Communication Structure and its Effects on Task Performance

The formal organization chart of a bureaucratic organization can be thought of as a network. It is a directed graph (a non-symmetric network) that records the social relation "reports to" (or, if you prefer the arrows to go downward, "is the boss of"). That social relation tends to channel a lot of the communications within an organization. For example, a lot of prescriptive information (i.e., do this, stop doing that) flows downward along the links. It's unusual, and can cause problems, when prescriptive information moves in a different pattern, such as from a boss to someone else's subordinate, or among peers, or from a subordinate to a supervisor. At the same time, a lot of descriptive ('this is the status of such-and-such project') flows up the links, often in the form of reports.

The formal organization also determines a lot of other communication as well. For example, most roles (jobs) within an organization are interlinked, forcing occupants of those roles to interact with others playing their own roles. For example, the personnel department, like the payroll department, generally has to interact with all employees. People in the marketing research department work closely with people in the marketing department, who also work with people in the new product development department.

In addition to these formally prescribed communications, there are also multitudes of informal communications, ranging from getting technical advice to sexual harassment. Some of these are affected by the official organizational structure (e.g., most communication occurs between people whose offices are within 50 feet of each other) and others are not.

Given that there are many ways to structure communication in an organization, the question arises how the pattern of communications within the organization affects the performance of the organization -- its ability to sell products, reduce costs, adapt to changes in environment, etc. In other words, what is the best communication structure?

One way to investigate this question is to perform controlled lab experiments. This is what Alex Bavelas and his student Harold Leavitt did at MIT in the late 40s and 50s.

The Leavitt Experiment

Basically, the experiment had 5 people play a game somewhat similar to Clue in which they have to solve a puzzle. At the start of the game, each person is given a key bit of information. In order to solve the puzzle, everyone's bit of information must be pooled together. The players communicate with each other, transmitting what information they have, until the puzzle is solved. Unlike Clue, the idea is for every single player to get the answer. The faster, more efficiently they can do it, the better.


Each person is put in a uniquely colored cubicle (this is before computers are invented). They are given colored stationary matching their cubicle. There are slots in the wall where they can send and receive messages.

At the start of each game, each person is given 5 symbols chosen from a set of six. The objective is to discover which symbol they all have in common. Each cubicle has 6 switches on the wall, labeled by the symbols. When a player learns the answer he (they were all men) flips the switch corresponding to the symbol he believes everyone has in common. The experimenters record the time when that happens. When all 5 subjects have flipped a switch, the experimenter calls a halt to the game and records whether they got it right.

If the game has not yet been halted, a subject could change his answer if he likes. Subjects are free to write anything they like on the their messages, and to send as many or as few as they like.

The cubicles do not all contain the same number of slots. Some cubicles might have just one slot, which would mean that the subject in that cubicle could only message one person (whoever is at the other end of the tube). The slots serve to restrict communications into certain patterns. Five separate patterns were tested: the star (wheel), the Y, the chain (line), and the circle. The subjects were not told what pattern they were in, or even that they were in a pattern.

The same set of subjects played the game in the same positions 15 times.


Time. The star and Y were considerably faster, on average, than the chain and circle.

Messages. The star and Y used the least number of messages. The chain was next, then the circle (which used quite a bit more).

Errors. An error was defined as the throwing of an incorrect switch before the end of a game. The star, the Y and chain made the fewest errors, while the circle made the most (however, the circle had the most error corrections).

Satisfaction. The subjects in a the circle network enjoyed themselves the most, followed by the chain, the Y and finally the star.

Leadership. The probability of opining that the group had a leader went up in the order: circle, chain, Y, and star. In addition, agreement as to who was the leader increased in the same order (it was 100% in the case of the star).

Improvement. Circle people were very likely to say that they could have done things more efficiently and that was missing was "a system". Star people did not feel they could improve much.


Which structure should have been the fastest? Theoretically, the star can solve the puzzle in a minimum of 5 time units, the Y in 4, the chain in 5, and the circle in just 3. How do you figure this? Consider the star:

At time 1, persons 2, 3, 4, and 5 can send their information to person 1, so at the end of time 1, person 1 knows all the information and therefore (presumably) the answer. At time 2, person 1 can send the answer to person 2. At time 3, person 1 can send the answer to person 3. At time 4, person 1 can send the answer to person 4. And finally, at time 5, person 1 can send the answer to person 5. So, given that you can't send stuff simultaneously to more than one person at a time, it takes a minimum of five units of time for this configuration to solve the puzzle.

Now consider the Y. At time 1, persons 2, 3 and 4 send their info to person 1, while person 5 sends to person 4. At time 2, person's 1 and 4 exchange what they know, so know both of them know everything. At time 3, person 1 sends the answer to person 2, while person 4 sends the answer to person 5. At time 4, person 1 sends the answer to person 3. So it takes the Y only 4 units of time, at minimum.

Now, the circle is more complicated to figure out, so I will use multiple figures to explain it. At time 1, persons 5 and 2 send to person 1, person 1 sends to person 2, and persons 4 and 3 exchange their information, so that at the end of time 1, here is what each person knows:

End of Time 1

(The numbers in the parentheses indicate whose information a person has at the end of the round. Person 1 has {1,5,2} which means they have obtained the information that 2 and 5 had, plus their own.) Now, in Time 2, person 1 sends their info to person 5, person 4 also sends to person 5,  person 3 and person 2 exchange information. At the end of Time 2, person 5 knows the answer, but no one else does:

End of Time 2

In time 3, persons 3 and 4 exchange their information (so both know the answer), and persons 1 and 2 exchange their information, so each of them knows the answer. Since person 5 already knew the answer, the round is over.

So the circle should have been the fastest. However, the actual experimental results were exactly the opposite. Now, it's easy to think: 'big deal: people are not computers. they don't necessarily do things in the mathematically most efficient way.' But if that were all there was to it, none of the structures would have performed consistently better than the others. There is clearly SOME effect of structure on performance, just not the one we expected.

So what is the relationship? According to Bavelas and Leavitt, it's centralization. The more centralized a structure is, the better it performs. They use "centralization" to refer to the distance nodes are from the most central node, who acts as an information integrator. The closer everyone is to that integrator, the faster the puzzle is solved. Of course, channeling all information to a single integrator is not the only possible strategy for solving problems. But it is a reasonable strategy that is easy to implement and which works well with simple problems.

Another feature of centralized systems is that the most central node is clearly more central than all the other nodes. This makes clear who is the leader, and also makes the funnel-everything-to-an-integrator strategy more obvious. So centralized systems don't waste time searching for a strategy nor vying for leadership: they just do it.

Later research, however, has shown that centralization is not always optimal. The next table shows under which conditions and criteria centralized vs decentralized systems are best.

Variable Simple Task Complex Task
Fewest messages: centralized centralized
Least time: centralized decentralized
Least errors: centralized decentralized
Most satisfaction: decentralized decentralized

Centralized systems don't work as well as decentralized systems with complex tasks because some problems are too big for an individual to handle: the whole idea is to use the entire organization as a distributed processing unit to solve problems that no one person could possibly handle. This is similar to the classic argument for the superiority of capitalism (decentralized) over communism (centralized). 

Also, with large systems (many nodes) central nodes can be overwhelmed with communications. In addition, in such systems, most of the network remains idle while waiting for information to filter back from the center.

Positional Analysis

Leavitt also analyzed the data by position in the network -- i.e., by node. He found that:

  1. The most central positions send the most messages
  2. They enjoy their jobs more
  3. The most central positions send fewer organizational messages and more informational/solution messages. (and in highly centralized networks, the most central nodes send few informational messages vs solution messages)


An interesting implication of the Bavelas-Leavitt research is that one path to leadership is centrality. In the past, many people have assumed that leadership is a personality trait that a person is born with or at least develops over a long period of time. Since in the experiment people are placed in the central position randomly, it is apparent that there are time when it is positional centrality that determines leadership, and not any enduring personality characteristic. 

Structure and Cognition

In these experiments, it was initially thought that the winning networks would be those whose pattern allowed the spread of information in the minimum time. This is a structural characteristic. However, it turned out that another structural characteristic, centralization, seemed to have more effect.

One possible reason for this is that in the centralized systems, the number of possible patterns of communication was much smaller. People were more or less forced to adopt a certain strategy for solving the problem. In contrast, for the circle, there were many many possibilities, only a few of which worked well. Even if they all worked well, it was much harder for people to choose one strategy and stick to it. It is often the case in organizations that a satisfactory strategy that is easy to find, implement and stick to is superior to an optimal strategy that is hard to find, hard to implement, and hard to stick with. 

It is also helpful if the strategy that a structure pushes people towards is one that people are naturally positively disposed towards. For example, people readily understand leadership. It is much harder to understand the system which, in the circle, would actually lead to much faster performance than the integrator strategy.

This strategy of choosing a satisfactory rather than optimal solution is known as satisficing, and is an important concept in organizational theory. It is part of a larger conception of organizations as systems that overcome human cognitive limitations -- a condition known as bounded rationality. The idea of bounded rationality is that people are intendedly rational, but they can't really be rational because they can't consider all the possibilities. There isn't enough time, nor information, nor the brains needed to sort it all out. 

Copyright 1996 Stephen P. Borgatti Revised: October 08, 1997 Go to Home page