3)); 12,346 (col (4)); 2,039,812 (col (5)); 13,026 (col (6))
    - Pairs: 421,217 (col (1)); 2,495 (col (2)); 526,824 (col (3)); 2,857 (col (4)); 559,031 (col (5)); 2,901 (col (6))

### Summary conclusions (from section 5, concluding material included in the source)
- Co-authorship network of IMF Working Papers (1990–2017) features:
  - Increasing number of co-authors per paper over time (text).
  - Large majority of authors published only one paper during the period; very few authors wrote more than 20 (text).
  - Increasing diversity: women and non-U.S. and non-European authors increased their shares in total working paper output (text).
  - Most papers deal with macroeconomics, international economics, financial economics, and public economics; recent increases in topics including climate change and gender (text).
- Network composition and structure:
  - IMF staff dominate in number of authors and working papers, followed by academics, central bankers, and staff from other international organizations (text).
  - Institution-level recurrent links: World Bank, U.S. and U.K. universities, other European universities, advanced economy central banks (text).
  - Within-IMF department collaborations are more common than cross-department or external collaborations (text).
  - Network exhibits "small world" properties: many authors with few direct co-authors but short indirect co-authorship chains and significant clustering (text).
- Main empirical findings on co-authorship formation:
  - Proximity in the existing co-authorship network strongly increases likelihood of collaboration, especially for first-time collaborations (text).
  - Higher average productivity in a pair increases likelihood of starting and continuing collaborations (text).
  - Authors with established networks are less likely to seek new collaborations, but pairs with differing network sizes are more likely to start and maintain collaborations (text).
  - Same-department employment and same-region citizenship increase both starting and maintaining collaborations; greater overlap in broadly defined research areas does not (text).

*Source: https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020144-print-pdf.pdf*

### References

### References

### Research on scientific collaboration networks and co-authorship
- Adams, Jonathan (2012). “The Rise of Research Networks”.Nature490, pp. 335–336.
- Blondel, Vincent D, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre (2008). “Fast Unfolding of Communities in Large Networks”.Journal of Statistical Mechanics: Theory and Experiment2008(10).
- Goyal, Sanjeev, Marco J van der Leij, and Jos ́e Luis Moraga-Gonz ́alez (2006). “Economics: An Emerging Small World”.Journal of Political Economy114(2), pp. 403–412.
- Newman, Mark E J (2001). “The Structure of Scientific Collaboration Networks”.PNAS98(2), pp. 404–409.
- Liben-Nowell, David and Jon Kleinberg (2007). “The Link-Prediction Problem for Social Networks”.Journal of the American Society for Information Science and Technology58(7), pp. 1019–1031.
- Fafchamps, Marcel, Marco J van der Leij, and Sanjeev Goyal (2010). “Matching and Network Effects”.Journal of the European Economic Association8(1), pp. 203–231.
- Wuchty, Stefan, Benjamin F Jones, and Brian Uzzi (2007). “The Increasing Dominance of Teams in Production of Knowledge”.Science316(5827), pp. 1036–1039.
- Frenken, Koen, Jarno Hoekman, Suzanne Kok, Roderik Ponds, Frank van Oort, and Joep van Vliet (2009). “Death of Distance in Science? A Gravity Approach to Research Collaboration”. Innovation Networks, Understanding Complex Systems. Ed. by Andreas Pyka and Andrea Scharnhorst. Berlin/Heidelberg: Springer-Verlag. Chap. 3, pp. 43–57.

### Empirical studies on co-authorship, productivity, and citation patterns
- Hollis, Aidan (2001). “Co-authorship and the Output of Academic Economists”.Labour Economics8(4), pp. 503–530.
- Medoff, Marshall H (2003). “Collaboration and the Quality of Economics Research”.Labour Economics10(5), pp. 597–608.
- McDowell, John M and Michael Melvin (1983). “The Determinants of Co-Authorship: An Analysis of the Economics Literature”.Review of Economics and Statistics65(1), pp. 155–160.
- Ductor, Lorenzo (2015). “Does Co-authorship Lead to Higher Academic Productivity”.Oxford Bulletin of Economics and Statistics77(3), pp. 385–407.
- Ductor, Lorenzo, Marcel Fafchamps, Sanjeev Goyal, and Marco J van der Leij (2014). “Social Networks and Research Output”.Review of Economics and Statistics96(5), pp. 936–948.
- Borjas, George J and Kirk B Doran (2015). “Which Peers Matter? The Relative Impacts of Collaborators, Colleagues, and Competitors”.Review of Economics and Statistics97(5), pp. 1104–1117.
- Colussi, Tommaso (2018). “Social Ties in Academia: A Friend is a Treasure”.Review of Economics and Statistics100(1), pp. 45–50.
- Kuld, Lukas and John O’Hagan (2018). “Rise of Multi-authored Papers in Economics: Demise of the ‘Lone Star’ and Why?”Scientometrics114, pp. 1207–1225.
- Henriksen, Dorte (2016). “The Rise in Co-authorship in the Social Sciences (1980-2013)”.Scientometrics107(2), pp. 455–476.
- Rath, Katharina and Klaus Wohlrabe (2016). “Recent Trends in Co-authorship in Economics: Evidence from RePEc”.Applied Economics Letters23(12), pp. 897–902.
- Sommer, Vera and Klaus Wohlrabe (2017). “Citations, Journal Ranking and Multiple Authorships Reconsidered: Evidence from Almost One Million Articles”.Applied Economics Letters24(11), pp. 809–814.
- Sutter, Matthias and Martin Kocher (2004). “Patterns of Co-authorship among Economics Departments in the USA”.Applied Economics36(4), pp. 327–333.
- Hsieh, Chih-Sheng, Michael K ̈onig, Xiaodong Liu, and Christian Zimmermann (2018). “Superstar Economists: Coauthorship Networks and Research Output”. CEPR Discussion Paper DP13239.
- Azoulay, Pierre, Joshua S Graff Zivin, and Jialan Wang (2010). “Superstar Extinction”.Quarterly Journal of Economics125(2), pp. 549–589.
- Hamermesh, Daniel S (2013). “Six Decades of Top Economics Publishing: Who and How?”Journal of Economic Literature51(1), pp. 162–172.
- —  (2018). “Citations in Economics: Measurement, Uses, and Impacts”.Journal of Economic Literature56(1), pp. 115–56.
- Chai, Sen and Richard B Freeman (2019). “Temporary Colocation and Collaborative Discovery: Who Confers at Conferences”.Strategic Management Journal40(13), pp. 2138–2164.
- Boudreau, Kevin J, Tom Brady, Ina Ganguli, Patrick Gaule, Eva Guinan, Anthony Hollenberg, and Karim R Lakhani (2017). “A Field Experiment on Search Costs and the Formation of Scientific Collaborations”.Review of Economics and Statistics99(4), pp. 565–576.
- Campos, Raquel, Fernanda Leon, and Ben McQuillin (2018). “Lost in the Storm: The Academic Collaborations That Went Missing in Hurricane Isaac”.Economic Journal128(610), pp. 995–1018.
- Colussi, Tommaso (2018). “Social Ties in Academia: A Friend is a Treasure”.Review of Economics and Statistics100(1), pp. 45–50.

### Diversity, ethnicity, and gender in collaboration
- AlShebli, Bedoor K, Talal Rahwan, and Wei Lee Woon (2018). “The Preeminence of Ethnic Diversity in Scientific Collaboration”.Nature Communications9(5163).
- Freeman, Richard B and Wei Huang (2015). “Collaborating with People Like Me: Ethnic Coauthorship within the United States”.Journal of Labor Economics33(S1), S289–S318.
- Ductor, Lorenzo, Sanjeev Goyal, and Anja Prummer (2018). “Gender and Collaboration”. Cambridge Working Papers in Economics 1820.

### Citation studies, databases, and evaluation of research quality
- Aizenman, Joshua, Hali Edison, Larissa Leony, and Yi Sun (2011). “Evaluating the Quality of IMF Research: A Citation Study”. IEO Background Paper BP/11/01.
- Mart ́ın-Mart ́ın, Alberto, Enrique Orduna-Malea, Mike Thelwall, and Emilio Delgado L ́opez-C ́ozar (2018). “Google Scholar, Web of Science, and Scopus: A Systematic Comparison of Citations in 252 Subject Categories”.Journal of Informetrics12(4), pp. 1160–1177.
- Card, David and Stefano DellaVigna (2013). “Nine Facts about Top Journals in Economics”.Journal of Economic Literature51(1), pp. 144–161.
- Fawcett, Tom (2006). “An Introduction to ROC Analysis”.Pattern Recognition Letters27(8), pp. 861–874.
- Ravallion, Martin and Adam Wagstaff (2012). “The World Bank’s Publication Record”.Review of International Organizations7(4), pp. 343–368.
- IEO (2011). “Research at the IMF: Relevance and Utilization”. Evaluation Report.

### Methodology, measurement, and structural studies
- Catalini, Christian (2018). “Microgeography and the Direction of Inventive Activity”.Management Science64(9), pp. 4348–4364.
- Kuld, Lukas and John O’Hagan (2018). “Rise of Multi-authored Papers in Economics: Demise of the ‘Lone Star’ and Why?”Scientometrics114, pp. 1207–1225.
- Medoff, Marshall H (2003). “Collaboration and the Quality of Economics Research”.Labour Economics10(5), pp. 597–608.
- McDowell, John M and Michael Melvin (1983). “The Determinants of Co-Authorship: An Analysis of the Economics Literature”.Review of Economics and Statistics65(1), pp. 155–160.

*Source: wpiea2020144-print-pdf - References*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020144-print-pdf.pdf_