NBER WORKING PAPER SERIES

(CO-)WORKING IN CLOSE PROXIMITY:

KNOWLEDGE SPILLOVERS AND SOCIAL INTERACTIONS

Maria P. Roche
Alexander Oettl
Christian Catalini

Working Paper 30120

http://www.nber.org/papers/w30120

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
June 2022

The authors thank Karen Houghton and her team for providing access to data. We thank
Annamaria Conti, Joachim Henkel, Matt Higgins, Bill Kerr, Rem Koning, Olav Sorenson, Pian
Shu, Peter Thompson, Stefan Wagner, John Walsh, Martin Watzinger as well as seminar
participants at Boston, Cornell, EPFL, Georgetown, Georgia Tech, Harvard, London Business
School, Max Planck Institute, Rice, UCLA, the EGOS Colloquium, and AOM meetings for
helpful comments. Thank you to Sonit Bafna for help with space measures. Karen Oettl, thank
you for being our lead generator. We gratefully acknowledge funding from the Kauffman Junior
Faculty Fellowship and from the Harvard Business School Division of Research and Faculty
Development. All errors and omissions are our own. The views expressed herein are those of the
authors and do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.

© 2022 by Maria P. Roche, Alexander Oettl, and Christian Catalini. All rights reserved. Short
sections of text, not to exceed two paragraphs, may be quoted without explicit permission
provided that full credit, including © notice, is given to the source.
(Co-)Working in Close Proximity: Knowledge Spillovers and Social Interactions
Maria P. Roche, Alexander Oettl, and Christian Catalini
NBER Working Paper No. 30120
June 2022
JEL No. M13,O3,O33,R12

ABSTRACT

We examine the influence of physical proximity on between-startup knowledge spillovers at one

of the largest technology co-working hubs in the United States. Relying on the random
assignment of office space to the hub's 251 startups, we find that proximity positively influences
knowledge spillovers as proxied by the likelihood of adopting an upstream web technology
already used by a peer startup. This effect is largest for startups within close proximity of each
other and quickly decays: startups more than 20 meters apart on the same floor are
indistinguishable from startups on different floors. The main driver of the effect appears to be
social interactions. While startups in close proximity are most likely to participate in social co-
working space events together, knowledge spillovers are greatest between startups that socialize
but are dissimilar. Ultimately, startups that are embedded in environments that have neither too
much nor too little diversity perform better, but only if they socialize.

Maria P. Roche Christian Catalini

Morgan Hall MIT Sloan School of Management
Soldiers Field 100 Main Street, E62-480
Boston, MA 02163 Cambridge, MA 02142
mroche@hbs.edu catalini@mit.edu

Alexander Oettl
Scheller College of Business
Georgia Institute of Technology
800 West Peachtree Street NW
Atlanta, GA 30308
and NBER
alexander.oettl@scheller.gatech.edu
1 Introduction
The COVID-19 pandemic affected close to all facets of life, perhaps none more greatly than
work life. As offices around the world closed and work shifted to Zoom, the trend towards
remote work was greatly accelerated. While remote work has allowed many organizations to
continue their day-to-day operations, there is preliminary evidence that the lack of physical
proximity has altered the interactions and collaborations that normally would have taken
place (Yang et al., 2021). Although the importance of place for innovation, entrepreneurship,
and firm performance has been well established in the literature (e.g., Rosenthal and Strange
2004; Michelacci and Silva 2007; Samila and Sorenson 2011; Glaeser et al. 2015), less is known
about the level at which geographic proximity matters most. On the one hand, cities may
serve as the appropriate focus for understanding the dynamics of labor markets, but it may
be at much smaller scales that more nuanced interpersonal interactions – especially those
that produce knowledge spillovers – take place.

While physical proximity is one of the more salient dimensions of distance that has been
shown to impact knowledge exchange among collaborators (Allen, 1977; Cowgill et al., 2009),
numerous other distances also facilitate/impede knowledge exchange and learning. For
example, social (Blau, 1977; McPherson and Smith-Lovin, 1987), product-market (Wang
and Zhao, 2018; Alcácer et al., 2015; Saxenian, 1996), and knowledge-space (Cohen and
Levinthal, 1990; Lee, 2019; Lane et al., 2020) distances have all been shown to impact the
ability or willingness to exchange knowledge. Although recent work stresses the importance
of taking such factors into account when aiming to optimize peer effects (Carrell et al., 2013;
Chatterji et al., 2019; Hasan and Koning, 2019), less is known about how the similarity and
dissimilarity of startups impacts knowledge transfer in physical proximity. On the one hand,
similar startups, due to their common ground and shared understanding, may more effectively
exchange knowledge. Conversely, however, the value of this more efficacious knowledge
transfer may be diminished due to redundant knowledge that both parties already possess.

1
Thus, being geographically proximate may be most advantageous to dissimilar startups who
can benefit most from diverse and novel knowledge.1

In this paper we build upon prior research by applying a micro-geographic lens to deepen our
understanding of the relationship between physical proximity and knowledge exchange between
early-stage entrepreneurial firms (startups). Shedding light on how startups interact with their
environment is of particular importance given that dependence on external resources (e.g.,
compute power, labor platforms, manufacturing, knowledge, etc.) has become increasingly
crucial for startups (Conti et al., 2021).2 In particular, we examine how geographic distance
impacts knowledge spillovers amongst nascent startups located within the same building – a
startup co-working space – and further document the role that differences among startups play
in modulating the effect of distance. Our results indicate a more nuanced role of proximity
in fostering knowledge spillovers across nascent firms. We find that physical proximity is less
important in promoting knowledge exchange amongst similar startups, but, in turn, more
crucial for startups that are dissimilar.

The setting for our study is one of the largest technology co-working spaces in the United
States. The building consists of five floors, covering 9,300 m2 (100,000 sq.ft.). One challenge
in examining the relationship between location and knowledge spillovers is that startups
and individuals may choose to locate in areas where knowledge exchange is already likely
to be high. To deal with this potential endogenous location choice, we rely on the random
assignment of office space to the hub’s 251 startups. We measure knowledge spillovers as
the instance of adopting a component of a peer startup’s technology stack. Using floor plans
to measure geographic distance, we find that close physical proximity greatly influences the
likelihood of these knowledge spillovers. This effect, however, quickly decays with distance
1
The greatest advancements in understanding the importance of similarity and physical proximity has
been evaluated at the individual level. However, due to methodological challenges it has still been difficult to
separately distinguish the effects of proximity and homophily on outcomes (Angst et al., 2010; Conley and
Udry, 2010).
2
The size of the mean high-tech startup has decreased to approximately two employees over the past two
decades (Ewens and Marx, 2017; Kaplan et al., 2009).

2
where startup startups that are more than 20 meters (66 feet) away are no longer influenced
by each other. Strikingly, being located more than 20 meters apart, but on the same floor
does not appear to differ from being located on a different floor altogether. In addition, we
find that when startups overlap with common areas at the hub (e.g., kitchens), the distance of
influence increases, revealing the important role that these spatial features play in extending
geographic reach and in promoting knowledge spillovers.

Why do these micro-distances matter? As suggested by Tortoriello et al. (2015), frequent

and repeated interactions may help promote fine-grained information sharing and allow for a
better understanding of a neighbor’s knowledge and skill. Via its impact on the likelihood
and frequency of interacting with others, physical proximity may thereby play an especially
fundamental role in not only enabling access and awareness of distinct knowledge pieces
(Borgatti and Cross, 2003), but also for the integration and internal use of externally sourced
information. This is provided that “(...) interpersonal channels are more effective in forming
and changing attitudes toward a new idea, and thus in influencing the decision to adopt
or reject a new idea (Rogers 2010, p.36)”. Therefore, to understand the possible dynamics
underlying knowledge spillovers at short distances, we examine the role of social interactions
in explaining the relationship between physical proximity and knowledge exchange. To do so,
we exploit event check-in data that provides information on the temporal overlap of startup
members at events where we would expect social interactions to occur. Our results indicate
that proximity predicts joint attendance of these events – joint socializing – and that startups
who co-attend these events produce the largest technology adoption peer effects when they
are dissimilar from one another.

The broader innovation literature stresses the importance of external knowledge in promoting
innovation and startup performance (Cohen and Levinthal, 1990; Chesbrough, 2012). Because
external knowledge provides unique insights previously unavailable to the startup (Zahra
and George, 2002; Laursen and Salter, 2006) and provides access to information from a

3
wide range of skills and experiences, it aids in maximizing a startup’s capacity for creativity,
knowledge-generation, and effective action (Reagans and Zuckerman, 2001; Aggarwal et al.,
2020). Building on this research, we further examine the impact of a startup’s environment
on early-stage startup performance (raising a seed round or receiving more than $1MM in
funding). We find that startups embedded in environments that have neither too much nor
too little diversity perform better, but only if they engage in social interactions.

Overall, our findings contribute to prior research in important ways. First, we provide a
better understanding of a fundamental decision early stage, high tech ventures face: building
their technology infrastructure. Especially in our context of technology-enabled high-growth
entrepreneurship, the adoption and integration of upstream production technologies, the
startup’s technology stack, may be considered comparable to supplier choice in more traditional
industries – a crucial decision, which tends to imply significant path dependency (Arthur,
1994; Murray and Tripsas, 2004; Alcácer and Oxley, 2014; Fang et al., 2020). Second, where
prior research has emphasized the role of a startup’s formal, structural features, such as its
size, age and prior social ties in the entrepreneurial process (Elfenbein et al., 2010; Hasan and
Koning, 2019), our analyses yield unique insights into knowledge exchange and integration of
startups by highlighting the role of diversity among exchange partners and the critical role of
proximity. We underscore that understanding which startups and how they respond to their
peer startups matters for designing effective environments for early stage startups. Finally, we
speak to the literature examining accelerators, bootcamps, incubators and other interventions
targeted at early stage entrepreneurs (e.g., Hassan and Mertens 2017; Cohen et al. 2019;
Lyons and Zhang 2018) by examining an additional type of entrepreneurial workplace that
has received limited attention so far in the literature: the co-working hub (Howell, 2022).3

Taken together, this paper informs our understanding of the scale at which knowledge
3
We intentionally use the term hub as described in e.g., Schilling and Fang (2014), since - similar to hubs
“who have significantly more connections than does the average member”(p.974) in an interpersonal network -
co-working spaces are designed to create more connections between entities in a shared environment.

4
spillovers among small, nascent firms take place. We thereby highlight important nuances in
terms of the benefits accruing from physical proximity depending on how different exchange
partners are from each other. Importantly, we observe that physical proximity is most helpful
for supporting knowledge exchange among startups that are otherwise distant. A feasible
explanation for our findings is that spatial proximity increases the likelihood and frequency of
social interaction, which facilitates the integration of diverse knowledge. As such, our results
carry fundamental implications for the design of work spaces that cross the boundaries of
collaboration, may they be of physical or virtual nature, for innovation and entrepreneurial
communities.

This paper is structured as follows. In the next section, we briefly discuss findings established
in the existing literature. The third section describes the empirical estimation strategy and
data sources. In section four, we present our main results, provide suggestive evidence in
support of social interactions as a feasible mechanism, and unveil potential consequences
of knowledge spillovers from proximate, but different peer for performance outcomes. We
conclude this paper with a discussion of our findings, including limitations, and broader
implications for designing collaborative work environments and for developing technologies
that mimic co-location.

2 Background
2.1 Physical Proximity and Knowledge Spillovers
The diffusion of ideas has been found to be highly localized (Allen, 1977; Arzaghi and
Henderson, 2008). In theory, the assumption pervades that knowledge (especially more
tacit know-how) transfers via face-to-face interaction between individuals (Gaspar and
Glaeser, 1998; Jacobs, 1969; Moretti, 2004; Rosenthal and Strange, 2001). Empirical research
supports this idea with results indicating that the extent to which physical proximity explains
information flows can depend on as little as a few hundred meters in certain circumstances
(Catalini, 2018; Cowgill et al., 2009; Kerr and Kominers, 2015; Reagans et al., 2005).

5
 One important environment where many interactions occur and information exchange takes
place on a daily basis is the workplace. As such, the workplace represents a setting for
unexpected influences, and for the serendipitous flow of information and ideas. Here the
physical layout of the workplace can play a critical role with early research dating back to
Allen (1977) showing the importance of proximity in determining and shaping workplace
interactions. Studies have tested the link between proximity and interpersonal interactions
in the context of, e.g., science (Boudreau et al., 2017; Catalini, 2018), options exchange
(Baker, 1984), technology companies (Cowgill et al., 2009), e-commerce (Lee, 2019), and first
responders (Battiston et al., 2020) finding that physical proximity has important implications
for sharing information among collaborators and co-workers.

The importance of (work)place for knowledge diffusion also has strong implications for
nascent startups. Generally, entrepreneurs gain information from a variety of sources, though
one particularly important channel is through fellow entrepreneurs (Nanda and Sørensen,
2010; Lerner and Malmendier, 2013). This is provided that entrepreneurs predominately
operate in fast-paced and uncertain environments, making local search (Cyert et al., 1963)
based on experimentation and frequent adjustments (Lippman and McCall, 1976; Gavetti
and Levinthal, 2000; Gans et al., 2019) a crucial component in the early stages of a venture.

We propose that an important channel through which physical proximity may enable
knowledge exchange across startups is through the creation of increased opportunities for
social interaction. Social interaction represents an important integration mechanism which
enables better understanding of others’ specific background, challenges and language. This
understanding facilitates the processing of external knowledge and the development of
absorptive capacity (Todorova and Durisin, 2007; Dingler and Enkel, 2016), which influences
the decision to adopt or reject a new idea (Rogers, 2010). Moreover, frequent interaction
with partners may help establish emotional closeness, intimacy and trust (Granovetter, 1973);
all of which facilitate knowledge exchange and integration.

6
2.2 The Interplay of Physical Proximity with Non-Geographic

Similarity
While physical proximity has been shown to be an important condition for knowledge
exchange, other dimensions of proximity/similarity have also been shown to impact knowledge
transfer. For example, social (e.g., Blau 1977; McPherson and Smith-Lovin 1987; Hasan and
Koning 2019), product-market (e.g., Wang and Zhao 2018; Alcácer et al. 2015; Saxenian 1996),
and knowledge-space (e.g., Cohen and Levinthal 1990; Lee 2019) proximity are important
facilitators of knowledge spillovers as established by the literature. The level of which two
entities are similar (or different) along these dimensions plays a crucial role in governing
exchange between actors (Granovetter, 1973; McPherson and Smith-Lovin, 1987; Singh, 2005),
in reducing or creating barriers for knowledge spillovers (Marshall, 1890; Stefano et al., 2017;
Saxenian, 1996), in influencing the ability to absorb (Cohen and Levinthal, 1990), and the
amount of non-redundant and relevant information available between actors (Azoulay et al.,
2019; Burt, 2004; Oh et al., 2006; Schilling and Fang, 2014; Rogers, 2010). What remains
to be understood is how these other forms of proximity interact with physical proximity.
Since it is particularly challenging to integrate distant knowledge, it is possible that via its
impact on frequent and repeated interactions, which help establish trust (Granovetter, 1973)
and an understanding of others’ skills (Rogers, 2010), physical proximity aids especially in
connecting those that are otherwise different. If this is the case, we should detect a positive
interaction between physical proximity and diversity in predicting the likelihood of integrating
technologies from peer startups.

3 Empirical Strategy and Data

3.1 Estimation Strategy
Estimating the role of physical proximity on knowledge spillovers – peer technology adoption
– not only requires data at a highly granular geographic level, but is also likely to yield biased

7
estimates of the effect size. Specifically, as has been well documented in the context of
individual-level peer effects by Manski (1993), these biases may be driven by issues of
endogenous sorting, contextual effects, and other correlated effects. On the one hand,
technology adoption could be a function of characteristics of the group (e.g., industry type)
where startups that would use similar input factors like to locate close to each other. On the
other hand, startups that are in physical proximity often experience similar social phenomena
which could drive exposure to certain input factors. To deal with such endogenous geographic
clustering, we rely on the random assignment of office space to the hub’s 251 startups, while
to deal with contextual contaminants we specifically examine startup i’s decisions to adopt
relevant input factors that are already being used by startup j. Table 1 shows that pairwise
characteristics do not correlate with physical proximity, serving as a validation of our random
room assignment assumption (and confirmed by multiple senior staff at the co-working
space).4

<Insert Table 1 here>

To operationalize knowledge spillovers, we focus our attention on a fundamental decision

nascent startups have to make pertaining to their web infrastructure that entails considerable
path-dependency (Arthur, 1994; Murray and Tripsas, 2004; Alcácer and Oxley, 2014): web
technology stack choices. Specifically, we examine a) the count of web technologies startupi
adopts that startupj has already adopted, and b) the probability that startupi adopts a web
technology that startupj has already adopted. Using the unique startup dyad as our unit of
analysis, we estimate the following specification using OLS:

Yij = γln(distanceij ) + Xij + θi + φj + η (1)

where Yij represents our web technology adoption measures, Xij is a vector of dyad-specific
controls, and θi and φj are Roomi × Startupi and Roomj × Startupj fixed effects, respectively.
4
Please refer to Table A1 of the Appendix for further robustness checks.

8
The inclusion of the startup-room specific fixed effects allows us to hold all time-invariant
individual startup characteristics constant so that estimation of γ solely arises from dyad-
level variation in distance. The nature of our error term, η, is more complicated. First, if
geographic proximity affects web technology adoption decisions, then the outcomes of all
startups in close proximity will be correlated. We resolve this standard clustering problem by
clustering at the floor-neighborhood level (15 clusters) to account for correlated outcomes
in close proximity.5 Second, because of the dyadic nature of our data, it is insufficient
to solely engage in 2-way clustering at the separate startupi and startupj level.6 As an
example, the dyad startupi -startupj will also be correlated with the dyads startupi -startup0j
since a common component of startup i’s web technology adoption decisions will also create
correlation across all of startup i’s web technology decisions from each other dyad alter.
However, dyad startupi -startupj will also be correlated with dyads startupj -startup0i , that is,
any dyad that shares a common connection, i.e., has either startupi or startupj in common.
To correct for these two issues we follow recent work (Aronow et al., 2017; Cameron and
Miller, 2014; Carayol et al., 2019; Harmon et al., 2019) and produce dyadic-robust standard
errors using the floor-neighborhood locations of startups i and j as the levels of clustering.

In alternate analyses we estimate the following specification:

Yij = βCloseij + Xij + θi + φj + η (2)

where Closeij is equal to 1 if startups i and j are in the first quartile of the distanceij
distribution and 0 otherwise and further extend our analysis by interacting variables with
Closeij .
5
Based on the spatial layout of the co-working building, we attain these floor-neighborhoods by splitting
each floor into four quadrants (with the exception of the smaller fifth floor which we split into three).
6
In this 2-way setup, we would allow arbitrary correlation between the dyad startupi -startupj and all
other dyads startupi -startupj 0 .

9
3.2 Data Sources and Construction
The data for our study were collected at one of the five largest technology co-working spaces
in the United States (in 2016). Designated as a startup hub where new ventures work side by
side, the building consists of five floors, 9,300 m2 (100,000 sq.ft.) and 207 rooms. The data
cover a period of 30 months from August 2014 – January 2017, during which 251 unique
startups had rented an office in the co-working space. For our analyses, we only examine
interactions between startups on the same floor resulting in 10,840 unique startup dyads.
Note that the co-working hub is relatively specialized in digital technologies, fintech, software
development, and marketing tech.

Approximately 35 percent of the startups ceased operations or left the co-working space
each year, which according to senior administrators at the co-working space, typically occurs
either because startups fail, grow out of the space, or occasionally fall stagnant and do not
want to pay for an office when they can work from home.7 As such, startups leave the
co-working hub in two ways: either by not renewing their membership or by outgrowing their
office space. The vacant office spaces are then assigned to startups based off a wait-list.8
Startups on the wait-list are prioritized as follows: technology startups over service providers,
and local vs. non local startups.

The layout of the floors we examine (floors two - five), is depicted in Figure 1.9 We
measure the distance between rooms from available floor plans using space syntax software
(Bafna, 2003; Kabo et al., 2014, 2015).10 One useful feature of space syntax software is
that it calculates distances between rooms as people would walk rather than the shortest
7
Outgrowing the office space is a celebrated event at the co-working hub akin to a graduation. During
the time covered by our data, only eight startups moved out because they “graduated” from (outgrew) the
building. While outside options for these startups surely exist, we can interpret our estimates as causal
conditional on remaining in the co-working space.
8
One threat to our assumption of exogeneity is the possibility that some startups may wish to remain on
the wait-list in hopes of securing a space they believe to be “better.” We do not detect this phenomenon in
our data nor did the co-working space administrators observe this taking place.
9
We exclude the ground level since the work space on this floor is a) open space and b) the work stations
are allocated to individuals and not complete startup entities (so called “hotdesks”).
10
Using this software, distance is measured by steps. One step is the equivalent of roughly 1.42m.

10
euclidian distance on a plane or “as the crow flies”. For each room dyad we calculate the
shortest walking distance. The variable Close is an indicator equal to one if the shortest
distance between startupi and startupj located on the same floor is within 20 meters; the 25th
percentile of pair-wise distances between all rooms).11 We flag dyads for whom the shortest
paths between rooms directly pass through a common area (Common Area). Common areas
are the kitchens and zones in front of the elevator on each floor as well as the open sitting
space on the second floor.

<Insert Figure 1 here>

Our main outcome variable of interest is new web technology adoption, which serves as our
proxy for knowledge spillovers (Fang et al., 2020). Prior studies have predicted a nascent
firm’s inherent propensity to adopt as a function of organizational factors and traits such
as size, structure, and resources (Fichman, 2004) and highlights that, especially for new
web-tech based ventures, technology choice is a fundamental decision (Kapoor and Furr,
2015) as it sets the building block(s) for the future. To construct this variable, we exploit a
novel data set (www.builtwith.com), covering over 25,000 web technologies (e.g., analytics,
advertising, hosting, and CMS) that tracks how technology usage of startups change on a
weekly basis (Koning et al., 2019). Builtwith is used by large and small companies alike to
learn about the adoption of software components used to build web applications. The set of
elements used to develop a web applications are colloquially known as a “technology stack”
(and often shortened to “tech stack”). In the Appendix Table A3, we provide examples of
the “tech stack” corresponding to three startups in our sample. As the table displays, there
is much variation between startups in terms of technology categories used, but also variation
of software components used within those categories.

From this website we collect information on the web technology usage of the startups in
our sample, including the exact date of implementation and abandonment. Web technologies
11
For a summary and description of all variables used in the dyadic model, please refer to Table A2 of the
Appendix.

11
are the markup languages and multimedia packages computers use to communicate and can
be thought of as tools at a startup’s disposition to ensure the functionality and efficiency
of their websites. Functionalities include interacting with users, connecting to back-end
databases, and generating results to browsers, which are updated continuously. When
choosing web technologies and “tech stacks” there are different aspects developers need to
consider. These are, e.g., the type of project, the team’s expertise and knowledge base, time
to market, scalability, maintainability, and overall cost of development. As an example, in
the subcategory of the Analytics and Tracking category, Error Tracking, at the time of our
study, the three most prominent technologies were Rollbar (used by Salesforce, Uber, and
Kayak), Bugsnag (used by Airbnb, Lyft, and Mailchimp), and Honeybadger (used by Ebay,
Digitalocean, and Heroku). Each technology has their unique advantages and disadvantages,
that may only become apparent after learning about peers’ experience using them. Similarly,
peers can share their experience applying other tools or combinations, specifically in terms of
if there was a notable boost in user attraction, conversion, sales, functionality, security or
efficiency in running the website. These aspects do not necessarily become palpable until
implemented on the website, but have implications that span across various layers of the
startup, including HR, finance, marketing, and management. Since implementation entails
costs associated with labor, user turnover and embeddedness with other existing technologies
reducing these types of frictions should come at the benefit of the startup.12

We construct two measures for technology adoption. The first is the number of technologies
startupi adopts from startupj (ln(AdoptCountij + 1)). An adopted technology is a technology
used by startupi in the focal period that startupi had not implemented in any previous period,
but startupj had already put to use. The second measure is 1(AdoptT echij ), which equals
one if startupi adopts a technology from startupj . The control variable Pre-period Technology
Overlap corresponds to the percentage of technologies startupi has adopted from startupj
12
In Figure A1, we present a histogram of the distribution of the number of technologies used by each
startup.

12
before both of the two startups are active at the co-working hub. We include this variable in
order to control, as far as possible, for the fact that some technologies may be adopted as
packages.

For each of the startups, we conducted extensive web-searches to find detailed information
regarding startups’ characteristics, such as industry and business models. For industry classi-
fication, we follow the industry categories found on AngelList (angellist.com) and BuiltWith.
The individual industries are Administration&Management, Data, Design&Development,
Digital, Education, Energy&Construction, Entertainment, Finance&Legal, Healthcare, Mar-
keting&PR, Real Estate, Retail, Science&Technology, Security, and Software&Hardware. For
our analyses we use each venture’s primary industry (the most prominent on their websites),
since many operate in more than one. The variable Same Industry equals one if startupi and
startupj operate in the same primary industry. Similarly, the variables Both B2B Companies
and Both B2C Companies indicate if startupi ’s and startupj ’s main customers are other
businesses (B2B) or individual consumers (B2C).13

We additionally identified startup age as a startup’s tenure at the co-working hub and
the gender composition of startups using information provided by the co-working space. As
derived from the entry date into the co-working space, |agei -agej | reflects the absolute value
of the age difference between startupi and startupj . The variable Both Majority Female flags
startup dyads where team members in both startupi and startupj are predominately female
(over 50 percent female). We have additional information on the CEOs/heads of each startup,
which we use to identify whether a startup is led by a woman (Female CEO) or not. We
determined the gender of founders conducting extensive web searches on the startups as well
as by comparing first names with lists provided by the US Census for most common names
13
We recognize that firms that operate in the same industry or that focus on the same customer type may
potentially operate quite differently and employ distinct business models. As such, we may not entirely capture
the level of competition between dyad members in the same industry. However, we are still reassured to observe
meaningful co-variation between technology adoption and operating in the same industry. Consequently,
we should view these effects as lower bounds given the potential measurement error (which thus introduces
attenuation bias) making it more difficult to detect an effect.

13
by sex.14

To capture differences in performance outcomes, we construct two measures using infor-

mation provided by the co-working space and AngelList. These two outcomes are based on
prior literature (Nanda and Sørensen, 2010; Ewens and Marx, 2017) and capture financial
performance of startups. One is raising a seed round, and the other is raising financial capital
in excess of US$ 1 million.

We further exploit a joint-event hosted at the co-working space on a weekly basis to analyze
the impact of proximity on the propensity of the entrepreneurs in our sample to interact.
This joint event is a lunch (open to the public; the price for non-members is $10) organized
by the co-working space every Friday at noon. The average number of people who attend
the lunch is approximately 250 every week. This shared meal is intended to give members
the opportunity to “network with other startups” and to “meet, greet and chowdown.” The
co-working space keeps track of the exact order individuals (both members and non-members)
enter to attend the lunch. For a period of time (January 2016 - December 2016), we identify
the number of lunches hosted at the co-working space that at least one team member of
startupi and startupj both attend (# Event Bothij Attend ). The average is 0.27. We further
exploit the order of entry to create an indicator equal to one if at least one team member of
startupi and startupj appear within 1, 2, 5, 10, or 25 people in line for the lunch (1(Ever
within X people in line)).

3.3 Descriptive Statistics

As displayed in Table 2, on average, each startup is at risk of spillovers from 53 other
startups. The average distance between room dyads is approximately 32 meters and the
average room size is ca. 27 m2 (288 sq.feet). Twenty-eight percent of the rooms (by floor)
are located close to each other and 38 percent of the shortest paths between two rooms pass
through a common area. Of the 251 startups, 12 percent are predominately female and 24
14
https://www2.census.gov/topics/genealogy/1990surnames

14
percent are considered to be successful startups. On average, the startups in our sample have
been at the co-working space for approximately one year. The use of web technologies is
highly skewed, ranging from a minimum of 0 to a maximum of 255. In Table 2, the variable
Min. Technology Usage (Max. Technology Usage) displays the minimum (maximum) amount
of technologies a startup ever hosted while at the co-working space. Over time, the startups
in our sample adopt about 7.33 technologies on average, 53 percent adopt at least one new
technology.

The main focus of our analyses is on startup dyads. A key component is thereby the
characteristics both startups have in common. Of the startup-dyads in the co-working hub, 11
percent operate in the same industry, 48 and 11 percent both have a B2B and B2C business
model respectively. The percentage of startup-dyads where the majority of team members
are female is 1.3 percent (N = 138), and eight percent of the startup-dyads are considered
successful. The average age difference between startups in a dyad is 7.30 months.

4 Results
For the purpose of this study, we operationalize the distinct proximity dimensions as follows.
Physical Proximity is measured using the geographic distance (in meters) between rooms on
one floor. Social Proximity captures when both startups possess a salient characteristic that
only a minority of the startups in the co-working space have. We identify socially proximate
startups as those where both startups are majority female. We measure Knowledge-Space
Similiarity using the pre-period technology tech-stack overlap between focal startupi and
startupj . The indicator is equal to one if startupi and startupj have high overlap in their
pre-period tech-stack (above the 75th percentile). In this paper, Productx-Market Proximity
captures when the consumers of two startups’ products are similar. We measure product-
market proximity by using a combination of two startup characteristics: a) industry, and b)
business model. Two startups are proximate in their product-market if they either operate in
the same industry or have the same business model.

15
4.1 Baseline Results: Physical Proximity
Table 3 presents the results from assessing the effect of distance on the amount of peer
technology adoption (ln(AdoptCountij + 1)) using a standard OLS model and using a linear
probability model to estimate the likelihood of adopting a technology from a peer startup
1(AdoptTechij ). In the full model (Columns 2 and 4), using startup-x-room fixed effects
and controlling for industry, business model, gender, age and pre-period technology overlap,
we find that the doubling of distance between two dyads reduces both the amount of peer
technology adoption by 3.5% and the likelihood of any peer technology adoption by 1.7%,
with both point estimates significant at the 1% level. As seen, the magnitude and statistical
significance of the effect remains largely unchanged with the inclusion of additional controls.15

<Insert Table 3 here>

We next loosen the (log)linearity assumption of distance on technology adoption by breaking

our distance measure into quartiles and estimate equation (1) using these indicators rather
than the continuous measure of distance. Figure 2 displays these regression results graphically.
We construct our omitted category as startups that are on different floors allowing us to
estimate the full set of (same-floor) distance quartiles. The results obtained from this approach
suggest that startup startups located within 20 meters of each other are those most influenced
by each other. Being more distant, however, greatly reduces the influence of peers. Put
differently, for technology adoption influence, startup pairs that are not within 20 meters of
each other on the same floor behave as if they were on different floors altogether.

<Insert Figure 2 here>

Having identified that the distance effect is strongest for the most proximate startups,
we create an indicator equal to one (Close) that flags dyads located within 20 meters of
15
Please refer to Tables A4 and A5 of the Appendix for models excluding controls.

16
each other (and equal to zero for all other dyads) and use this measure for the remainder of
our results. In Table 3, Columns 5-8, we display our findings from estimating equation (1)
using this more nuanced classification of distance. The results indicate that close proximity
positively influences the likelihood of adopting an upstream (production) technology also used
by a peer startup. We find that being in close proximity is associated with a three percentage
point higher probability of adopting a peer technology ( = 0.025, dyad and floor-neighborhood
cluster-robust standard errors 0.011). This finding remains robust to including different
covariates. As displayed in Columns 5 and 6, applying an OLS model and estimating the
count of adopted peer technologies (ln(AdoptCountij + 1)) provides a similar result. In
the full model (Column 6), the point estimate on the coefficient for close proximity is 0.048
(cluster-robust standard errors 0.015). This implies that a switch to a room in close proximity
would translate into a five percent increase in the number of peer technologies adopted from
the mean.

For robustness and to ensure that the results we present are not due to spurious correlations,
we utilize a randomization inference method suggested by Athey and Imbens (2017) and
Young (2019) using a Monte Carlo simulation (1,000 runs). In this simulation, we randomly
draw closeness (with replacement) for each dyad and then estimate the likelihood of adopting
a technology as a function of this random closeness variable. The placebo treatment effect
results attained from the simulation are presented in Figure 3.16 In line with our findings,
only 2 of the simulated Monte Carlo draws (from 1,000) had a coefficient greater than the
point estimate of our main results (=0.022), resulting in a randomized inference p-value
of 0.002 - strongly rejecting our null of no relationship between proximity and technology
adoption.

<Insert Figure 3 here>

16
As expected from this randomization exercise, the mean correlation is close to 0, and 5% of the results
were significant at the 5% level.

17
 An additional feature of the physical layout of the office space are common areas provided
by the co-working space, such as kitchens on each floor. To examine the extent to which
common areas may help extend the effect of proximity and the precise spatial distances this
applies to, we again break our distance measure into quartiles (recall that Close corresponds
to the first quartile) and interact these quartiles with the CommonArea dummy (using
CommonArea × 4th distance quartile as the omitted category).17 The results are displayed
in Figure 4, which reveals two things. First, being close (first quartile of distance) to a startup
increases technology adoption likelihood independent of whether or not the two startups
pass through a common area. Second, and more interestingly, the likelihood of technology
adoption for a peer in the second quartile (between 21 and 30 meters apart) also is greater
but this effect only activates for startup dyads that pass through a common area. In other
words, it appears that these common areas extend the co-location premium to startups that
are more distant from one another.

<Insert Figure 4 here>

4.2 Interplay of physical proximity with other proximity dimen-

sions
We now turn to the results on the interplay between physical proximity and other proximity
dimensions.

4.2.1 Interplay with social proximity

We first examine how social proximity – the gender composition of the startup dyads –
may influence the effect of physical proximity on peer technology adoption. In the case of
our setting, female startups represent a minority group. As suggested by Reagans (2011),
17
Please refer to Table A6 of the Appendix for the results from estimating equation (1) including a variable
equal to one that indicates if the shortest path between startupi and startupj is across a common area
(Common Area). As shown, common area overlap is associated with a higher likelihood of technology adoption.
The interaction of common area overlap with an indicator equal to one if startups are located within 20
meters from each other (Close) is negative, yet not statistically significant (p-value>0.1).

18
demographic characteristics that define minority status are more likely to be salient. Salience
is important because entities are more likely to identify with a salient characteristic, and
identification with a characteristic generates positive affect for in-group members (Hogg and
Turner, 1985; Grieve and Hogg, 1999). As shown in Table 4, Column 1, we find that dyads
where both startups are predominately female overcome the distance discount suggesting
that these startups rely on alternate mechanisms to overcome the negative effects of distance
or, as a minority within the co-working space, may have different networking behavior (Kerr
and Kerr, 2018).

4.2.2 Interplay with product-market proximity

In Table 4, Column 2, we present the results including an interaction of physical and
product-market proximity in order to gauge the role of competition-based dynamics. The
main effect of physical proximity, Close – which reflects the benefits of proximity for startup
dyads in different product-markets – increases the likelihood of peer technology adoption by
3.7%. The interaction between product-market and physical proximity, however is negative
and reduces the aforementioned proximity benefits by 2.3 percentage points (or over 60%
of the total effect, 2.3/3.7). This indicates that physical and product-market proximity
are substitutes and that being physically close is most beneficial to startup dyads that are
dissimilar.

<Insert Table 4 here>

4.2.3 Interplay with knowledge-space similarity

In Table 4, Column 3, we present the results including an interaction of physical and
knowledge-space proximity in order to evaluate the role of information-based dynamics.
For simplicity, we count a dyad as similar along the knowledge-space dimension if their
pre-period technology overlap is over 0.27.18 As seen earlier across a number of other
proximity dimensions, the interaction between technology overlap/similarity and physical
18
The 75th percentile of this variables distribution

19
proximity is negative, implying that being physically close is less valuable for startups that
are already proximate in knowledge/technology space. We omit our pre-period technology
overlap measure in Column 3 as it is highly correlated with the knowledge-space similarity
measure.

4.2.4 Interplay with diversity

Thus far, the results suggest that proximity along non-geographic dimensions may substitute
for being physically close. This points to possible advantages of co-location for facilitating
knowledge spillovers among startups that are otherwise dissimilar. To test this, we create
a composite variable called Diverse that is equal to one if a startup dyad differs along the
social, product-market, and knowledge space dimensions, 0 otherwise. As displayed in Table
4, Column 4, we find that being physically close matters most for knowledge exchange that
leads to integration of new technologies among otherwise distant startups. This may indicate
that the advantages of close physical proximity lie in supporting more exploratory search by
better enabling access to different and non-obvious sources of knowledge (Fleming, 2001). In
contrast to the exploitation of more proximate knowledge, the exploration of new information
– an important feature of innovation – typically entails substantial search costs (especially
with regard to speed), risk taking, and experimentation (March, 1991). Shorter distances
and more immediate feedback may reduce such barriers to both more efficiently transmit and
adopt distant knowledge.

4.3 The role of social interactions

One potential explanation for our previous set of results is that physical proximity shapes
the social interactions of individuals (Battiston et al., 2020; Hasan and Bagde, 2015; Allen,
1977; Lane et al., 2020). To explore the likelihood of this mechanism in the co-working hub
context, we further exploit a joint event – a lunch – hosted at the co-working space on a
weekly basis. Table 5, Columns 1 and 2, present the results using the number of lunches (#
Event) hosted at the co-working space that at least one team member of startupi and startupj

20
both attend (Bothij Attend ). Columns 3 and 4 present the results using an indicator equal
to one if both ever attended one together. Since common areas seem to extend the effect of
proximity, we include this variable in our model. The main result reinforces a result shown
throughout: proximity matters. Startup dyads that are within 20 meters are more likely to
attend a lunch together and attend more lunches together than dyads that are further apart.
Passing through a common area also increases the likelihood of jointly socializing. Further,
startups that are different are less likely to socialize, i.e., jointly attend these events together,
yet being close has no differential impact on socializing for startups that are different. In
other words, the extent to which a startup is different from the focal startup has no bearing
on the likelihood of socializing when they are both close.

We further provide evidence for the effect of proximity on socializing by exploring the
extent to which the two startups went to the event together. To do so we create an indicator
equal to one if at least one team member of startupi and startupj appears within five people
in the check-in line for the event (1(within 5 people in line)).19 We present the results from
estimating the effect of room proximity on check-in line proximity in Columns 5-6, Table 5.
Similar to our results using the number of events both attended, we see a positive impact
of close room proximity on checking-in together. Here, however, we observe homophilous
behavior, wherein startup pairs that are different/diverse are less likely to attend the event
together. While we do not want to overstate this result, given the interaction’s marginal
significance (at conventional levels), we do want to draw attention to this discrepancy and
seemingly contradictory finding: startups that are close and different receive more knowledge
spillovers from each other yet startups that are close are less likely to socialize (co-attend
events) with startups different from them. Of greatest interest then, is to examine the impact
of diverse proximity on knowledge spillovers when the startups do socialize. We explore this
19
In the Appendix, Table A7, we further create indicators equal to one if at least one team member of
startupi and startupj appear within 1, 2, 5, 10, or 25 people in line for the lunch (1(Ever within X people in
line)). The results indicate that close room proximity (within 20 meters) only increases check-in line proximity
for the group of people within 1-5 individuals from each other at check-in and not for those individuals further
away in line.

21
next.

<Insert Table 5 here>

4.4 Proximity, socializing, and diversity

We next combine physical proximity, socializing, and diversity and examine their joint
relationship with technology adoption. As in earlier tables, the outcome 1(AdoptTechij )
equals one if startupi adopted at least one new technology from startupj . Close equals to
one if startupi and startupj are located within 20 meters (14 steps; the 25th percentile of
pair-wise distances between all rooms) of each other on the same floor. The variable (#
Event Bothij Attend equals one if least one team member of startupi and startupj both
attend a lunch hosted at the co-working space. The indicator 1(within 5 people in line)
equals to one if at least one team member of startupi and startupj appear within 5 people
in line for the lunch. Diverse is an indicator equal to one if the startup dyads differ along all
non-geographic proximity dimensions we in examine and equal to zero otherwise. We control
for age differences, pre-period technology overlap, and the passing through a common area
en route between startupi and startupi . We include startupi x room fixed effects. Standard
errors are robust to dyadic clustering at the floor-neighborhood level. As displayed in Table
6, Column 1, social activity – measured by number of mutually-attended events and check-in
line proximity – predicts technology adoption alongside physical proximity. In Column 2, we
present the result of interacting our measure of social activity with our measure for diversity.
The coefficient suggests that although diversity alone does not predict technology adoption
(as was also shown in Table 4 Column 4), the more socializing diverse startup dyads engage
in, the greater the likelihood of technology adoption.

Next, we form all pair-wise combinations of our proximity and diverse measures in order
to more effectively evaluate their combined effect. These are 1) far and similar (Close=0 &
Diverse =0 ); 2) far and different (Close=0 & Diverse =1 ); 3) close and similar (Close=1 &
Diverse =0 ); and 4) close and different (Close=1 & Diverse =1 ). As displayed in Column 3

22
(similar and far serving as the omitted category), technology adoption is especially strong
among dyads that are close and different, even when controlling for social activity. In Column
4, we examine how dyad properties amplify the benefits of socializing. Dyads that socialize,
are in close physical proximity, and are different, experience that largest boost to technology
adoption particularly relative to those dyads that are similar.

<Insert Table 6 here>

4.5 Performance
The notion that peers influence performance has been demonstrated in a host of different
environments such as retail (Chan et al., 2014a,b), finance (Hwang et al., 2019) and science
(Oettl, 2012; Catalini, 2018). The idea being that sharing knowledge, helping, and setting
expectations (e.g., Mas and Moretti 2009; Herbst and Mas 2015; Housman and Minor 2016)
enhances performance. Moreover, the broader innovation literature stresses the importance
of external knowledge in promoting innovation and performance (Cohen and Levinthal, 1990;
Chesbrough, 2012). External knowledge introduces novelty with respect to the knowledge
available inside a startup (Zahra and George, 2002; Laursen and Salter, 2006), and access to
information from a wide range of skills and experiences aids in maximizing a group’s capacity
for creativity, knowledge-generation, and effective action (Reagans and Zuckerman, 2001;
Aggarwal et al., 2020). Diversity of external knowledge sources (in our case peer startups)
thereby increases the amount of novel information pieces a startup has access to.

To provide more insight into the potential role of the immediate environment for startup
performance, we move our analysis away from the startup-dyad level and aggregate to
the startup-level. We then estimate the probability of achieving two important startup
performance milestones as a function of the diversity of the micro-environment (startups
located within 20m of each other) and the extent to which startups engage in social events.
Following prior literature, we use indicators identifying startups that raise seed funding and
raise funding in excess of US$ million as measures for new venture financial performance

23
(e.g., Hochberg et al. 2007; Nanda and Rhodes-Kropf 2013).

In Figure 5, we display results from estimating the relationship between the likelihood
of raising a seed round and raising funding in excess of US$ million as a function of the
aggregate diversity indicator of startups within 20 meters of the focal startup interacted
with an indicator equal to one if the focal startup engages in the lunches hosted at the
co-working space (Social=1 ). We thereby control for the following startup characteristics:
size, gender, remoteness20 of the location and age. We further include floor fixed effects and
cluster standard errors on the floor-neighborhood level. We break our diversity measure into
quintiles and the plot the corresponding coefficients (with 95% confidence intervals). Results
suggest that startups located within a balanced environment (middle level of diversity) and
that engage in social activity are most likely to receive seed funding and funding in excess
of US$ 1 million. The corresponding regression results can be found in Appendix Table A8.
This highlights the importance of not only bringing people together, but socializing with each
other for promoting better startup performance outcomes. Moreover, our results provide
suggestive evidence that striking a balance between diversity and similarity is especially
crucial.

<Insert Figure 5 here>

5 Discussion and Conclusions

For Katalin Karikó, whose research was critical for creating mRNA-based tharapeutics that
do not induce an antiviral immune response, there was a long period when it seemed that her
research on messenger RNA would never get funding. Her approach was so different from that
of colleagues she struggled to find support. An encounter at the copy machine – a common
area – with Drew Weissman brought a new perspective and idea for a potential application
laying the foundation for the successful COVID-19 vaccines made by Pfizer-BioNTech and
Moderna.
20 1
P
We calculate Remotenessi = N j distanceij to control for the general location of a startup.

24
 Stories like these have executives pondering what the future of work entails in balancing
the flexibility and productivity enhancing benefits of working from home with the creativity-
generating potential of serendipitous encounters that are most commonly formed via face-to-
face interactions. We contribute to this discussion in three important ways by examining
how physical environments provide knowledge spillovers at the micro-geographic level for
knowledge workers and entrepreneurs. First, our findings indicate that knowledge spillovers,
and more specifically the type that lead to the integration of external knowledge, occur at
very short distances. We show that in one of the largest entrepreneurial co-working spaces in
the US, startups are influenced by peer startups that are within a distance of 20 meters and
no longer at greater distances – even if they are located on the same floor. While the focus of
our study has been on deepening our understanding of inter-startup knowledge spillovers, the
same mechanisms can be conceptually extended to examine within-organizational knowledge
spillovers as in Allen (1977).

Second, we contribute to the literature examining physical proximity and knowledge

exchange by incorporating additional dimensions of similarity/diversity and examining their
interdependencies. In doing so, we find support for the idea that particularly the integration
of external, diverse knowledge is facilitated through physical proximity. We thereby provide
evidence for heterogeneity in the effect of physical distance on knowledge integration depending
on similarity along other dimensions, highlight the importance of engaging in social activities,
and directly respond to the call for a better understanding of structures and processes adopted
by startups to facilitate or impede learning (Alcácer and Oxley, 2014). This finding not only
presents a possible avenue to reconcile Marshall-Arrow-Romer specialization externalities
(Romer, 1986) and Jacobs-style diversification externalities (Jacobs, 1969), but also may
serve as guidance in the design of workplaces that promote knowledge exchange between
non-collaborating entities – may they be research groups, teams or startups.

Third, we provide insight on how micro-environments can be leveraged to enhance startup

25
performance. Our findings suggest that environments that strike a balance between diversity
and similarity can contribute to achieving important startup milestones. However, our results
suggest an important caveat. This boost to performance only occurs if startups socially
engage with their environment.

We acknowledge that our paper is not without limitations. For one, we restrict our analysis
to only one co-working space. In this case we are trading-off a higher level of generalizability
for richer data. Furthermore, the sample of startups we observe are primarily digital and
web-based. These are the types of nascent startups that may benefit the most from integrating
new knowledge. However, both in terms of current startup industry trends and technology
sophistication, the findings we present should nonetheless be fairly representative for the
population of startups working in similar co-working spaces around the world. Furthermore,
we restrict our focus to one type of decision entrepreneurs make as a proxy for knowledge
integration: web technology adoption. We use this measure since, on the one hand, choices
regarding the technology of a startup are especially fundamental for startups (Murray and
Tripsas, 2004), and on the other hand, because we can clearly identify the time these changes
were implemented and the technology was integrated into a startups tech stack.

Taken together, our findings provide fundamental insights for the design of workplaces that
support knowledge production, entrepreneurship, and innovation. We highlight important
trade-offs and stress that understanding which startups and how they respond to their peers
matters for creating effective environments for early stage ventures. Where physical structure
may lay the groundwork for exchange to take place, other factors may determine who benefits
more from presented opportunities.

References
Aggarwal, V. A., Hsu, D. H. and Wu, A. (2020), ‘Organizing knowledge production teams within
firms for innovation’, Strategy Science 5(1), 1–16.
Alcácer, J., Dezső, C. and Zhao, M. (2015), ‘Location Choices under Strategic Interactions’, Strategic
Management Journal 36(2), 197–215.
Alcácer, J. and Oxley, J. (2014), ‘Learning by Supplying’, Strategic Management Journal 35(2), 204–

26
 223.
Allen, T. J. (1977), Managing the flow of technology: Technology transfer and the dissemination of
technological information within the R&D organization, Cambridge: MIT Press.
Angst, C. M., Agarwal, R., Sambamurthy, V. and Kelley, K. (2010), ‘Social contagion and information
technology diffusion: The adoption of electronic medical records in u.s. hospitals’, Management
Science 56(8), 1219–1241.
Aronow, P. M., Samii, C. and Assenova, V. A. (2017), ‘Cluster–Robust Variance Estimation for
Dyadic Data’, Political Analysis 23(4), 564–577.
Arthur, W. B. (1994), Increasing Returns and Path Dependence in the Economy, University of
Michigan Press.
Arzaghi, M. and Henderson, J. V. (2008), ‘Networking off Madison Avenue’, Review of Economic
Studies 75(4), 1011–1038.
Athey, S. and Imbens, G. W. (2017), The Econometrics of Randomized Experiments, Vol. 1,
Amsterdan: Elsevier, pp. 73–140.
Azoulay, P., Fons-Rosen, C. and Graff Zivin, J. S. (2019), ‘Does science advance one funeral at a
time?’, American Economic Review 109(8), 2889–2920.
Bafna, S. (2003), ‘Space Syntax: A Brief Introduction to Its Logic and Analytical Techniques’,
Environment and Behavior 35(1), 17–29.
Baker, W. E. (1984), ‘The social structure of a national securities market’, American Journal of
Sociology 89(4), 775–811.
Battiston, D., Blanes i Vidal, J. and Kirchmaier, T. (2020), ‘Face-to-Face Communication in
Organizations’, The Review of Economic Studies pp. 1–69.
Blau, P. M. (1977), ‘A Macrosociological Theory of Social Structure’, American Journal of Sociology
83(1), 26–54.
Borgatti, S. P. and Cross, R. (2003), ‘A relational view of information seeking and learning in social
networks’, Management Science 49(4), 432–445.
URL: http://www.jstor.org/stable/4133949
Boudreau, K. J., Brady, T., Ganguli, I., Gaule, P., Guinan, E., Hollenberg, A. and Lakhani, K. R.
(2017), ‘A Field Experiment on Search Costs and the Formation of Scientific Collaborations’, The
Review of Economics and Statistics 99(4), 565–576.
Burt, R. S. (2004), ‘Structural Holes and Good Ideas’, American Journal of Sociology 110(2), 349–
399.
Cameron, A. C. and Miller, D. L. (2014), Robust Inference for Dyadic Data, Unpublished manuscript,
University of California-Davis.
Carayol, N., Bergé, L., Cassi, L. and Roux, P. (2019), ‘Unintended Triadic Closure in Social
Networks: The Strategic Formation of Research Collaborations between French Inventors’,
Journal of Economic Behavior & Organization 163, 218–238.
Carrell, S. E., Sacerdote, B. I. and West, J. E. (2013), ‘From Natural Variation to Optimal Policy?
The Importance of Endogenous Peer Group Formation’, Econometrica 81(3), 855–882.
Catalini, C. (2018), ‘Microgeography and the Direction of Inventive Activity’, Management Science
64(9), 4348–4364.
Chan, T. Y., Li, J. and Pierce, L. (2014a), ‘Compensation and Peer Effects in Competing Sales
Teams’, Management Science 60(8), 1965–1984.
Chan, T. Y., Li, J. and Pierce, L. (2014b), ‘Learning from Peers: Knowledge Transfer and Sales
Force Productivity Growth’, Marketing Science 33(4), 463–484.
Chatterji, A., Delecourt, S., Hasan, S. and Koning, R. (2019), ‘When Does Advice Impact Startup
Performance?’, Strategic Management Journal 40(3), 331–356.

27
Chesbrough, H. (2012), ‘Open innovation where we’ve been and where we’re going’, Research
Technology Management 55(4), 20–27.
Cohen, S. L., Bingham, C. B. and Hallen, B. L. (2019), ‘The Role of Accelerator Designs in Mitigating
Bounded Rationality in New Ventures’, Administrative Science Quarterly 64(4), 810–854.
Cohen, W. M. and Levinthal, D. A. (1990), ‘Absorptive Capacity: A New Perspective on Learning
and Innovation’, Administrative Science Quarterly pp. 128–152.
Conley, T. G. and Udry, C. R. (2010), ‘Learning about a new technology: Pineapple in ghana’,
American Economic Review 100(1), 35–69.
Conti, A., Peukert, C. and Roche, M. (2021), Beefing IT Up for Your Investor? Open Sourcing and
Startup Funding: Evidence from Github, Working Paper 22-001, Harvard Business School.
Cowgill, B., Wolfers, J. and Zitzewitz, E. (2009), Using Prediction Markets to Track Information
Flows: Evidence from Google, Berlin: Springer.
Cyert, R. M., March, J. G. et al. (1963), ‘A Behavioral Theory of the Firm’, Englewood Cliffs:
Prentice Hall 2(4), 169–187.
Dingler, A. and Enkel, E. (2016), ‘Socialization and Innovation: Insights from Collaboration across
Industry Boundaries’, Technological Forecasting and Social Change 109, 50–60.
Elfenbein, D. W., Hamilton, B. H. and Zenger, T. R. (2010), ‘The Small Firm Effect and the
Entrepreneurial Spawning of Scientists and Engineers’, Management Science 56(4), 659–681.
Ewens, M. and Marx, M. (2017), ‘Founder Replacement and Startup Performance’, SSRN .
Fang, T. P., Wu, A. and Clough, D. R. (2020), ‘Platform diffusion at temporary gatherings: Social
coordination and ecosystem emergence’, Strategic Management Journal pp. 1–40.
Fichman, R. G. (2004), ‘Going beyond the dominant paradigm for information technology innovation
research: Emerging concepts and methods’, Journal of the association for information systems
5(8), 11.
Fleming, L. (2001), ‘Recombinant Uncertainty in Technological Search’, Management Science
47(1), 117–132.
Gans, J. S., Stern, S. and Wu, J. (2019), ‘Foundations of Entrepreneurial Strategy’, Strategic
Management Journal 40(5), 736–756.
Gaspar, J. and Glaeser, E. L. (1998), ‘Information Technology and the Future of Cities’, Journal of
Urban Economics 43(1), 136–156.
Gavetti, G. and Levinthal, D. (2000), ‘Looking Forward and Looking Backward: Cognitive and
Experiential Search’, Administrative Science Quarterly 45(1), 113–137.
Glaeser, E. L., Kerr, S. P. and Kerr, W. R. (2015), ‘Entrepreneurship and Urban Growth: An
Empirical Assessment with Historical Mines’, The Review of Economics and Statistics 97(2), 498–
520.
Granovetter, M. S. (1973), ‘The Strength of Weak Ties’, American Journal of Sociology 78(6), 1360–
1380.
Grieve, P. G. and Hogg, M. A. (1999), ‘Subjective Uncertainty and Intergroup Discrimination in
the Minimal Group Situation’, Personality and Social Psychology Bulletin 25(8), 926–940.
Harmon, N., Fisman, R. and Kamenica, E. (2019), ‘Peer effects in legislative voting’, American
Economic Journal: Applied Economics 11(4), 156–80.
Hasan, S. and Bagde, S. (2015), ‘Peers and Network Growth: Evidence from a Natural Experiment’,
Management Science 61(10), 2536–2547.
Hasan, S. and Koning, R. (2019), ‘Prior Ties and the Limits of Peer Effects on Startup Team
Performance’, Strategic Management Journal 40(9), 1394–1416.
Hassan, T. A. and Mertens, T. M. (2017), ‘The Social Cost of Near-Rational Investment’, American
Economic Review 107(4), 1059–1103.

28
Herbst, D. and Mas, A. (2015), ‘Peer Effects on Worker Output in the Laboratory Generalize to the
Field’, Science 350(6260), 545–549.
Hochberg, Y. V., Ljungqvist, A. and Lu, Y. (2007), ‘Whom You Know Matters: Venture Capital
Networks and Investment Performance’, The Journal of Finance 62(1), 251–301.
Hogg, M. A. and Turner, J. C. (1985), ‘Interpersonal Attraction, Social Identification and Psycho-
logical Group Formation’, European Journal of Social Psychology 15(1), 51–66.
Housman, M. and Minor, D. B. (2016), Workplace Design: The Good, the Bad and the Productive,
Working Paper 16-147, Harvard Business School, Working Paper.
Howell, T. (2022), ‘Coworking spaces: An overview and research agenda’, Research Policy
51(2), 104447.
Hwang, B.-H., Liberti, J. M. and Sturgess, J. (2019), ‘Information Sharing and Spillovers: Evidence
from Financial Analysts’, Management Science 65(8), 3624–3636.
Jacobs, J. (1969), The Economy of Cities, New York: Vintage Books.
Kabo, F., Hwang, Y., Levenstein, M. and Owen-Smith, J. (2015), ‘Shared paths to the lab: A
sociospatial network analysis of collaboration’, Environment and Behavior 47(1), 57–84.
Kabo, F. W., Cotton-Nessler, N., Hwang, Y., Levenstein, M. C. and Owen-Smith, J. (2014),
‘Proximity effects on the dynamics and outcomes of scientific collaborations’, Research Policy
43(9), 1469–1485.
Kaplan, S. N., Sensoy, B. A. and Stromberg, P. E. R. (2009), ‘Should Investors Bet on the Jockey or
the Horse? Evidence from the Evolution of Firms from Early Business Plans to Public Companies’,
The Journal of Finance 64(1), 75–115.
Kapoor, R. and Furr, N. R. (2015), ‘Complementarities and competition: Unpacking the drivers of
entrants’ technology choices in the solar photovoltaic industry’, Strategic Management Journal
36(3), 416–436.
Kerr, S. P. and Kerr, W. R. (2018), Immigrant Networking and Collaboration: Survey Evidence
from CIC, Chicago: University of Chicago Press.
Kerr, W. R. and Kominers, S. D. (2015), ‘Agglomerative Forces and Cluster Shapes’, Review of
Economics and Statistics 97(4), 877–899.
Koning, R., Hasan, S. and Chatterji, A. (2019), Experimentation and Startup Performance: Evidence
from A/B testing, Working Paper 26278, National Bureau of Economic Research.
URL: http://www.nber.org/papers/w26278
Lane, J. N., Ganguli, I., Gaule, P., Guinan, E. and Lakhani, K. R. (2020), ‘Engineering Serendipity:
When does knowledge sharing lead to knowledge production?’, Strategic Management Journal .
Laursen, K. and Salter, A. (2006), ‘Open for innovation: the role of openness in explaining innovation
performance among uk manufacturing firms’, Strategic management journal 27(2), 131–150.
Lee, S. (2019), ‘Learning-by-Moving: Can reconfiguring spatial proximity between organizational
members promote individual-level exploration?’, Organization Science 30(3), 467–488.
Lerner, J. and Malmendier, U. (2013), ‘With a Little Help from My (Random) Friends: Success and
Failure in Post-Business School Entrepreneurship’, The Review of Financial Studies 26(10), 2411–
2452.
Lippman, S. A. and McCall, J. J. (1976), ‘The Economics of Job Search: A Survey’, Economic
Inquiry 14(2), 155–189.
Lyons, E. and Zhang, L. (2018), ‘Who does (not) benefit from entrepreneurship programs?’, Strategic
Management Journal 39(1), 85–112.
Manski, C. F. (1993), ‘Identification of Endogenous Social Effects: The Reflection Problem’, The
Review of Economic Studies 60(3), 531–542.
March, J. G. (1991), ‘Exploration and Exploitation in Organizational Learning’, Organization

29
 Science 2(1), 71–87.
Marshall, A. (1890), Principles of Economics, London: Macmillan.
Mas, A. and Moretti, E. (2009), ‘Peers at Work’, American Economic Review 99(1), 112–45.
McPherson, J. M. and Smith-Lovin, L. (1987), ‘Homophily in Voluntary Organizations: Status
Distance and the Composition of Face-to-Face Groups’, American Sociological Review 52(3), 370–
379.
Michelacci, C. and Silva, O. (2007), ‘Why so many local entrepreneurs?’, The Review of Economics
and Statistics 89(4), 615–633.
Moretti, E. (2004), ‘Workers’ Education, Spillovers, and Productivity: Evidence from Plant-level
Production Functions’, American Economic Review 94(3), 656–690.
Murray, F. and Tripsas, M. (2004), ‘The Exploratory Processes of Entrepreneurial Firms: The Role
of Purposeful Experimentation’, Advances in Strategic Management 21, 45–76.
Nanda, R. and Rhodes-Kropf, M. (2013), ‘Investment Cycles and Startup Innovation’, Journal of
Financial Economics 110(2), 403–418.
Nanda, R. and Sørensen, J. B. (2010), ‘Workplace Peers and Entrepreneurship’, Management Science
56(7), 1116–1126.
Oettl, A. (2012), ‘Reconceptualizing Stars: Scientist Helpfulness and Peer Performance’, Management
Science 58(6), 1122–1140.
Oh, H., Labianca, G. and Chung, M.-H. (2006), ‘A Multilevel Model of Group Social Capital’, The
Academy of Management Review 31(3), 569–582.
Reagans, R., Argote, L. and Brooks, D. (2005), ‘Individual experience and experience working
together: Predicting learning rates from knowing who knows what and knowing how to work
together’, Management Science 51(6), 869–881.
Reagans, R. and Zuckerman, E. W. (2001), ‘Networks, diversity, and productivity: The social capital
of corporate R&D teams’, Organization Science 12(4), 502–517.
Rogers, E. M. (2010), Diffusion of innovations, New York: Simon and Schuster.
Romer, P. M. (1986), ‘Increasing Returns and Long-Run Growth’, Journal of Political Economy
94(5), 1002–1037.
Rosenthal, S. S. and Strange, W. C. (2001), ‘The Determinants of Agglomeration’, Journal of urban
economics 50(2), 191–229.
Rosenthal, S. S. and Strange, W. C. (2004), Chapter 49 - Evidence on the Nature and Sources of
Agglomeration Economies, in J. V. Henderson and J.-F. Thisse, eds, ‘Cities and Geography’,
Vol. 4 of Handbook of Regional and Urban Economics, Elsevier, pp. 2119 – 2171.
Samila, S. and Sorenson, O. (2011), ‘Venture Capital, Entrepreneurship, and Economic Growth’,
The Review of Economics and Statistics 93(1), 338–349.
Saxenian, A. (1996), Regional Advantage, Cambridge: Harvard University Press.
Schilling, M. A. and Fang, C. (2014), ‘When hubs forget, lie, and play favorites: Interpersonal
Network Structure, Information Distortion, and Organizational Learning’, Strategic Management
Journal 35(7), 974–994.
Singh, J. (2005), ‘Collaborative Networks as Determinants of Knowledge Diffusion Patterns’,
Management Science 51(5), 756–770.
Stefano, G., King, A. and Verona, G. (2017), Too Many Cooks Spoil the Broth? Geographic Concen-
tration, Social Norms, and Knowledge Transfer, Vol. 36 of Advances in Strategic Management,
Emerald Publishing Limited, pp. 267–308.
Todorova, G. and Durisin, B. (2007), ‘Absorptive Capacity: Valuing a Reconceptualization’, Academy
of Management Review 32(3), 774–786.
Tortoriello, M., McEvily, B. and Krackhardt, D. (2015), ‘Being a catalyst of innovation: The role of

30
 knowledge diversity and network closure’, Organization Science 26(2), 423–438.
URL: https://pubsonline.informs.org/doi/abs/10.1287/orsc.2014.0942
Wang, S. and Zhao, M. (2018), ‘A Tale of Two Distances: A Study of Technological Distance,
Geographic Distance and Multilocation Firms’, Journal of Economic Geography 18(5), 1091–1120.
Yang, L., Holtz, D., Jaffe, S., Suri, S., Sinha, S., Weston, J., Joyce, C., Shah, N., Sherman, K.,
Hecht, B. and Teevan, J. (2021), ‘The effects of remote work on collaboration among information
workers’, Nature Human Behaviour .
URL: https://doi.org/10.1038/s41562-021-01196-4
Young, A. (2019), ‘Channeling Fisher: Randomization Tests and the Statistical Insignificance of
Seemingly Significant Experimental Results’, The Quarterly Journal of Economics 134(2), 557–
598.
Zahra, S. A. and George, G. (2002), ‘Absorptive capacity: A review, reconceptualization, and
extension’, The Academy of Management Review 27(2), 185–203.

31
32
Figure 1: Floor plan of the co-working space
Notes: This figure displays the floor-plans of the co-working hub we examine. The legend and scale can be found on the bottom right corner of the figure.
 0.04

0.03

0.02

0.01
Coefficient

Null
0
0

-0.01

-0.02

-0.03
1st quartile 2nd quartile 3rd quartile 4th quartile Different Floor
(omitted)

Figure 2: Quartile plots

Notes: This figure displays the results from estimating equation (1) using a quartile regression. We thereby split
our distance measure into quartiles instead of using a continuous measure of distance. Our omitted category
consists of distances among startup dyads that span more than one floor.

33
 50 40
Kernel Density Estimate
20 30
10
0

-.04 -.02 0 .02 .04

b
kernel = epanechnikov, bandwidth = 0.0018

Figure 3: Randomized Inference using Monte Carlo Simulation

Notes: This figure presents the kernel density distribution of coefficients from simulated Monte Carlo draws (1,000
runs). In the simulation, we randomize closeness between each dyad and subsequently estimate the likelihood of
adopting a technology as a function of closeness (Close) using the simulated strata. The vertical line indicates
the point estimate of our main results (β = 0.022). Only 2 of the simulated Monte Carlo draws (from 1,000) had
a coefficient greater than the point estimate of our main results, resulting in a randomized inference p-value of 0.002.

34
 Do Not Pass Through a Common Area Pass Through a Common Area

0.10

0.08

0.06

0.04

0.02

0.00

-0.02

-0.04
1st quartile 2nd quartile 3rd quartile 4th quartile
(ommited)

Figure 4: Common area quartile plots

Notes: This figure displays the results from estimating equation (1) using a quartile regression and including an
interaction with the CommonArea dummy. We thereby use CommonArea× 4th distance quartile as the omitted
category.

35
 Seed Funding: Predictive Margins of Social $1M+ Funding: Predictive Margins of Social
quantile quantile
Social=0 Social=1 Social=0 Social=1
.4

.2
Linear Prediction

Linear Prediction
.2

.1
0
0

-.1
-.2

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Diversity Quantiles Diversity Quantiles

Prob(Seed): Differences between Social and Not Social Prob($1M+): Differences between Social and Not Social
.4

.2
Contrasts of Linear Prediction

Contrasts of Linear Prediction

.2

.1
0

0
-.2

-.1
-.4

-.2

1 2 3 4 5 1 2 3 4 5
Diversity Quantiles Diversity Quantiles

Figure 5: How a startup’s socializing and the diversity of proximate startups predicts raising
funding
Notes: This figure displays margins plots for the results from estimating the likelihood of raising a seed round
(left)/$1M+ or more (right) as a function of the aggregate diversity index of startups within 20 meters of the
focal startup interacted with an indicator equal to one if the focal startup engages in social events (Social=1 ).
We thereby control for startup characteristics (industries, age, size) and the number of startups in the immediate
environment. 95% confidence intervals are displayed.

36
 Table 1: Pairwise characteristics do not predict geographic proximity - OLS Regressions

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable ln(distanceij )
(1) (2)
Same Industry 0.000 0.001
(0.023) (0.023)
Both B2B Companies 0.029 0.030
(0.041) (0.039)
Both B2C Companies 0.030 0.030
(0.045) (0.044)
Both Majority Female 0.015 0.015
(0.126) (0.124)
Both Successful 0.021 0.022
(0.059) (0.058)
|agei -agej | 0.001 0.001
(0.003) (0.003)
Pre-period Technology Overlap −0.074
(0.082)
Firmi X Room Fixed Effects X X
Firmj X Room Fixed Effects X X
Observations 10840 10840
R2 0.12 0.12
Notes: This table displays the results from OLS regressions predicting physical distance
between two firms as a function of firm-dyad characteristics. These variables (indicated by
Both and Same) equal one if both firmi and firmj operate in the same industry, both have a
B2B (B2C) business model, are both predominately female, and are both successful. The
variable |age i-age j| represents the absolute age difference in months between firmi and firmj .
Pre-period Technology Overlap presents the share of firmi ’s technologies also used by firmj in
the previous period. Standard errors (in parentheses) are robust to dyadic clustering at the
floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

37
 Table 2: Summary Statistics

Firm level (N = 251) mean sd min p25 p50 p75 max

Age (in months) 12.24 9.59 0 3 11 20 29
Room size (in sq.feet) 271.18 315.82 50 134 143 255 1878
Room size (in m2 ) 25.20 29.34 4.64 12.45 13.29 23.70 174.50
Female CEO (= 0/1) 0.12 0.32 0 0 0 0 1
B2B Company (= 0/1) 0.74 0.44 0 0 1 1 1
B2C Company (= 0/1) 0.39 0.49 0 0 0 1 1
Successful (= 0/1) 0.24 0.43 0 0 0 0 1
Min. Technology Usage 33.15 33.15 0 0 28 54 168
Max. Technology Usage 51.06 49.70 0 0 43 79 255
Dyad level (N = 10840) mean sd min p25 p50 p75 max
Adopted a Technology (= 0/1) 0.53 0.50 0 0 1 1 1
Number of Adopted Technologies 7.33 10.49 0 0 2 12 76
Distance (in m2 ) 32 15.20 4.30 20 30 44 77
Close (= 0/1) 0.28 0.45 0 0 0 1 1
Common Area (= 0/1) 0.38 0.48 0 0 0 1 1
Pre-period Technology Overlap (%) 0.14 0.18 0 0 0 0.27 0.85
Same Industry (= 0/1) 0.11 0.31 0 0 0 0 1
Both B2B Companies (= 0/1) 0.48 0.50 0 0 0 1 1
Both B2C Companies (= 0/1) 0.11 0.31 0 0 0 0 1
Both Female (= 0/1) 0.013 0.11 0 0 0 0 1
Age Difference (in months) 7.30 7.28 0 1 5 12 29
Both Successful (= 0/1) 0.08 0.27 0 0 0 0 1
Non-geographically distant (= 0/1) 0.38 0.49 0 0 0 1 1
Notes: This table displays summary statistics for the startups operating at the co-working space we examine. We report summary
statistics both on the firm and dyad level. Please refer to Table A1 in the Appendix for a description of the variables displayed.

38
 Table 3: Physical proximity positively affects peer technology adoption

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable ln(AdoptCountij + 1) 1(AdoptTechij ) ln(AdoptCountij + 1) 1(AdoptTechij )
mean 1.275 0.531 1.275 0.531
(1) (2) (3) (4) (5) (6) (7) (8)

ln(distanceij ) -0.043∗∗∗ -0.035∗∗∗ -0.019∗∗∗ -0.017∗∗∗

(0.017) (0.010) (0.007) (0.005)
Close 0.057∗∗ 0.048∗∗∗ 0.025∗∗ 0.022∗∗∗
(0.026) (0.015) (0.011) (0.007)
Same Industry 0.021 0.005 0.021 0.005
(0.029) (0.013) (0.029) (0.013)
Both B2B Companies -0.034 -0.007 -0.034 -0.007
(0.022) (0.011) (0.022) (0.011)
Both B2C Companies 0.030 0.005 0.029 0.004
(0.029) (0.008) (0.029) (0.008)
Both Majority Female -0.102∗ 0.013 -0.103∗ 0.012
(0.057) (0.027) (0.057) (0.028)

39
|agei -agej | -0.006∗∗∗ -0.001∗∗ -0.006∗∗∗ -0.001∗
(0.002) (0.000) (0.001) (0.001)
Pre-period Technology Overlap 3.624∗∗∗ 1.007∗∗∗ 3.624∗∗∗ 1.007∗∗∗
(0.146) (0.066) (0.145) (0.065)
Firmi X Room Fixed Effects X X X X X X X X
Firmj X Room Fixed Effects X X X X X X X X
Observations 10840 10840 10840 10840 10840 10840 10840 10840
R2 0.80 0.86 0.79 0.83 0.80 0.86 0.79 0.83
Notes: This table displays the results from OLS regressions predicting technology adoption as a function of physical distance
(proximity) and other dyad characteristics. The outcome ln(AdoptCountij + 1) is the natural logarithm of the number of new
to f irmi technologies f irmi adopts from f irmj . The outcome 1(AdoptTechij ) equals one if f irmi adopted at least one new
technology from f irmj . Distance is captured using the natural logarithm of step distance between two firms (ln(distanceij )).
Close equals to one if f irmi and f irmj are located within 20 meters (the 25th percentile of pair-wise distances between all
rooms) of each other on the same floor. The variables denoted by Both and Same equal one if both firmi and firmj operate
in the same industry, both have a B2B (B2C) business model, and are both predominately female. The variable |age i-age j|
represents the absolute age difference in months between firmi and firmj . Pre-period Technology Overlap presents the share of
firmi ’s technologies also used by firmj in the previous period. We include firmi x room and firmj x room fixed effects. Standard
errors (in parentheses) are robust to dyadic clustering at the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
 Table 4: Proximity and Diversity

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable 1(AdoptTechij )
mean 0.531
(1) (2) (3) (4)

Close 0.024∗∗∗ 0.037∗∗∗ 0.031∗∗∗ 0.014∗∗

(0.007) (0.007) (0.012) (0.006)
Both Female 0.018
(0.016)
Close x Female −0.089∗∗∗
(0.016)
Same Product Market 0.013∗∗∗
(0.005)
Close x Same Product Market −0.023∗∗∗
(0.008)
High Tech-Stack Overlap 0.209∗∗∗
(0.027)
Close x High Tech-Stack Overlap −0.027∗∗
(0.014)
Diverse −0.001
(0.007)
Close x Diverse 0.029∗∗∗
(0.005)
Pre-period Technology Overlap 1.007∗∗∗ 1.006∗∗∗ 1.011∗∗∗
(0.066) (0.065) (0.063)
|age i-age j| −0.001∗∗ −0.001∗∗ −0.001∗ −0.001∗∗
(0.001) (0.000) (0.000) (0.001)
Firmi X Room Fixed Effects X X X X
Firmj X Room Fixed Effects X X X X
Dimension Social Product-Market Knowledge Composite Index
Observations 10840 10840 10840 10840
R2 0.8305 0.8306 0.8063 0.8306
Notes: This table displays the results from linear probability models predicting technology adoption as a
function of physical proximity (close) and the interaction with other proximity dimensions. Diverse is an
indicator equal to one if the firm dyads differ along all non-geographic proximity dimensions we in examine.
The outcome 1(AdoptTechij ) equals one if f irmi adopted at least one new technology from f irmj . Close
equals to one if f irmi and f irmj are located within 20 meters (the 25th percentile of pair-wise distances
between all rooms) of each other on the same floor. The variables denoted by Both and Same equal one if
both firmi and firmj operate in the same product market, or both predominately female. High Tech-Stack
Overlap denotes dyads that have a pre-period tech-stack overlap of over 0.27, which represents the 75th
percentile. We include controls for age differences and firmi X room fixed effects as well as the share of firmi ’s
technologies also used by firmj in the previous period in columns 1, 2, and 4. Standard errors (in parentheses)
are robust to dyadic clustering at the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

40
 Table 5: Joint Attendance and Checkin-line Proximity - OLS Regressions

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable # Event Bothij Attend 1(Event) 1(w/in 5 people in line)
mean 0.27 0.11 0.06
(1) (2) (3) (4) (5) (6)
Close 0.036∗∗ 0.039∗ 0.010∗ 0.009∗ 0.017∗∗∗ 0.023∗∗∗
(0.018) (0.022) (0.005) (0.046) (0.006) (0.009)
Common Area 0.025∗∗ 0.024∗∗ 0.010∗ 0.010∗ 0.013∗∗∗ 0.013∗∗∗
(0.010) (0.011) (0.005) (0.029) (0.005) (0.005)
Diverse -0.028∗∗∗ -0.013∗∗∗ -0.009∗∗∗
(0.008) (0.003) (0.003)
Close x Diverse -0.010 0.003 -0.018∗
(0.021) (0.010) (0.011)

Firmi X Room Fixed Effects X X X X X X

Firmj X Room Fixed Effects X X X X X X
Observations 10840 10840 10840 10840 10840 10840
R2 0.5443 0.5444 0.5141 0.5142 0.3525 0.3532
Notes: This table displays the results from OLS regressions predicting the number of lunches hosted at
the co-working space that at least one team member of f irmi and f irmj both attend (# Event Bothij
Attend and the likelihood of attending (1(Event)). The indicator 1(w/in 5 people in line) equals to one
if at least one team member of f irmi and f irmj ever appear within 5 people in line for the lunch. The
variable Common Area equals one if the shortest path between f irmi and f irmj passes through a common
area. Common areas are the kitchens and zone in front of the elevator on each floor as well as the open
sitting space provided on the second floor. We include firmi x room and firmj x room fixed effects. Diverse
is an indicator equal to one if the firm dyads differ along all non-geographic proximity dimensions we in
examine. Standard errors (in parentheses) are robust to dyadic clustering at the floor-neighborhood level. ∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

41
 Table 6: Proximity, Socializing, and Diversity - OLS Regressions

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable 1(AdoptTechij )
mean 0.531
(1) (2) (3) (4)
Close 0.023∗∗∗ 0.022∗∗∗
(0.009) (0.009)
# Events 0.043∗∗∗ 0.037∗∗∗ 0.043∗∗∗
(0.007) (0.007) (0.007)
Diverse −0.002
(0.006)
# Events x Diverse 0.044∗∗∗
(0.006)
Close = 1 & Diverse = 1 0.041∗∗∗
(0.007)
Close = 0 & Diverse = 1 −0.002
(0.007)
Close = 1 & Diverse = 0 0.013∗
(0.008)
# Events x (Close = 0 & Diverse = 0) 0.034∗∗∗
(0.006)
# Events x (Close = 0 & Diverse = 1) 0.077∗∗∗
(0.010)
# Events x (Close = 1 & Diverse = 0) 0.041∗∗∗
(0.010)
# Events x (Close = 1 & Diverse = 1) 0.093∗∗∗
(0.021)

Pre-prd. Tech. Overlap, Age Diff., Common Area X X X X

Firmi X Room Fixed Effects X X X X
Firmj X Room Fixed Effects X X X X
Observations 10840 10840 10840 10840
R2 0.8325 0.8330 0.8326 0.8325
Notes: This table displays the results from OLS regressions predicting technology adoption as a function of physical distance
(proximity) and other dyad characteristics. The outcome 1(AdoptTechij ) equals one if f irmi adopted at least one new
technology from f irmj . Close equals to one if f irmi and f irmj are located within 20 meters (the 25th percentile of
pair-wise distances between all rooms) of each other on the same floor. The variable # Event Bothij Attend equals the
number of lunch hosted at the co-working space that at least one team member of f irmi and f irmj both attend. Diverse
is an indicator equal to one if the firm dyads differ along all non-geographic proximity dimensions we in examine and zero
(Diverse = 0 ) otherwise. In Columns 3-4, we include categories that indicate whether a dyad is 1) far and similar (Close=0
& Diverse =0 ); 2) far and different (Close=0 & Diverse =1 ); 3) close and similar (Close=1 & Diverse =0 ); and 4) close
and different (Close=1 & Diverse =1 ). In Column 3, the omitted category is Close=0 & Diverse = 0. The variables
|age i-age j|, Pre-period Technology Overlap and Common Area are included. Variables including “&” denote categories.
We include firmi x room and firmj x room fixed effects. Standard errors (in parentheses) are robust to dyadic clustering at
the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

42
 Appendix

(Co-)Working in Close Proximity: Knowledge

Spillovers and Social Interactions

A1
 Figure A1: Technology Adoption Counts - Histogram

Notes: This figure displays the relative distribution of technology adoption (techcount) by the startups in our sample.

A2
Table A1: Pairwise characteristics do not predict geographic proximity - OLS Regressions

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable Close
(1) (2)
Same Industry −0.001 −0.002
(0.021) (0.022)
Both B2B Companies −0.023 −0.023
(0.029) (0.028)
Both B2C Companies −0.005 −0.005
(0.032) (0.032)
Both Majority Female 0.022 0.021
(0.102) (0.100)
Both Successful −0.024 −0.025
(0.035) (0.034)
|age i-age j| −0.000 −0.000
(0.001) (0.001)
Pre-period Technology Overlap 0.054
(0.076)
Firmi X Room Fixed Effects X X
Firmj X Room Fixed Effects X X
Observations 10840 10840
R2 0.10 0.10
Notes: This table displays the results from OLS regressions predicting that two firms are
located within 20m as a function of firm-dyad characteristics. These variables (indicated by
Both and Same) equal one if both firmi and firmj operate in the same industry, both have a
B2B (B2C) business model, are both predominately female, and are both successful. The
variable |age i-age j| represents the absolute age difference in months between firmi and
firmj . Pre-period Technology Overlap presents the share of firmi ’s technologies also used by
firmj in the previous period. Standard errors (in parentheses) are robust to dyadic clustering
at the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

A3
 Table A2: Variable Description
Variable Description

Outcome Variables
ln(Distanceij ) The distance between f irmi and f irmj in steps (log transformed). One step
corresponds to 1.8 meters.
ln(AdoptCountij + 1) The number of technologies f irmi adopts from f irmj (log transformed and
normalized). An adopted technology is a technology used by f irmi in the
focal period that f irmi had not implemented in any previous period, but
f irmj had.
1(AdoptT echij ): Equals one if f irmi adopts a technology from f irmj .
# Event Bothij Attend The number of events hosted at the co-working space at least one person
working for of f irmi and f irmj both attend.
1(Ever within X people in line) Equals one if at least one team member of f irmi and f irmj appear within
X (1, 2, 5, 10, 25) people in line for an event hosted at the co-working space.
Dyad-Level Independent Variables
Close Equals to one if f irmi and f irmj are located within 20 meters (14 steps; the
25th percentile of pair-wise distances between all rooms) of each other on the
same floor.
Common Area Equals one if the shortest path between f irmi and f irmj passes through
a common area. Common areas are the kitchens and zone in front of the
elevator on each floor as well as the open sitting space provided on the second
floor. Please refer to Figure 1 for a visual depiction of the location of these
areas.
Same Industry Equals to one if f irmi and f irmj operate in the same industry. We follow the
classification of industries provided by AngelList and BuiltWith. The individ-
ual industries are Administration&Management, Data, Design&Development,
Digital, Education, Energy&Construction, Entertainment, Finance&Legal
Healthcare, Marketing&PR, Real Estate, Retail, Science&Technology, Secu-
rity, Software&Hardware. For our analyses we use each firm’s primary indus-
try, since many operate in more than one. We determined this by conducting
extensive web searches on the startups in our sample.
Pre-period Technology Overlap Percentage of same technologies f irmi and f irmj used in the period prior to
the focal period.
Both Majority Female Equals to one if the team members in both f irmi and f irmj are predomi-
nately female (over 50 percent). We determined the gender of founders con-
ducting extensive web searches on the startups as well as by comparing first
names with lists provided by the US Census for most common names by sex
(https://www2.census.gov/topics/genealogy/1990surnames).
Both B2B Companies Equals to one if f irmi ’s and f irmj ’s main customers are other businesses.
Both B2C Companies Equals to one if f irmi ’s and f irmj ’s main customers are individual con-
sumers.
Both Successful Equals to one if f irmi and f irmj have received a TAG40 award, have received
the Village Verified certificate, have raised a seed round or have ever raised a
VC seed investment.
Diverse Equals to one if a startup dyad differs along the social, product-market and
knowledge dimensions. For simplicity, we count a dyad as different along the
knowledge space dimension if their pre-period technology overlap is below the
mean.
|agei -agej | The age difference between f irmi and f irmj (derived from date of entry at
the co-working space).

A4
 Table A3: Example Tech-Stacks (only including first 90 in category alphabetic order)
Firm A: Firm B: Firm C:
hosts a platform for sellers to execute digital selling tasks, communicate with buyers, runs a marketing platform used to target companies, decision-makers, and accelerate
Technology Category created and hosts an automated scheduling tool. Founded in 2013. and get coaching. Founded in 2012. pipeline speed. Founded in 2014.
A/B Testing Mixpanel; Optimizely; KISSmetrics Optimizely Mixpanel, Optimizely
Ad Analytics AdStage; Advertising.com; ContextWeb
Ad Exchange AppNexus Facebook Exchange FBX; BlueKai; IponWeb BidSwitch; Eyeota
Ad Network Bizo Burst Media; Tribal Fusion; AdRoll; Twitter Ads
Ad Server AppNexus Openads/OpenX
Ads AdsNative; LinkedIn Ads; DoubleClick.Net; AppNexus Segment Pixel DoubleClick.Net Index Exchange; Adap.TV; Yahoo Small Business; Yield Manager; SpotXchange; DoubleClick.Net;
Aggregate Knowledge; Rubicon Project; eXelate; AppNexus Segment Pixel; Tapad; Simpli.fi; Ad
Tech Japan AOL; Adbrain; RUN Ads; Arbor Marketplace; Uberflip
Advertiser Tracking Bizo Insights Bombora; Krux Digital; comScore; Datalogix
Affiliate Programs adingo
Analytics Google Universal Analytics; Google Analytics Classic Google Universal Analytics Google Universal Analytics; Lotame Crowd Control; Twitter Website Universal Tag; Tynt Tracer;
Facebook Signal; Facebook Pixel; Marketo Real Time Personalization
Animation GSAP
Application Performance New Relic; Heap New Relic; Google Analytics VisiStat; Google Analytics
Audience Measurement FullStory FullStory; Shareaholic
Audience Targeting Turn; DemDex
Business Email Hosting Google Apps for Business; UserVoice Mail; Intercom Mail Google Apps for Business Google Apps for Business
Call Tracking CallRail CallRail
CAPTCHA Are You a Human
Charting, UI D3 JS
Campaign Management MailChimp SPF
Cloud Hosting, Cloud PaaS Amazon Amazon Google Cloud; Google
Content Delivery Network CloudFront; Twitter CDN; Bootstrap CDN; AJAX Libraries API; CDN JS; GStatic Google Static Akamai; CDN Bootstrap CDN; GStatic Google Static Content; AJAX Libraries API; CloudFront; Max CDN; CDN
Content JS; jQuery CDN; Amazon S3 CDN
Compatibility Modernizr; html5shiv Modernizr; html5shiv
Comment System Disqus
Content Curation Taboola
Conversion Optimization Google Conversion Tracking; Twitter Analytics; LinkedIn Insights; Bing Universal Event Tracking LinkedIn Insights Twitter Analytics; Twitter Conversion Tracking; Google Conversion Tracking; BrightFunnel; G2
Crowd Conversion
Cookie Sync Adobe Audience Manager Sync
CRM Zendesk Salesforce SPF; Zendesk
Data Management Platform Segment BlueKai DMP
Dedicated Hosting Rackspace
Demand-side Platform The Trade Desk; DoubleClick Bid Manager
Dynamic Creative Optimization Pubmatic
Enterprise DNS Amazon Route 53; Microsoft Azure DNS Wistia; Amazon Route 53 Amazon Route 53
Error Tracking Sentry; Bugsnag; Rollbar

A5
Feedback Forms and Surveys Wufoo; Intercom Intercom Contact Form 7
Feeds RSS RSS; Pingback Support; Really Simple Discovery; Live Writer Support Really Simple Discovery; RSS; Live Writer Support; Pingback Support
Fonts Font Awesome; Google Font API Font Awesome; Google Font API; jQuery Form
Framework Ruby on Rails Token; Heroku Proxy; AMP Project; Handlebars Ruby on Rails Token Ruby on Rails Token
JavaScript jQuery; Marionette.js; Moment JS; Backbone.js jQuery; CryptoJS CryptoJS; Froogaloop; Angular JS v1; Raven JS; punycode; Isotope; jQuery Waypoints; Webpack;
Facebook Graph API; jQuery; lodash; Moment JS; jQuery Masonry
Lead Generation LiveRamp; Insightera
Lightbox Magnific Popup
Live Chat SnapEngage SnapEngage; Drift
Live Stream YouTube
Marketing Automation Pardot; Captora; Bizible; Hubspot Pardot; Rapleaf; Marin Software; Bizible; OwnerIQ, Terminus; Marketo
Marketing Platform Pardot Mail
Media SundaySky
Multi-Channel Crosswise
Open Source, Blog WordPress WordPress
Plugins Yoast WordPress SEO Plugin; Yoast SEO Premium; Disqus Comment System for WordPress Yoast Plugins; Yoast WordPress SEO Plugin; Yoast SEO Premium; jQuery UI; Click To Tweet for
WordPress; Lazy Load for WordPress; SiteOrigin Panels; WP Super Cache
Payments Processor Stripe
Programming Language Ruby on Rails PHP PHP; Ruby on Rails
Retargeting / Remarketing Twitter Ads AdRoll Perfect Audience; Facebook Custom Audiences; Google Remarketing
Root Authority Comodo SSL GeoTrust SSL; LetsEncrypt; DigiCert SSL
Server Ubuntu
Site Optimization CrazyEgg
Site Search Algolia Sitelinks Search Box Sitelinks Search Box
Social Management Facebook Domain Insights
SSL SSL by Default; Heroku SSL; GoDaddy SSL Comodo PositiveSSL; SSL by Default RapidSSL; SSL by Default
Standard SPF SPF SPF; DMARC
Tag Management Google Tag Manager Google Tag Manager
Toolbar Hello Bar Hello Bar
Transactional Email Sendgrid Mandrill
US hosting Amazon Virginia Region Amazon Virginia Region
Video Analytics Mux Vidyard
Video Players MediaElement.js
Web Master Google Webmaster Google Webmaster
Web Server Cowboy; nginx nginx Apache; nginx
Widgets Typekit; Wordpress Plugins; Twemoji; Gravatar Profiles Wordpress Plugins
Wildcard Comodo PositiveSSL Wildcard Typekit; Twemoji; Ipify; Lever; ZeroClipboard; TurboLinks; Facebook Sharer; Pinterest
 Table A4: Distance negatively affects peer technology adoption - OLS Regressions

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable ln(AdoptCountij + 1)
(1) (2) (3) (4) (5) (6)

ln(distanceij ) −0.073∗∗ −0.044 −0.056∗∗ −0.043∗∗∗ −0.043∗∗ −0.035∗∗∗

(0.037) (0.043) (0.027) (0.017) (0.018) (0.010)
Same Industry 0.056 0.021
(0.034) (0.029)
Both B2B Companies 0.004 −0.034
(0.028) (0.022)
Both B2C Companies 0.007 0.030
(0.031) (0.029)
Both Female −0.095 −0.102∗
(0.098) (0.057)
|age i-age j| −0.004∗∗∗ −0.006∗∗∗

A6
(0.001) (0.002)
Pre-period Technology Overlap 3.624∗∗∗
(0.146)
Firmi Fixed Effects X
Firmj Fixed Effects X
Firmi X Room Fixed Effects X X X
Firmj X Room Fixed Effects X X X
Observations 10840 10840 10840 10840 10840 10840
R2 0.00 0.35 0.44 0.80 0.80 0.86
Notes: This table displays the results from OLS regressions predicting technology adoption as a function of physical distance (proximity) and other
dyad characteristics. The outcome ln(AdoptCountij + 1) is the natural logarithm of the number of new to f irmi technologies f irmi adopts from
f irmj . Distance is captured using the natural logarithm of step distance between two firms (ln(distanceij)). Close equals to one if f irmi and f irmj
are located within 20 meters (14 steps; the 25th percentile of pair-wise distances between all rooms) of each other on the same floor. The variables
denoted by Both and Same equal one if both firmi and firmj operate in the same industry, both have a B2B (B2C) business model, and are both
predominately female. The variable |age i-age j| represents the absolute age difference in months between firmi and firmj . Pre-period Technology
Overlap presents the share of firmi ’s technologies also used by firmj in the previous period. Standard errors (in parentheses) are robust to dyadic
clustering at the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
 Table A5: Distance negatively affects peer technology adoption - LPM Regressions

Unit of Analysis Firmi -Firmj Dyad

Dependent Variable 1(AdoptTechij )
mean 0.531
(1) (2) (3) (4) (5) (6)

ln(distanceij ) −0.030∗∗ −0.022 −0.024∗ −0.019∗∗∗ −0.019∗∗∗ −0.017∗∗∗

(0.014) (0.022) (0.014) (0.007) (0.007) (0.005)
Same Industry 0.015 0.005
(0.016) (0.013)
Both B2B Companies 0.004 −0.007
(0.014) (0.011)
Both B2C Companies −0.002 0.005
(0.012) (0.008)
Both Female 0.015 0.013
(0.019) (0.027)
|age i-age j| −0.001 −0.001∗∗

A7
(0.000) (0.000)
Pre-period Technology Overlap 1.007∗∗∗
(0.066)
Firmi Fixed Effects X
Firmj Fixed Effects X
Firmi X Room Fixed Effects X X X
Firmj X Room Fixed Effects X X X
Observations 10840 10840 10840 10840 10840 10840
R2 0.00 0.37 0.42 0.79 0.79 0.83
Notes: This table displays the results from predicting the likelihood of technology adoption as a function of physical distance (proximity) and other
dyad characteristics. The outcome 1(AdoptTechij ) equals one if f irmi adopted at least one new technology from f irmj . Distance is captured using
the natural logarithm of step distance between two firms (ln(distanceij)). Close equals to one if f irmi and f irmj are located within 20 meters (14
steps; the 25th percentile of pair-wise distances between all rooms) of each other on the same floor. The variables denoted by Both and Same equal one
if both firmi and firmj operate in the same industry, both have a B2B (B2C) business model, and are both predominately female. The variable
|age i-age j| represents the absolute age difference in months between firmi and firmj . Pre-period Technology Overlap presents the share of firmi ’s
technologies also used by firmj in the previous period. We include firmi X room fixed effects. Standard errors (in parentheses) are robust to dyadic
clustering at the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
 Table A6: Common-area overlap increases technology adoption

Dependent Variable 1(AdoptTechij )

(1) (2)
∗∗
Close 0.029 0.032∗∗∗
(0.012) (0.012)
Common Areaij 0.010∗ 0.011∗∗
(0.005) (0.005)
Close × Common Areaij −0.036
(0.027)
Firmi X Room Fixed Effects X X
Firmj X Room Fixed Effects X X
Observations 10840 10840
R2 0.79 0.79
Notes: This table displays the results from OLS regressions the likelihood of technology adoption as a
function of physical proximity and common areas. The variable Common Area equals one if the shortest
path between f irmi and f irmj passes through a common area. Common areas are the kitchens and zone
in front of the elevator on each floor as well as the open sitting space provided on the second floor. We
include firmi X room fixed effects. Standard errors (in parentheses) are robust to dyadic clustering at the
floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

A8
 Table A7: Joint Attendance and Checkin-line Proximity - OLS Regressions

Unit of Analysis Firmi -Firmj Dyad

Bothij Attend 1(Ever within X people in line)

Dependent Variable # Events 1(Event) 1 person 2 people 5 people 10 people 25 people

(1) (2) (3) (4) (5) (6) (7)

Close 0.240∗ 0.010∗ 0.064∗ 0.091∗ 0.093∗∗ 0.010 −0.027

(0.142) (0.006) (0.035) (0.047) (0.046) (0.039) (0.036)
Common Areaij 0.147∗∗ 0.010∗ 0.019 0.061∗ 0.057∗ 0.047 0.056∗∗∗
(0.064) (0.005) (0.040) (0.036) (0.029) (0.041) (0.007)

A9
Firmi X Room Fixed Effects X X X X X X X
Firmj X Room Fixed Effects X X X X X X X
Observations 10840 10840 1398 1398 1398 1398 1398
R2 0.47 0.51 0.42 0.45 0.48 0.51 0.47
Notes: This table displays the results from OLS regressions predicting the number of lunches (likelihood of attending at least two lunches) hosted at the
co-working space that at least one team member of f irmi and f irmj both attend (# Event Bothij Attend /1(Event)). The indicator 1(Ever within X
people in line) equals to one if at least one team member of f irmi and f irmj appear within 1, 2, 5, 10, or 25 people in line for the lunch conditional on
jointly attending the event. The variable Common Area equals one if the shortest path between f irmi and f irmj passes through a common area. Com-
mon areas are the kitchens and zone in front of the elevator on each floor as well as the open sitting space provided on the second floor. We include firmi
X room fixed effects. Standard errors (in parentheses) are robust to dyadic clustering at the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
 Table A8: Socializing, Diversity Quintiles and Financial Performance Outcomes

Funding raised Seed >1M+

(1) (2)
Social = 0 × Diversity Quantiles = 2 −0.045 −0.031
(0.073) (0.046)
Social = 0 × Diversity Quantiles = 3 0.017 0.003
(0.062) (0.039)
Social = 0 × Diversity Quantiles = 4 −0.078 0.010
(0.070) (0.044)
Social = 0 × Diversity Quantiles = 5 −0.104 −0.033
(0.070) (0.044)
Social = 1 × Diversity Quantiles = 1 −0.035 −0.049
(0.068) (0.043)
Social = 1 × Diversity Quantiles = 2 −0.131 −0.063
(0.085) (0.054)
Social = 1 × Diversity Quantiles = 3 0.155∗∗ 0.107∗∗
(0.074) (0.047)
Social = 1 × Diversity Quantiles = 4 −0.034 0.069
(0.101) (0.064)
Social = 1 × Diversity Quantiles = 5 −0.152 −0.010
(0.102) (0.064)
Room Size 0.000 0.000
(0.000) (0.000)
Female CEO −0.128 −0.047
(0.082) (0.052)
Remoteness −0.004 −0.006
(0.008) (0.005)
Age 0.004 0.004∗∗
(0.003) (0.002)
No. Firms 0.003 −0.000
(0.004) (0.002)
Floor FE X X
Observations 248 248
R2 0.10 0.11
Notes: This table displays the results from OLS regressions predicting the likelihood of raising a seed round
(Seed ) and $ million or more (>1M+) as a function of the aggregate diversity of firms within 20 meters of the
focal firm interacted with an indicator equal to one if the focal firm engages in social events (Social=1 ). The
aggregate diversity index ins split into quintiles. We thereby control for firm characteristics (industries, age,
size) and the number of firms in the immediate environment. Standard errors (in parentheses) are robust to
dyadic clustering at the floor-neighborhood level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

A10