ECON 3123, Week 12: 1 Hierarchies As Decentralized Information Processors
ECON 3123, Week 12: 1 Hierarchies As Decentralized Information Processors
Hongyi Li
October 2016
(1)
+ a3
(2)
+ a4 .
(3)
From arithmetic class, we know that the order in which addition takes place doesnt
matter: we could calculate (((a1 + a2 ) + a3 ) + a4 ) or (a1 + (a2 + (a3 + a4 ))) or ((a1 +
a2 ) + (a3 + a4 )) and get the same result each time. The main thing is that three addition
operations have to be performed.
Now, suppose that an individual is tasked to add four numbers together, and he can only
perform one addition per day. How long would he take to complete the task?
Three days, of course.
Now, if delays are costly, this might not be ideal; perhaps, decisions have to be made quickly
and delayed decisions result in reduced profits. Is there a way to speed up information
processing?
Yes. Consider the following setup:
There are three agents, x, y, z. Each takes one day to complete an addition operation.
x is in charge of adding a1 and a2
1
The organization has to pay each agent within the organization w in wages per day.
Consider the two alternative organizational structures that we introduced in the first
example: (i) a one-agent structure, versus (ii) a two-layer hierarchy with three agents.
Structure (i) results in a net payoff to the organization of w c per day. (The
one agent incurs a total daily wage of w, and each task is delayed for three days,
resulting in a per-day delay cost of 3c/3 = c.)
Structure (ii) results in a net payoff to the organization of 3w 2c
3 per day. (The
three agents incur a total daily wage of 3w, and each task is delayed for two days,
resulting in a per-day delay cost of 2c/3 = c.)
So, a one-layer hierarchy with a single agent is preferred to a two-layer, threeagent hierarchy whenever w is sufficiently large compared to c (specifically, whenever
2w > 3c ).
The take-away point here is that large hierarchies may reduce delay, but are more costly
to maintain. Consequently, organizations with less information-processing demands may
optimally choose to maintain less complicated hierarchies and tolerate some additional
delay.
2.1
may be more efficient: each worker learns to solve a subset of all problems, and whenever he
encounters a problem that he cannot solve, he can pass the problem on to another agent who
has learned how to solve some of the problems that the production worker doesnt know. We
may think of this agent as a manager.
However, communication is time-consuming, in the following way. Each manager starts with a
unit measure of time. Listening to the production worker describe a fraction u of the workers
unit measure of problems takes up fraction uh of the managers time.
The advantage of such an arrangement is that it economizes on learning: if a manager helps multiple production workers, then the production workers do not have to
independently learn the problems that the manager specializes in solving; instead, only
one individual (the manager) has to learn how to solve those problems. This saves on
learning costs.
We may think of h as a function of the availability of communication technology; an
improvement in available communication technology would result in a decrease in h.
We may think of such an arrangement as a hierarchy, with managers at the top and production
workers at the bottom.
In particular, we are interested in the optimal hierarchical shape, i.e. the optimal ratio of
managers to production workers.
Let m be the number of managers. This would mean that for each manager, there
are N/m production workers assigned to that manager. That is, whenever each of
the production workers encounters a problem he cannot solve, he will pass it on to his
manager to solve.
We call N/m the span of control associated with the hierarchy. Were interested in the
determinants of the optimal span of control.
Before we go on, an important point to note: it is always optimal (from the point of view of
maximizing the organizations profits) for the workers to learn the most common problems,
i.e. each worker learns all problems of type x below some threshold x
. The reason is that,
by learning the most common problems, production workers avoid having to pass on tthese
frequently encountered problems to their managers, and thus to economize on communication
costs.
Consider an organization with N production workers and m managers, where every production worker learns the z most common problems, i.e. every problem with x < z. We assume
that each manager must learn how to solve every problem, i.e. zm = 1 (remember that the
hardest problem has type x = 1), so that all problems end up being solved (either by a
production worker or by his manager). Finally, suppose that each worker (both production
workers and managaers are paid a wage w.
We may think of z as the degree of autonomy or decentralization within the organization.
An organization with high z is one where each production worker has a lot of autonomy,
i.e., each worker solves most problems on his own without having to rely on his manager.
(4)
Were interested in choosing m and z to maximize net profit (all other parameters are fixed.)
Notice that we are ignoring all incentive conflicts within the organization.
To solve this problem, note that
Each managers span of control is limited by the time he spends communicating. Remember that given the fraction z of problems that each worker learns how to solve, he
will have to communicate the remaining problems to the manager, which then takes up
h(1F (z)) of the managers time. Given that each manager has N/m subordinates (and
must spend h(1 F (z)) of his time on each), it must be that N
m h(1 F (z)) 1. So, to
minimize the wages paid and learning costs for managers, we minimize the number of
managers required (for given z) by choosing m = N h(1 F (z)).
Thus the organizations profit function becomes (remember that zm = 1)
= N N (cz + w) N h(1 F (z)(c + w)
(5)
(6)
The organization thus chooses z to maximize this expression, i.e. the optimal value of z
is such that the first-order condition is satisfied:
(7)
c = hf (z )(c + w).
(8)
i.e.
We wont directly obtain an expression for the optimal value of z, which we denote as z .
However, well look at how z depends on the other parameters.
2
0
Before diving in, notice that z
2 = h(c + w)f (z) < 0, which ensures that the first-order
approach is valid (i.e. the first-order condition gives us a maximum profit).
dz
dc
using the
(9)
We already know that the denominator of this expression is negative. The numerator is also
c
easy to sign: 2 /cz|z=z = N (c + hf (z )) = N (1 + c+w
) < 0. (Here we use the fact,
following from the first-order condition, that hf (z ) = c+w .) This implies that dz
dc < 0.
This suggests that an improvement in information technology (a decrease in c) will
increase worker autonomy. The intuition is this. An increase in worker autonomy has
two effects, one positive and one negative: (i) an increased cost of learning, versus
6
(10)
We already know that the denominator of this expression is negative. The numerator is also