Slack Time and Innovation
Ajay Agrawal, Christian Catalini, Avi Goldfarb, Hong Luo∗
December 30, 2016
Abstract
Traditional innovation models assume that ideas are developed as long as their benefit is
greater than their cost. So, when the opportunity cost of time is lowered, such as during a
holiday from work, more projects are developed, but they have a lower expected benefit. We
develop a simple theoretical framework that allows innovators to choose the level of effort they
apply towards the development of their idea in a context where the returns to effort are greater
for higher quality ideas. We also allow for lower team coordination costs during periods when
individuals have overlapping slack time. Under these conditions, slack time may lead to not only
lower but also higher quality projects. We test the model’s predictions using data on 165,410
projects posted on Kickstarter during college breaks. Consistent with the model’s predictions,
during university breaks more projects are posted and the increase is largest for projects of either
very high or very low quality. Furthermore, projects posted during breaks are more complex,
with larger teams and more diverse skills.
JEL Classifications: L26, O31, J22.
Keywords: crowdfunding, entrepreneurship, slack time, teamwork, low-opportunity cost time,
internet.
∗
We thank the review team, Michael Ewens, Lee Fleming, Joshua Gans, Scott Stern, and seminar participants
at the NBER Changing Financing Market for Innovation & Entrepreneurship Conference, the 2014 AEA Meetings,
the 2014 REER Conference at Georgia Tech, Duke University, Johns Hopkins University, University of Maryland,
National University of Singapore, Singapore Management University, Stanford University, Harvard Business School,
ZEW Conference 2014, UC Berkeley, the EEI Conference at Queen’s School of Business, World Bank, MIT, and
University of Toronto for comments. We thank SSHRC, the Centre for Innovation and Entrepreneurship at the
University of Toronto, and the Edward B. Roberts (1957) MIT Fund Grant for Technology-Based Entrepreneurship for
financial support. Errors remain our own. Ajay Agrawal (University of Toronto): ajay.agrawal@rotman.utoronto.ca;
Christian Catalini (MIT): catalini@mit.edu; Avi Goldfarb (University of Toronto): agoldfarb@rotman.utoronto.ca;
Hong Luo (HBS): hluo@hbs.edu.
“It’s no coincidence that Microsoft and Facebook both got started in January. At Harvard that
is (or was) Reading Period, when students have no classes to attend because they’re supposed to
be studying for finals.” - Paul Graham, Y-Combinator Founder, quoted in “How to Get Startup
Ideas”
1
Introduction
Companies and governments encourage innovation because it enhances productivity and competitiveness. However, innovation is costly. It requires time, effort, and other resources. In equilibrium,
under common innovation model assumptions, the marginal benefit of one more innovation would
just equal the marginal cost of producing it. As such, when innovators have more slack time, they
will innovate more as lowering the opportunity cost of time shifts the supply curve to the right.
Traditional models also predict that the additional innovations are marginal ones.
This raises a puzzle. Why do firms, such as Google, LinkedIn, 3M, and Wella, which already
provide innovation-oriented work environments, also provide slack time to encourage more innovation if doing so only facilitates the development of marginal ideas? How should such policies
be designed? Why do we see a recent shift from spread-out forms of individual slack time (e.g.,
Google’s 20% time) to more structured programs that provide teams of employees with longer,
continuous blocks of time off from regular commitments (e.g., Google’s “Area 120”)?
We develop a simple theoretical framework to capture the differences between periods of high
versus low opportunity cost time (work weeks versus break weeks). Low opportunity cost time may
induce: 1) lower quality ideas to be developed (a selection effect); 2) more effort to be applied for
any given idea quality (an effort effect); and 3) an increase in the use of teams because scheduling
is less constrained and communication is less delayed (a coordination effect). As a result, the effect
of breaks on the outcome distribution is ambiguous because on the one hand the selection effect
may induce lower quality ideas, but on the other hand the effort and coordination effects may lead
to more high-quality, complex ideas.
We explore this model using data from Kickstarter, the world’s leading rewards-based crowd-
1
funding platform.1 We leverage variation in the availability of slack time generated by school breaks
across locations with top U.S. colleges. Relative to more fragmented, constrained spare time during the semester, breaks provide students with longer blocks of time to work on non-school related
projects. Crowdfunding platforms have become increasingly important for entrepreneurial endeavors in a variety of segments, ranging from the arts to technology. Further understanding of the
determinants of the creative supply on these platforms is in and of itself interesting. Additionally,
our results generate useful implications for the design of “slack time” policies in firms and other
organizations.
Our sample includes all projects launched on Kickstarter between April 2009 and April 2015.
We combine these data with information on the exact timing of school breaks in locations with top
US colleges. Our unit of analysis is the city-week. We compare the number of projects launched in
a given city during break weeks versus work weeks. The regressions control for city and week fixed
effects to isolate confounding factors such as general time trends, seasonality, and different baseline
levels of innovative activity across different locations.
We first document the positive impact of breaks on the quantity of projects posted. In particular,
the regression results show that the number of crowdfunding campaigns launched increases during
school breaks by up to 46%, depending on the specification. We provide a series of empirical
results in support of the causal interpretation that the additional projects are driven by an increase
in the availability of slack time. Specifically, we show that when top engineering schools are on
break, we see a positive effect on technology projects but not art projects, and vice versa when
art and design schools are on break. We also examine days when universities are closed for snow,
providing an exogenous increase in slack time. Significantly more projects are posted on snow days.
Furthermore, we report results suggesting that the increase in postings and funding is unlikely to
be driven by increased capital availability due to students providing more funds while on break.
We also show that relative to work weeks, break weeks are associated with significantly more
projects at the very low end of the value distribution, fewer projects in the middle, and more
1
Crowdfunding is the practice of funding a project by raising small amounts of money from a large number of people
by leveraging a digital platform. It is a form of early-stage capital for many technology companies and has become an
important source of funding for the arts (Agrawal, Catalini, and Goldfarb 2013, Mollick and Nanda 2016, Belleflamme,
Lambert, and Schwienbacher 2014).
2
projects at the high end of the distribution. In terms of welfare, it is significant that a substantial
proportion of the latter category of extra projects are relatively large, high-value projects. The
empirical patterns also support the model predictions about the differences in effort levels: when
the opportunity cost of time is lower, more effort is allocated to ideas, especially high-value ones.
Lower opportunity cost time may also ease coordination among individuals working as a team.
In an extension of the baseline model, we explore the consequences of easier coordination on the
types of projects being developed, with the premise that complex projects may benefit disproportionately from teams with different skills and perspectives. We test and find support for this
prediction using data on the number of individuals involved in the projects and their skills.
Last, we exploit a Kickstarter policy change that increased the posting requirement for projects
in the Design and Technology categories (but not in other categories). This difference-in-differences
empirical strategy allows us to test the robustness of our baseline model to a change in the minimum
effort needed to post an idea. Results are consistent with the model prediction: after the change, the
relative share of top-valued and team projects in the affected categories increases disproportionately
during the breaks.
Our paper provides a formal framework for analyzing the relationship between the opportunity
cost of time and innovative outputs. Our results highlight that a sufficient amount of time (in
contiguous blocks, in particular) is critical for implementing high-potential and complex projects,
and that “overlapping slack” is important when ideas benefit from team collaboration and diverse
skills. At the same time, it cautions that too many marginal, low-potential projects might also be
generated when slack is introduced. Therefore, complementary policies mitigating such incentives
- such as stricter screening mechanisms - are likely to be valuable in conjunction with slack.
The paper proceeds as follows. The rest of this section summarizes the related literature.
Section 2 presents a formal model that derives the differences in the quantity and type of projects
developed during break and work periods. Section 3 describes the data and our empirical strategy.
Section 4 reports the regression results, and Section 5 concludes.
3
1.1
Related Literature
This paper links the nascent but rapidly growing literature on crowdfunding to the more established
literature on innovation. Mollick (2014) provides an early study documenting the funding dynamics,
and Agrawal, Catalini, and Goldfarb (2013) discuss a broad range of topics from the viewpoints of
both entrepreneurs and financiers, in particular on information-related frictions. A key finding of the
literature is that funding increases with accumulated capital. The crowd views accumulated capital
as a signal of quality, and this may lead to herding (Agrawal, Catalini, and Goldfarb 2015, Zhang
and Liu 2012, Kuppuswamy and Bayus 2013). One exception to this finding is Burtch, Ghose,
and Wattal (2013), who show that public-goods concerns can counteract herding effects. Mollick
and Nanda (2016) examine the “wisdom of the crowd” and show that overall the crowd appears to
select similar projects to experts. Also related to information, Burtch, Ghose, and Wattal (2015)
show that privacy concerns play a role in funding decisions. Instead of informational issues given
existing projects, our paper focuses on the determinants of the supply of projects and their quality.
Whereas the literature in management, economics, and sociology on innovation is extensive,
there have only been a few studies focusing explicitly on the impact of the opportunity cost of
time on inventive outcomes. Davis, Davis, and Hoisl (2014), using survey data on more than 3,000
inventions in Germany, find that leisure-time inventions by employees are more likely to focus on
conceptually-based problems, rely on interactions with people outside the organization, and focus on
smaller R&D projects. Harhoff and Hoisl (2007) find instead that leisure-time patented inventions
are more likely to be of higher value. A relevant stream of work in the management literature
studies the costs and benefits of organizational slack on innovation, often measuring slack in terms
of financial resources. In his seminal work on organizational design, Galbraith (1974) highlights the
key trade-offs that the creation of slack resources entails for a firm. Later proponents for slack argue
that it fosters a culture of experimentation and supports creativity and innovation (e.g., Bourgeois
(1981) and Singh (1986)). The opposing view argues that slack is an unnecessary cost that detracts
from the overall value of a company (e.g., Leibenstein (1969)). Using a combined measure derived
from two interview questions covering the extent of slack in terms of time and financial resources,
Nohria and Gulati (1996) show an inverse U-shaped relationship between slack and innovation in
4
a sample of 178 departmental managers from two large multinational corporations.2
Catalini (2016) also studies the role of low-opportunity cost time on invention, in his case from
a microgeographic perspective: co-location, by lowering search and execution cost for proximate
scientists, influences both the rate and direction of inventive activity.
Additionally, our results broadly resonate with studies on the impact of organizational design
on innovation such as the degree of decentralization versus centralization (Argyres and Silverman
2004, Lerner 2007, Arora, Belenzon, and Rios 2013) and the identification of constraints preventing
individuals from launching entrepreneurial projects (e.g., Chatterji and Bennett (2016)).
Finally, our focus on time relates to studies in sociology on time as social input. For example,
Perlow (1999) documents the potential negative impact of interactive activities for engineers because
they fragment individual concentration. Winship (2009) studies scheduling conflicts as barriers to
social relations. Young and Lim (2014) argue that time is a network good, the value of which
depends on an individual’s ability to coordinate that time with others.
2
Theoretical Framework
This section develops a simple model to explore the impact of breaks on innovation. Relative to
work periods, break periods are different in two important ways: 1) the opportunity cost of time
is lower, and 2) the coordination cost among collaborators is lower because people have more open
schedules and communication is less delayed. In the following, we first develop a baseline model
that considers only the first difference in the opportunity cost of time. We later extend the model
to allow for the use of teams and incorporate the second difference in coordination costs.3
2.1
Basic setup
One idea arrives in each period. The realized value of a project, v, depends on two factors: the
idea’s intrinsic quality, q, and the amount of effort (or, equivalently, time in our context) devoted
2
A number of studies in the product development literature highlight the impact of deadlines and scheduling on
new product development (Richtner and Ahlstrom 2006, Richtner, Ahlstrom, and Goffin 2014).
3
In the Appendix, we also provide an extension to a two period model in which projects arriving during the
work period can be “shelved” to the break. This model does not generate new predictions, but it proves useful in
interpreting the empirical results.
5
to developing the project, e. We assume that v = q · e. Thus, the realized value of a project
increases with both the idea’s intrinsic quality and the amount of effort. The innovator’s payoff
from developing an idea at a certain effort level is assumed to be:
πw = qe − c(e; w),
(1)
where c(e; w) is the cost of effort, with w ∈ {0 = break; 1 = work} indicating the time period.
Because the time cost is higher during work weeks, c′ (e; 0) < c′ (e; 1).
To obtain analytical results, we assume that idea quality q follows an exponential distribution
2
and that the cost function is convex in effort: c(e; w) = cw e2 , with c0 < c1 .4 The payoff in
equation (1) can be re-written as:
πw = qe − cw
e2
.
2
(2)
The first-order condition yields the optimal effort level as e∗w = q/cw . Thus, given any quality level,
less effort is spent when time is more precious; while given the same time cost, more effort is spent
when the idea is better. Inserting e∗w back into equation (2), we obtain the innovator’s payoff at
∗ (q) =
the optimal level as πw
q2
2cw .
The innovator needs to exert a minimum amount of effort, denoted as ē, to pass the posting
requirement of the funding platform. Denote the innovator’s payoff at the minimum required effort
2
level as π̄w (q) = qē − cw ē2 .
Taking into account the minimum effort requirement, the innovator’s payoff can be written as:
π̃w (q) =
where
✶{q≥cw ē} 2cq w + ✶{q<cw ē} (qē − cw ē2 ) if develop the project
2
2
0
,
(3)
if drop the project
✶{·} is an indicator function. If the innovator develops a project, she obtains πw∗ (q) =
q2
2cw
when the quality of the idea is sufficiently high (when q ≥ cw ē). In this quality range, the optimal
amount of effort is above the minimum requirement. For ideas of relatively low quality (when
4
We use an exponential distribution because it captures two features of our context: the support is positive, and
the distribution is skewed to the low end (that is, lower-quality ideas are more likely than higher-quality ones).
6
2
q < cw ē), the innovator’s payoff is π̄w (q) = qē − cw ē2 , because the optimal effort is below the
required level. In this range, the innovator over-develops the project (relative to the desired amount
of effort given the quality of the idea) to meet the minimum requirement and obtains a payoff that
is below the desired level. Despite this, the innovator still develops the project if π̄w (q) > 0, which
is the normalized outside option.
Figure 1a illustrates the innovator’s payoffs separately for break and work periods using a
numerical example. Given any idea quality q, the innovator’s payoff for a break is higher than
a work period. This is because the time cost is lower during breaks and, at the same time, the
project’s value is higher as more time is spent on it. It is straightforward to show that the innovator
develops an idea only if its quality is above a certain threshold and that the threshold for breaks
is less stringent.5 Thus, we have Prediction 1 (see Appendix A for the proof of this and other
results):6
Prediction 1 (Difference in quantity). More ideas are developed during a break period than a work
period.
The value of a developed project is:
ṽw (q) = ✶{q≥cw ē}
Notice that ideas with qualities q <
cw ē
2
q2
+ ✶{ cw ē <q<cw ē} qē.
2
cw
(4)
are dropped. Figure 1b illustrates the value of developed
projects as a function of idea quality. It highlights two consequences of a lower time cost on the
overall value of projects. First, because time is less costly, the innovator is less selective about
screening out marginal ideas. Thus, the value of some projects developed during breaks is lower
than any project developed during a work period. We call this the “selection effect.” Second,
also because time is less costly, the innovator works more on any given idea. This “effort effect”
increases a project’s value v, given the same intrinsic quality q.
2
5
The thresholds are determined by π̄w (q) = 0 ⇔ qē − cw ē2 = 0 ⇔ q̄w = cw2 ē . Thus, q̄0 < q̄1 as c0 < c1 .
Because it is empirically impossible to identify different arrival rates, we normalize them to be the same for
the two periods in the model. Allowing for different arrival rates affects only the first prediction because the other
predictions are normalized statements. Intuitively, the arrival rate is likely to be higher during the break because
there is more time to think about ideas. This would make the quantity of ideas developed during the break even
higher.
6
7
The comparison of the expected value of projects between a work and a break period is ambiguous because the two effects discussed above push it in opposite directions. We therefore compare
the distributions directly. Figure 1c plots the conditional densities of developed projects by their
value v using the same numerical example in Figure 1, with q following an exponential distribution. Compared to a work period, there are disproportionately more projects at the low end of
the distribution during a break because the innovator is less selective about screening out marginal
projects. At the same time, there are also more projects at the high end of the distribution because
the required idea quality to achieve the same project value is lower during a break due to the effort
effect.
Prediction 2 (Difference in distribution of project value, conditional on development). Relative
to a work period, projects developed during the break are both more likely to be at the low end of the
c1 ē2
2 )
distribution (v <
and at the high end (v ≥ c1 ē2 ).
The above result hinges on the assumption that more effort makes projects better. As illustrated
in Figure 1b, the innovator can achieve the same value v with a lower-quality idea during the break
because she can devote more time to the project. Prediction 3 builds on this observation.
Prediction 3 (Difference in effort). For projects with the same value v, the amount of effort spent
on projects developed during a break is greater than (or equal) to the amount of effort spent on
projects developed during a work period. Furthermore, the difference in effort is greater for projects
of higher value v.
During our sample period, Kickstarter increased the posting requirements for projects in certain
categories. We will explain the policy change in detail in Section 4.4, but in the context of our
model, this policy translates into an increase in the minimum effort requirement ē. Overall, a
more stringent posting requirement should increase the quality thresholds for both break and work
periods and, as a result, fewer projects should be developed. The relative size of the reductions,
however, is ambiguous. On the one hand, the increase in the threshold per se is greater for a work
period (i.e.,
∂ q̄1
∂ē
=
c1
2
>
∂ q̄0
∂ē
=
c0
2 ),
which is intuitive because increasing the effort requirement
is more taxing when time is more precious. On the other hand, because the distribution of idea
8
quality q is likely to be skewed to the left, the number of projects dropped may be greater during
breaks if these periods are disproportionately characterized by lower quality projects. The following
result shows that if the distribution of q is not too skewed to the left, the reduction in quantity is
smaller for breaks.
Prediction 4 (Effect of increasing minimum effort requirement ē on relative quantities). When
the distribution of idea quality, q, is not too skewed to the left (when the mean of the exponential
distribution is greater than
c1 ē
2 ),
the reduction in quantity due to an increase in the minimum effort
requirement is smaller for a break than for a work period.
For both break and work periods, the expected value of projects developed should be higher
after an increase in the posting requirement. The relative size of the increases is again ambiguous,
depending on the parameter values and the distribution of q. However, because this policy shock
does not affect projects of high quality, the share of top-valued projects developed during a break
(relative to the total number of projects developed during both periods) should increase. That is:
Prediction 5 (Effect of increasing minimum effort requirement ē on relative project value). The
share of top-valued projects developed during breaks (relative to the total number of projects developed
for both periods) increases after an increase in the minimum effort requirement.
2.2
An Extension: Team and Project Complexity
Projects may be carried out by teams for at least two reasons: 1) with multiple people sharing the
workload, a team may relax the capacity constraint of a single person, and 2) a complex project may
require people with different, complementary skills (or perspectives on a problem). The downside
of a team, however, is that it requires coordination, and this coordination requirement may be
especially high for complex projects. In this section, we extend the baseline model to allow for the
use of teams and incorporate a second difference between break and work periods: coordination
costs among team members are lower during breaks.
In addition to the baseline setup, this extended model assumes the following. First, a project
can be developed solo, by a team of two, or dropped. Second, in team projects, each individual’s
9
payoff consists of half of the value of the project, minus the individual time cost and a coordination
cost. For simplicity, the coordination cost is normalized to be zero for simple projects regardless of
the time period. Complex projects involve a positive coordination cost, and this cost is greater for
work periods. Third, whereas in simple projects teaming up does not bring any direct benefit to
the project’s value, in complex projects there is a positive effect (e.g., from complementary skills
and perspectives), and this benefit is more likely to be pronounced as idea quality increases.
Appendix A2 provides details on the extended model and the decision rules separately for
simple and complex projects. The basic trade-offs are as follows. For simple projects, there are no
coordination costs, and the benefit from using a team is limited to the ability to share the workload.
In our model, this benefit matters only for ideas of relatively low quality (for which the minimum
effort requirement is too high given the idea’s quality). For complex projects, while teaming up
incurs non-trivial coordination costs, it also introduces two benefits: workload sharing and a direct
increase in project value. The second, direct benefit of teaming up is more relevant for ideas of
relatively high quality because for these ideas such a benefit is likely to outweigh coordination costs.
As in simple projects, teams also help share the workload when the minimum effort requirement
is too high relative to the idea’s quality. However, for complex projects we may not see the use of
teams for this range of ideas if coordination costs are too high.
A key observation is that, relative to work periods, breaks offer a relative advantage for developing complex projects because teamwork is easier. Such an advantage becomes more salient as
the discrepancy between coordination costs becomes larger. In contrast, this relative advantage of
breaks is much less relevant for simple projects as they require little coordination. Thus, we have
the following prediction:
Prediction 6 (Difference in project complexity, conditional on development). When the coordination cost for complex projects is sufficiently low during breaks but sufficiently high during work
periods, the probability of complex projects, conditional on development, is higher during a break
than a work period.
An important variable observable in the data is whether a project is developed by a team rather
than a single person. In general, it is ambiguous whether the conditional probability of using a team
10
should be higher during breaks, in part because teaming up also helps share the workload, and this
is particularly useful for work periods when time is more constrained. Since breaks are particularly
useful for developing complex projects, the comparison is clearer when separating simple from
complex projects:
Prediction 7 (Difference in probability of team projects, conditional on development). When
the coordination cost for complex projects is sufficiently low during breaks but sufficiently high
during work periods compared to simple projects that require little coordination, complex projects
are relatively more likely to be developed by a team during a break than a work period.
3
Data and Empirical Strategy
Our empirical setting is Kickstarter, the world-leading reward-based crowdfunding platform. We
collect data for the 165,410 US-based projects that attempted to raise money on the platform
between April 2009 and April 2015. Project-level information that is publicly available includes
the time each project was posted, total funds raised, descriptive information of the project, and the
background of the entrepreneur(s). We do not have comprehensive data on the timing of individual
contributions within a project nor the location of the funders.7
Kickstarter requires projects to state a funding goal in advance. If a project fails to achieve
its funding goal, then the capital is returned to funders and the entrepreneur does not receive
any money. We label projects that achieve their funding goal as “successful” and ones that do
not as “failed.” Our sample projects collectively raised $1.4B by both successful projects (44%
of the projects, 89% of the raised capital) and failed projects (56% of the projects, 11% of the
raised capital, though it is eventually returned). As previously documented (Agrawal, Catalini,
and Goldfarb 2013), the distribution of capital is highly skewed: the top 1% (10%) of projects
accounts for $590M ($1B), or 42% (76%) of the capital.
Kickstarter identifies a city for each project, based on the location reported by the entrepreneurs.
It provides a smaller geographic measure than a core-based statistical area (CBSA) and a larger
7
The campaign data are obtained from Kickspy, one of the top websites that monitors live Kickstarter activity
over time. This data source is often used in papers studying crowdfunding because it is comprehensive in terms of
the variables it captures.
11
one than a Census Place. We use city as the location measure in our analysis because it does not
involve further aggregation or disaggregation of the core dependent variables. Our sample projects
span 10,091 US cities, taken from Kickstarter’s list on its website.
Kickstarter is widely used in many college towns. For example, Boulder Colorado, Provo Utah,
and Ann Arbor Michigan are among the top 10 places for technology projects per capita. In light
of the prominence of locations with colleges on Kickstarter, our empirical analysis exploits the
week-by-week variation in the extent of slack time in these places. We manually collect data on
school breaks between 2009 and 2015 for the top 200 US colleges as defined by US News & World
Report.8 This information is publicly available through posted academic calendars. We consider
a location to have a school break in a given week if there is a top 200 college within five miles of
the city center and the college has a break that week. In addition to scheduled breaks, we also
manually collect data on snow days from school websites, Twitter, and online news reports.
Table 1 presents descriptive statistics at the city-week level for our sample. Approximately
3.3% of our observations are city-weeks where at least one of the top 200 colleges is on break, and
Table 2 compares key outcome variables between break weeks and work weeks, providing model-free
evidence for some of the basic predictions of our model.
During our study period, the average city-week had 0.052 projects launched, with slightly more
than half (0.029) failing to reach their goal and the remainder (0.023) successfully reaching it. In
most city-weeks, no projects are launched. Relative to work weeks, the number of projects per
week is significantly higher for breaks (0.39 versus 0.04).9
We use three different proxies for effort in our analysis: the length of the text (in words) used
to describe a project, the number of frequently asked questions (FAQs) the entrepreneur lists on
the page, and the number of project updates posted. Relative to work weeks, the descriptive
statistics show that projects posted during the breaks have longer descriptions, more FAQs, and
more updates.
8
Source: http://colleges.usnews.rankingsandreviews.com/best-colleges/rankings/national-universities/data (accessed September 2013).
9
In addition, the descriptive statistics show that the maximum number of successful projects for any US city in a
single week is 45 (Los Angeles, CA), whereas the maximum number of failed projects is 79 (also Los Angeles). In a
single week, cities are able to attract as much as $21.8M in successful funds, with an average per city-week of $445
and a standard deviation of $23,633.
12
We capture team projects by identifying the number of team members listed on a project’s
biography pages. Overall, 3% of the projects are posted by teams (usually consisting of two people).
Relative to work weeks, a significantly greater percentage of projects posted during the breaks use
teams (12.8% versus 2.7%). Using biographies, we also identify the skills mentioned by creators in
their profiles (e.g., programming, photography, etc.). We use this measure to build two proxies for
project complexity: the number of unique skills listed, and the number of unique skills divided by
the size of the team (to avoid a mechanical relationship between team size and the total number
of unique skills). Among projects posted, the mean number of skills is 1.5 and the mean number
of skills per team member is 1.2. These values are higher for projects posted during breaks.
Overall, the raw data are consistent with the basic predictions of our model: break weeks are
characterized by significantly more projects, more effort being devoted to the posted projects, an
increase in project complexity and the use of teams. However, because these differences might
be driven by a number of confounding factors such as time trends, seasonality, and differences in
the baseline level of crowdfunding activity across cities, we control for these differences using the
empirical strategy described below.
3.1
Empirical Strategy
Our econometric framework is straightforward: we exploit variation across cities in the timing of
breaks at local colleges and universities to estimate how an increase in slack time influences the
number and type of projects posted. The unit of analysis is a city-week. City fixed effects are
included to control for underlying differences across cities that are constant over time, and week
fixed effects are added to control for changes in the Kickstarter environment over time. We focus
on a linear model with fixed effects to document the underlying relationships in a direct and easily
interpretable manner:
Yct = γAllBreaksct + µc + ψt + ǫct ,
13
where Yct is the outcome variable such as the number of projects posted on the platform in city c
during week t,10 AllBreakct is a dummy that equals one if there is a top 200 college in the focal city
on break in the focal week, µc and ψt are city and week fixed effects, and ǫct is an idiosyncratic error
term. There are few city-level measures that change at the week level. Unsurprisingly, given city
and week fixed effects, results do not change when including controls such as weekly temperature
and annual CBSA-level demographics. Thus, we focus on the more parsimonious specification and
do not include additional covariates. Because the fixed effects completely capture cities in which we
never see any crowdfunding activity, we remove these cities from the analysis without any empirical
consequences. Robust standard errors are clustered at the city level for all regressions.
4
Results
We first present the regression results on the quantity of projects posted. The result, consistent
with the raw data, shows that a substantially greater number of projects are posted during break
weeks. We follow this main result with evidence supporting a casual interpretation. We then
report results on the impact of breaks on the distribution of project value because the welfare
implications are different depending on whether these extra projects are mostly marginal ones with
limited potential or if they include large, high-value projects. Third, we examine the differences
in project complexity and the likelihood of team projects during break weeks versus work weeks.
The results are informative about the role breaks play in reducing coordination costs. Finally,
we exploit a Kickstarter policy change that increased the posting requirement in some but not all
project categories to further test our interpretation of data and mechanisms.
4.1
Quantity of Ideas Developed (Prediction 1)
Column 1 of Table 3 shows that, controlling for week and city fixed effects, the number of crowdfunding campaigns launched increases during school breaks by 0.024 (significant at the 1% level). The
10
Other dependent variables include the total amount raised by the projects posted in city c during week t, the
average values of certain characteristics over all projects posted in city c during week t, such as the amount of effort
invested into a project (description length, number of FAQs, number of updates), the number of unique skills, and
team size.
14
magnitude of the increase is substantial relative to the average number of campaigns per city-week
of 0.052 (first row in Table 1), implying an increase of 46%. In the Appendix, we show robustness to
focusing on cities with more Kickstarter activity, yielding larger coefficients but smaller percentage
increases (20-30%).
In terms of the total amount of funds raised, we use funding (not logged) as the dependent
variable in Column 2 of Table 3. The estimated coefficient on school breaks is also positive but
not significantly different from zero. This is likely because the distribution of funding is highly
skewed. Thus, we use log(Funding+1) as the dependent variable in Column 3. The result shows
a positive and significant correlation between total funding and college breaks, with an increase of
about 1.5%.11
Evidence for a causal interpretation
The mechanism behind Prediction 1 of our model suggests that the positive correlation between
the quantity of projects and breaks is causal: Breaks enable more projects to be developed because
the opportunity cost of time is lower. By including city and week fixed effects, our basic empirical
strategy helps isolate potential spurious correlation due to general time trends, seasonality, or timeconstant differences in innovative activity across cities. In the rest of this subsection, we provide
additional evidence in support of a causal interpretation of this relationship. In later subsections,
our results on outcome variables other than quantity provide further support for the mechanisms
introduced in the model.
First, we exploit variation across types of universities and Kickstarter categories to examine
whether the spike in activity is consistent with the type of human capital involved. We do this
because, although we obtain variation in breaks using university-level data, we measure activity at
the city level. Thus, demonstrating that the city-level effect (e.g., more technical projects posted
on Kickstarter) is consistent with local university-level human capital (e.g., break week for an
11
As with many other quasi-experimental regression papers (e.g., Athey and Stern (2002) and Simcoe and Waguespack (2011)), the R-squared in the analysis in this table is low. This is not surprising given that city fixed effects
are differenced out rather than estimated and that there are many reasons why people post projects on Kickstarter
besides having time during college breaks. Key for our conclusions is that our coefficient estimates have statistical
power and magnitudes of economic importance.
15
engineering school as opposed to an art school) provides further evidence that is consistent with
our interpretation. In Column 1 of Table 4, we use only projects in art, and in Column 2 we focus on
projects in technology.12 Breaks at top art, design, film, and theater schools are positively correlated
with art projects but not technology projects. Conversely, breaks at top engineering schools are
positively correlated with technology but not art projects, consistent with our expectation that
technical orientation plays a key role in these types of projects. Relatedly, in Column 3 of Table
4, we run our main specification only for projects where we are able to identify the presence of a
student on the team using the biographies posted on Kickstarter. Whereas the effect is smaller
in magnitude (possibly because of our inability to capture all student projects), it is positive and
statistically significant.
Second, we examine the impact of an unexpected increase in the amount of slack time using
college closings due to heavy snow. Column 4 of Table 4 shows that snow breaks are also associated
with a significant and positive increase in the quantity of projects. In addition, Column 5 uses the
city-year as the unit of observation and regresses the number of projects on the total number of
snow days in that city-year, controlling for city and year fixed effects. Given that the number of
days that a college cancels classes due to snow varies exogenously from year to year and city to
city, the positive and significant coefficient is consistent with such random increases in slack time
having a positive causal impact on the number of projects produced.
Third, if breaks provide college students with more time to browse Kickstarter and fund projects,
including those by local entrepreneurs, then our result may be driven by an increase in the supply
of local capital. To assess this alternative explanation, we examine the timing of funding associated
with projects posted in the week preceding a break. In particular, we measure whether funding
for those projects increases during the break. For consistency, we specifically look at 8-14 days
after posting. If the effect we measure is due to increased funding during a break, then we would
expect to observe an increase in funding during break weeks for projects posted just prior to the
break. However, results reported in Columns 6 and 7 of Table 4 indicate that funding associated
with projects started during a work week does not rise during break weeks even though fundraising
12
Our sample projects span the main 13 categories defined by Kickstarter, including Art, Comics, Dance, Design,
Fashion, Film & Video, Food, Games, Music, Photography, Publishing, Technology, and Theater.
16
efforts continue during those weeks.
It is important to highlight that the breaks we explore in our main analysis are scheduled
and expected, and therefore people might shelve ideas and wait until they have more time during
breaks.13 While it is empirically difficult to capture shelving, for a subset of projects we are able
to collect information on when the creator started working on the idea. Excluding these projects
that have been clearly shelved does not affect our results.14
We also do not observe strong evidence for shelving in the data. Figure 2 plots the regression
results that include dummies for the weeks before and after a college break period in our main
specification, and the baseline is any week more than five weeks away from a college break. The
figure shows no decrease in activity immediately before or after the break relative to the baseline
level. In addition, we also do not observe a significant decreasing trend prior to the breaks. Even
though we cannot fully rule out shelving, the lack of any pre-trend at least provides no evidence for
a natural way we would expect shelving to happen; that is, if people shift ideas anticipating breaks,
we would expect that there is a greater drop immediately before the breaks (because one does not
lose much by delaying just a few days) than semester weeks that are farther ahead of the breaks.
It is important to note that shelving per se does not invalidate the causal impact of breaks. The
notion that people “need” to delay until breaks to work on certain projects implies that a sufficient
amount of slack time is critical to materialize these ideas. In other words, without breaks, some of
these projects may not be generated at all.15
13
Notice that snow breaks, despite being unexpected, do not provide a good test for the presence of shelving because
they take place in the same context, where individuals might shelve projects for anticipated breaks.
14
In Appendix Table B-6, when we run our analysis excluding the subset of projects that were clearly shelved
(Column 3), our estimate is identical to the one for the full sample. Furthermore, the effect of breaks on shelved
projects (Column 2) is very small in magnitude. Shelved projects also seem to appear across all weeks of the breaks
(Column 8) in a fairly homogeneous way, i.e., they do not seem to disproportionally appear at the beginning of a
break (as one would expect if they were driving the results). Furthermore, even when looking at heterogeneity across
breaks of different length and within weeks of the same break, excluding shelved projects does not alter our results
(Columns 7 versus 9).
15
An important caveat is that the exact impact of breaks on the quantity and quality of projects would be different
with and without the ability to shelve projects. However, assigning a direction or magnitude to this difference is not
obvious. In an extension of the baseline model (see Appendix A3 in the paper for details), we allow projects to be
shelved with a discount factor on the project’s value. The tradeoff is that if a project is delayed until a break, the
time cost to implement it is eventually smaller but at the same time the project may lose some of its value (e.g., due
to competing ideas, lack of patience, forgetfulness, loss of momentum). The only situation in which there is some
real tension is when the discount factor is in an intermediate range. In this range, the entrepreneur would choose to
immediately develop very good ideas and delay relatively poor ones (as ideas of the lowest value are dropped). This
is because when the idea quality is high, the entrepreneur would not want to sacrifice its potential by delaying it,
17
4.2
Project Value Distribution and Effort (Predictions 2 & 3)
In the previous section, we show that when college students are on break in a city, there are substantially more projects posted and more funds raised on Kickstarter. The increase in quantity
may have different welfare implications depending on the quality of these additional projects. Prediction 2 shows that relative to work weeks, during breaks we should observe disproportionately
more projects at both the low end and the high end of the value distribution. This prediction arises
from the lower opportunity cost time during breaks which induces: 1) more marginal projects to
be pursued because time is less precious (a selection effect), and 2) more effort to be allocated to
the development of an idea, which in turn increases the project’s value (an effort effect).
To examine the impact of breaks on the distribution of project value, we divide projects into 15
bins based on the amount of funds raised. The bins are defined by year to account for the growth
of the platform. Table 5 reports the results from a regression in which the dependent variable is
the number of projects in a specific bin, the unit of analysis is a city-week bin, and fixed effects for
weeks and cities are included as in previous tables. The results are consistent with Prediction 2:
relative to work weeks, breaks are characterized by significantly more projects of the lowest value
(Bin 1), fewer projects in the middle of the distribution (Bins 2 to 6), and more projects again on
the right tail of the distribution (Bins 7 to 15). Furthermore, the effects become increasingly larger
and more significant as we move away from the middle of the distribution.16
Recall that the model assumes that more effort makes projects better, which explains why we
observe a positive effect of breaks on the right tail of the project value distribution. It is therefore
important to see whether the data are consistent with this mechanism. In particular, Prediction 3
shows that we should observe more effort during breaks and that the difference in effort should be
larger for projects of higher value. This prediction is confirmed by the results reported in Table 6.
First, Columns 1 to 3 show that the average values of the three different measures of effort (length
of project description, number of FAQs, and number of updates) are all significantly higher for
and the potential is likely to be large compared to the benefit from the time cost savings. This behavior would push
against us finding that a substantial amount of the additional projects posted during breaks are better, high-value
projects. In the alternative two scenarios, one choice always dominates the other regardless of the intrinsic quality
of ideas.
16
In Appendix Table B-4, we show similar results when segmenting the distribution, further allowing for more bins.
18
projects posted during breaks. Second, when we separate projects into different bins of the value
distribution, results show that for all measures of effort, the difference between break and work
weeks generally increases as we move from the lowest bins (Columns 4, 7, and 10) to the highest
ones (Columns 6, 9, and 12).17
To summarize, the results of this section are consistent with the model’s prediction that we
should observe disproportionately more projects on both tails of the value distribution. From a
welfare perspective, the right tail is particularly interesting because these projects are more likely
to have a disproportionate impact on the economy. The results on effort seem to corroborate the
mechanism we hypothesized for the creation of these high-impact projects, i.e., entrepreneurs are
able to devote more effort to the execution of their ideas when time is less costly.
4.3
Teams and Project Complexity (Predictions 6 & 7)
Lower opportunity cost time may also make it easier for team members to coordinate their schedules
and execute on their ideas. This relative advantage of breaks should have a disproportionate impact
for complex projects, especially if these projects are more likely to benefit from different perspectives
and skills.
Prediction 6 of our model shows that when the coordination cost for complex projects is sufficiently high during work weeks but sufficiently low during breaks, we should observe an increase
in the share of complex projects during breaks.18 As for the likelihood of teams, the theory is
ambiguous because teaming up can also help share the workload across multiple people, and this
effect could be more valuable when time is more constrained (i.e., during work weeks). However, if
we contrast complex projects to simpler ones, we should be more likely to observe a team forming
around a complex project during a break than during a work week (Prediction 7).
The results in this section are consistent with these predictions. As explained above, we use
two proxies for project complexity: the number of unique skills on a project (as captured from
entrepreneur biographies), and the number of unique skills divided by team size. Columns 1 and
17
In the Appendix, we further focus on the right tail of the distribution by limiting the analysis to successful
projects (Table B-10): again, as project value increases, effort levels during the breaks increase too.
18
We turn to Predictions 4 and 5 in Section 4.4 in examining the impact of an increase in minimum effort.
19
2 of Table 7 show that for both measures, projects during breaks appear to be significantly more
complex than projects posted during work weeks.
Second, consistent with the raw data, the regression results show that, on average, projects
posted during breaks are significantly more likely to be carried out by teams than projects posted
during work weeks (Column 3 of Table 7).
Last, we further separate the sample into relatively complex projects versus simpler projects
based on whether a project is above or below the median of our measures of complexity. Consistent
with Prediction 7, the results show that for relatively complex projects (Columns 5 and 6), the
likelihood of teams is significantly higher during breaks. In contrast, for simpler projects (Columns
7 and 8), there is no significant difference between break and work periods in the likelihood of using
a team.
4.4
Effects of Increasing the Minimum Effort Requirement (Predictions 4 & 5)
In May 2012, Kickstarter implemented a policy that disproportionately increased requirements
to post projects in the Design and Technology categories (e.g., preparing a manufacturing plan,
building a working prototype, describing them clearly on the site). We explain the policy change
in greater detail below. In the context of our model, the new policy increases the minimum effort
required to develop a project. Whereas it is ambiguous whether the relative share of projects posted
during breaks should increase after the policy change (Prediction 4), the model predicts that we
should observe an increase in the share of top-valued projects (Prediction 5).
In response to a series of high-profile Design and Technology projects that raised a significant
amount of capital and then failed to deliver the promised products in the anticipated amount of time
(some delivered very late and others failed to deliver at all), Kickstarter increased the requirements
for posting projects in those two categories. One of the primary criticisms was that in most of
these cases the entrepreneur raised capital and promised funders a product without any experience
or preparation for production or distribution. Therefore, in May 2012 Kickstarter revised its rules
to address accountability concerns, explaining in a note that it now requires project creators to
“provide information about their background and experience, a manufacturing plan (for hardware
20
projects), and a functional prototype. [Kickstarter] made this change to ensure that creators have
done their research before launching and backers have sufficient information when deciding whether
to back these projects.”19 The policy was further reinforced in September 2012, when Kickstarter
reiterated the need for a prototype and made it clear that product simulations and renderings were
not enough for posting.20
The regression results in Table 8 contrast the relative changes in a number of key outcome
variables for break weeks versus work weeks before versus after the policy shock for projects in
the Design and Technology categories compared to the other 11 categories. First, for the relative
changes in the quantity of projects, we use a regression in which the dependent variable is the
ratio between the number of projects in Design and Technology and the number of projects in all
other categories (Column 2 of Table 8). The coefficient of the interaction term between breaks
and the period following the policy change is positive and significant. This suggests that when
comparing Design and Technology to the other categories, after the shock breaks are associated
with an increase in the relative number of projects. Column 4 of Table 8 shows a similar result for
the relative amount of funding.21
Second, for project value, Column 6 of Table 8 reports results from a regression in which
the dependent variable is the ratio between the share of projects with large target amounts (over
$30,000) in Design and Technology and the share of such projects in the other categories. Consistent
with Prediction 5, the coefficient of the interaction term shows that, after the policy change, relative
to other categories, the share of large-goal projects significantly increased during break weeks in
the Design and Technology categories.
Finally, Column 8 of Table 8 shows that after the policy change, the relative likelihood of using
teams during breaks increases disproportionately for Design and Technology projects relative to
other projects. Intuitively, this result is consistent with breaks facilitating coordination in teams. A
more stringent posting requirement makes teams relatively more desirable because multiple people
19
https://www.kickstarter.com/blog/accountability-on-kickstarter (Accessed on March 12, 2015)
https://www.kickstarter.com/blog/kickstarter-is-not-a-store (Accessed on March 12, 2015)
21
Columns 1 and 3 provide a baseline level comparison between break and work periods for the relative number
of projects (or the relative amount of total funds) in Design and Technology categories with respect to all the other
categories.
20
21
can share the workload. Compared to work weeks, innovators are better able to respond to the
requirement change by working in teams during break weeks because coordination costs are lower.
5
Conclusion
We explore how the opportunity cost of time influences the quantity, quality, and type of innovation.
We develop a simple theoretical framework to capture the differences between periods of high versus
low opportunity cost of time (work weeks versus break weeks). Lower opportunity cost time during
break weeks may induce: 1) lower quality ideas to be developed (a selection effect); 2) more effort
to be applied for any given idea quality (an effort effect); and 3) an increase in the use of teams
because scheduling is less constrained (a coordination effect). As a result, the effect of an increase
in slack time on inventive outcomes is ambiguous because the selection effect may induce more low
quality ideas, whereas the effort and coordination effect may lead to more high quality, complex
ideas.
Our model, together with the empirical support from data on crowdfunding activity and college
breaks, generates the following results. First, during breaks, more projects are posted on Kickstarter
in the immediate region near to the colleges. Second, not all of these projects are marginal ones. In
fact, we observe an increase on both tails of the project value distribution. Higher value projects
seem to be linked to the ability of entrepreneurs to dedicate more effort to their ideas when the
opportunity cost of time is low. Finally, overlapping slack time appears to ease team coordination,
leading to more complex projects and more projects being developed by teams rather than lone
innovators during break weeks.
A key limitation of our paper is that the results are obtained from an idiosyncratic setting
(crowdfunding) and population (college students). Within that setting, our estimated effects are
large: 46% more projects are posted during breaks. We are also only able to imperfectly measure
project value, complexity, and the shelving of ideas in expectation of a future break.
That said, we interpret the empirical results as providing support for the insights of the model.
These insights are general and inform policies targeted at encouraging innovation and experimentation within firms and other organizations.
22
First, low opportunity cost time may help develop high potential ideas. Second, “overlapping
slack” might be more valuable for innovating on complex ideas than individual slack because of
the way it allows teams to coordinate around their development. Anecdotally, Google, which used
to provide 20% slack to its employees to work on novel ideas, recently restructured the policy as
“Area 120.”22 According to Google’s CEO, Area 120 “is giving people a chance at 20% time more
formally” by giving them longer spells of continuous time (up to multiple months), as well as the
possibility to build teams across the organization and work together at the same time on ideas with
potential. Third, in the absence of other, complementary incentives, slack time may induce more
marginal, low-value projects. This suggests that companies interested in increasing experimentation
may need to combine these policies with incentives targeted at avoiding marginal contributions.
In the case of LinkedIn’s [in]cubator program, not only are projects screened by top executives
before entry, but the company has adopted a milestone-based approach to ensure the quality of the
projects at different stages in their evolution.23 In this way, the insights from the model, supported
by the empirical analysis, can help us understand the opportunities and challenges in company
policies around slack time.
22
Source:
http://www.forbes.com/sites/miguelhelft/2016/05/19/google-ceo-sundar-pichai-confirms-area-120corporate-incubator/#409a3dc44f73.
23
According to the company, ideas must first be developed into prototypes and clear two rounds of judging, the
final round run directly by Founder Reid Hoffman and CEO Jeff Weiner. Given the stakes and visibility of these
judging rounds, it is clear that employees have strong incentives to perform. Only projects showing promise at 30
days graduate to become 60-day projects, and only those that have made sufficient additional progress at Day 60 are
allowed another 30-day extension. Source: https://blog.linkedin.com/2012/12/07/linkedin-incubator.
23
References
Agrawal, A., C. Catalini, and A. Goldfarb (2013): Some Simple Economics of Crowdfunding,
in Innovation Policy and the Economy, Volume 14 - Josh Lerner and Scott Stern editors. NBER,
University of Chicago Press.
(2015): “Crowdfunding: Geography, Social Networks, and the Timing of Investment
Decisions -,” Journal of Economics and Management Strategy (forthcoming).
Argyres, N., and B. Silverman (2004): “R&D, organization structure, and the development of
corporate technological knowledge,” Strategic Management Journal, 25, 929–958.
Arora, A., S. Belenzon, and L. A. Rios (2013): “Make, buy, organize: The interplay between
research, external knowledge, and firm structure,” Strategic Management Journal, 35, 317–337.
Athey, S., and S. Stern (2002): “The Impact of Information Technology on Emergency Health
Care Outcomes,” RAND Journal of Economics, 33(3), 399–432.
Belleflamme, P., T. Lambert, and A. Schwienbacher (2014): “Crowdfunding: Tapping the
right crowd,” Journal of Business Venturing, 29(5), 585–609.
Bourgeois, L. J. (1981): “On the measurement of organizational slack,” Academy of Management
Review, 6, 29–39.
Burtch, G., A. Ghose, and S. Wattal (2013): “An Empirical Examination of the Antecedents
and Consequences of Contribution Patterns in Crowd-Funded Markets,” Information Systems
Research, 24(3), 499–519.
(2015): “The Hidden Cost of Accommodating Crowdfunder Privacy Preferences: A Randomized Field Experiment,” Management Science (forthcoming).
Catalini, C. (2016): “Microgeography and the Direction of Inventive Activity,” MIT Working
Paper.
Chatterji, A., and V. Bennett (2016): “Why Don’t Startups Start?: Evidence from a Nationally Representative Survey,” Working Paper.
Davis, L. N., J. D. Davis, and K. Hoisl (2014): “Leisure Time Invention,” Organization Science,
24(5), 1439–1458.
Galbraith, J. (1974): “Organization Design: An Information Processing View,” Interfaces, 4,
28–36.
Harhoff, D., and K. Hoisl (2007): “Institutionalized incentives for ingenuity—Patent value and
the German Employees’ Invention Act,” Research Policy, 36, 1143–1162.
Kuppuswamy, V., and B. L. Bayus (2013):
“Crowdfunding Creative Ideas:
the Dynamics of Projects Backers in Kickstarter,”
SSRN Working Paper,
http://papers.ssrn.com/sol3/papers.cfm?abstract id=2234765.
Leibenstein, H. (1969): “Organizational or Frictional Equilibria, X-efficiency, and the Rate of
Innovation,” Quarterly Journal of Economics, 83, 600–623.
24
Lerner, J, W. J. (2007): “Innovation and incentives: evidence from corporate R&D,” Review of
Economics and Statistics, 89, 634–644.
Mollick, E. R. (2014): “The Dynamics of Crowdfunding: An Exploratory Study,” Journal of
Business Venturing, 29(1), 1–16.
Mollick, E. R., and R. Nanda (2016): “Wisdom or Madness? Comparing Crowds with Expert
Evaluation in Funding the Arts,” Management Science, 62, 1533–1553.
Nohria, N., and R. Gulati (1996): “Is Slack Good or Bad for Innovation?,” Academy of Management Journal, 39(5), 1245–1264.
Perlow, L. A. (1999): “The time famine: Toward a sociology of work time,” Administrative
science quarterly, 44, 57–81.
Richtner, A., and P. Ahlstrom (2006): “Influences on organizational slack in new product
development projects,” International Journal of Innovation Management, 10, 375–406.
Richtner, A., P. Ahlstrom, and K. Goffin (2014): “Squeezing R&D: A Study of Organizational Slack and Knowledge Creation in NPD, Using the SECI Model,” Journal of Product
Innovation Management, 31(6), 2168–1290.
Simcoe, T., and D. Waguespack (2011): “Status, Quality and Attention: What’s in a (Missing)
Name?,” Management Science, 57(2), 274–290.
Singh, J. V. (1986): “Performance, slack and risk-taking in organizational decision making,”
Academy of Management Journal, 29, 562–585.
Winship, C. (2009): “Time and Scheduling,” in Oxford Handbook of Analytical Sociology, ed. by
P. Hedstr´’om, and P. Bearman. Oxford University Press.
Young, C., and C. Lim (2014): “Time as a Network Good: Evidence from Unemployment and
the Standard Workweek,” Sociological Science, 1, 10–27.
Zhang, J., and P. Liu (2012): “Rational Herding in Microloan Markets,” Management Science,
58(5), 892–912.
25
6
Tables and Figures
Figure 1: A Numerical Illustration of the Model
(a) Innovator’s payoff
(b) Value of developed projects
60
120
Break week
Work week
Break
Work
50
100
40
project value, v
innovator payoff, :
80
30
20
60
40
10
20
0
-10
0
1
c0 ē
2
2
c1 ē 3 c0 ē
2
4
5
0
6 c1 ē
7
8
9
10
0
1
2
3
4
idea quality, q
5
6
7
8
9
10
idea quality, q
(c) Density distribution of project value, conditional on
development
2
Break week
Work week
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
1
1.5
2
2.5
3
3.5
4
4.5
5
log(project value, v)
Notes: (a) plots the innovator’s payoff from developing a project as a function of the idea’s intrinsic quality, q. (b) plots the
value of developed projects as a function of the idea’s intrinsic quality, q. (c) plots the probability density distributions of v
conditional on development. The parameter values of the numerical example are: c0 = 1; c1 = 2; and ē = 3. In addition, for
(c), idea quality q is assumed to follow an exponential distribution with mean 1.5.
26
Figure 2: The Number of Projects per Week is Higher During Breaks
Notes: Estimated coefficient for weeks before and after the school breaks. Dependent variable is the number of projects
created in the focal city-week. Regression includes city and week fixed effects. Error bars represent 95% confidence intervals
based on robust standard errors clustered at the city level.
27
Table 1: Summary Statistics
Projects
Total Successful Projects
Total Failed Projects
N
3,178,665
3,178,665
3,178,665
Mean
.052
.023
.029
SD
.761
.398
.425
Min
0
0
0
Max
114
45
79
Funding
Total Successful Funding
Total Failed Funding
3,178,665
3,178,665
3,178,665
444.686
393.86
50.821
23633.61
23149.05
1784.375
0
0
0
21.8M
21.8M
62,3041
All Breaks
Snow Closing
3,178,665
3,178,665
.033
.001
.178
.034
0
0
1
1
Table 2: Comparing Break and Work Periods
Projects
Funding
Log Funding
Difference
.356
3700.27
.613
SE
.002
74.33
.004
p-value
0
0
0
28
N
Break
104,440
104,440
104,440
Mean
Break
.396
4023.38
.755
N
Work
3,074,225
3,074,225
3,074,225
Mean
Work
.040
323.11
.141
Table 3: Impact of Breaks on the Quantity of Projects
(1)
Projects
(2)
Funding
(3)
Log Funding
All Breaks
0.0238***
(0.0083)
20.9730
(259.5038)
0.0155***
(0.0039)
Observations
R-squared
Number of Cities
3,178,665
0.164
10,091
3,178,665
0.003
10,091
3,178,665
0.787
10,091
VARIABLES
Notes: The dependent variables are, respectively, the number of projects in a city-week, the total amount of funding, and
log(Funding +1). City and week fixed effects are included in all regressions. Robust standard errors are clustered at the city
level. *** p < 0.01, ** p < 0.05, * p < 0.1.
Table 4: Impact of Breaks on the Quantity of Projects: Robustness Checks
(1)
Artistic
Projects
(2)
Technology
Projects
(3)
Student
Projects
(4)
All
Projects
All Breaks
0.0020
(0.0013)
0.0003
(0.0004)
0.0034**
(0.0017)
0.0495***
(0.0121)
Break at Top Engineering School
-0.0007
(0.0032)
0.0053*
(0.0028)
0.0024**
(0.0011)
0.0003
(0.0009)
VARIABLES
Break at Top Art, Design,
Film, Theatre School
Snow Closing
Observations
R-squared
Number of Cities
1,604,469
0.001
10,091
1,604,469
0.000
10,091
3,178,665
0.003
(5)
All
Projects
(Annual Data)
0.0403**
(0.0178)
1.2407***
(0.4758)
2,966,754
0.008
10,091
50,455
0.013
10,091
(6)
Funding
Days 8-14
(7)
Log Funding
Days 8-14
96.3529
(88.8682)
0.0256
(0.0162)
1,321,921
0.000
10,091
1,321,921
0.002
10,091
Notes: The dependent variable is Column (1) is the number of art-intensive projects, and in Column (2) it is the number of
technology-intensive projects (available May 2012 to April 2015). In Column (3) the dependent variable only includes projects
where the creators explicitly mention in their biography their student affiliation. Columns (4) and (5) use data on snow
closing (available until December 2014). The unit of analysis for Column (5) is the city-year level. Columns (6) and (7) use
data on weekly investments (available from October 2012 to April 2015). City and week fixed effects are included in all
regressions (with the exception of Column (5), which uses Year Fixed Effects because of the different unit of analysis). Robust
standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
29
Table 5: Impact of Breaks on Project Value Distribution
(1)
Projects
VARIABLES
All Breaks * Bin 1 (Lowest Value)
All Breaks * Bin 2
All Breaks * Bin 3
All Breaks * Bin 4
All Breaks * Bin 5
All Breaks * Bin 6
All Breaks * Bin 7
All Breaks * Bin 8
All Breaks * Bin 9
All Breaks * Bin 10
All Breaks * Bin 11
All Breaks * Bin 12
All Breaks * Bin 13
All Breaks * Bin 14
All Breaks * Bin 15 (Highest Value)
Observations
Number of Cities
R-squared
0.0091***
(0.0019)
-0.0115***
(0.0024)
-0.0018
(0.0012)
-0.0015
(0.0011)
-0.0013
(0.0012)
-0.0012
(0.0008)
0.0003
(0.0008)
0.0007
(0.0005)
0.0025***
(0.0008)
0.0024**
(0.0010)
0.0028**
(0.0014)
0.0050**
(0.0020)
0.0053**
(0.0024)
0.0061**
(0.0029)
0.0063*
(0.0034)
47,679,975
10,091
0.031
Notes: The dependent variable is the number of projects in a specific bin of project value, and the unit of analysis is a
city-week-bin. We divide projects into 15 bins based on the amount of funds they raise, from the lowest (Bin 1) to the highest
(Bin 15). Bins are defined by year to account for the growth of the platform. City and week fixed effects are included the
regression. Robust standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
30
Table 6: Differences between Breaks and Work Periods in (Endogenous) Effort Provision
(1)
Mean Desc.
Length
(2)
Mean
FAQs
(3)
Mean
Updates
(4)
0-33 Mean
Desc. Length
(5)
33-66 Mean
Desc. Length
(6)
66-100 Mean
Desc. Length
(7)
0-33 Mean
FAQs
(8)
33-66 Mean
FAQs
(9)
66-100 Mean
FAQs
(10)
0-33 Mean
Updates
(11)
33-66 Mean
Updates
(12)
66-100 Mean
Updates
All Breaks
1.6298***
(0.5042)
0.0093***
(0.0019)
0.0185**
(0.0072)
0.8673**
(0.4007)
1.2517***
(0.4561)
1.9980***
(0.5826)
0.0054***
(0.0010)
0.0047***
(0.0012)
0.0071***
(0.0022)
0.0075***
(0.0028)
0.0144***
(0.0048)
0.0263***
(0.0089)
Observations
R-squared
Number of Cities
3,178,665
0.466
10,091
3,178,665
0.079
10,091
3,178,665
0.203
10,091
3,178,665
0.158
10,091
3,178,665
0.166
10,091
3,178,665
0.152
10,091
3,178,665
0.016
10,091
3,178,665
0.027
10,091
3,178,665
0.038
10,091
3,178,665
0.042
10,091
3,178,665
0.081
10,091
3,178,665
0.102
10,091
VARIABLES
Notes: The dependent variables are proxies for effort on the platform. In Columns (1) and (4)-(6) we use the mean length of the project description; in Columns (2)
and (7)-(9) the mean number of FAQs; in Columns (3) and (10)-(12) the mean number of project updates to the crowd. In Columns (4) to (12), we progressively
divide the projects into three segments based on the bins defined in Table 5. For example, the “0-33” category includes the bottom five bins containing projects that
raised the lowest amount of funds, and “66-100” is the highest. City and week fixed effects are included in all regressions. Robust standard errors are clustered at the
city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
31
Table 7: Differences between Breaks and Work Periods in Project Complexity and the Share of Teams
(1)
Mean
Skills
(2)
Mean
Skills PP
(3)
Mean
Team Size
(4)
Share Team in
Complex (Skills)
(5)
Share Team in
Simple (Skills)
(6)
Share Team in
Complex (Skills PP)
(7)
Share Team in
Simple (Skills PP)
All Breaks
0.0066***
(0.0018)
0.0046***
(0.0016)
0.0024*
(0.0014)
0.0009**
(0.0004)
0.0004
(0.0004)
0.0009**
(0.0004)
0.0007
(0.0005)
Observations
R-squared
Number of Cities
3,178,665
0.350
10,091
3,178,665
0.322
10,091
3,178,665
0.458
10,091
3,178,665
0.049
10,091
3,178,665
0.065
10,091
3,178,665
0.107
10,091
3,178,665
0.164
10,091
VARIABLES
Notes: For Columns (1) and (2), the dependent variables are the mean level of measures of project complexity for all projects in a given city-week. The two measures
we use are the unique number of skills (“Skills”) and the unique number of skills per team member (“Skills PP”). For Columns (3)-(7), the dependent variable is the
share of team projects among all projects for a given city-week. We define complex projects as those that are above the median in terms of our two complexity
measures (and simple projects as those that are below the median). City and week fixed effects are included in all regressions. Robust standard errors are clustered at
the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
Table 8: Effects of Increasing the Minimum Effort Level
VARIABLES
All Breaks
(1)
Projects
Ratio
(2)
Projects
Ratio
(3)
Funding
Ratio
(4)
Funding
Ratio
(5)
Large Projects
Ratio
(6)
Large Projects
Ratio
(7)
Team Size
Ratio
(8)
Team Size
Ratio
0.0000
(0.0002)
-0.0002
(0.0002)
0.0006**
(0.0003)
0.0004
(0.0003)
-0.0006**
(0.0003)
0.0024***
(0.0006)
0.0002
(0.0002)
-0.0008***
(0.0002)
0.0024***
(0.0006)
0.0006
(0.0005)
-0.0012*
(0.0006)
0.0045***
(0.0013)
3,178,665
0.025
10,091
3,178,665
0.025
10,091
3,178,665
0.014
10,091
3,178,665
0.015
10,091
3,178,665
0.002
10,091
3,178,665
0.002
10,091
3,178,665
0.019
10,091
3,178,665
0.019
10,091
All School Breaks *
After the Change
Observations
R-squared
Number of Cities
Notes: For Columns (1) and (2), the dependent variable is the ratio between the number of projects in the Design and
Technology categories and the number of projects in all the other categories for a given city-week. The dependent variables of
the other columns are similarly defined based on the total amount of funds, the share of projects with large target amounts
(over $30,000), and the share of team projects. “After the Change” indicates time periods after the May 2012 policy change.
City and week fixed effects are included in all regressions. Robust standard errors are clustered at the city level. *** p < 0.01,
** p < 0.05, * p < 0.1.
32
Appendix A: Proofs of Theoretical Results
A1. The baseline model
Proof of Prediction 1. The innovator’s payoff from developing the project (Equation (3)) is an
2
increasing function of q. Because 2cq w > 0 regardless of q, the development decision is determined
2
by π̄w (q) = 0. The quality thresholds are thus qē − cw ē2 = 0 ⇔ q̄w = cw2 ē . The probability of
developing a project is then Pr(q ≥ cw2 ē ) = 1 − F ( cw2 ē ), where F is the cumulative probability
distribution of idea quality q. This probability is a decreasing function of cw , and therefore the
probability of developing a project during the break is higher than that a work period.
2
Proof of Prediction 2. At the low end, no projects are developed in either period if v < c02ē .
2
2
However, when c02ē < v < c12ē , no projects are developed during a work period but they are
developed during the break. Thus, Pr(v < v̂|developed; break) ≥ Pr(v ≤ v̂|developed; work) for
2
any v̂ ≤ c12ē , because the latter is always zero in this range while the former is positive or zero.
At the high end, ignoring the minimum effort requirement for the moment, the project’s value
takes the form q · e∗w (q) = q · cqw . Thus, given project value v̂, the corresponding idea quality is
√
√
v̂cw and the corresponding effort level is √cv̂w . The conditional probability of projects with value
greater than v̂ is therefore Pr(v ≥ v̂|developed) =
above conditional probability can be written as
√
c0 v̂
c ē
−λ( 0
2 )
e
e−λ
>
√
c1 v̂
c ē
−λ( 1
2 )
e
e−λ
when v̂ ≥
√
ē2 ( c
1+
4
√
c0
)2
√
1−F ( cw v̂)
ē .
)
1−F ( cw
2
√
cw v̂
ē
−λ( cw
2 )
e
e−λ
With an exponential distribution, the
, where 1/λ is the mean of the distribution.
.
Notice that c1 ē2 is the project value at which the desired effort level for a work week equals ē.
So the projects’s value derives from q · cqw when v is greater than this value for both break and work
periods (and thus the previous analysis holds
even
after taking into consideration the minimum
√
√
ē2 ( c1 + c0 )2
, c1 ē2 } = c1 ē2 because c0 < c1 . Therefore, we
effort requirement). Furthermore, max{
4
can conclude that Pr(v ≥ v̂|developed; break) ≥ Pr(v ≥ v̂|developed; work) for any v̂ ≥ c1 ē2 .
Proof of Prediction 3. We have three scenarios. First, for any v̂ ≥ (c1 ē)ē, the project’s
√ value
takes the form q · e∗w (q) = q · cqw for both periods. Thus, the corresponding idea quality is v̂cw and
√
the corresponding effort level is √cv̂w . Because c1 > c0 , the effort level during the break is higher.
√ √c −√c
The difference in effort level is v̂( √1c0 c1 0 ), and it is a positive function of v̂.
Second, for
c1 ē2
2
≤ v̂ < (c1 ē)ē, the effort level for projects developed during a work week is ē,
√
while that for the break is max{ √cv̂0 , ē}. Thus, effort level during the break is weakly higher, and
the difference in effort level is weakly increasing in v̂.
2
Third, for v̂ < c12ē , we cannot make such a comparison because no projects are developed
during a work week.
Proof of Prediction 4. The probability of developing an idea is Pr(q ≥ q̄w = cw2 ē ) = 1 − F ( cw2 ē ).
The change in this probability due to an increase in ē is −f ( cw2 ē ) c2w , which is negative. In order
for the reduction to be smaller for the break, we need to have ∂(−f ( cw2 ē ) c2w )/∂cw < 0. With an
A-1
exponential distribution, the secondary derivative is:
∂(−f (
cw ē cw
cw ē 1
cw ē cw
λcw ē
) )/∂cw = ∂(−λe−λ 2
)/∂cw = −λe−λ 2 ( −
)
2 2
2
2
4
Thus, for the above equation to be negative, we need to have λ < cw2 ē and thus the mean of the
exponential distribution (i.e., 1/λ) to be greater than cw2 ē . Given that we have two discrete values
of cw , it is sufficient (though not necessary) to take the greater of the two (i.e., c12ē ) for the above
condition to hold.
Proof of Prediction 5. First, consider a generic ratio r = a/c, where 0 < a < c. The difference
′)
results from reducing the denominator from c to c′ is ca′ − ac = a(c−c
cc′ . This difference is positive
and increases with the numerator a. From Prediction 2, we know that there are more projects
with v ≥ c1 ē2 during the break than a work period. Then, the share of top projects developed
during the break relative to the total number of projects (break and work periods combined) is
greater than the share of top projects developed during a work period relative to the same total
number of projects. Thus, following the relationships derived from a generic ratio stated above, as
the denominator (the total number of projects) decreases after the posting requirement becomes
more stringent, the relative share of top-value projects developed during the break increases.
A2. Extension of the baseline model—allowing teams
In this section, we extend the baseline model to allow for team projects. Either one or two people
can work on the project. If solo, the individual innovator’s payoff is the same as in the baseline
model. If with a team, each team member’s payoff consists of half of the value of the project, minus
her own time cost and a coordination cost.24
We separate projects into two types: d ∈ {0 = simple; 1 = complex}, and let the unconditional
probability of a complex project be ρ and that for a simple project be 1 − ρ. For simplicity, we
assume that project complexity is independent of the the distribution of an idea’s intrinsic quality q.
We normalize the coordination cost to be zero for simple projects. For complex projects, teaming
up incurs a positive coordination cost and such a cost is smaller during the break than a work
period; that is, δ0 < δ1 .
Finally, for simple projects, teaming up per se generates no direct benefit to the project’s value,
while for complex projects, teaming up brings direct benefit to the project (e.g., from complementary skills and perspectives), and this benefit is more pronounced for ideas with greater quality.
Thus, the innovator’s payoff from developing a project at effort level e is:
(
2
qe − cw e2
if solo
πw =
,
1
e2
2 q(1 + dγ) · 2e − cw 2 − δw d if team
where w ∈ {0 = break; 1 = work}, d ∈ {0 = simple; 1 = complex}, and γ indicates the direct
benefit to a project’s value. We assume that the direct benefit from using teams is not too high
24
In this simple extension, we abstract away from the scenario in which each individual observes a separate idea
and decides whether to develop one of the two ideas jointly or develop the ideas separately. The simple extension is
as if assuming that the second team member is unlikely to develop any projects on his own if not asked by the first
person.
A-2
q
ē2
(i.e., γ < cw ēc2w−2δ
− 1) to rule out the scenario in which using a team dominates developing the
w
project solo regardless of idea quality.
Taking into account the minimum effort requirement ē, we can write the innovator’s payoff
functions as follows (analogous to Equation (3) in the baseline model):
if solo
✶{q≥cw ē} 2cq w + ✶{q<cw ē} (qē − cw ē2 )
2
q 2 (1+dγ)2
ē
1
π̃w (q) =
cw ē
✶{q≥ 2(1+dγ)
− δw d) + ✶{q< cw ē } ( 2 q(1 + dγ)ē − cw 8 − δw d) if team .
}(
2cw
2(1+dγ)
2
2
0
if drop
(5)
Consider only simple projects for the moment. Using a team does not incur any coordination
costs. As there is no direct benefit to the project’s value, the benefit from using a team comes
from sharing the workload. In our simple setup, this benefit matters only for ideas of relatively low
qualities (for which the minimum effort requirement is too high relative to the idea’s quality). The
following result shows that for both break and work periods, a simple project is developed solo if
the idea’s quality q is sufficiently high, by a team if q is in an intermediate range, and dropped if q
is sufficiently low. Notice that allowing a team to share the workload lowers the quality threshold
of developing a project relative to that in the baseline model ( cw4 ē versus cw2 ē ).
Result 1 (Decision rules on “solo” versus “team” for a simple project). A simple project is developed
solo if q ≥ cw ē, by a team if cw4 ē ≤ q < cw ē, and dropped otherwise.
Proof. For a simple project, d = 0. Both payoffs (solo or team) are monotone increasing functions
of q. When q > cw ē, developing the project solo generates the same payoff as that from using
a team. We let the innovator choose solo because there may still be a small coordination cost
working with a team. When q < cw ē, the innovator is better off working in a team as multiple
people share the workload, resulting in a smaller time cost for each individual member. The payoff
from developing a project using a team is greater than zero if and only if q > cw4 ē , which is the
2
solution of ( 21 qē − cw ē8 ) = 0
For complex projects, while coordination costs for teams are non-zero, teaming up introduces
two benefits: sharing the workload, and an increase in project value. This second, direct benefit
becomes more relevant as the idea quality q increases. Similar to simple projects, teams may also
help share the workload when the minimum effort requirement is too high relative to the idea’s
quality. However, for complex projects we may not see the use of teams for this range of ideas if
coordination costs are too high. Formally, we have the following decision rule for complex projects:
Result 2 (Decision rules on “solo” versus “team” for a complex project). For a complex project,
2
(a) when the coordination cost is sufficiently high (i.e., δw > cw8ē (1 + γ)2 ), the project is
developed by a team if q ≥ q̂w (all thresholds are defined below), solo if cw2 ē ≤ q < q̂w , and dropped
otherwise.
2
(b) when the coordination cost is not too high (i.e., δw < cw8ē (1 + γ)2 ), the project is developed
team ≤ q < q̌ , and dropped otherwise.
by a team if q ≥ q̂w , solo if q̌w ≤ q < q̂w , team again if q̄w
w
Proof. For a complex project, d = 1. Again, both payoffs (solo or team) are monotone increasing
functions of q. First, define the following thresholds:
A-3
• The value of q that makes the innovator indifferent between
qusing a team or not at the higher
2δw cw
)+✶
end of the quality distribution: q̂w = ✶
(γ 2 +2γ)cw ē2 (
(γ 2 +2γ)cw ē2 (q̃w,1 ),
γ 2 +2γ
{δw <
{δw >
}
}
2
2
p
1
2 − c2 ē2 (γ 2 + 2γ)) is the greater of the two solutions
2c
δ
(1
+
γ)
(c
ē+
where q̃w,1 = (1+γ)
w
w w
2
w
2
2
2
2
2
w ē
, the payoff from using a team and
−δw . When δw > (γ +2γ)c
that make qē−cw ē2 = q (1+γ)
2cw
2
the payoff from developing the project solo intersect when the desired effort levels for both
2
2
w ē
are above the minimum requirement (ē). When δw < (γ +2γ)c
, the two payoffs intersect
2
when the desired effort level when using a team is above ē while that for solo is below ē.
• The value of q that makes the innovator indifferent between
p using a team or not at the low
1
end of the quality distribution: define q̌w = (1+γ)2 (cw ē − 2cw δw (1 + γ)2 − c2w ē2 (γ 2 + 2γ)),
2
which is the smaller of the two solutions that equate qē − cw ē2 =
q 2 (1+γ)2
2cw
− δw .
• The value of q that makes the innovator indifferent between developing the
√ project using
cw ē
2δw
team
w δw
a team and dropping it: q̄w
= ✶{δ < cw ē2 } ( (1+γ)4 + (1+γ)ē ) + ✶{δ ≥ cw ē2 } ( 2c
1+γ ). When
w
cw ē2
8 ,
8
w
8
the payoff from using a team equals zero when the desired effort level is below ē;
δw <
otherwise, the payoff equals zero when the desired effort level is above ē.
2
We have two scenarios. First, when the coordination cost is sufficiently high (i.e., δw > cw8ē (1 +
2
γ) , see the derivation of this threshold value in the last paragraph of this proof), sharing the
workload alone is not sufficient to justify the use of teams. The only cases in which teams are
worth the high coordination cost are when the direct benefit is high enough, which is when the
idea’s quality is sufficiently high. Thus, the project is developed by a team if q ≥ q̂w , solo if
cw ē
2 ≤ q < q̂w , and dropped otherwise (the thresholds are defined as above).
2
Second, when the coordination cost is not too high (i.e., δw < cw8ē (1 + γ)2 ), it may also be
worthwhile using a team simply to share workload. Thus, project is developed by a team if q ≥ q̂w ;
team ≤ q < q̌ ; and dropped otherwise.
solo if q̌w ≤ q < q̂w ; team again if q̄w
w
Finally, to derive the separating threshold for these two scenarios, note that the value of q
at which the desired effort level when using a team equals the minimum required effort level is
smaller than the value of q at which the innovator is indifferent between dropping the project and
cw ē
< cw2 ē ). Thus, the threshold value of coordination cost δw that
developing it solo (that is, 2(1+γ)
separates the two scenarios of whether a team is used again at the relatively low end of the quality
distribution is when the optimal payoff from using a team at q = cw2 ē is equal to zero. This yields
2
the threshold value of cw8ē (1 + γ)2 .
A key interesting insight from the above results is that complex projects are likely to benefit
more from breaks because coordination is easier than during a work week. In contrast, for simple
projects, the relative advantage of breaks is not relevant since no coordination is required. The
probability of complex projects conditional on development is however not necessarily higher for
breaks because the low opportunity cost of time during breaks alone would also allow a greater
number of simple projects to be developed. Prediction 6 in the paper provides a sufficient condition
under which the likelihood of complex projects is greater during breaks. This sufficient condition
is satisfied when the coordination cost for a complex project is sufficiently high during a work week
but sufficiently low during breaks.
A-4
2
2
Proof of Prediction 6. When δ1 > c18ē (1 + γ)2 and δ0 < c08ē , the probabilities of complex
projects (conditional on development) during a work and a break week are (recall that ρ is the
unconditional probability that an idea is complex):
ρ(1 − F ( c12ē ))
,
(1 − ρ)(1 − F ( c14ē )) + ρ(1 − F ( c12ē ))
Pr(Complex|work; developed) =
Pr(Complex|break; developed) =
team )
ρ(1 − F (q̄w
,
c0 ē
(1 − ρ)(1 − F ( 4 )) + ρ(1 − F (q̄0team )
2δ0
c0 ē
+ (1+γ)ē
(defined in Result 2).
where q̄0team = (1+γ)4
Pr(Complex|break; developed) > Pr(Complex|work; developed) when the following holds:
c0 ē
2δ0
(1 − F ( (1+γ)4
+ (1+γ)ē
)
(1 − F ( c12ē ))
<
.
c1 ē
c0 ē
(1 − F ( 4 ))
(1 − F ( 4 ))
(6)
With q following an exponential distribution, Equation (6) holds as long as δ0 < 18 ((1 + γ)ē2 c1 +
2
2
2
γc0 ē2 ). This latter condition is satisfied when δ0 < c08ē as c08ē < c18ē < 18 ((1 + γ)c1 ē2 + γc0 ē2 ).
An important variable observable in the data is whether a project is carried out by a team. It is
ambiguous whether the likelihood of team projects (conditional on development) should be higher
during breaks, because teaming up also helps share the workload and that is particularly useful
for work periods when the opportunity cost of time is higher. Given that the relative advantage
of breaks in easier coordination among multiple people is more salient for complex projects, we
should observe that compared to simple projects that require no coordination, complex projects
are relatively more likely to be developed by a team during a break than during a work period.
Prediction 7 in the paper shows that this statement is true when the coordination cost for a complex
project is sufficiently high during a work week but sufficiently low during breaks.
Proof of Prediction 7. For simple projects, the difference in the probability of using a team
between break and work weeks is:
=
Pr(Team|break; simple & developed) − Pr(Team|work; simple & developed)
c0 ē
)
4
c0 ē
1−F ( 4 )
F (c0 ē)−F (
−
c1 ē
)
4
c1 ē
1−F ( 4 )
F (c1 ē)−F (
With q following an exponential distribution,
c0 ē
)
4
c0 ē
1−F ( 4 )
F (c0 ē)−F (
c0 ē2
8 .
The difference
Pr(Team|break; complex & developed) − Pr(Team|work; complex & developed)
1−F (q̂0 )+(F (q̌0 )−F (q̄0team,
1−F (q̄0team, d=1 )
d=1
)
−
1−F (q̂1 ))
c ē
1−F ( 12 )
(7)
increases with c. Thus, Equation (7) is
negative; i.e., the probability of team projects is smaller during breaks.
2
For complex projects, consider the scenario when δ1 > c18ē (1+γ)2 and δ0 <
in the probability of using a team between break and work weeks is:
=
.
,
(8)
where the thresholds are all defined in Result 2. During work periods, because coordination cost is
too high, teams are used only at the high end of the quality distribution to take advantage of the
A-5
direct benefit to project value. During break periods, because the coordination cost is sufficiently
small, teams are used both at the high end and at the relatively low end in order to take advantage
of workload sharing.
We can show that the second term in Equation (8), Pr(Team|work; complex & developed) =
1−F (q̂1 ))
c1 ē , is a decreasing function of δ1 . That is, the higher the coordination cost, the likelihood
1−F (
2
)
of using a team is smaller during a work period. At the limit, this likelihood goes to zero. Thus,
given any value of δ0 (thus, fixing Pr(Team|break; complex & developed)), there exists a δ̂1 such
that Equation (8) is positive for all δ1 > δ̂1 ; that is, the probability of team projects is greater
during breaks than work periods.
A3. Extension of the baseline model—allowing for shelving
In this section, we extend the basic setup to allow the innovator to shelve an idea and develop it
later. For an idea arriving during a work week, the innovator can choose to develop it, to shelve
and develop it during the break, or to drop the idea. If developing the idea, the innovator faces the
same payoff as in the baseline model. If they wait until the break, then the value of the project
is discounted to v = βqe, where 0 < β ≤ 1, perhaps due to obsolescence of ideas, competing ideas
that emerge during the waiting period, etc. However, the cost of time during the break is lower.
Note that for ideas arriving during a break week, the innovator’s problem stays the same as the
baseline model as immediate development dominates shelving. This is because the cost of time is
higher later while the value of the project is discounted. Thus, we focus on the innovator’s decision
during a work week.
The innovator’s payoff from developing a project at effort level e is:
(
2
if develop immediately
qe − c1 e2
.
π1 =
e2
βqe − c0 2 if shelve and develop later during the break
Relative to immediate development, the benefit from delaying is that time is less costly during
the break, while the downside is that the project may lose some of its value. Taking into account
the minimum effort requirement ē, we can write the innovator’s payoff as follows (analogous to
Equation (3) in the baseline model):
π̃1 (q) =
✶{q≥c1 ē} 2cq 1 + ✶{q<c1 ē} (qē − c1 ē2 )
if develop immediately
β 2 q2
ē2
✶{q≥ c0 ē } 2c0 + ✶{q< c0 ē } (βqē − c0 2 ) if shelve and develop later during the break
2
β
2
β
0
if drop
The following results shows that when the project does not lose too much of its value by waiting
(that is, β is sufficiently large), shelving the project dominates immediate development regardless
of the idea’s intrinsic quality q. In contrast, when the project loses too much value from waiting
(that is, β is sufficiently low), immediate development dominates shelving the project. Only when
β is in an intermediate range does the innovator face meaningful tradeoffs between the two choices.
The innovator is better off with immediately developing the project for very good ideas and better
off with shelving when the idea quality is not too high. This is because when the idea quality is
very good, the innovator would not want to sacrifice its potential, which is high compared to the
A-6
benefit from saving the time cost. When the idea’s quality is at an intermediate range, the benefit
from saving the time cost would outweigh the cost of discounting the value of the idea.
Result 3 (Decision rules on shelving). For ideas arriving during a work week:
q
• When β > cc01 , shelving dominates immediate development regardless of the idea’s quality q.
0 ē
and drops it otherwise.
Thus, the innovator shelves the project if q ≥ c2β
q
p
ē
}. The innovator
• When cc10 < β < cc01 , define q̃ = min{ βē2 (c0 + c0 (c0 − β 2 c1 )), (c1 −c0 ) 2(1−β)
develops the idea immediately if q ≥ q̃, shelves the idea if
c0 ē
2β
< q < q̃, and drops it otherwise.
• When β < cc10 , immediate development dominates shelving for values of q with positive payoffs.
Thus, the innovator develops the idea immediately if q ≥ c12ē and drops it otherwise.
Proof. When comparing only the innovator’s optimal payoffs (without considering the minimum
effort requirement),
shelving is better than immediate development regardless of q if and only
q
c0
if β >
c1 (i.e., the project does not lose too much of its value by waiting). This is because
q
2
2
2
q
β q
c0
>
⇒
β
>
2c0
2c1
c1 . Taking into consideration the requirement ē, we have the following three
scenarios:
q
The first scenario is when β > cc01 . From above, we know that the innovator’s optimal payoff
from shelving is everywhere higher than that from immediate development. At the same time, in
this case, the slope of the innovator’s payoff from shelving and exerting ē is smaller (i.e., flatter)
than that from immediately developing the project at ē. This is true because the intersection point
1 −c0 )ē
of these two payoffs, (c2(1−β)
, is to the right of cβ0 ē , which is the value of q at which the optimal
level of effort with shelving is the same as ē.25 As a result, shelving is better than immediate
0 ē
and drop it
development regardless of q, and the decision rule is to shelve the idea if q ≥ c2β
otherwise.
q
The second scenario is when cc10 < β < cc01 . From above, we know that the innovator’s optimal
payoff from immediate development is everywhere higher than that from shelving the project.
However, in this case, we can show that the slope of the innovator’s payoff from shelving and
exerting ē is smaller (i.e., flatter) than that from immediately developing the project at ē, and the
value of q at which they intersect generates a positive payoff for shelving. This implies that there
is a range of value of q for p
which shelving generates a higher payoff than immediate development.
ē
ē
Denote q̃ = min{ β 2 (c0 + c0 (c0 − β 2 c1 )), (c1 − c0 ) 2(1−β)
}, where the former is the value of q
at which the payoff from shelving and optimal effort level is the same as that from immediately
developing the project at ē,26 and the latter is the intersecting point for the innovator’s payoff
from shelving and exerting ē and that from immediately developing the project at ē. Therefore,
the decision rule in this case is to develop the project immediately if q ≥ q̃, shelve the project if
c0 ē
2β < q < q̃, and drop the idea otherwise.
0
to hold, we need β > c02c+c
, which is satisfied when β >
1
q
4c2
c0
0
0
> c02c+c
⇔ cc01 > c2 +c2 +2c
The last condition holds because
⇔ (c1 − c0 )2 > 0.
c1
1
0 c1
0
1
p
2 2
2
2
26 β q
= qē − c1 ē2 ⇒ (βq − cβ0 ē )2 = βē 2 c0 (c0 − β 2 c1 ) ⇒ q = βē2 (c0 + c0 (c0 − β 2 c1 )).
2c0
25
ē
For (c1 − c0 ) 2(1−β)
>
c0 ē
)
β
A-7
q
c0
c1
because
q
c0
c1
>
2c0
.
c0 +c1
The final scenario is when β < cc01 . In this case, developing the project immediately dominates
shelving for all values of q that generate a positive payoff for the former. Thus, the innovator
chooses to develop the idea immediately if q ≥ c12ē and drops the idea otherwise.
A-8
Appendix B: Additional Tables and Figures
Figure B-1: The Number of Projects per Week is Higher During Breaks (Poisson)
Notes: Estimated coefficient for weeks before and after the school breaks. Dependent variable is the number of projects
created in the focal city-week. Poisson regression includes city and week fixed effects. Error bars represent 95% confidence
intervals based on robust standard errors clustered at the city level.
B-1
Table B-1: Robustness to Sample Definition by City Size
(1)
Projects in All Cities
(2)
Projects in Top 10% Cities
(3)
Projects in Top 1% Cities
All Breaks
0.0238***
(0.0083)
0.0745***
(0.0282)
0.1811*
(0.1084)
Observations
R-squared
Number of Cities
3,178,665
0.164
10,091
402,885
0.150
1,279
34,650
0.265
110
VARIABLES
Notes: The dependent variable is the number of projects. Column (1) replicates the baseline estimation in Table 3. Columns
(2) and (3) respectively only use top cities in terms of crowdfunding activity (ranking is based on the total number of
projects). City and week fixed effects are included in all regressions. Robust standard errors are clustered at the city level.
*** p < 0.01, ** p < 0.05, * p < 0.1.
Table B-2: Robustness to Sample Definition by Enrollment in Schools with Breaks Data
(1)
Enrollment 0-100
(2)
Enrollment 0-75
(3)
Enrollment 0-50
(4)
Enrollment 0-25
All Breaks
0.0238***
(0.0083)
0.0240**
(0.0102)
0.0092**
(0.0038)
0.0088
(0.0054)
Observations
R-squared
Number of Cities
3,178,665
0.164
10,091
3,115,035
0.170
9,889
3,051,405
0.393
9,687
2,987,775
0.408
9,485
VARIABLES
Notes: The dependent variable is the number of projects. Column (1) replicates the baseline estimation in Table 3. Columns
(2)-(4) gradually exclude cities associated with colleges with the largest number of enrolled students. City and week fixed
effects are included in all regressions. Robust standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, *
p < 0.1.
Table B-3: Robustness to Sample Definition by County-Level College Enrollment and Population
(1)
Below Median
College Students
(2)
Above Median
College Students
(3)
Below Median
Population
(4)
Above Median
Population
All Breaks
-0.0084
(0.0069)
0.0257***
(0.0090)
-0.0068
(0.0050)
0.0274***
(0.0095)
Observations
R-squared
Number of Cities
1,589,175
0.621
5,045
1,589,490
0.142
5,046
1,589,490
0.622
5,046
1,589,175
0.141
5,045
VARIABLES
Notes: The dependent variable is the number of projects. Columns (1) and (2) separate the sample by whether the associated
counties have above versus below the median number of college students according to the American Community Survey
(ACS). Columns (3) and (4) separate the sample by whether a city’s population is above or below the median. City and week
fixed effects are included in all regressions. Robust standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, *
p < 0.1.
B-2
Table B-4: Impact of Breaks on Project Value Distribution (20 Bins)
(1)
Projects
VARIABLES
All Breaks * Bin 1 (Lowest Value)
All Breaks * Bin 2
All Breaks * Bin 3
All Breaks * Bin 4
All Breaks * Bin 5
All Breaks * Bin 6
All Breaks * Bin 7
All Breaks * Bin 8
All Breaks * Bin 9
All Breaks * Bin 10
All Breaks * Bin 11
All Breaks * Bin 12
All Breaks * Bin 13
All Breaks * Bin 14
All Breaks * Bin 15
All Breaks * Bin 16
All Breaks * Bin 17
All Breaks * Bin 18
All Breaks * Bin 19
All Breaks * Bin 20 (Highest Value)
Observations
R-squared
Number of Cities
0.0128***
(0.0024)
-0.0097***
(0.0020)
-0.0055***
(0.0014)
-0.0017*
(0.0009)
-0.0007
(0.0009)
-0.0010
(0.0008)
-0.0007
(0.0007)
-0.0017**
(0.0008)
0.0001
(0.0007)
0.0006
(0.0006)
0.0008*
(0.0005)
0.0019***
(0.0006)
0.0021***
(0.0008)
0.0017**
(0.0008)
0.0023*
(0.0013)
0.0042***
(0.0016)
0.0038**
(0.0016)
0.0046**
(0.0022)
0.0042**
(0.0021)
0.0052*
(0.0028)
63,573,300
0.024
10,091
Notes: The dependent variable is the number of projects in a specific bin of project value, the unit of analysis is a
city-week-bin. We divide projects into 20 bins based on the amount of funds they raise, from the lowest (bin 1) to the highest
(bin 20). Bins are defined by year to account for the growth of the platform. City and week fixed effects are included the
regression. Robust standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
B-3
Table B-5: Length of the Break and Week of the Break
VARIABLES
All Breaks
(1)
Projects
(2)
Projects
0.0238***
(0.0083)
1 Week Break
0.0841
(0.0530)
0.0421**
(0.0165)
0.0665*
(0.0382)
0.0200**
(0.0079)
2 Weeks Break
3 Weeks Break
> 4 Weeks Break
1 Week Break
0.0839
(0.0530)
0.0424**
(0.0188)
0.0418**
(0.0169)
0.0697*
(0.0371)
0.0650
(0.0404)
0.0021
(0.0082)
0.0220***
(0.0083)
2 Weeks Break, Week 1
2 Weeks Break, Week 2
3 Weeks Break, Week 1
3 Weeks Break, Week 2+
> 4 Weeks Break, Week 1
> 4 Weeks Break, Week 2+
Observations
R-squared
Number of Cities
(3)
Projects
3,178,665
0.164
10,091
3,178,665
0.164
10,091
3,178,665
0.164
10,091
Notes: The dependent variable is the number of projects. Column (1) replicates the baseline estimation in Table 3. Column
(2) separates breaks by their length. Column (3) separately investigates the first and the second (or more) week for breaks of
different lengths. City and week fixed effects are included in all regressions. Robust standard errors are clustered at the city
level. *** p < 0.01, ** p < 0.05, * p < 0.1.
B-4
Table B-6: Shelving
VARIABLES
All Breaks
(1)
All
Projects
(2)
Clearly
Shelved
(3)
Excluding
Clearly Shelved
0.0447***
(0.0129)
0.0012**
(0.0005)
0.0435***
(0.0126)
1 Week Break
2 Weeks Break
3 Weeks Break
> 4 Weeks Break
(4)
All
Projects
(5)
Clearly
Shelved
(6)
Excluding
Clearly Shelved
0.0708
(0.0477)
0.0365*
(0.0188)
0.0662*
(0.0385)
0.0442***
(0.0128)
0.0021
(0.0026)
0.0012
(0.0010)
0.0013
(0.0022)
0.0012**
(0.0005)
0.0688
(0.0462)
0.0353*
(0.0184)
0.0650*
(0.0370)
0.0430***
(0.0124)
1 Week Break
B-5
2 Weeks Break, Week 1
2 Weeks Break, Week 2
3 Weeks Break, Week 1
3 Weeks Break, Week 2+
> 4 Weeks Break, Week 1
> 4 Weeks Break, Week 2+
Observations
R-squared
Number of Cities
2,754,843
0.164
10,091
2,754,843
0.011
10,091
2,754,843
0.164
10,091
2,754,843
0.164
10,091
2,754,843
0.011
10,091
2,754,843
0.164
10,091
(7)
All
Projects
(8)
Clearly
Shelved
(9)
Excluding
Clearly Shelved
0.0706
(0.0476)
0.0296
(0.0226)
0.0435**
(0.0175)
0.0724
(0.0460)
0.0632*
(0.0365)
0.0288***
(0.0112)
0.0459***
(0.0131)
0.0021
(0.0026)
-0.0007
(0.0013)
0.0032*
(0.0019)
0.0021
(0.0032)
0.0008
(0.0021)
0.0023*
(0.0013)
0.0011**
(0.0005)
0.0686
(0.0461)
0.0303
(0.0226)
0.0404**
(0.0168)
0.0703
(0.0443)
0.0623*
(0.0352)
0.0265**
(0.0104)
0.0448***
(0.0128)
2,754,843
0.164
10,091
2,754,843
0.011
10,091
2,754,843
0.164
10,091
Notes: The dependent variable in Columns (1), (4), and (7) is all projects. In Columns (2), (5), and (8), the dependent variable only captures projects that we
identified as clearly shelved by their creators based on the project page (data is available from the beginning of the sample to July 2014). Columns (3), (6), and (9)
remove from all projects the projects that are clearly shelved (i.e., the dependent variable is always the difference between the previous two columns). City and week
fixed effects are included in all regressions. Robust standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
Table B-7: Prediction 3 (Only Successful Projects)
(1)
0-33 Mean
Desc. Length
(2)
33-66 Mean
Desc. Length
(3)
66-100 Mean
Desc. Length
(4)
0-33 Mean
FAQs
(5)
33-66 Mean
FAQs
(6)
66-100 Mean
FAQs
(7)
0-33 Mean
Updates
(8)
33-66 Mean
Updates
(9)
66-100 Mean
Updates
All Breaks
0.6184*
(0.3370)
0.7593*
(0.4289)
1.6403***
(0.5855)
0.0028***
(0.0009)
0.0023**
(0.0010)
0.0057***
(0.0021)
0.0106***
(0.0041)
0.0082
(0.0060)
0.0296***
(0.0094)
Observations
R-squared
Number of Cities
3,178,665
0.064
10,091
3,178,665
0.063
10,091
3,178,665
0.059
10,091
3,178,665
0.011
10,091
3,178,665
0.013
10,091
3,178,665
0.018
10,091
3,178,665
0.045
10,091
3,178,665
0.047
10,091
3,178,665
0.047
10,091
VARIABLES
Notes: The dependent variables are the mean level of measures of the amount of effort devoted to a project in a city-week.
Only successful projects (funds raised exceeding the target amount) are used in these regressions. For each measure of effort,
projects are divided into three segments based on the bins defined in Table 5. For example, “0-33” category includes the
bottom five bins containing successful projects that raised the lowest amount of funds and “66-100” the highest. City and
week fixed effects are included in all regressions. Robust standard errors are clustered at the city level. *** p < 0.01, **
p < 0.05, * p < 0.1.
Table B-8: Complexity and Week of the Break
VARIABLES
(1)
Mean Skills per Team Member
Week 1
0.0053
(0.0034)
0.0023
(0.0033)
-0.0010
(0.0042)
0.0058***
(0.0018)
Week 2
Week 3
Week 4+
Observations
Number of Cities
R-squared
3,178,665
10,091
0.322
Notes: The dependent variable is the mean level of our proxy for project complexity (the number of unique skills per team
member). The regression separately investigates the first, second, third, or fourth (or more) week within a break. For
example, “week 2” is identified by breaks that are at least two weeks long. City and week fixed effects are included in all
regressions. Robust standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
B-6
Table B-9: Snow Breaks and Project Value Distribution
(1)
Projects
VARIABLES
Snow Breaks * Bin 1
Snow Breaks * Bin 2
Snow Breaks * Bin 3
Snow Breaks * Bin 4
Snow Breaks * Bin 5
Snow Breaks * Bin 6
Snow Breaks * Bin 7
Snow Breaks * Bin 8
Snow Breaks * Bin 9
Snow Breaks * Bin 10
Snow Breaks * Bin 11
Snow Breaks * Bin 12
Snow Breaks * Bin 13
Snow Breaks * Bin 14
Snow Breaks * Bin 15
Observations
Number of Cities
R-squared
0.0063
(0.0040)
-0.0132***
(0.0045)
-0.0055*
(0.0031)
-0.0088**
(0.0040)
-0.0016
(0.0024)
-0.0033
(0.0028)
0.0022
(0.0027)
0.0012
(0.0021)
0.0049*
(0.0028)
0.0086**
(0.0039)
0.0097***
(0.0034)
0.0054
(0.0034)
0.0113
(0.0075)
0.0081
(0.0050)
0.0080
(0.0062)
44,501,310
10,091
0.032
Notes: The dependent variable is the number of projects in a specific value bin, the unit of analysis is a city-week-bin, and the
fixed effects for weeks and cities are included. The 15 bins are defined in the same way as in Table 5, and the regression
controls for the interactions between the other breaks and the bins (not reported for space constraints). Robust standard
errors are clustered at the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
B-7
Table B-10: Snow Breaks and Complexity
VARIABLES
Snow Breaks
Other Breaks
Observations
R-squared
Number of Cities
(1)
Mean Skills
(2)
Mean Team Size
(3)
Mean Skills per Team Member
-0.0006
(0.0076)
0.0064***
(0.0018)
0.0019
(0.0051)
0.0022
(0.0015)
-0.0014
(0.0064)
0.0047***
(0.0016)
2,966,754
0.356
10,091
2,966,754
0.455
10,091
2,966,754
0.326
10,091
Notes: The dependent variable is the mean level of our measures of complexity (Columns 1 and 2), and the mean team size
(Column 2). Robust standard errors are clustered at the city level. *** p < 0.01, ** p < 0.05, * p < 0.1.
B-8