ArticlePDF Available
Face-to-Face Communication
in Organisations
Diego Battiston
, Jordi Blanes i Vidaland Tom Kirchmaier§
November 3, 2017
Abstract
We study how communicating in person (in addition to electronically) affects team
productivity in organisations. Understanding this relation empirically has proven elu-
sive due to measurement and endogeneity issues. We exploit a unique natural experi-
ment in an organisation where workers must transmit complex electronic information
to their teammates. For exogenous reasons, workers can sometimes also communicate
face-to-face. We show that productivity is higher when face-to-face communication is
possible, and that this effect is stronger for urgent and complex tasks, for homoge-
neous teams, and during high pressure conditions. We highlight the opportunity costs
of face-to-face communication and their dependence on organisational slack.
JEL classification: D23, M11.
Keywords: Teamwork, Face-to-Face Communication, Organisations.
We thank Tore Ellingsen, Miguel Espinosa, Mitch Hoffman, Luis Garicano, Alan Manning, Ignacio
Palacios-Huerta, Veronica Rappoport, Yona Rubinstein and Catherine Thomas for valuable insights. Special
thanks also to Chief Constable Ian Hopkins QPM, and to Steven Croft, Peter Langmead-Jones, Duncan
Stokes, Ian Wiggett and many others at the Greater Manchester Police for making this project possible. No
financial support was received for this paper.
Department of Economics and Centre for Economic Performance, London School of Economics, London
WC2A 2AE, United Kingdom. Email: d.e.battiston@lse.ac.uk.
Corresponding author, Department of Management and Center for Economic Performance, London
School of Economics, Houghton Street, London WC2A 2AE, United Kingdom; and Center for Economic
and Policy Research, 33 Great Sutton Street, London EC1V 0DX, United Kingdom. Email: j.blanes-i-
vidal@lse.ac.uk.
§Copenhagen Business School, Solbjerg Pl. 3, 2000 Frederiksberg, Denmark; and Centre for Economic
Performance, London School of Economics, London WC2A 2AE, United Kingdom.
Email: t.kirchmaier@lse.ac.uk.
1
1 Introduction
Workers in teams typically need to communicate effectively with each other, especially when
dealing with tasks that are urgent and complex. While a lot of attention has been devoted
to understanding the effects of team incentives (Burgess et al. 2010, Bandiera et al. 2013,
Friebel et al. 2017) or team composition (Hamilton et al. 2003, Mas and Moretti 2009)
on performance, the central issue of team communication has been empirically neglected.
A necessary step in this direction consists of understanding the causal relation between
(access to more) communication and team productivity inside organisations. Unfortunately,
even this first step has been impeded by measurement and endogeneity concerns. There are
well-known difficulties in gaining access to data on the internal operations of organisations,
especially when these are sophisticated enterprises. Yet, without unusually rich data it is
not possible to measure communication between teammates. Secondly, the organisational
communication infrastructure is typically the result of an efficiency-maximising decision
process, prompting often insurmountable endogeneity concerns.
This paper overcomes these issues by taking advantage of an extremely rich dataset
and a unique natural experiment in a large and complex public sector organisation. In our
setting, individuals working in teams are always able to communicate electronically. Some
teams, exogenously chosen by a computerised system allocating tasks to workers, can also
communicate in person. Therefore, our experiment is best interpreted as identifying the
value of communicating face-to-face, in addition to electronically.
Our paper has three objectives. Firstly, we provide the first evidence on a causal link
between the ability to communicate face-to-face and team productivity inside organisations.
Secondly, we document substantial heterogeneity in the size of this relation. In particular,
the ability to communicate face-to-face is more valuable to teams that are demographically
homogenous, have experience of working together, face high pressure, and deal with urgent
and information-intensive tasks. In contexts where encouraging face-to-face communication
is costly, this finding suggests that managers should condition such investments on the
nature of the tasks, workers and production environments. Thirdly, we seek to understand
and measure the operational costs of communication. In our context, these costs arise from
workers being slower to undertake new tasks when they spend time communicating face-to-
face on existing tasks. By contrast, we find no displacement of attention away from other
tasks that workers are contemporaneously handling.
This Study The setting is the branch in charge of answering 999 calls and allocating
officers to incidents in the Greater Manchester Police. An incoming call is answered by a
call handler, who describes the incident in the internal computer system. When the handler
2
officially creates the incident, its details become available to the radio operator responsible
for the neighbourhood where the incident occurred. The radio operator then allocates a
police officer on the basis of incident characteristics and officer availability. The main metric
of performance is the time that it takes for the operator to allocate an officer1. Often, delays
result from the radio operator’s need to gather additional information, which she can do
through a variety of channels including communicating with the call handler in person.
To identify the importance of face-to-face communication we exploit both a natural ex-
periment and highly detailed information throughout the production process. In the Greater
Manchester Police, handlers and operators are spread across four rooms, each in a separate
part of Manchester. Each room contains the radio operators responsible for the surrounding
neighbourhoods as well as a subset of the call handlers, who can take calls from anywhere in
Manchester. This arrangement implies that, for some incidents, an operator reads the infor-
mation inputted in the system by a handler located in the same room. For other incidents,
the information will instead have been entered by a handler based in another location. A
direct consequence of co-location is that it allows the two teammates, handler and operator,
to communicate face-to-face if they wish to do so.
We first exploit the fact that the computerised queuing system matching incoming
calls to newly available handlers creates exogenous variation in the co-location of handler
and operator. Our baseline finding here is that allocation time is 2% faster when handler
and operator work in the same room2. This improvement is not at the expense of observable
dimensions of the quality of the allocation, such as the seniority of the officer sent. We also
show that proximity within the room is important - the effect of co-location is twice as high
when handler and operator are sitting close together. In fact, allocation time is lower even
when the same pair of workers are located inside the room closer together. This last finding
rules out unobservable characteristics in the match between handler and operator (correlated
with co-location) as the explanation for the baseline findings. We provide additional evidence
in this respect with a placebo test that exploits an organisational restructure that altered
the regular workplaces of handlers and operators.
Having identified the causal effect of co-location on productivity, we proceed to estab-
lish face-to-face communication as the primary explanatory mechanism. Unsurprisingly, our
organisation did not record any information transmitted through informal in-person interac-
1We describe this measure in detail in Section 2. There, we also list its advantages and potential lim-
itations and explain why the organisation assigned high importance to this measure during our sample
period.
2Although not large, this effect compares well with typical annual productivity increases in the public
sector (Simpson, 2009). Another comparison is with the effect of introducing team performance pay in the
field experiment of Friebel et al. (2017), which they find to be 3%. The effect in our study is twice as large
for urgent and information-intensive tasks, among others. At the police force level the baseline effect adds
up to approximately 900 hours per month, a substantial magnitude.
3
tions between co-workers, and therefore we are not able to use these informal messages here.
Instead, we provide a set of complementary tests. Firstly, we use several proxies to show
that the quality of the handler’s electronic communication is not higher when a co-located
operator will be reading the incident’s description. Secondly, we show that operators do
not assign higher priority to co-located incidents, at the expense of other contemporaneous
incidents. These two findings are counter to the most natural channels (alternative to face-
to-face communication) through which co-located teammates could increase productivity.
We further distinguish between different mechanisms by examining the behaviour of
the handler after officially creating the incident. Under the face-to-face communication
mechanism, the handler then spends time talking to the operator, which temporarily pre-
vents her from being available to take new calls. Alternative mechanisms, such as better
electronic communication by the handler or higher operator effort, do not naturally have
that prediction. We show that handlers spend more time ’unavailable’ to take new calls
following the creation of co-located incidents, and we interpret this as strong evidence that
they are communicating with their operators in these incidents.
The second objective of the paper is to uncover conditions under which face-to-face
communication is particularly important. We find first that co-location increases produc-
tivity more for incidents that are more information-intensive. This is reassuring, in that
it is consistent with the notion that having access to an additional communication chan-
nel is valuable particularly when more information needs to be transmitted. The effect of
co-location is also higher for more intrinsically urgent incidents, as well as during periods
when operators face a higher incident workload. These last two findings are consistent with
each other, in that they both suggest that operators facing higher time pressure benefit most
from being able to gather information through an additional quick, informal channel. Lastly,
we investigate the characteristics of the teams associated with a higher effect of co-location
on productivity. We provide three separate but mutually consistent results: teams of the
same gender, similar age, and with a longer history of working together benefit more from
co-location. Together, the three findings indicate that the ability to communicate face-to-
face benefits more teams that are more cohesive, because of either demographic traits or a
common, shared, experience.
The third objective of the paper is to identify and highlight the opportunity costs of
face-to-face communication. As mentioned earlier, we do not find that operators distort
their attention towards co-located incidents and at the expense of other contemporaneous
incidents. Negative spillovers of this type do not therefore seem to be present in our setting.
However, we do find that handlers spend more time unavailable to take new calls after
creating co-located incidents. This clearly imposes a delay on incoming calls whenever the
queue of incoming calls is not empty. In other words, communicating face-to-face has an
4
opportunity cost whenever the organisation has no slack. We provide a simple theoretical
framework and a set of tests to quantify this cost in our organisation. Empirically, we
find that the cost is very small, relative to the benefits of face-to-face communication. As
expected, the cost is however higher when the number of on-duty handlers is low relative to
the number of incoming calls (i.e. when there is less organisational slack).
Contribution This paper provides, we believe, the first causal evidence on the relation
between (face-to-face) communication and team productivity inside organisations. Of course,
the study involves a particular setting and production technology. As such, the implications
are stronger for high pressure environments such as the healthcare professionals assessing
and treating patients in emergency rooms, or the frontline staff and their supervisors in air
traffic control, the military, and other time-critical settings.
More generally, the results on the contingent value of face-to-face communication have
broader applicability. For instance, the results regarding the urgency and information-
intensity of tasks indicate the type of production environments where investments encour-
aging communication are likely to be particularly valuable. Equally significant is the finding
that homogeneous teams benefit more from being able to talk to each other. We briefly
mention in the conclusion a number of policy prescriptions based on this finding.
Lastly, the insights on the opportunity costs of communication are of general validity.
Of course, increasing communication in the workplace is likely to be associated in many
contexts with fixed costs (such as capital, estate or traveling costs) that we do not analyse
here. The cost that we focus on is based on the insight that every second spent communi-
cating cannot be devoted to other activities. This is a general observation, as is the idea
that this opportunity cost depends on the alternative use of workers’ time and therefore on
the amount of slack in the organisation. While our setting is not unusual in the existence
of this trade-off, it is unusual in that the highly structured nature of the production process
and the granularity of the dataset allow us to estimate it empirically.
Related Literature Despite its importance, field evidence on communication in organi-
sations is scant. Gant et al. (2002) argue that the adoption of innovative HRM practices
induces more communication among co-workers. Palacios-Huerta and Prat (2012) use email
exchanges to generate a measure of the relative importance of individual managers. Bloom
et al. (2014) investigate whether firms adopting technologies such as data intranets altered
their spans of control and autonomy levels. None of these papers explore effects on produc-
tivity, as we do.
By contrast, a large body of work investigates whether communication affects team per-
formance in laboratory experiments. Early research, typically by psychologists, focused on
5
the shape of the communication networks (Bavelas and Barrett 1951, Leavitt 1951, Guetzkow
and Simon 1955). Later on, Weber and Camerer (2003) study how productivity-enhancing
languages emerge and are disrupted during mergers. Cooper et al. (1992) and Blume and
Ortmann (2007) show that pre-play communication about strategies increases efficiency in
weak-link coordination games. An advantage of laboratory experiments is that informal
messages between subjects can be observed, something that is much more difficult in real
organisations3. It is of course unclear how results from the laboratory extrapolate to the
field.
The experimental variation in this paper relates to the co-location of teammates. This
suggests a link with Catalini (2016), who uses the relocation of departments in a French
university to analyse how search costs and monitoring costs vary with physical proximity
between academics4. Another related paper is Bloom et al. (2015), who find that working
from home increased productivity in a Chinese call centre. The contrast with our finding
that co-location increases productivity is likely the result of the many differences between
the two settings. A very important one is the complexity of the production process. While
the simple individual production of Bloom et al. (2015) can be easily monitored and co-
ordinated remotely, we show that, in organisations requiring tight co-ordination between
colleagues, working in the same place may have significant advantages, especially when tasks
are relatively information-intensive.
Plan We describe the institutional setting in Section 2. We introduce the data and the
empirical strategy in Section 3. We present the main results of the paper in Section 4. In
Section 5, we provide evidence in support of the face-to-face communication mechanism.
Section 6 explores the heterogeneity of the main results. In Section 7, we provide a cost-
benefit analysis of the face-to-face communication effect. Section 8 concludes.
2 Institutional Setting
We exploit a natural experiment in the Operational Communications Branch (OCB) of the
Greater Manchester Police (GMP). The OCB is the unit in charge of answering 999 calls
from members of the public and managing the allocation of officers to the corresponding
incidents. Figures 1 and 2 provide a simplified visualisation of this production process.
3In our study, we can (partially) observe the electronic messages between teammates, but not the face-
to-face messages.
4A large body of work examines the relation between geographical proximity, assumed to facilitate face-to-
face interactions, and the diffusion and generation of knowledge (Jaffe et al. 1993, Thompson and Fox-Kean
2005). A challenge here is to disentangle geographical distance from other factors, such as knowledge or
social distance, correlated with it. In addition, the typical bird’s-eye view of these papers does not allow for
the isolation of mechanisms explaining why geographical proximity matters.
6
Call Handler Emergency calls requesting the police are allocated to call handlers using
a standard computerised queuing system. A result of the system is that any handler can
respond to calls from any Manchester location.
The handler questions the caller, assigns an opening code and a grade level, and records
any information deemed relevant. The grade level can range from one to three and, very
coarsely, determines the official urgency of an incident. The opening code describes, horizon-
tally and at a fairly detailed level, the type of issue that the incident relates to (neighbour
dispute, disturbance in licensed premises, etc.). The description of the incident will include
information on the individuals involved, their states of mind, the existence of prior history
between these individuals and the likelihood of further incidents in the near future5.
All the information above is recorded in GMPICS, a specialised IT package used
throughout the GMP to create, record and manage incidents6. The handler ticks a box
in GMPICS to officially create the incident, and then indicates her status as ’not ready’
(which allows the handler, among other things, to step away from his desk), or instead
’ready to receive new calls’. Under the ’ready’ status, a call can arrive at any point and
must immediately be answered by the handler. Once an incident has been created, the
handler cannot keep adding details to it.
Radio Operator When an incident is created, it immediately appears on the computer
screen of the radio operator overseeing the Manchester subdivision where the incident oc-
curred. The allocation of incidents to radio operators is deterministic, since at any point in
time there is only a single operator in charge of a specific subdivision (a corollary of this is
that handlers do not decide to which operator they assign an incident). Radio operators are
in charge of processing the information inputted by the handler and allocating police officers
to incidents, on the basis of incident characteristics and officer availability.
Lacking a direct link with the caller, the radio operator has to rely on the information
recorded by the handler in GMPICS. It is, however, often the case that additional information
is needed before an officer can be allocated. For instance, written descriptions of incidents
are regarded by radio operators as lacking sufficient emotional content, which makes it harder
to understand the state of mind of the victim and the impact that the incident has had on it.
5The language used in these descriptions is highly efficient, as it includes a large number of official
and unofficial abbreviations for features of incidents that appear repeatedly. For instance, official abbre-
viations include A/ABAN (apparently abandoned) and NFA (no fixed abode). Unofficial but widely used
abbreviations include XXX (very drunk). Despite this, the written descriptions inevitably fail to perfectly
communicate the full richness of the information gathered by the call handler.
6Our personal conversations with multiple handlers, radio operators and their supervisors indicate that
GMPICS is widely regarded as an efficient system. GMPICS was developed in-house and incrementally over
more than two decades. OCB staff receive extensive training and accumulate considerable expertise in its
use.
7
Similarly, a full characterisation of the physical surroundings where the incident occurred, or
of the complex relationships between the people involved are often difficult to communicate
in writing. A complete picture of the incident is often necessary to efficiently match incidents
with officers, advise the attending officer of important details that she may find at the scene,
or even understand the level of priority that the incident merits7.
The additional information can be acquired by conducting targeted searches on specific
individuals or addresses in the GMP databases, asking the call handler or contacting the
initial caller directly. Typically, the allocation of an officer will be delayed until the radio
operator can gather this information.
Teamwork In this paper our definition of a team comprises the combination of the call
handler and the radio operator. While officially equal in rank, the positions of call handler
and radio operator are associated with different status within the OCB. This stems from the
fact that the job of radio operator is both more complex and more stressful, as it involves
carrying out a variety of tasks in parallel and bearing the ultimate responsibility for the
outcomes of incidents. The decision-making authority of radio operators is also wider. For
instance, they can overrule the code and grade allocated by the handler (although this is in
practice rare). Accordingly, radio operators earn a higher salary and have on average more
experience in the OCB. Many in fact transferred into radio operations from the call handling
desk, a move widely seen in the organisation as a promotion.
Face-to-Face Communication When a radio operator regards the electronic description
of an incident insufficient, an efficient and fast way to gather this information is to ask the
handler in person8. Alternatively, it is often the handlers who decide to complement the
written description with additional information delivered face-to-face. When handler and
operator are communicating in person, the handler will need to be in ’not ready’ status, as
she may otherwise be forced to abruptly end the conversation when a new call arrives.
Our conversations with members of the OCB suggest that they attach several advan-
7Regarding the optimal matching between incidents and officers, note for instance that some incidents
can be responded alternatively by sworn police officers or by PCSOs (police community support officers)
and the likelihood that the more extensive legal powers and expertise of police officers may be needed
is decision-relevant information. Similarly, incidents involving vulnerable individuals require officers with
specialist training, which makes it critical to understand the condition of the caller and other individuals
affected. More generally, certain officers are particularly well-suited to dealing with specific types of incidents
or individuals.
8Sending an electronic message to the handler is possible in the GMPICS Communicating on the phone
is theoretically possible but in practice unlikely, as a handler in status ’ready to take new calls’ cannot be
contacted on the phone without first alerting the handler’s supervisor. On the other hand, a handler can
easily switch status from ’ready’ to ’not ready’ if an operator approaches in person with the need to clarify
some doubt.
8
tages to face-to-face communication: firstly, it is a highly efficient channel, in that it allows
for rapid, short exchanges that provide immediate feedback to both teammates. Secondly,
non-verbal cues can help to communicate fuzzy concepts that in writing would require lengthy
descriptions. Thirdly, it is a more natural vehicle for the use of colloquialisms that can suc-
cinctly and effectively communicate characteristics of an incident including the physical or
mental condition of the individuals involved. For a variety of reasons (including both the
potential for misunderstanding and the possibility of future audits of the official GMPICS
descriptions) these colloquialisms are less likely to be used in written communication9.
Co-Location In the period between November 2009 and January 2012, OCB staff were
spread across four buildings or ’rooms’, each in a different part of Manchester: Claytonbrook,
Leigh, Tameside and Trafford. Every room accommodated the radio operators overseeing
the surrounding subdivisions (Figure 3 displays the areas overseen from each of the four
locations). As discussed earlier, call handlers were not geographically specialised. However,
for historical reasons they were also dispersed across the four locations. This assignment
meant that radio operators would sometimes be reading the descriptions of incidents created
by same room handlers, while on other occasions the handlers were based in a different part
of Manchester.
In January 2012, a major reorganisation of the OCB reassigned all handlers to a single
location (Trafford), while radio operators were divided between Claytonbrook and Tameside.
This put an end to the natural experiment that we study here.
Measures of Performance As is the case with other public sector organisations (Dewa-
tripont et al., 1999), objectives in the GMP are multifaceted and often vague. The prevention
of harm or damage to property, the satisfaction and reassurance of the public, and the appli-
cation of sufficient but proportionate force are all important objectives that escape precise
measurement. Capturing every one of these objectives with explicit measures of performance
is therefore an impossible task. Our first measure of performance is the allocation time of
an incident: the time elapsed between its creation by the call handler and the allocation of
an officer by the radio operator. We also study the effect of distance on response time: the
time between creation and the officer reaching an incident’s scene10.
9Note that face-to-face communication has several important features here. Firstly, it is two-way. Sec-
ondly, it is oral. Thirdly, handler and operator are able to observe each others’ faces. Because communication
by phone is not a realistic alternative in our setting, we are unable to precisely disentangle the features of
face-to-face communication that are responsible for the increase in productivity.
10These two measures are strongly correlated, since response time is equal to allocation time plus the
officer’s travel time. It is worth noting that better information on the part of the radio operator could affect
travel time also. Imagine, for instance, a radio operator deciding whether to allocate the closest officer, or
an officer who is further away but has a specialised skill. Better information could reveal that the incident
9
The two measures that we use are undoubtedly partial. They do not capture, for
instance, any notion of whether the ’right’ officer was allocated to an incident, or whether
the attending officer was in possession of all the relevant information prior to arrival. They
also do not indicate whether or not excessive or insufficient resources were allocated to resolve
an incident.
The two measures are nevertheless very important for the organisation that we study,
for two main reasons. The first reason is that the GMP is partly evaluated on the basis
of these variables. Specifically, nation-wide numerical targets for maximum allocation and
response times were introduced by the UK Home Office in 200811. The second reason is that
these measures are regarded as important determinants of the public’s satisfaction. UK-wide
survey evidence suggests that response time is one of the most important variables predicting
citizens’ satisfaction with the police forces (Dodd and Simmons, 2002/03).
Table 1 provides direct evidence of this in our setting. In the GMP, a subset of callers
is regularly questioned about their satisfaction with the treatment they received, after their
incident has been closed. We obtained these surveys and linked the response time in our
dataset with the answers to the two most important questions (our dataset is described in
detail in the next section). Table 1 shows that there are very strong correlations between
these variables. For instance, in incidents where police response time was below the maximum
target prescribed by the Home Office, satisfaction was .14 standard deviations higher. Callers
were also more likely to report that their opinion of the police had improved. The effects of
response time on satisfaction are not linear, but instead concentrated at the top end of the
response time distribution (Figure 4)12.
Overall, there is substantial evidence that the leadership of the GMP internalised the
need for minimising allocation and response times. One example can be found in the GMP
Incident Response Policy manual April 2011. Allocation and response times are the only
tactical performance measures mentioned in the manual. In particular, this indicates that13:
does not require the specialised skill, and that the officer with the shorter travel time can be safely allocated.
11For Grade 1 crimes, for instance, these targets were for a maximum of two minutes and fifteen minutes
for allocation time and response time, respectively. The equivalent targets for Grade 2 (respectively Grade 3)
were 20 and 60 minutes (respectively 120 and 240 minutes). While these targets were nominally scrapped in
June 2010, police forces continued to regard them as objectives and to believe that they were being informally
evaluated on this basis (Curtis, 2015). Information on response times was also frequently discussed in the
reports produced by the HMIC (the central body that in the UK regulates and monitors police forces). For
an example, see HMIC (2012).
12While we do not claim that these coefficients can be interpreted as causal effects, they suggest at the
very least the type of evidence on which the GMP based their decisions. Unfortunately, we are unable to
use the victim satisfaction variables as dependent variables in the main analysis of the paper. The number
of survey responses is relatively low and it mostly falls outside our baseline sample period.
13Additional examples include the following. The launching in April 2010 of a website where the public
could access up-to-date statistics on response times, separately for each of the twelve divisions (Pilling,
2010). Secondly, the fact that throughout our sample period every report by the GMP to the Manchester
City Council Citizenship and Inclusion Overview and Scrutiny Committee provided detailed statistics on
10
The OCB will produce daily reports regarding graded response performance. This
will include the % of incidents resourced within target and the % attended within
target for each division. This will enable ongoing analysis of the accuracy of the
resource management of that BCU.
3 Empirical Strategy
In this section we present and discuss the dataset and main variables of the paper. We also
first explain the empirical strategy to estimate the effect of co-location on performance, and
then justify it with a set of balancing tests. Establishing such a causal effect is not an easy
task. In addition to exploiting the idiosyncratic allocation of incidents to handlers, which we
outline in this section, we will need to consider the possibility that co-location represents a
proxy for unobserved characteristics of the handler, or handler/operator pair. We postpone
the discussion of these confounding effects, together with the tests that we use to evaluate
them empirically, to Section 4.
Dataset Our baseline dataset contains every incident reported through the phone to the
GMP between November 2009 and December 2011. We restrict our attention to incidents
where the handler allocated the call a grade below or equal to three, therefore transferring
responsibility to a radio operator rather than to a divisional commander. For every incident
we observe, among others, the allocation and response time, the location of the incident, the
grade and (horizontal) opening code, the identity of the call handler and radio operator, and
the desk position from which the handler took the call. The dataset was made available to
us under a strict confidentiality agreement.
Table 2 provides basic summary statistics for the main variables in our study. Note
first that our sample size is very large, as it includes close to one million incidents. In around
one in four observations the handler and operator are in the same room. The performance
variables are highly right-skewed. For response, for instance, the median time is 19 minutes,
while the average time is more than four times larger14.
We find that there is considerable gender and age variation among handlers and op-
erators. Consistently with our earlier discussion of the differences in status, operators are
significantly older than handlers. They are also more likely to be female, likely the result of
females being more likely to regard the OCB as a long-term career choice.
response times and, if these were deemed unsatisfactory, a list of reasons for the failure.
14The maximum value is more than 15 days, likely the result of some error in the classification of the
incident. The fact that the left hand side variables in our regressions are in logarithmic form should dampen
the effect of outlying observations. Nevertheless, in Appendix Table A10 we show that our baseline estimates
are robust to the exclusion of these outliers.
11
Intuition of Empirical Strategy The computerised queuing system allocating calls to
handlers works as follows. As calls come in, they join the back of a call queue. The system
matches the call at the front of the queue with the next handler that becomes available. If
the call queue is empty and several handlers start to become available, they form their own
queue. The system then matches the handler at the front of the handler queue with the next
incoming call. The system creates exogenous variation in the co-location of the handler and
operator involved in an incident. We visualise this notion in Figures 5A and 5B where, for
simplicity, we assume that there are only two locations (Trafford and Leigh), rather than
four.
Assume that, within a relatively narrow time horizon, two calls (one from Trafford,
one from Leigh) reach the queuing system, and that two handlers (one based in Trafford, the
other in Leigh) become available. The exact timing at which handlers become available is
the result of a large number of factors, including the length of their previous calls, the time
at which the calls started, the existence and length of ’not ready’ periods etc. Similarly, the
exact order at which the calls arrive is the result of many factors, including the times at
which the incidents occurred, the delay in dialling 999 and the further delay in opting for a
police service and being transferred to the GMP. These factors are arguably orthogonal to
the factors determining the order at which handlers become available. It follows that two
handlers that are on duty during the same time period should be equally likely to be the
one assigned to an incoming call. If, as in Figure 5A, the handlers are assigned calls from a
subdivision that their room oversees, they will be co-located with the radio operators with
whom they have to communicate electronically. For arguably exogenous reasons, they may
instead be assigned a call (and have to communicate with an operator) from a different area
of Manchester. We capture this variation with the dummy variable SameRoom, which is
the main independent variable in our study.
We have just argued that, conditional of the exact time period at which a call arrives,
on duty handlers should be equally likely to be assigned that call. In practice, some rooms
(for instance Trafford) are bigger than others (e.g. Leigh) and therefore contain a larger
number of handlers. This implies that the likelihood of SameRoom = 1 will be mechanically
higher if the call originates in a Trafford neighbourhood, relative to a Leigh neighborhood.
Calls originating from Trafford and Leigh may also have different characteristics, which
could independently affect their average allocation and response times. Therefore, our claim
regarding the exogeneity of the variable SameRoom is only conditional on hour (i.e. year X
month X day X hour of day) and (handler and operator) room fixed effects15.
15In most regressions, we use hour fixed effects to condition for the exact time period at which a call
arrives. Our findings are qualitative unchanged if we instead control for the half-hour or quarter-hour period
(see, for instance, Appendix Table A7).
12
Estimating Equation Our baseline estimating equation is:
yi=βSameRoomj(i)k(i)+θt(i)+λj(i)+µk(i)+πg(i)+γh(i)+Xi+i(1)
where yiis a measure of OCB performance for incident i. Throughout our paper, allocation
and response times are measured in log form, both for ease of interpretation of the coeffi-
cients and in the presence of right-skewness to minimise the effect of outlying observations.
Consistently with our earlier discussion, we control for θt(i)(the fixed effect for the hour tat
which the incident arrived) and λj(i)and µk(i)(the fixed effects for the rooms jand kfrom
which the incident was handled and dispatched). Our main independent variable of interest
is the dummy SameRoomj(i)k(i), which takes value 1 when rooms jand kcoincide.
We also control in our baseline specification for πg(i)and γh(i)(the fixed effects for
the individual handler gand operator hassigned to the incident) and by other incident
characteristics (such as the assigned grade) included in the vector Xi. These latter controls
are not essential for identification, but should contribute to the reduction of the standard
errors. We cluster these standard errors at the operator room and year/month level. In
Appendix Tables A1 and A2 we show that the baseline findings are robust to the inclusion
or exclusion of additional controls and to alternative clustering choices.
Balancing Tests Our first set of tests examines the balance of incident (grade, location
of the incident scene), worker (gender, age, location of the desk, current workload) and
room time-varying (measures of current average workload) variables across the co-location
of handler and operator. To perform these tests, we regress each variable on SameRoom,
after controlling for hour and room fixed effects. These standard balance regressions are
essentially variations of equation (1), with incident characteristics on the left hand side
and without the right hand side non-essential controls. To ease interpretation, non-binary
dependent variables are standardised.
The results in Figure 6, where we label each row in the left axis by the regression
dependent variable, plot the estimated confidence intervals of SameRoom. To illustrate the
need for our empirical strategy, we report for every variable the estimates of two regressions:
with and without the hour and room controls. We find first that SameRoom is (uncondition-
ally) strongly correlated with incident characteristics: the estimates are large and most are
statistically significant. The introduction of the hour and room controls, however, greatly
decreases both the standard errors and the estimates, which then become extremely small in
magnitude. For instance, among the non-binary variables all the estimated coefficients imply
an effect of SameRoom lower than .005 standard deviations of the dependent variable16.
16Appendix Figure A1 shows that it is the room controls that are critical to the empirical strategy. Failing
to control for the hour of the incident does not lead to a stronger correlation between SameRoom and the
incident characteristics.
13
We also find that, after including the hour and room controls, only two of the sixteen
coefficients are statistically different from zero at the 5% level. Although higher than one-
in-twenty, we regard this ratio as remarkably low considering that the regressions are run on
close to one million observations. Note further that the significant coefficient associated with
the Grade 1 regression is both small in magnitude and negative, suggesting that same room
incidents are more likely to be allocated a low priority by the handler, and should therefore
have higher allocation and response times.
The variables in Figure 6 do not include the incident opening code, an important determinant
of allocation and response times. The opening code is captured empirically by a large
set of dummy variables that are mechanically correlated with each other, which creates
a mechanical correlation on the results of balance regressions based on equation (1). We
therefore switch the dependent and independent variables, and estimate:
SameRoomj(i)k(i)=αi+θt(i)+λj(i)+µk(i)+i(2)
where αiare the fixed effects for the incident opening code. We find that the F-statistic of
joint significance of these effects is 1.15 (P-value = .30), suggesting that SameRoom and
the opening code dummies are conditionally uncorrelated. Overall, we interpret the results
of estimating (2) and the regressions of Figure 6 as consistent with our assumption that
co-location between the handler and operator of an incident is conditionally orthogonal to
incident, handler, operator and room time-varying characteristics.
4 Baseline Results
In this section we present and interpret the baseline results of the paper. We then use a
number of tests to confirm that these estimates can indeed be interpreted as the causal
effect of co-location on performance, rather than the result of co-location being a proxy
for unobserved determinants of allocation and response time. We also explore whether the
quality of the response is different for co-located incidents. The section concludes with an
investigation of potential spillovers onto other (contemporaneous) incidents assigned to the
radio operator.
Baseline Estimates Our baseline regressions are variations of equation (1). In the first
two columns of Table 3 we find that allocation and response time are approximately 2%
faster on average when handler and operator are located in the same room. At the mean
(respectively, median) of the independent variable, this 2% translates into 76 seconds (re-
spectively, 5.4 seconds) saved in terms of allocation time. For response time these savings
are of 104 and 20 seconds, evaluated at the mean and median respectively. Aggregated over
all the incidents in a month, the savings amount to approximately 900 hours.
14
We also investigate whether these times are ’on target’. Throughout our sample period,
it was an explicit objective of the UK Home Office that allocation and response times should
typically be below certain levels17. As a result, the GMP recorded information on whether
the target maximum time was exceeded for an incident. We use these dummies as dependent
variables and find in Columns 3 and 4 that the likelihood of being on target is higher when
SameRoom = 1. For instance, the coefficient in Column 3 indicates that the likelihood of
missing the allocation target decreases by .4 percentage points (around 2% of the mean of
.25), when handler and operator are co-located.
Lastly, we find in Column 5 no evidence of co-location affecting the likelihood that
incidents classified as crimes are cleared by the GMP18.
Estimates by Distance Inside the Room Table 3 has established that co-location of
handler and operator is associated with higher performance, relative to them working in
rooms in separate areas of Manchester. We now investigate whether performance improves
as distance decreases even when handler and operator are already working in the same room.
In addition to providing richer evidence on the functional form of the relation between
proximity and teamwork performance, within-room variation allows the introduction of han-
dler/operator pair fixed effects in the regression. We argue in the next subsection that the
introduction of these controls strengthens the credibility of our claim regarding the causal
interpretation of the estimates.
The assignment of desks to workers was as follows. Inside a room, a fixed desk would be
earmarked for the radio operator overseeing a specific subdivision. Handlers, on the other
hand, were free to work from any remaining and available desk. To measure the within-
room distance between desks, we use yearly-updated floorplans of the four OCB rooms (see
Figure 7 for an example)19. We set distance to zero if handler and operator are not in the
17See Section 2 for details about these targets. The fact that the ’on target’ dummies are affected by
co-location confirms that the results are not disproportionately due to extreme values of the allocation and
response time distributions.
18The absence of a statistically significant effect on the likelihood of clearing the crime may be due to
the fact that our sample size is much smaller in this regression, since only around 16% of incidents are
crimes. Nevertheless, it is surprising given the findings of Blanes i Vidal and Kirchmaier (2017) that a faster
response time increases the likelihood of clearing the crime. In that paper, the identification strategy exploits
discontinuities in distance across locations next to each other but on different sides of division boundaries.
In the current paper, co-location between handler and operator would likely not be a valid instrument for
response time. The exclusion restriction is unlikely to be satisfied because co-location could affect clearance
likelihood through many channels in addition to faster response times.
19The floorplans are unfortunately not to scale, which prevents us from measuring distance in metric units
and is likely to introduce measurement error in the within-room distance variable. Instead, desks are depicted
in the floorplans in a matrix (x, y) format. Our measure is therefore the euclidean distance between desks
inside this matrix. D=p[(yRO yH)2+ (xRO xH)2], where yRO is the position of the radio operator
along the row dimension and the other coordinates are defined accordingly. As an example, two adjacent
desks in the same row or column are at a distance of one, while the distance between two diagonally-adjacent
15
same room, and add the interaction of distance and the same room variable to our baseline
specification.
We provide two types of evidence. In Table 4 distance is measured parametrically,
in logs. In Figure 8 we instead split distance into four categories of approximately equal
sample size, and plot the interactions of SameRoom with these dummies. The estimates
from both specifications indicate that teammates that sit closer together are more productive.
In the parametric estimation, a 10% decrease in within-room distance is associated with a
2.6% increase in the effect of SameRoom on allocation time. The non-parametric evidence
is perhaps more informative. We find that incidents assigned to workers separated by a
distance lower than 2 (e.g. diagonally adjacent desks at most) are on average allocated and
responded 4% faster. The effect decreases monotonically with distance and becomes zero
when handler and operator are separated by a distance higher than 420.
A question evident in Figure 8 is why should the benefits of proximity be so local in
our context, to the point where being on the other side of the room is equivalent to being on
the other side of Manchester. While we cannot provide a definitive answer, our conversations
with GMP staff have pointed to the fact that some handlers’ supervisors (labelled to us as
’old-school’) discourage the communication between handlers and operators. This is because
these supervisors feel mostly responsible for managing the flow of incoming calls and therefore
view conversations that occupy the handlers’ time (even if they benefit the rapid allocation
of officers) as hindering that objective. These attitudes often make handlers unwilling to
attract attention by stepping far away from their desks.
Establishing a Causal Interpretation Our preferred interpretation of the findings in
Table 3 is that: (a) being physically closer allows teammates to communicate face-to-face,
and (b) in settings where information is complex and must be processed relatively quickly,
this additional communication channel is performance-improving. An alternative interpre-
tation is that call handlers may be better informed or motivated to deal with incidents
originating in the geographical area that surrounds their workplace. To understand this
potential confounding effect, note in Figures 3 and 5A that SameRoom = 1 when a handler
based in a location is allocated an incident from the geographical area surrounding that
location. If handlers are more effective at dealing with cases that occur closeby, the findings
in Table 3 may reflect proximity to the incident scene, rather than to co-location with the
co-worker.
A second alternative interpretation is that co-location may be a proxy for some un-
desks is 2=1.4.
20To interpret this, note that two desks that are three positions apart along both the row and the column
dimension are separated by an euclidean distance of 4.2. Two desks separated by three positions along one
dimension and two positions along the other are at a distance of 3.6.
16
observed dimension of similarity between teammates. In an extreme example, imagine that
workers communicate through room-specific language, which makes electronic communica-
tion with individuals outside one’s room less efficient. This would be the case if, for instance,
there are strong local dialects and the workers in a room are drawn from the neighbourhoods
surrounding that room. In that case, co-location would represent a proxy for the ease of elec-
tronic communication between teammates, as opposed to providing a performance-improving
additional communication channel.
In Columns 3 and 4 of Table 4 we find evidence that is inconsistent with the two
alternative interpretations above. We add a set of handler/operator pair fixed effects to the
baseline regressions, and estimate the effect of distance within the room on performance.
Because handlers and operators do not typically change workplace, the introduction of pair
fixed effects effectively absorbs the same room variable.
We find that the same pair of workers operating from the same room are more pro-
ductive when their desks are closer together. The estimated coefficients are in fact almost
identical to those in Columns 1 and 2, without the pair fixed effects. These effects absorb
any time-invariant characteristics of the match between handler and operator (including the
match between the handler and the location of the incident). The robustness to their in-
clusion therefore confirms that it is the location of the handler relative to the operator that
causes the estimated Table 3 decreases in allocation and response times21.
A second strategy to evaluate the above is to perform a placebo test using the post-2012
information. As we mentioned in Section 2, the 2012 reorganisation of the OCB relocated
all the call handlers to Trafford, while the radio operators were split between Claytonbrook
and Tameside. Therefore, handlers and operators never shared a room after 2012. Using the
information on the workplaces of handlers and operators just before the reorganisation, we
can construct ’placebo same room’ variables taking value one when an incident is allocated
to a pair of teammates that used to be co-located22 . In the estimation of (1) we now
interact the same room variable with dummies for each of the five semesters comprising our
21A potential caveat here is of course that handlers choose daily the desks where they sit, conditional
on these desks being unoccupied. Therefore, within-room distance between handler and operator cannot
be considered random. This would be problematic to the extent that it is correlated with time-varying
characteristics of their match. For instance, it may be that handlers choose to sit next to operators with
whom they have worked on more incidents in the past (if these seats are available). While this is a theoretical
possibility, we note two things. Firstly, handlers and operators who have worked together on more incidents
in the past are empirically not more likely to sit closer to each other (Appendix Table A11). Secondly, the
effect of within-room distance on allocation time is robust even after controlling for the interaction of the
handler/operator pair and the year/semester pair (Appendix Table A12). In fact the estimates are very
similar, if anything larger. Of course, the introduction of such a large number of fixed effects implies that
this regression is highly demanding, as most of the variation in within-room distance is absorbed.
22Following the reorganisation radio operators remained in their previous roles in terms of the subdivisions
for which they dispatched officers. Therefore, a post-2012 handler-operator match continues to capture
accurately whether the handler is assigned a case from the geographical area around her pre-2012 workplace.
17
baseline period (the last semester of 2009 includes only two months, since the data starts in
November). We then use the post-2012 data to estimate (1) again, interacting the placebo
same room variable with semester dummies. The coefficients are displayed in Figure 9.
We find that the same room variable is essentially zero for every semester of the post-
2012 period, while it is negative for most of the baseline period. Note in particular the
large difference in the estimates between late 2011 and early 2012. This difference suggests
that the same pairs of workers that were able to deliver higher performance when jointly
assigned to an incident ceased to do so when they stopped being co-located. The evidence
in Figure 9 reinforces the conclusion that it is indeed distance between co-workers, rather
than unobservables correlated with distance, that improves allocation and response times23 .
Effects on the Type of Officer Sent We now study whether the faster allocation and
response times associated with co-located incidents are at the expense of other dimensions of
the quality of the response. As we argued in Section 2, these are typically difficult to measure
empirically. One aspect that we can observe in our dataset is the rank and experience of
the officer that was sent to the incident. Officers with the rank of ’response officer’ are
trained (and accumulate on-the-job experience) specifically to deal with incidents that the
police is alerted to. Neighbourhood officers are instead in charge of patrolling but can be
called to attend certain types of incidents, for instance if response officers are temporarily
unavailable. If the likelihood of sending an officer with the rank of ’response officer’ is lower
for co-located incidents, a faster response time might be interpreted as being at the expense
of lower ’quality’.
In Column 1 of Table 5 we find, however, that this is not the case. In Column 2, we
regress the officer’s number of years in the force on the same room dummy, and again find
no correlation. We conclude that, to the extent that we can measure quality, there is no
evidence that co-location is associated with both a faster response and a worse response.
Spillovers to Other (Contemporaneous) Incidents We now investigate the existence
of potential spillovers from same room incidents into other contemporaneous incidents. Ra-
dio operators typically have open (i.e. yet to be allocated) several incidents at the same
time. Theoretically same room incidents can generate both positive and negative spillovers.
Positive spillovers will occur, for instance, when the time and effort that the operator saves
23Interestingly, Figure 9 also suggests that the effect of co-location on productivity may have been increas-
ing over time. In November 2009 a major reorganisation had taken place that created a Manchester-wide
handling system and split the roles of handler and operator. Workers may have taken time to adapt to
their new roles, and to fully exploit the sources of higher productivity in the new setting. In particular, the
coefficients of Figure 9 are consistent with workers learning about the performance-improving potential of
co-location over time. We return to this issue in Section 6, where we investigate whether individuals that
have worked together on more incidents in the past benefit more from co-location.
18
on a same room incident (as a result of being able to gather information more efficiently) is
redistributed to other contemporaneous incidents. Negative spillovers are equally plausible.
One potential channel would be operators assigning higher priority to incidents that have
been created by co-located handlers. If that was the case, the improvement in performance
for same room incidents that we document in Tables 3 and 4 would be, at least partially, at
the expense of other contemporaneous incidents, as attention is diverted away from them.
To study whether spillovers are in fact present in our setting we first replicate our
baseline specification and use as independent variable of interest the percentage of incidents
assigned to the operator that, in the period surrounding the index incident, are same room
incidents. Positive spillovers should lead to a negative coefficient for this variable because, if
same room incidents are easier to deal with, a higher share of those will allow for more time
and effort being available for the index incident. Negative spillovers would instead imply
that valuable attention or resources are diverted away from the index incident when other
incidents are handled in the same room, leading to higher allocation and response times, and
a positive coefficient in this regression24.
We find in Table 6 no evidence of either positive or negative spillovers. Given the
uncertainty about the time horizon on which spillovers might occur, we calculate the inde-
pendent variable at the 60, 30, and 15 minutes time horizon. We find in every case that a
higher share of same room incidents does not translate into different performance for other
contemporaneous incidents.
We perform a second exercise by ordering the incidents assigned to each operator
according to the time at which they were created. We then create leads and lags for the
four incidents that, for a given operator, immediately precede and follow a same room
incident25. The estimated coefficients in Figure 10 are inconsistent with the existence of
negative spillovers, since none of the lag and lead coefficients are positive and statistically
different from zero. One of the eight coefficients is negative, providing at most weak evidence
of some positive spillovers. Overall, we interpret Figure 10 as suggesting, consistently with
Table 6, that the improvement in performance of same room incidents is neither at the
expense nor to the benefit of other contemporaneous incidents.
24We use the baseline sample for this exercise, since in principle spillovers could occur both to same room
and to non-same room incidents. In Appendix Table A6 we restrict the sample to including only non-same
room incidents and find very similar effects.
25Because incidents are not dispatched immediately, a same room incident could create spillovers to other
incidents that were assigned to the same operator earlier in time.
19
5 Mechanism
The findings above have established the existence of a causal relation between co-location
and performance. Our preferred explanation is that co-location permits face-to-face inter-
actions which communicate relevant details about incidents. In this section we first discuss
alternative mechanisms, and then provide evidence that is consistent with the face-to-face
communication mechanism but inconsistent with these alternative mechanisms.
Alternative Mechanisms The first alternative channel consists of the handler exerting
more effort in the transmission of the GMPICS electronic information under co-location.
The second alternative channel is similar: the operator might exert more effort in the in-
terpretation of this information, and the subsequent allocation of an officer, for co-located
incidents. A third potential channel would be the preferential allocation of scarce resources,
such as police officers, to co-located incidents and in detriment of other incidents. We do
not consider this third channel here because the evidence in Section 4 showing the lack of
negative spillovers is inconsistent with it.
We can think of two plausible reasons why workers may exert more effort under co-
location, even in the absence of face-to-face communication26. The first reason would be
some type of silent psychological effect leading to higher priority assigned to incidents that
will be read, or were written, by a same room co-worker. The second potential reason would
be handler and operator exerting silent visual peer pressure on each other, similarly to the
visual pressure among supermarket cashiers identified by Mas and Moretti (2009).
We regard this second reason as unlikely, in particular with regards to the handler
exerting peer pressure on the operator, as several features of the institutional setting are
inconsistent with it. Firstly, while handler and operator are ’teammates’, they are not
actually ’peers’. As discussed in Section 2, operators are both more senior and uniquely
responsible for the allocation of the incident, which makes it improbable that they may feel a
lot of pressure from handlers. There is in fact little scope for handlers to even be aware of the
allocation and response times of the incidents that they created, unless they actively search
for them in the GMPICS system. Furthermore, the cognitive and desk-bound activities
of the operator are difficult to monitor visually, especially relatively to manual tasks like
supermarket item checking. For instance, an operator may appear busy by virtue of looking
26Note that face-to-face communication could lead to the higher motivation of its receiver (in this case,
the radio operator). Storper and Venables (2004) argue persuasively that face-to-face communication can
serve as a signal about the importance of a task, thereby stimulating a ’psychological rush’ that leads to
greater and better efforts. In our context, it is possible that discussing an incident in person may induce the
operator to devote more time and effort to it, and this channel is not incompatible with the higher ability
to deal with the incident resulting from a richer information set. Similarly, to the extent that the act of
communicating face-to-face itself requires effort by the handler, it is by construction correlated with it.
20
at her computer screen, while in fact paying little attention to her work. In addition, there
are significant physical barriers (computer monitors, desk screens...) between the workers
in the rooms of our setting. These barriers make it impossible to observe the behaviour of
all but the closest co-workers, unless a handler actively stands up from her desk. While it
is possible in theory for a handler to stand up and watch over the operator’s shoulder in
silence, we think that is an unlikely possibility.
Evidence on the Handler’s Effort Mechanism The first alternative mechanism con-
sists of the handler communicating better electronically. We now test whether there is any
evidence of the handler being more precise and thorough in the electronic communication
of co-located incidents. We have three good measures of this communication. The first one
is the handler’s creation time: the time elapsed between the handler answering the call and
the creation of the incident in the GMPICS system. Remember that this creation time takes
place before the radio operator is informed of the incident’s existence (see Figure 2). We
expect that a more thorough and precise electronic communication will require more time
devoted to writing the description of the incident, and probably also to the elicitation of
the information from the caller. In Column 1 of Table 7 we however replicate our baseline
specification using creation time as dependent variable, and find that it is unaffected by
co-location.
As complementary measures of the quality of the electronic communication, we use
the number of characters and number of words in the first line of the description of the
incident27. Unsurprisingly, these two variables are very correlated with each other, even after
conditioning on the baseline set of controls (Appendix Table A8). They are also strongly
correlated with the creation time, suggesting that, despite their coarseness, there is valuable
information in them. In Columns 2 and 3 of Table 7 we find that these variables are not
different for co-located incidents.
To conclude, we find no evidence that the electronic information inputted by handlers
is better or worse for co-located incidents, relative to other incidents. Therefore, higher effort
on the handler’s part and the resulting better electronic communication does not appear to
be an important mechanism in our setting28.
27Unfortunately, due to a combination of technical challenges and the extreme confidentiality of this
information, we were not able to obtain the full content of these descriptions. The first line of the incident
description consists of a maximum of 210 characters, and serves as a quick summary of the nature of the
incident. When operators have more than one incident open at one time, they typically only see the first
line of this description, which then plays a role similar to the subject of an email in an inbox.
28Table 7 also suggests that co-located handlers do not devote less time and effort to the electronic
communication, in the expectation of complementing the information face-to-face. One explanation of this
lack of substitution may be the fact that an electronic ’paper trail’ needs to be established by the handler,
so that other staff members can access that information in the future and the handling of the incident is not
criticised during later audits.
21
Evidence on the Face-to-Face Communication Mechanism The mechanisms out-
lined above entail different predictions about the behaviour of the handler after the incident
has been created, in particular with respect to the likelihood that the handler is ’not ready’
to take a new call. Consider first the alternative mechanism whereby the operator exerts
more effort for co-located incidents. Handlers are continually monitored by their supervisors,
and are expected to remain at their desks unless there is a reason to leave them. Therefore,
any handler exerting visual pressure on an operator would typically be doing so from her
desk, an activity that is perfectly compatible with being available to take a new call. Simi-
larly, the notion that operators are psychologically prone to exert more effort for co-located
incidents does not require any change in behaviour on the handler’s part. In particular, it
does not require handlers being more or less willing to take new calls after creating co-located
incidents.
Face-to-face communication, on the other hand, is an activity that typically requires
the handler’s full attention. Being in ’ready’ status while talking to an operator risks having
to either ignore an incoming call (an offence so serious that it is likely to trigger disciplinary
action) or abruptly cut short the discussion of important details. Therefore, a prediction of
the face-to-face communication channel is that, following the creation of co-located incidents,
handlers will be more likely to be in ’not ready’ status. This prediction is not shared by
alternative plausible channels.
In Column 4 of Table 7 we replicate the baseline specification using the length of
the ’Not Ready’ interval following an incident as the dependent variable. The SameRoom
coefficient is 2.5% and statistically significant, suggesting that handlers step away from
their desks (or remain on their desks while being unavailable) for longer periods following
co-located incidents. In Column 5 of Table 7, we repeat this exercise using as dependent
variable a dummy for whether the handler signals her immediate availability to take new
calls or instead takes some ’not ready’ time at all. Again, we find that the likelihood of not
being immediately available is higher for co-located incidents. Of course, the organisation did
not record informal communication exchanges between co-workers, and therefore we cannot
directly observe these exchanges here. In the absence of such direct evidence, we interpret
the estimates in Table 7 as strong evidence of face-to-face communication being the main
mechanism through which co-location improves performance29 .
29A potential explanation for the effect of co-location on performance that we have not mentioned up to
this point is as follows. When the two teammates are within close proximity of each other as the call handler
takes a call, the radio operator overhears the exchange with the caller and starts preparing her reaction even
before the handler has officially created the incident. This would still represent in-person communication,
although of a different kind than the one that we have been discussing throughout. We have however strong
reasons to discard this explanation. Firstly, the effects are present even when the two teammates sit relatively
far apart, such as at two positions away along the row dimension and three along the column dimension.
Secondly, the noise levels in these rooms are incompatible with the ability to overhear or signal across more
22
6 Heterogeneity
In this section we identify characteristics of incidents, teammates and the working environ-
ment that are associated with a higher effect of co-location on performance. We regard this
exercise as one of the main contributions of the paper. As discussed in the introduction, a
better understanding of the specific circumstances in which face-to-face communication has
the highest impact can help guide the communication-enhancing investments by managers.
Characteristics of Incidents We first examine whether the effects from Table 3 are
stronger for some types of incidents, relative to others. We focus on two particularly relevant
characteristics of incidents: their urgency and the complexity of the information required
to understand and describe them. The main hypothesis is that if co-location improves
performance because it enables face-to-face communication, we should find a stronger effect
for complex incidents where a lot of information must be transmitted. In addition to being
intuitive, this hypothesis is consistent with the vast literature arguing that human production
is at a lower risk of being substituted by technology for (cognitive) non-routine tasks, relative
to routine tasks (Acemoglu and Autor, 2011).
We also study empirically the relation between the urgency of an incident and the
effect of co-location on performance. In principle, it is unclear what the sign of this relation
should be. On the one hand, the ability to communicate information quickly might be more
valuable and therefore used more often when an allocation decision needs to be done faster.
On the other hand, in very urgent incidents (e.g. a serious crime in progress) the operator
may not want to wait for many nuanced details and will instead allocate an officer as quickly
as possible. If that is the case, more urgent incidents will be associated with a lower effect
of co-location on allocation time.
Both theoretical concepts, ’urgency’ and ’information intensity’, have elusive empirical
counterparts. The information intensity of incidents is difficult to measure because we un-
fortunately lack access to complete characterisations of the features of every incident in our
dataset. We also lack the full GMIPCS descriptions recorded by handlers, although of course
any classification of an incident reliant on the actions taken by its call handler would risk
confusing the diligence or ability of the handler with the intrinsic features of the incident.
To overcome the measurement challenges above we use information based on generic
incident types to create an indirect measure of information intensity, as follows. We first
classify each incident according to its opening code/grade combination. We then regress
creation time (the time elapsed between the handler answering the call and the creation
than the very shortest distances. Thirdly and most importantly, this potential alternative mechanism is
unable to explain the evidence in this subsection, whereby the call handler takes longer to be available for
the next call following a co-located incident.
23
of the incident) on every one of the resulting 144 dummies. The fitted values from this
regression, which constitute our measure of (predicted) information intensity, capture how
long on average it takes for handlers to extract information from the caller and record it
in GMPICS, for every incident type. Although the measure is undoubtedly coarse, our
interpretation is that incident types with high average creation time should be those where
the amount and complexity of information is typically the largest. We construct our measure
of (predicted) urgency in an equivalent way, this time regressing allocation time on the 144
incident type dummies (naturally, lower average allocation time is interpreted as higher
urgency).
We interact our measures of information intensity and urgency with the same room
dummy in the baseline regression. For ease of interpretation, these measures are entered as
above-median dummies. The estimates are displayed in Table 8. We find first that incident
types of high average information intensity are associated with a higher effect of co-location
on performance30. We also find (weaker) evidence on the urgency of incidents exacerbating
the effect of co-location. In particular, the estimate for the interaction with urgency is
negative, although statistically significant only in the allocation time regression31.
We interpret the estimates from Table 8 as indicating that co-location does not increase
performance for non-urgent, non-complex incidents. It, however, decreases allocation time
(respectively, response time) by 4% (respectively, 2.7%) for incidents that are above-median
both in their urgency and their information intensity. The estimate on the interaction with
information intensity is, in particular, consistent with the notion that co-location enables an
additional communication channel, leading to higher performance for incidents when a lot
of communication is necessary.
Characteristics of the Working Environment In our second heterogeneity exercise, we
study whether co-location improves performance more when workers have to deal with more
incidents. Our main interest is in the workload of the operator, because it is for operators
that a high number of incoming incidents in their subdivision can start to accumulate,
exerting competing demands on their attention. Our hypothesis is that, if co-location allows
operators to quickly resolve any doubt through face-to-face communication, it should be
more valuable when the time and effort of the operator are scarce, that is, in periods of
30This finding is robust to measuring information intensity with quintiles (Appendix Figure A2) and
in parametric (log) format (Appendix Table A4). It is also robust to building the information intensity
prediction exclusively with out-of-sample (i.e. pre-November 2009 and post-January 2012) observations
(Appendix Table A3).
31Both effects become statistically stronger if information intensity is measured parametrically (Appendix
Table A4). However, we find that the effect of co-location does not vary when we use a simpler and coarser
measure of the urgency of an incident: its grade. Although the effect is stronger for Grade 1 incidents,
relative to Grade 2 and Grade 3, the differences are not statistically significant (see Appendix Table A9).
24
higher workload32.
Our measure of the operator’s workload is the number of incidents created in the
subdivision that the operator is overseeing during the hour of the index incident (note that
there is a single operator responsible, at any one time, for a subdivision). For ease of
interpretation, we enter this measure in the baseline regression as an above-median dummy,
both by itself and interacted with the same room variable.
The results are displayed in Table 9. We first find that allocation and response times are
slower when the operator is busier, as expected. Our main interest is in the estimate of the
interaction between the same room variable and the high operator workload dummy, which
we find to be negative and statistically significant. The estimated coefficients indicate that
co-location reduces allocation time (respectively, response time) by 1.1% (respectively, .8%)
during periods of low operator workload, but 2.9% (respectively, 2%) during periods of high
workload. This finding lends support to our hypothesis that the benefit of communicating
personally with the handler is higher when the operator is more pressured for time and needs
to gather information more quickly33.
Characteristics of the Workers We now examine whether the effect of co-location on
performance is stronger when the teammates share the same age and gender, and have
worked together more often in the past. This may be the case for two reasons. Firstly,
workers of a similar background (or more familiar with each other) may be more likely to
initiate the face-to-face communication exchanges that transmit information regarding an
incident. This is because they may be more likely to sit close to each other, or, conditional
on the within-room distance, they may be more likely to leave their desk and talk to each
other. Secondly, in-person communication may also be more efficient among these types
of workers34. Alternatively, homogenous teams may be so efficient at communicating elec-
32By contrast, our understanding of the institutional environment is that the notion of being ’pressured for
time’ is less meaningful for handlers. Handlers deal with incidents sequentially and share the responsibility
of responding to incoming calls with a large number of colleagues (since every handler can handle incidents
from every Manchester area). Together with the fact that handlers are not responsible for the allocation of
officers to incidents, this implies that we do not have a strong hypothesis about the relation between our
measure below of handler workload and the effect of co-location on performance
33Our measure of the handler workload is very coarse, mostly because as discussed earlier, the notion that
handlers are busier in some periods relative to others is not clear-cut. We use the (above-median dummy of
the) number of incoming calls during the index hour, divided by the number of available handlers. Because
this variable is defined at the Manchester-wide level, it is absorbed in the baseline regression by the hour
fixed effect. We find in Table 9 that the coefficient on the interaction with the same room variable is smaller
in magnitude and only weakly statistically significant.
34Storper and Venables (2004) discuss how the transmission of uncodifiable information (at which face-to-
face communication excels) depends on a ’communication infrastructure’ that is specific to a sender-receiver
pair. This infrastructure is likely improved through learning by doing, leading to more efficient face-to-face
communication as the teammates accumulate experience with each other. It is also likely more efficient
among demographically proximate teammates. Alternatively, we could interpret demographic proximity as
25
tronically that additional in-person communication is more valuable when the team is not
homogenous.
In Table 10 we display estimates of our baseline specification, where we add a same
gender dummy, the (log of the) difference in age, and the (log of the) number of past
incidents in which handler and operator worked together. We further interact these variables
with the same room variable. To isolate the effect of the handler/operator pair experience,
the specification controls for the individual experiences of handler and operator and their
interactions with the same room variable.
Our main finding is that the estimates for the three interactions of interest are statisti-
cally significant and of the expected sign. For instance, the effect of co-location is 1.6% higher
when handler and operator share the same gender. A 10% increase in the age difference (re-
spectively, number of past interactions) between handler and operator decreases the effect
of co-location on performance by 2.5% (respectively, it increases it by 2.1%). These findings
are consistent with the notion that face-to-face communication, and therefore co-location,
leads to higher performance among co-workers that know and understand each other bet-
ter35. On the other hand, the non-significant interactions with individual experience suggest
that, unless it is specific to the teammate in this particular incident, individual experience
does not by itself allow workers to exploit better the potential advantages of co-location.
7 The Cost of Face-To-Face Communication
In this section we provide a measure of the operational costs of communication36. In Section
4 we found no evidence of negative spillovers to other incidents being handled contempora-
neously by the operator. On the other hand, Section 5 has shown that handlers spend 2.5%
more time unavailable to take new calls following the creation of co-located incidents. This
unavailability imposes a cost on the organisation, as it contributes to incoming calls being
answered with a longer delay. We now provide a framework to measure the opportunity cost
of the time spent in face-to-face communication, so that it can be compared to its benefit37.
a proxy for the existence of friendship ties between two co-workers (Bandiera, Barankay and Rasul, 2010). If
workers are more willing to and effective at communicating face-to-face with their friends, a similar prediction
for the relation between demographic proximity and the effect of co-location on performance would arise.
35We find qualitatively similar results when age and past interactions are measured as above-median
dummies (see Appendix Table A5). While not the focus of this paper it is interesting to note that, even when
teammates are not co-located, a similar age and a longer experience with each other are still associated with
higher performance (although this is not the case for the same gender variable). A potential explanation
of these estimates is that, given the complexity of the information that must often be transmitted, even
electronic communication is more efficient among these types of teammates.
36Building communication channels between workers may entail fixed investments, and we abstract from
the cost of these investments here.
37Note that, to the extent that communication takes time and that time cannot be devoted to other
activities, the type of cost that we measure here is present in every organisation where communication takes
26
We then compute this cost in our organisation.
Theoretical Framework We formalise the process by which calls to the police arise, join
the call queue and are answered. Assume a population of individuals (of normalised size
1) who can potentially call the police. Every individual can be in one out of three states:
dormant (waiting for an incident to happen), in the call queue, or on the phone with the
handler. xi,i= 1,2,3 denotes the share of individuals in each state. H < 1 handlers are on
duty to answer calls.
Transitions between states are as follows. Dormant callers join the queue at an exoge-
nous rate aper unit of time. All callers must spend at least one unit of time there before
being assigned a handler. When a call is being answered, it terminates (and the caller re-
joins the dormant pool) at a constant rate υ(with 1being the average duration of calls).
The number of handlers that become available to take new calls per unit of time is therefore
υx3. The total number of calls answered per unit of time is then min{υx3, x2}, since it is
limited both by the number of newly available handlers and by the size of the call queue.
Using this simple framework we can show that the size of the call queue evolves over
time depending on the difference between the inflow (the number of dormant individuals who
encounter an incident) and the outflow (the number of queued calls answered by handlers):
x2
t=a(1 x2x3)min{υx3, x2}(3)
Similarly, x3
t=min{υx3, x2} − υx3(4)
If υx3< x2, then it must be that all handlers are busy and x3=H. Combining
equations (3) and (4) and assuming a steady state, we compute the time in the queue for
incoming calls, q, as:
q=((1H)
υH 1
aif H < a
a+υ+
1 if Ha
a+υ+
(5)
This simple framework generates the following predictions. First, incoming calls are
answered immediately when there are many handlers (Hhigh), few dormant calls become
actual calls (alow) and calls are brief (υhigh). Secondly, q
(1)>0 so an increase in average
call length leads to longer queuing times. Lastly, this effect is lower when the number of
handlers is higher, 2q
(1)H <0. We can interpret an increase in Has the increase in
organisational slack, as the same amount of incoming work is divided over a higher number
of workers. Therefore, this model predicts that an increase in slack both decreases queuing
times and reduces the effect of higher average call duration.
place.
27
Computing the Opportunity Cost of Face-To-Face Communication Section 5 pro-
vided evidence of an increase in ’not ready’ time following the creation of co-located incidents.
This is equivalent in our framework to an increase in the duration of the call, as it mechani-
cally prevents handlers from relieving the pressure in the call queue. We now use information
on all calls (not just the ones that led to the creation of incidents) to relate call duration,
the number of calls and the number of on-duty handlers to the average time spent in the call
queue. The resulting coefficients allow us to understand the opportunity cost of an additional
second spent dealing with a previous call. We estimate:
qi=α+γni(τ) + δhi(τ) + βdi(τ) + i(6)
where qiis the (log of the) queuing time of incoming call i,niand hiare the (log of)
number of calls and on-duty handlers in a time window before i, and diis the (log of ) average
duration of answered calls in the same time window.
Table 11 Panel A shows that the estimated elasticity of average call duration on queuing
time ranges from .58 to .96. We can compute the effect that an increase in the duration of
a single call jhas on the queuing time of future calls as follows. First, note that such an
increase has an effect on the queuing time of a single future call ithat can be computed
as ˆ
βexp(qi)
T Di, where exp(qi) is the queuing time of iand T Diis the total duration of the calls
preceding i(which include j). Aggregating over the Kcalls that follow j, we can write the
overall effect of an increase in j’s duration as ˆ
βPj+K
i=j+1
exp(qi)
T Di.
The statistic ˆ
βPj+K
i=j+1
exp(qi)
T Dican be interpreted as the opportunity cost (in terms of
additional queuing time of future calls i=j+ 1 . . . K) of increasing the duration of call jby
one second. This statistic can be computed directly from our dataset, using the elasticity
estimated in Table 11 and information on the queuing time of every call, together with the
duration of the calls preceding it. Using a time window of 60 minutes to define the calls
affected by the increase in the duration of a preceding call, we calculate it as 0.13 seconds. In
Table 7 we estimated that co-located incidents increase ’not ready’ time by 2.5%. Evaluated
at the mean of ’not ready’ time (66 seconds), co-located incidents are therefore associated
with a cost of 0.13 ×2.5% ×66 = 0.21 seconds.38 In our organisation, this is arguably a
small cost, when compared with the decreases in allocation and response times of 76 and
104 seconds respectively that we estimated in Section 439.
Motivated by our theoretical framework, we expect the opportunity cost of face-to-face
communication to be lower when organisational slack, as captured by the relation between
38We repeat this exercise for time windows of 15, 30 and 120 minutes and we estimate the cost in 0.20,
0.18 and 0.22 seconds respectively.
39The benefits and costs associated with co-location affect different types of calls. The costs are for the
average call, including those which do not lead to incidents and those leading to incidents that are not
deemed to merit a response within four hours of the incident creation. The benefits are instead concentrated
on the calls that are deemed important enough to be assigned to a radio operator.
28
on duty handlers and incoming calls, is higher. In Panels B and C we repeat the exercise
in Panel A for the subsamples of calls with high and low organizational slack. Consistently
with the prediction that increasing a call’s duration is less costly when the relative number
of handlers is higher, we find a higher elasticity in Panel C (high slack) and a lower in Panel
B (low slack). Replicating the analysis above, we calculate costs associated with co-location
of 0.15 (respectively 0.31) seconds, for periods of low (respectively high) slack.
Overall, our analysis highlights the importance of measuring the opportunity cost of
the time engaged in face-to-face communication, as well as the dependence of this cost on
the slack characterising the organisation. In our setting, we find this cost to be much lower
than the benefit.
8 Conclusion
This paper has provided evidence of a causal relation between co-location and performance,
in a teamwork setting characterised by the communication of complex information. A series
of additional tests point towards face-to-face communication as the most important mech-
anism. We have also provided additional evidence on the heterogeneity of the main result
and highlighted that face-to-face communication has opportunity costs, as well as benefits.
We are not aware of any existing study studying these questions, especially one that is
comparable in terms of the detail of analysis and the credibility of the estimated effects.
One immediate policy prescription for the specific organisation that we study is in terms
of supervisors’ awareness of the benefits of communication between co-workers. Discussions
between handler and operator following the creation of incidents were not encouraged and
were even frowned upon by some supervisors. Because the cost of communication is orders
of magnitude smaller than the benefit, one implication is that, in our specific context, there
may be too little communication among co-located workers rather than too much. This
indicates that a change of norms and culture to encourage more communication could be
efficiency-enhancing. More generally, however, the fact that the cost of communication is
not zero indicates that the limitations on the information sets of decision-makers highlighted
by Hayek (1945) and Arrow (1974) are unlikely to be fully overcome.
Our findings provide direct guidance to managers organising the geographical distribu-
tion of activities. Most directly, the evidence casts doubt on the appropriateness of telecom-
muting policies in settings where workers must communicate complex information to each
other. Our results further suggest that telecommuting may be particularly unsuitable (and
co-location of teammates particularly valuable) when activities are informationally demand-
ing, workers are homogenous and likely to be busy, and teams are likely to be stable.
There may be additional implications for recruitment policy. A large literature in
29
organisational behaviour is concerned with the advantages and challenges of diversity in
the workplace (Shore et al., 2009). In economics, a parallel body of work has studied
the differences in productivity between homogeneous and heterogeneous teams (Hamilton
et al. 2012, Hjort 2014, Lyons 2016), a question of clear recruitment policy implications.
Our results indicate that the relative benefits of homogeneity depend on the geographical
configuration of activities. In particular, a more homogeneous organisation is most valuable
when workers are likely to be based in the same physical space.
Our results also identify a distinct driver of firm-specific human capital accumulation
(Topel, 1991), with implications for staff turnover and team-rotation policies. Consistently
with Hayes et al. (2006) and Jaravel et al. (2016), we find in Section 6 that workers
accumulate human capital that is specific to a particular co-worker. Importantly, our finding
is however that this capital is most valuable (or more rapidly accumulated) among co-located
workers. It follows that managers should be wary of the team disruption induced by turnover
particularly when the team members work in close proximity.
30
FIGURES
31
32
33
FIGURE 6: Balance of Incident, Worker and Room
Characteristics on Same Room
Each row in the figure displays the results of two regressions, where the row variable is the dependent
variable and Same Room is the independent variable. The first regression includes no controls and the
second regression controls for Year X Month X Day X Hour of Day, Radio Operator Room and Call
Handler Room. The displayed 95% confidence intervals are for the coefficient of the Same Room variable.
Non-binary dependent variables are standardised. Standard errors are clustered at the Year X Month X
Radio Operator Room level. Grade 1, Grade 2, Handler Female and Operator Female are the only dummy
variables. Handler’s Desk Dist. Centre is the euclidean distance between the handler’s desk and the centre
of the room. Hourly Incidents per Handler in Room is the number of incidents created during the hour of
the index incident, divided by the number of handlers working during that hour. A similar definition
applies to Hourly Incidents per Operator in Room. Hourly Incidents of Handler is the number of incidents
created by the handler in charge of the index incident, during the hour of creation. Hourly Incidents of
Operator is the number of incidents allocated by the operator in charge of the index incident, during the
hour of the creation of the incident.
34
35
36
37
38
TABLES
TABLE 1: CORRELATIONS BETWEEN
ALLOCATION/RESPONSE TIME
AND VICTIM SATISFACTION MEASURES
(1) (2)
Dep. Variable Victim Victim
Satisfaction Change in
Score Opinion of Police
Log Allocation Time -.038*** -.023***
(.006) (.004)
Log Response Time -.055*** -.035***
(.009) (.006)
On Target Allocation .095*** .051***
(.019) (.012)
On Target Response .14*** .082***
(.028) (.016)
Observations 9617 7827
This table displays estimates of OLS regressions of measures of caller satisfaction on allocation and
response time. Every coefficient is a different regression. The variables in the columns are the
dependent variables and the variables in the rows are the independent variables. Victim satisfaction
score is the answer by the caller to a survey ranking how satisfied she is with the police dealing
with the incident. The score takes values between 1 (Very Dissatisfied) and 8 (Very Satisfied), but
has been standardised. Victim change in opinion of police can take values -1, 0 or 1, depending
on whether the opinion has worsened, remained the same or improved. All regressions also include
indicators for Call Source, Year X Month X Day, Hour of Day, Division, Grade and Opening Code.
Standard errors are clustered at the Division X Year level.
39
TABLE 2: SUMMARY STATISTICS
Mean Median SD Min Max
Allocation Time (min.) 64.124 4.583 276.568 0 21331.78
On Target Allocation .748 1 .434 0 1
Response Time (min.) 87.484 19.933 311.166 .05 21391.92
On Target Response .877 1 .328 0 1
Creation Time (min.) 3.889 2.85 4.946 0 219.533
Grade 1 .197 0 .398 0 1
Grade 2 .432 0 .495 0 1
Same Room .229 0 .42 0 1
Distance inside Room 4.34 4.243 1.782 .5 11.885
Handler Female .27 0 .444 0 1
Operator Female .498 0 .5 0 1
Handler’s Age 38.406 38 11.471 19 64
Operator’s Age 45.15 46 8.243 19 66
This Table reports summary statistics for the baseline sample (N=957137). An observation is an incident.
Allocation time is the time between the creation of the incident by the call handler and the allocation of
a police officer by the radio operator. Response time is the time between creation of the incident and the
police officer arriving at the scene. On target allocation (respectively, response) is a dummy taking value one
if the allocation time falls wihin the UK Home Office targets, which are 2, 20 and 120 minutes (respectively
15, 60 and 240 minutes) for Grades 1, 2 and 3. Creation Time is the time between the handler answering
the call and the creation of the incident in GMPICS. Grade 1 and Grade 2 are dummies for the grade of
the incident. Same Room is a dummy when handler and operator are located in the same room. Distance
inside the room is the euclidean distance between the handler and the radio operator desks. This variable is
defined in this table only when same room is equal to one (N=219184). Handler female and operator female
are dummy variables.
TABLE 3: BASELINE ESTIMATES
(1) (2) (3) (4) (5)
Dep. Log Alloc. Log Response On Target On Target Cleared
Variable Time Time Alloc. Response
Same Room -.02*** -.017*** .004*** .002*** -.001
(.004) (.003) (.001) (.001) (.003)
This table displays estimates of OLS regressions of five different performance measures on whether the call handler and the radio
operator are located in the same room. The sample includes all incidents received by the GMP between November 2009 and
December 2011 (N=957137). In Column (1) the performance variable is the log of the allocation time (i.e. the time between
the creation of the incident by the call handler and the allocation of a police officer by the radio operator). In Column (2) the
performance variable is the log of the response time (i.e. the time between the creation of the incident and the police officer arriving
at the scene). In Columns (3) and (4) the dependent variables are dummy variables taking value one if allocation and response times
fall within the UK Home Office targets, respectively. The target response times for Grades 1, 2 and 3 are 15, 60 and 240 minutes,
respectively. The target allocation times are 2, 20 and 120 minutes. In Column (5) the dependent variable is a dummy taking value
one if the crime was cleared. In Column (5) the sample includes only incidents that the police classified as crimes (N=156550). All
regressions include indicators for Grade, Call Source, Year X Month X Day X Hour of Day, Radio Operator Room, Call Handler
Room, Radio Operator and Call Handler. Standard errors are clustered at the Year X Month X Radio Operator Room level.
40
TABLE 4: HETEROGENEITY OF SAME ROOM
BY DISTANCE INSIDE ROOM
Individual F.E. Pair F.E.
(1) (2) (3) (4)
Dep. Variable Log Alloc. Log Response Log Alloc. Log Response
Time Time Time Time
Same Room -.049*** -.035*** - -
(.012) (.01) - -
Same Room .026*** .018*** .027*** .017**
X Log Distance (.009) (.007) (.01) (.008)
This table displays estimates of OLS regressions of allocation time and response time on whether the call handler
and the radio operator are located in the same room, interacted with the distance between their desks when they
are in the same room. The sample includes all incidents received by the GMP between 2009 and 2012 (N=957137).
The distance between their desks is calculated as the euclidean distance in the floorplans provided by the GMP.
All regressions include indicators for Grade, Call Source, Year X Month X Day X Hour of Day, Radio Operator
Room X Year and Call Handler Room X Year. Columns (1) and (2) also include Radio Operator and Call Handler
Identifiers. Columns (3) and (4) include Radio Operator/Call Handler Pair Identifiers. Standard errors are clustered
at the Year X Month X Radio Operator Room level.
TABLE 5: INVESTIGATING EFFECTS
ON TYPE OF OFFICER SENT
(1) (2)
Dep. Variable Response Log Officer
Rank Experience
Same Room -.001 .002
(.001) (.002)
This table displays estimates of OLS regressions of measures of the type of officer sent on
the Same Room dummy. In Column (1) the dependent variable is a dummy for whether
the officer sent has the rank of response officer. In Column (2) the dependent variable
is the officer’s number of years in the GMP. All regressions also include indicators for
Call Source, Year X Month X Day X Hour of Day, Radio Operator Room, Call Handler
Room, Radio Operator and Call Handler. Standard errors are clustered at the Year X
Month X Radio Operator Room level.
41
TABLE 6: INVESTIGATING SPILLOVERS
TO OTHER INCIDENTS, BY SAME ROOM INCIDENTS
Spillovers by Same Room Incidents during Period:
60 min. 30 min. 15 min.
(1) (2) (3) (4) (5) (6)
Dependent LogAlloc LogResp LogAlloc LogResp LogAlloc LogResp
Variable Time Time Time Time Time Time
% Same Room .005 .004 .006 .007 .009 .007
Incidents Received (.005) (.004) (.006) (.004) (.007) (.005)
by Operator
This table investigates potential spillovers from Same Room incidents into other contemporaneous incidents. The dependent
variables in the OLS regressions are log of allocation time and log of response time. The independent variable is the percentage
of incidents during the index incident time period for which the call handler and the radio operator were located in the same
room, excluding the index incident. In Columns (1) and (2) the period comprises of 60 minutes (respectively, 30 minutes for
columns (3) and (4) and 15 minutes for columns (5) and (6)). All regressions include indicators for Grade, Call Source, Year X
Month X Day X Hour of Day, Radio Operator Room, Call Handler Room, Radio Operator and Call Handler. The regressions
also include indicators for whether there were no calls received by the Radio Operator during the time period. Standard errors
are clustered at the Year X Month X Radio Operator Room level.
TABLE 7: INVESTIGATING EFFECTS
ON OTHER ACTIONS BY THE HANDLER
(1) (2) (3) (4) (5)
Dep.Var. Log Creation Log Number Log Number Log Not Not
Time of Characters of Words Ready Ready>0
Same Room .00446 -.0004 -.00028 .02513*** .00443**
(.00326) (.00138) (.0015) (.00928) (.00201)
This table displays estimates of OLS regressions of three actions by the handler prior to creating the incident, on whether the call
handler and the radio operator are located in the same room. The sample includes all incidents received by the GMP between
November 2009 and December 2011. In Column (1) the dependent variable is the log of the creation time (i.e. the time between
the handler answering the call and the creation of the incident). In Column (2) the dependent variable is the number of characters
in the first line of the description of the incident (maximum number of characters = 210). In Column (3) the dependent variable
is the number of words in the first line of the description of the incident. In Column (4) the dependent variable is the log of the
not ready time following the creation of the incident. In Column (5) the dependent variable is a dummy for whether the not ready
time takes value bigger than zero. All regressions include indicators for Grade, Call Source, Year X Month X Day X Hour of Day,
Radio Operator Room, Call Handler Room, Radio Operator and Call Handler. Standard errors are clustered at the Year X Month
X Radio Operator Room level.
42
TABLE 8: HETEROGENEITY OF SAME ROOM
BY INCIDENT CHARACTERISTICS
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Same Room .001 -.001
(.008) (.006)
Same Room X Urgent -.019*** -.007
(.008) (.006)
Same Room X Information Intensive -.021*** -.02***
(.008) (.006)
This table displays estimates of OLS regressions of allocation and response time on the Same Room
dummy, interacted in measures of the urgency and information intensity of an incident. To compute
the information intensity variable we use the sample from 2008 to 2014 and regress the creation time
(i.e. the time between the handler answering the call and the creation of the incident) on the opening
code/grade indicators. We then assign to every opening code/grade incident type its predicted creation
time, and label an incident type as being information intensive if its predicted creation time is above
the median. To compute the urgency variable, we do a similar exercise using allocation time instead
of creation time. All regressions also include indicators for Call Source, Year X Month X Day X Hour
of Day, Radio Operator Room, Call Handler Room, Radio Operator and Call Handler, and opening
code/grade indicators. Standard errors are clustered at the Year X Month X Radio Operator Room
level.
43
TABLE 9: HETEROGENEITY OF SAME ROOM
BY WORKER WORKLOAD
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Same Room -.011* -.008*
(.006) (.004)
Same Room X High Operator Workload -.018** -.012*
(.008) (.006)
Same Room X High Handler Workload -.006 -.01*
(.008) (.006)
High Operator Workload .128*** .046***
(.005) (.004)
This table displays estimates of OLS regressions of allocation and response time on the Same Room
dummy, interacted with measures of the workload of the operator and handler. To compute the operator
workload measure, we use the number of incidents created in the operator’s subdivion during the index
hour. To compute the handler workload measure, we use the number of Manchester-wide incidents
during the index hour, divided by the number of handlers on duty during that hour. The variables in
the regression are dummies taking value one when the workload is above the sample median. We report
the uninteracted operator workload measure. The uninteracted handler workload measure is absorbed
by the Year X Month X Day X Hour of Day fixed effects. All regressions also include indicators for
Call Source, Year X Month X Day X Hour of Day, Radio Operator Room, Call Handler Room, Radio
Operator and Call Handler. Standard errors are clustered at the Year X Month X Radio Operator Room
level.
44
TABLE 10: HETEROGENEITY OF SAME ROOM
BY HANDLER-OPERATOR DEMOGRAPHIC DISTANCE
BY NUMBER OF PAST INTERACTIONS
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Same Room -.021 -.031*
(.023) (.018)
Same Room X Same Gender -.016** -.019***
(.008) (.006)
Same Room X Log Difference in Age .025*** .024***
(.005) (.004)
Same Room X Log Number Past Interactions -.021*** -.019***
(.005) (.004)
Same Room X Log Handler Experience -.004 -.003
(.004) (.003)
Same Room X Log Operator Experience .005 .009*
(.006) (.005)
Same Gender -.002 -.003
(.004) (.003)
Log Difference in Age .013*** .01***
(.003) (.002)
Log Number Past Interactions -.073*** -.061***
(.005) (.004)
Log Handler Experience .058*** .045***
(.009) (.007)
Log Operator Experience -.057 -.026
(.049) (.036)
This table displays estimates of OLS regressions of allocation and response time on the Same Room dummy, interacted with
whether the Radio Operator and the Handler are of the same gender, with the log of their difference in age, and with the
number of previous incidents in which they have worked together. All regressions include indicators for Grade, Call Source,
Year X Month X Day X Hour of Day, Radio Operator Room X Year, Call Handler Room X Year, Radio Operator and Handler.
All regressions also control for Handler Experience and Operator Experience and their interactions with Same Room. Standard
errors are clustered at the Year X Month X Radio Operator Room level.
45
TABLE 11: OPPORTUNITY COST
OF HIGHER CALL DURATION
(1) (2) (3)
Dep. Var. = 15 min. 30 min. 60 min.
Log Queuing Time Window Window Window
Panel A: All
Log Calls .843*** .832*** .734***
(.005) (.006) (.006)
Log Handlers -.881*** -.872*** -.773***
(.007) (.007) (.008)
Log Avg Call Duration .582*** .819*** .959***
(.008) (.01) (.011)
Panel B: Low Organisational Slack
Log Calls .301*** .35*** .379***
(.008) (.009) (.01)
Log Handlers -.308*** -.368*** -.418***
(.01) (.011) (.012)
Log Avg Call Duration .402*** .603*** .776***
(.009) (.011) (.014)
Panel C: High Organisational Slack
Log Calls 1.827*** 1.664*** 1.48***
(.018) (.02) (.021)
Log Handlers -1.653*** -1.514*** -1.326***
(.018) (.02) (.02)
Log Avg Call Duration .941*** 1.184*** 1.272***
(.013) (.016) (.018)
This table displays estimates of OLS regressions of queuing time on measures of organisational slack
and average call duration in the period preceding the start of the call. We estimate the effects
separately at 15, 30 and 60 minutes periods before the call. High organisational slack is defined as
periods during which the number of calls per handler was below the median. The sample includes all
calls received by the GMP during the second semester of 2011. N=909256 for panel A, N=455023 for
panel B and N=454233 for panel C. All regressions include an indicator for whether the call reached
the GMP through an emergency line.
46
TABLES AND FIGURES FOR ONLINE APPENDIX
FIGURE A1: Balance of Incident, Worker and Room
Characteristics on Same Room Incidents
Each row in the figure displays the results of two regressions, where the row variable is the dependent
variable and Same Room is the independent variable. The first regression includes only Year X Month X
Day X Hour of Day controls and the second regression includes only controls for Radio Operator Room
and Call Handler Room. The displayed 95% confidence intervals are for the coefficient of the Same Room
variable. Non-binary dependent variables are standardised. Standard errors are clustered at the Year X
Month X Radio Operator Room level. Grade 1, Grade 2, Handler Female and Operator Female are the
only dummy variables. Handler’s Desk Dist. Centre is the euclidean distance between the handler’s desk
and the centre of the room. Hourly Incidents per Handler in Room is the number of incidents created
during the hour of the index incident, divided by the number of handlers working during that hour. A
similar definition applies to Hourly Incidents per Operator in Room. Hourly Incidents of Handler is the
number of incidents created by the handler in charge of the index incident, during the hour of creation.
Hourly Incidents of Operator is the number of incidents allocated by the operator in charge of the index
incident, during the hour of the creation of the incident.
47
48
TABLE A1: ROBUSTNESS TO CONTROLS
(1) (Baseline) (3) (4) (5)
Log Allocation Time -.023*** -.02*** -.018*** -.019*** -.02***
(.004) (.004) (.004) (.004) (.004)
Log Response Time -.02*** -.017*** -.016*** -.016*** -.017***
(.003) (.003) (.003) (.003) (.003)
Hour F.E. Yes Yes Yes Yes Yes
Grade/Call Source F.E. Yes Yes Yes Yes Yes
Room F.E. Yes Yes No Yes Yes
Individual F.E. No Yes No Yes Yes
Room/Date F.E. No No Yes No No
Individual/Month F.E. No No Yes No No
Opening Code/Grade F.E. No No No Yes No
Handler Position F.E. No No No No Yes
This table displays estimates of OLS regressions of allocation time and response time on whether the call handler and the
radio operator re located in the same room. The sample is the basleine sample. Every coefficient is from a different regression.
Standard errors clustered at the Year X Month X Operator Room level.
TABLE A2: ALTERNATIVE CLUSTERING
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Panel A: Baseline
Same Room -.0201*** -.0172***
(.004) (.003)
Panel B: By Handler/Operator Pair
Same Room -.0201*** -.0172***
(.0041) (.0032)
Panel C: By Subdivision
Same Room -.0201*** -.0172***
(.0039) (.003)
This table displays estimates of OLS regressions of allocation and response time on the Same
Room dummy. All regressions include indicators for Grade, Call Source, Year X Month X Day
X Hour of Day, Radio Operator Room, Call Handler Room, Radio Operator and Call Handler.
Standard errors are clustered at the Year X Month X Radio Operator Room level.
49
TABLE A3: HETEROGENEITY OF SAME ROOM
BY INCIDENT CHARACTERISTICS
PREDICTION WITH OUT OF SAMPLE DATA
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Same Room -.001 -.002
(.009) (.006)
Same Room X Urgent -.015 -.006
(.009) (.007)
Same Room X Information Intensive -.024*** -.026***
(.01) (.007)
This table displays estimates of OLS regressions of allocation and response time on the Same Room
dummy, interacted in measures of the urgency and information intensity of an incident. To compute the
information intensity variable we use the post-2012 sample and regress the log of the handler’s time to
creation (i.e. the time between the handler answering the call and the creation of the incident) on the
opening code/grade indicators. We then assign to every opening code/grade incident type its predicted
time to creation, and label an incident type as being information intensive if its predicted time to creation
is above the median. To compute the urgency variable, we do a similar exercise using the log of the
allocation time instead of the log of the handler’s time to creation. All regressions also include indicators
for Call Source, Year X Month X Day X Hour of Day, Radio Operator Room, Call Handler Room, Radio
Operator and Call Handler, and opening code/grade indicators. Standard errors are clustered at the
Year X Month X Radio Operator Room level.
50
TABLE A4: HETEROGENEITY OF SAME ROOM
BY INCIDENT CHARACTERISTICS
INTERACTION WITH VARIABLES IN LOGS
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Same Room .081*** .066***
(.027) (.02)
Same Room X Non-Urgent .01*** .006***
(in Logs) (.003) (.002)
Same Room X Information Intensive -.079*** -.064***
(in Logs) (.021) (.016)
This table displays estimates of OLS regressions of allocation and response time on the Same Room
dummy, interacted in measures of the urgency and information intensity of an incident. To compute the
information intensity variable we use the 2008-2014 sample and regress the log of the handler’s time to
creation (i.e. the time between the handler answering the call and the creation of the incident) on the
opening code/grade indicators. We then assign to every opening code/grade incident type its predicted
time to creation, and use the variable in logs. To compute the urgency variable, we do a similar exercise
using the log of the allocation time instead of the log of the handler’s time to creation. All regressions
also include indicators for Call Source, Year X Month X Day X Hour of Day, Radio Operator Room,
Call Handler Room, Radio Operator and Call Handler, and opening code/grade indicators. Standard
errors are clustered at the Year X Month X Radio Operator Room level.
51
TABLE A5: HETEROGENEITY OF SAME ROOM
BY DEMOGRAPHIC DISTANCE (MEDIAN)
BY NUMBER OF PAST INTERACTIONS (MEDIAN)
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Same Room -.0107 -.0093
(.0093) (.0071)
Same Room X Same Gender -.0191*** -.0215***
(.008) (.0061)
Same Room X Difference in Age High .0129 .0125**
(.0079) (.006)
Same Room X Number Past Interactions High -.0005* -.0001
(.0003) (.0002)
Same Gender -.0027 -.0033
(.0041) (.003)
Difference in Age High .0166*** .0077
(.0063) (.0048)
Number Past Interactions High -.0338*** -.0274***
(.0049) (.0038)
This table displays estimates of OLS regressions of allocation and response time on the Same Room dummy, interacted with
whether the Radio Operator and the Handler are of the same gender, with their difference in age (measured as an above
median dummy), and with the number of previous incidents in which they have worked together (measured as an above median
dummy). All regressions include indicators for Grade, Call Source, Year X Month X Day X Hour of Day, Radio Operator Room
X Year, Call Handler Room X Year, Radio Operator and Handler. All regressions also control for Handler Experience and
Operator Experience. Standard errors are clustered at the Year X Month X Radio Operator Room level.
52
TABLE A6: INVESTIGATING SPILLOVERS
ON NON-SAME ROOM INCIDENTS, BY SAME ROOM INCIDENTS
Spillovers by Same Room Incidents during Period:
60 min. 30 min. 15 min.
(1) (2) (3) (4) (5) (6)
Dep. Var LogAlloc LogResp LogAlloc LogResp LogAlloc LogResp
Time Time Time Time Time Time
% Same Room .001 .001 -.001 .002 -.004 -.003
Incidents Rece (.006) (.005) (.007) (.005) (.008) (.006)
by Operator
This table investigates potential spillovers from Same Room incidents into non-Same Room incidents. The sample includes only
incidents where Handler and Operator were in different rooms (N=734767). The dependent variables in the OLS regressions
are log of the allocation time and log of the response time. The independent variable is the percentage of incidents during the
index incident time period for which the call handler and the radio operator were located in the same room, excluding the index
incident. In Columns (1) and (2) the period comprises of 60 minutes (respectively, 30 minutes for columns (3) and (4) and 15
minutes for columns (5) and (6)). All regressions include indicators for Grade, Call Source, Year X Month X Day X Hour of
Day, Radio Operator Room, Call Handler Room, Radio Operator and Call Handler. The regressions also include indicators for
whether there were no calls received by the Radio Operator during the time period. Standard errors are clustered at the Year
X Month X Radio Operator Room level.
TABLE A7: ROBUSTNESS TO CONTROLLING FOR
THE TIME PERIOD MORE PRECISELY
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Panel A: Baseline (60 minutes)
Same Room -.0201*** -.0172***
(.004) (.003)
Panel B: 30 minutes
Same Room -.0207*** -.0177***
(.004) (.003)
Panel C: 15 minutes
Same Room -.0198*** -.0179***
(.0041) (.0031)
This table displays estimates of OLS regressions of allocation and response time on the Same Room dummy. All regressions
include indicators for Grade, Call Source, Radio Operator Room, Call Handler Room, Radio Operator and Call Handler. In
Panel A we also include Year X Month X Day X Hour of Day. Panel B substitutes the Hour of Day by the half hour period.
Panel C substitutes by the 15 minute period. Standard errors are clustered at the Year X Month X Radio Operator Room
level.
53
TABLE A8: CORRELATION BETWEEN MEASURES
OF OTHER ACTIONS BY THE HANDLER
(1) (2) (3)
Dep. Variable Log Number Log Number Log Number
of Words of Characters of Characters
Log Time to Creation .076*** .076***
(.005) (.005)
Log Number of Words .906***
(0)
Pairwise Correlation .12 .14 .97
This table displays estimates of the conditional correlation among three actions by the handler during the creation
of the incident. The sample includes all incidents received by the GMP between 2008 and 2013 where the dependent
and independent variables are available (N=956440). The log of the handler’s time to creation is the time between
the handler answering the call and the creation of the incident. The number of characters is measured in the first line
of the description of the incident (maximum number of characters = 210). The number of words is also measured
in the first line of the description of the incident. All regressions include indicators for Grade, Call Source, Year
X Month X Day X Hour of Day, Radio Operator Room, Call Handler Room, Radio Operator and Call Handler.
Standard errors are clustered at the Year X Month X Radio Operator Room level. The unconditional correlation
coefficients are also reported.
TABLE A9: HETEROGENEITY OF SAME ROOM
BY INCIDENT GRADE
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Same Room X Grade 1 -.023*** -.013***
(.005) (.004)
Same Room X Grade 2 -.016*** -.016***
(.006) (.004)
Same Room X Grade 3 -.014 -.013*
(.009) (.007)
P-Value G1 6= G2 .336 .552
P-Value G1 6= G3 .412 .955
P-Value G2 6= G3 .885 .722
This table displays estimates of OLS regressions of allocation and response time on the Same
Room dummy, interacted in the Grade of an incident. All regressions also include indicators for
Grade, Call Source, Year X Month X Day X Hour of Day, Radio Operator Room, Call Handler
Room, Radio Operator and Call Handler. Standard errors are clustered at the Year X Month X
Radio Operator Room level.
54
TABLE A10: ROBUSTNESS TO EXCLUSION OF
OUTLYING OBSERVATIONS
(1) (2)
Dep. Variable Log Alloc. Log Response
Time Time
Panel A: Excluding .5%
Same Room -.0193*** -.0171***
(.0039) (.0029)
Panel B: Excluding 1%
Same Room -.0196*** -.0164***
(.0038) (.0028)
Panel C: Excluding 5%
Same Room -.0174*** -.0136***
(.0036) (.0026)
This table displays estimates of OLS regressions of allocation and response time on the Same Room dummy.
All regressions include indicators for Grade, Call Source, Radio Operator Room, Call Handler Room, Radio
Operator and Call Handler. In Panel A Column (1) (respectively, Column (2)) we drop from the baseline
sample the observations with the .5% highest values of allocation time (respectively, response time). In
Panels B and C we do the same for the 1% and 5% highest values. Standard errors are clustered at the Year
X Month X Radio Operator Room level.
TABLE A11: DISTANCE INSIDE ROOM
AND PAST INTERACTIONS HANDLER/OPERATOR
(1) (2)
Dep. Variable Log Distance Log Distance
Log Number Past Interactions .002 .005
(.003) (.005)
Pair Fixed Effects No Yes
This table displays estimates of OLS regressions of distance inside room on the number of past
incidents on which the handler and the operator worked together. The sample includes all incidents
received by the GMP between 2009 and 2012 for which handler and operator were based in the
same room (N=209180. The distance between their desks is calculated as the euclidean distance in
the floorplans provided by the GMP. All regressions include indicators for Grade, Call Source, Year
X Month X Day X Hour of Day, Radio Operator Room X Year and Call Handler Room X Year.
Column (1) also includes Radio Operator and Call Handler Identifiers. Column (2) also includes
Radio Operator/Call Handlers Pair Identifiers. Standard errors are clustered at the Year X Month
X Radio Operator Room level.
55
TABLE A12: HETEROGENEITY OF SAME ROOM
BY DISTANCE INSIDE ROOM
CONTROLLING FOR PAIR/SEMESTER
(1) (2)
Dep. Variable Log Allocation Log Response
Time Time
Same Room X Log Distance .032*** .019**
(.011) (.009)
This table displays estimates of OLS regressions of allocation time and response time on whether
the call handler and the radio operator are located in the same room, interacted with the distance
between their desks when they are in the same room. The sample includes all incidents received by
the GMP between 2009 and 2012. The distance between their desks is calculated as the euclidean
distance in the floorplans provided by the GMP. All regressions include indicators for Grade, Call
Source, Year X Month X Day X Hour of Day, Radio Operator Room X Year and Call Handler Room
X Year, and Radio Operator/Call Handler/Year/Semester Identifiers. Standard errors are clustered
at the Year X Month X Radio Operator Room level.
56
REFERENCES
Acemoglu, D., and Autor, D. (2011), “Skills, tasks and technologies: Implications
for employment and earnings”, Handbook of Labor Economics, 4, 1043-1171.
Arrow, K. E. (1974), “The limits of organization”, Norton.
Autor, D. H., Levy, F., and Murnane, R. J. (2003), “The skill content of re-
cent technological change: An empirical exploration”, The Quarterly Journal of Economics,
118(4), 1279-1333.
Bandiera, O., Barankay, I., and Rasul, I. (2010), “Social incentives in the work-
place”, The Review of Economic Studies, 77(2), 417-458.
Bandiera, O., Barankay, I., and Rasul, I. (2013), “Team incentives: evidence from
a firm level experiment”, Journal of the European Economic Association, 11(5), 1079-1114.
Bavelas, A. and Barrett, D. (1951), “An Experimental Approach to Organizational
Communication”, Personnel 27, 367-371.
Blanes i Vidal, J. and Kirchmaier, T. (2017), “The Effect of Police Response
Time on Crime Clearance Rates”, Review of Economic Studies, forthcoming.
Bloom, N., Garicano, L., Sadun, R., and Van Reenen, J. (2014), “The dis-
tinct effects of information technology and communication technology on firm organization”,
Management Science, 60(12), 2859-2885.
Bloom, N., Liang, J., Roberts, J., and Ying, Z. J. (2015), “Does working from
home work? Evidence from a Chinese experiment”, The Quarterly Journal of Economics,
qju032.
Blume, A. and Ortman, A. (2007), “The Effects of Costless Pre-Play Commu-
nication: Experimental Evidence from Games with Pareto-Ranked Equilibria”, Journal of
Economic Theory, 132, 274-290.
Burgess, S., Propper, C., Ratto, M., Scholder, K., von Hinke, S., and
Tominey, E. (2010), “Smarter Task Assignment or Greater Effort: the impact of incentives
on team performance”, The Economic Journal, 120(547), 968-989.
Catalini, C. (2016), “Microgeography and the direction of inventive activity”, Work-
ing paper.
Chan, D. C. (2016), “Teamwork and moral hazard: evidence from the emergency
department”, Journal of Political Economy, 124(3), 734-770.
Cooper, R. W., DeJong, D. V., Forsythe, R. and Ross, T. W. (1992), “Com-
munication in Coordination Games”, Quarterly Journal of Economics, 107, 739-771.
Curtis, I. (2015), “The use of targets in policing ”, available at: https://www.gov.uk/
government/uploads/system/uploads/attachmentdata/file/466058/ReviewTargets2015.pdf
Dewatripont, M., Jewitt, I., and Tirole, J. (1999), “The economics of career
concerns, part II: Application to missions and accountability of government agencies”, The
Review of Economic Studies, 66(1), 199-217.
Dodd, T., and Simmons, J. (Eds.) (2002/03), “British Crime Survey”, available
at http://webarchive.nationalarchives.gov.uk/20110220105210/rds.homeoffice.gov.uk/rds/
pdfs2/hosb703.pdf
Friebel, G., Heinz, M., Krueger, M., and Zubanov, N. (2017), “Team Incentives
and Performance: Evidence from a Retail Chain”, American Economic Review, 107(8),
2168-2203.
57
Gant, J., Ichniowski, C., and Shaw, K. (2002), “Social Capital and Organizational
Change in HighInvolvement and Traditional Work Organizations”, Journal of Economics and
Management Strategy, 11(2), 289-328.
Gibbons, R. and Roberts, J. (Eds.) (2013), “The Handbook of Organisational
Economics”, Princton University Press.
Guetzkow, H. and Simon, H. A. (1955), “The Impact of Certain Nets upon Orga-
nization and Performance in Task-Oriented Groups”, Management Science, 1, 233-250.
Hamilton, B. H., Nickerson, J. A., and Owan, H. (2012), “Diversity and pro-
ductivity in production teams.”, In Advances in the Economic Analysis of Participatory and
Labor-Managed Firms, Emerald Group Publishing Limited.
Hayek, F. A. (1945), “The Use of Knowledge in Society”, The American Economic
Review, 35(4), 519-530.
Hayes, R. M., Oyer, P., and Schaefer, S. (2006), “Coworker complementarity
and the stability of top-management teams”, Journal of Law, Economics, and Organization,
22(1), 184-212.
Hjort, J. (2014), “Ethnic divisions and production in firms”, The Quarterly Journal
of Economics, 129(4), 1899-1946.
Jaffe, A. B., Trajtenberg, M., and Henderson, R. (1993), “Geographic local-
ization of knowledge spillovers as evidenced by patent citations”, The Quarterly Journal of
Economics, 108(3), 577-598.
Jaravel, X., Petkova, N., and Bell, A. (2016), “Team-specific capital and innova-
tion”, Working paper.
Leavitt, H.J. (1951), “Some Effects of Certain Communication Patterns on Groups
Performance”, Journal of Abnormal Psychology, 46, 38-50.
Lyons, E. (2016), “Team production in international labor markets: Experimental
evidence from the field”, American Economic Journal: Applied Economics.
Mas, A., and Moretti, E. (2009), “Peers at work”, The American Economic Review,
99(1), 112-145.
Palacios-Huerta, I. and Prat, A. (1012), “Measuring the Impact Factor of Agents
within an Organization Using Communication Patterns”, Working paper.
Pilling, K. (2010), “Police to publish call response times on website”, available at:
http://www.independent.co.uk/news/uk/crime/police-to-publish-call-response-times-on-
website-1949967.html
Shore, L. M., Chung-Herrera, B. G., Dean, M. A., Ehrhart, K. H., Jung,
D. I., Randel, A. E., and Singh, G. (2009), “Diversity in organizations: Where are we
now and where are we going?”, Human Resource Management Review, 19(2), 117-133.
Simon, H. A. (1957), “Administrative behavior; a study of decision-making processes
in administrative organization.”, Macmillan.
Simpson, H. (2009), “Productivity in public services.”, Journal of Economic Surveys,
23(2), 250-276.
Storper, M., and Venables, A. J. (2004), “Buzz: face-to-face contact and the
urban economy”, Journal of Economic Geography, 4(4), 351-370.
Thompson, P., and Fox-Kean, M. (2005), “Patent citations and the geography of
knowledge spillovers: A reassessment”, American Economic Review, 450-460.
58
Topel, R. (1991), “Specific capital, mobility, and wages: Wages rise with job senior-
ity”, Journal of Political Economy, 99(1), 145-176.
Weber, R. A. and Camerer, C. F. (2003), “Cultural Conflict and Merger Failure:
An Experimental Approach”, Management Science, 49, 400-415.
59
... They find that home-working led to a 13% performance increase and that, after the experiment, over half of the workers chose to switch to home-working. Battiston et al. (2017) con-duct an experiment at the Operational Communications Branch of a police station. They conclude an opposite finding: face-to-face working increases productivity, and attribute the contrast to the complexity of the tasks in their experiment. ...
... From the survey of Upwork Future Workforce Report, Ozimek (2020) find that more hiring managers, including executives, VPs, and managers, report productivity increases from home working than report productivity decreases. While Bloom et al. (2015), Battiston et al. (2017) and Ozimek (2020) focus on a few occupations, the Covid-19 outbreak and the lockdowns in many countries have dramatically increased the prevalence of working from home in almost all occupations. Since the home-working style is likely to persist even after the Covid-19 pandemic ends, this paper, by looking at productivity changes by sectors (occupations and industries), provides relevant policy insights, such as the possible productivity outcomes after implementation of home working across different sectors, and a direct way to model supply shocks in multi-sectoral macroeconomic models. ...
Preprint
Full-text available
This paper examines the effects of pandemics on income inequality, specifically those pandemics that claimed more than 100,000 lives. Given that pandemics are events that rarely occur, we have use data spanning over the last 100 years (1915-2017) and relating to four pandemics. The study includes four countries that had income inequality data covering that period. Using panel data methods – Fixed Effects and Augmented Mean Group estimators – we found a significant effect of these pandemics on declining income inequality. The study argues that based on the characteristics of the COVID-19 pandemic, namely that fatalities are highly concentrated in older age groups, we can neither expect a labor scarcity nor a sharp decline in productivity; however, we could expect a reduction in consumption, the possibility of savings, high unemployment rates, and high public debt ratios. The ultimate effects of COVID-19 on inequality remain unclear so far, as some of its inherent characteristics push for an increase in inequality. In contrast, others push toward a narrowing of the income gap.
... In terms of the types of task that are most suited to homeworking Dutcher (2012) finds that for students working from home improves productivity of creative tasks but compromises productivity of dull tasks. Whilst not a study of homeworking Battiston et al. (2020) find that in the 999 call handling department of Greater Manchester Police worker productivity is enhanced through colocation. Here the ability to communicate face-to-face improves worker performance when workers perform inter-connected but discrete tasks -in this case taking the initial call (the handler), and directing the emergency response (the radio operator). ...
... Firms may counterbalance this be requiring homeworkers to be based in the office early in their job tenure (see for example Battiston et al. (2020) and Choudhury et al. (2020)). ...
Technical Report
Full-text available
Study on the potential impact of a step-change in homeworking, as induced by the COVID-19 pandemic, on agglomeration benefits in transport appraisal. Research into the unparalleled shift towards homeworking during the coronavirus (COVID-19) pandemic and how to consider this when completing transport appraisals.
... But we may also be interested in measuring endogenous peer e↵ects, i.e., the e↵ect of co-workers' current productivity on a worker's current productivity, especially if we believe that current and permanent productivity potentially represent di↵erent mechanisms, such as peer-pressure, learning, or free-riding (Kandel and Lazear, 1992;Mas and Moretti, 2009;Bandiera et al., 2013). Ultimately, we may be interested in obtaining measures of both e↵ects within a single context, which (again) has been shown to be a non-trivial task. 2 1 Studies which use direct measures of worker productivity include Hamilton et al. (2003), Falk and Ichino (2006), Mas and Moretti (2009), Bandiera et al. (2010), Guryan et al. (2009), De Grip and Sauermann (2012), De Grip et al. (2016, Horrace et al. (2016), Horton and Zeckhauser (2016), Steinbach and Tatsi (2021), Steinbach (2017), and Battiston et al. (2021). For low-skilled jobs, Cornelissen et al. (2017) find corroborative evidence for wages using representative register data. ...
Article
Full-text available
During the COVID-19 pandemic, traditional (offline) chess tournaments were prohibited and instead held online. We exploit this unique setting to assess the impact of remote–work policies on the cognitive performance of individuals. Using the artificial intelligence embodied in a powerful chess engine to assess the quality of chess moves and associated errors, we find a statistically and economically significant decrease in performance when an individual competes remotely versus offline in a face-to-face setting. The effect size decreases over time, suggesting an adaptation to the new remote setting.
Article
This study examines the productivity of working from home (WFH) practices during the COVID-19 pandemic. The results reveal that the mean WFH productivity relative to working at the usual workplace was about 60%–70%, and it was lower for employees and firms that started WFH practice only after the spread of the COVID-19 pandemic. However, there was a large dispersion of WFH productivity, both by individual and firm characteristics. Highly educated and high-wage employees tended to exhibit a small reduction in WFH productivity. The results obtained from the employee and employer surveys were generally consistent with each other.
Article
Under the hypothesis that workers become distracted and spend less time working during the COVID‐19 pandemic, we analyze, via a model of endogenous growth, the relationship between distracting activities and home office supply. We find a strong theoretical relation between the intertemporal elasticity of substitution of the time spent on distracting activities and the intertemporal elasticity of substitution of home office labor. We also analyze the effect that the intertemporal elasticity of substitution of home office labor has on the marginal productivity of effective labor and the growth rate of production.
Article
Full-text available
The Covid-19 health crisis has led to a substantial increase in work done from home, which shifts economic activity across geographic space. We refer to this shift as a Zoomshock. The Zoomshock has implications for locally consumed services; much of the clientele of restaurants, coffee bars, pubs, hair stylists, health clubs, and the like located near workplaces is transferred to establishments located near where people live. In this paper we measure the Zoomshock at a very granular level for UK neighbourhoods. We establish three important empirical facts. First, the Zoomshock is large; many workers can work-from-home and live in a different neighbourhood than they work. Second, the Zoomshock is very heterogeneous; economic activity is decreasing in productive city centres and increasing residential suburbs. Third, the Zoomshock moves workers away from neighbourhoods with a large supply of locally consumed services to neighbourhoods where the supply of these services is relatively scarce. We discuss the implications for aggregate employment and local economic recovery following the Covid-19 health crisis.
Preprint
Full-text available
The Covid-19 health crisis has led to a substantial increase in work done from home, which shifts economic activity across geographic space. We refer to this shift as a Zoomshock. The Zoomshock has implications for locally consumed services; the clientèle of restaurants, coffee bars, pubs, hair stylists, health clubs located near workplaces now demand those services near where they live. In this paper we measure the Zoomshock at a granular level for UK neighbourhoods. We establish three important empirical facts. First, the Zoomshock is large; many workers can work-from-home and live in a different neighbourhood than they work. Second, the Zoomshock is very heterogenous; economic activity is decreasing in productive city centres and increasing residential suburbs. Third, the Zoomshock moves workers away from neighbourhoods with a large supply of locally consumed services to neighbourhoods where the supply of these services is relatively scarce. We discuss the implications for aggregate employment and local economic recovery following the Covid-19 pandemic. Paper now available in Covid Economics, Issue 64, pp. 1-41.
Article
Full-text available
We estimate the causal effect of workplace–home commuting distance on inventor productivity. We construct a novel panel of U.S. inventors with precisely measured workplace–home distances and inventor-level productivity. Our identification strategy exploits firm relocations as exogenous variations in the commuting distance of the firm’s inventors. We find a significant negative effect from commuting distance on inventor productivity: every 10 km increase in distance is associated with a 5% decrease in patent counts per inventor–firm pair per year and an even greater 7% decrease in patent quality. The highest-performing inventors suffer the greatest from increased commuting distance. We discuss the implications of our findings in the light of recent trends around telecommuting and remote work during the COVID-19 pandemic.
Article
We conducted a field experiment in a sales firm to test whether improving knowledge flows between coworkers affects productivity. Our design allows us to compare different management practices and isolate whether frictions to knowledge transmission primarily reside with knowledge seekers, knowledge providers, or both. We find large productivity gains from treatments that reduced frictions for knowledge seekers. Workers who were encouraged to seek advice from a randomly chosen partner during structured meetings had average sales gains exceeding 15%. These effects lasted at least 20 weeks after the experiment ended. Treatments intended to change knowledge providers’ willingness to share information, in the form of incentives tied to partners’ joint output, led to positive—but transitory—sales gains. Directing coworkers to share knowledge raised average productivity and reduced output dispersion between workers, highlighting the role that management practices play in generating spillovers inside the firm.
Article
Full-text available
In a field experiment with a retail chain (1,300 employees, 193 shops), randomly selected sales teams received a bonus. The bonus increases both sales and number of customers dealt with by 3 percent. Each dollar spent on the bonus generates $3.80 in sales, and $2.10 in profit. Wages increase by 2.2 percent while inequality rises only moderately. The analysis suggests effort complementarities to be important, and the effectiveness of peer pressure in overcoming free-riding to be limited. After rolling out the bonus scheme, the performance of the treatment and control shops converges, suggesting long-term stability of the treatment effect.
Article
Coworkers are increasingly diverse in their nationality and skill sets. This paper studies the effect of diversity on how workers are organized using data from a field experiment conducted in an environment where diversity is pervasive. Findings show that team organization improves outcomes when workers are from the same country. The opposite is true when workers are nationally diverse. These results are more pronounced for teams of workers with specialized skills. Further investigation of the data suggests that nationally diverse teams have difficulty communicating.
Article
We present new evidence indicating that changing from a traditional human resource management (HRM) environment to an innovative one entails a change not only in formal work practices, but also in the informal networks and patterns of interaction among employees. We focus on differences in the social capital of these workplaces and measure differences in the structure of interactions and information transfer among employees across a sample of manufacturing lines with a common production technology and different HRM systems. We then consider the implications of these differences and show that the change from one form of workplace practices to the other is therefore not just a matter of paying for the direct costs of a new set of HRM practices. Rather, it would involve a disruptive overhaul in the entire network of interactions among all workers at the plant.
Article
We present evidence on social incentives in the workplace, namely on whether workers' behaviour is affected by the presence of those they are socially tied to, even in settings where there are no externalities among workers due to either the production technology or the compensation scheme in place. To do so, we combine data on individual worker productivity from a firm's personnel records with information on each worker's social network of friends in the firm. We find that compared to when she has no social ties with her co-workers, a given worker's productivity is significantly higher when she works alongside friends who are more able than her, and significantly lower when she works with friends who are less able than her. As workers are paid piece rates based on individual productivity, social incentives can be quantified in monetary terms and are such that (i) workers who are more able than their friends are willing to exert less effort and forgo 10% of their earnings; (ii) workers who have at least one friend who is more able than themselves are willing to increase their effort and hence productivity by 10%. The distribution of worker ability is such that the net effect of social incentives on the firm's aggregate performance is positive. The results suggest that firms can exploit social incentives as an alternative to monetary incentives to motivate workers.
Article
I investigate how teamwork may reduce moral hazard by joint monitoring and management. I study two organizational systems differing in the extent to which physicians may mutually manage work: Physicians are assigned patients in a “nurse-managed” system but divide patients between themselves in a “self-managed” system. The self-managed system increases throughput productivity by reducing a “foot-dragging” moral hazard, in which physicians prolong patient stays as expected future work increases. I find evidence that physicians in the same location have better information about each other and that, in the selfmanaged system, they use this information to assign patients.
Article
We use laboratory experiments to explore merger failure due to conflicting organizational cultures. We introduce a laboratory paradigm for studying organizational culture that captures several key elements of the phenomenon. In our experiments, we allow subjects in "firms" to develop a culture, and then merge two firms. As expected, performance decreases following the merging of two laboratory firms. In addition, subjects overestimate the performance of the merged firm and attribute the decrease in performance to members of the other firm rather than to situational difficulties created by conflicting culture.
Article
A body of literature suggests that ethnic heterogeneity limits economic growth. This article provides microeconometric evidence on the direct effect of ethnic divisions on productivity. In team production at a plant in Kenya, an upstream worker supplies and distributes flowers to two downstream workers, who assemble them into bunches. The plant uses an essentially random rotation process to assign workers to positions, leading to three types of teams: (i) ethnically homogeneous teams, and teams in which (ii) one or (iii) both downstream workers belong to a tribe in rivalry with the upstream worker’s tribe. I find strong evidence that upstream workers undersupply non-coethnic downstream workers (vertical discrimination) and shift flowers from non-coethnic to coethnic downstream workers (horizontal discrimination), at the cost of lower own pay and total output. A period of ethnic conflict following Kenya’s 2007 election led to a sharp increase in discrimination. In response, the plant began paying the two downstream workers for their combined output (team pay). This led to a modest output reduction in (i) and (iii) teams—as predicted by standard incentive models—but an increase in output in (ii) teams, and overall. Workers’ behavior before conflict, during conflict, and under team pay is predicted by a model of taste-based discrimination. My findings suggest that interethnic rivalries lower allocative efficiency in the private sector, that the economic costs of ethnic diversity vary with the political environment, and that in high-cost environments firms are forced to adopt “second best” policies to limit discrimination distortions. JEL Codes: D03, D22, D24, D61, D64, D74, F63, J24, J33, J71, 012, 014.