Contracting Theory by rfp70474


Contracting Theory document sample

More Info
									                Discussion of “Contracting Theory and Accounting”
                                            Robert P. Magee1
                                 Kellogg Graduate School of Management
                                         Northwestern University
                                        Evanston, IL 60208-2002

                                                   March 2001


Professor Lambert provides a very useful synthesis of the major issues in managerial accounting

and the insights that agency theory has provided on those issues. In this discussion, I highlight

some of the limitations of these models in examining accounting measurement questions.

Lambert calls for additional work in multiperiod contracting models, and I discuss some of the

modeling choices that must be made in such work. Finally, I present some thoughts on the

relation between contract models and empirical research.

JEL Classification:       M41, L14, D82

Keywords:        Contracting, Agency Theory, Managerial Accounting

 While retaining responsibility for the content of this discussion, I would like to acknowledge many helpful
conversations with my colleagues Ronald Dye, Frank Gigler and S. Sridhar.
       In "Contracting Theory and Accounting," Professor Lambert has focused on the area in

which contracting theory has had the greatest impact in our field – agency theory research in

managerial accounting. It is true that contracting theory has had an impact on other research

areas, e.g., in the market for audit services and in the design of covenants for financial

instruments, but the volume of research on managerial accounting issues far exceeds that in these

other areas. It is also true that there are other models of conflict and information, such as adverse

selection and informal contracting (e.g., “reputation”) models, but again the agency paradigm has

received the most attention. The model explicitly incorporates potential conflicts between two

(or more) parties, and it recognizes the role of information in managing these conflicts. As a

result, the agency paradigm has been very useful in providing an understanding of many

managerial accounting practices.

       Lambert's objective in the paper is to provide a document that would guide researchers

and doctoral students through the area's principal results and insights. In my opinion, he

achieves that objective. The paper covers many of the substantive topics in managerial

accounting and highlights the economic forces that drive the agency theory results on these

topics. There is enough detail to understand the results, but not so much as to let the reader get

bogged down.

       The paper can be divided roughly into thirds and, for the most part, I will try to follow

this organization. The first third of the paper concentrates on the classic agency model of

Holmström's [1979] paper and its multi-person equivalent of Blackwell's [1953] sufficiency

result for single-person information economics. As Lambert points out, this result has affected

our notions of controllability, confirmed our estimation of the benefits of flexible budgeting, and

refined our ideas about variance investigation and relative performance evaluation. One of the

most useful insights from this model is that, in designing a contract, actions you don't want the

agent to take are just as important as the action you do want the agent to take. The set of

allowable actions and the outcome function (or functions) are critical to the nature of an effective

contract. I will come back to this issue later in my review.

       Once I discovered Lambert's focus on managerial accounting, I put together a list of the

topics I would expect to see covered. These topics can be divided into two broad categories –

incentive function design issues (contract shape, contact length, variance investigation, budget-

based behavior and participation) and performance measure design issues (controllability, goal

congruence, cost allocation, transfer pricing, investment return, etc.). Models such as

Holmström's (and extensions discussed below) are pretty good at the first set of issues, but they

have a difficult time dealing with performance measurement problems that are fundamental to

managerial accounting. For instance, cost system design decisions require that costs of raw

materials, labor, machine time and other resources be determined and then aggregated into costs

of products. In an early agency theory accounting model, Demski [1972] portrays the principal's

problem as requiring a double choice – an accounting performance measure and an incentive

function based on that accounting performance measure. In fact, the paper's example holds the

incentive function fixed, at a 10% profit share, and focuses on the choice between direct costing

and full costing for the measurement of profit. But Holmström’s incentive function could be

viewed as operating directly on the information primitives (e.g., raw materials, labor, etc.), with

very little in the way of implications for composite performance measures like profit.

       One could adopt this information primitive perspective in looking at accounting

measurement issues. For instance, Demski [1981] states that the essential purpose of cost

allocation is to provide information about the amount of resource used and the agent's action,

rather than to form a composite performance measure. That is, the principal cares about square-

feet, not the allocation of occupancy cost based on square-feet. This interpretation works well

from an information perspective, but then why not use square-feet directly? It becomes difficult

to make a case for aggregating occupancy cost along with others to form a composite accounting

statistic like product cost or profit, because such aggregation typically reduces the efficiency of

the contract. As a result, it is not easy to examine issues like depreciation or overhead allocation.

As Lambert notes, the linear aggregations of Banker and Datar [1989] provide a step in an

interesting direction, but it’s still pretty difficult to get empirical predictions about accounting

measurement from this model.

        One of the shortcomings of Lambert’s review is that it doesn’t give much attention to

areas where the agency paradigm makes predictions that do not conform to observed practice.

My own experience with the agency model’s implications for managerial accounting has focused

on cost allocation, finding that optimal contracts can use revenues and costs in non-additive ways

that do not bear much resemblance to the procedures we commonly see in practice (Magee

[1988]). This type of result raises an interesting question. If the purpose of a model is to help

researchers understand the phenomena we observe in the world of contracting, what does it mean

when that model makes a prediction that bears little correspondence to observed phenomena? It

could be that empirical researchers have had no reason to look for the phenomenon, so we don't

know whether it's there or not. It might mean that practice is inefficient, or it might mean that

the model is incomplete in some significant way, or given the model’s sensitivity to assumptions,

it might mean that we have the wrong action set or utility function or outcome function or

information structure. Many elements of the agency model are not observable to the principal,

much less the researcher, leaving us in a position similar to the theoretical physicist who

hypothesizes the existence of a particle with no way to verify it. Like the physicist, we might

hope that future technologies would allow such observations, but we would also have to

recognize that observability would change the fundamentals of the agency problem. A more

cautionary interpretation of Lambert’s lack of attention to this issue is that our editorial processes

screen out such papers, and they receive little attention.

        The middle section of Lambert’s paper examines the multi-action, multi-outcome, “LEN”

models (linear contract, negative exponential utility, and normal outcome distributions) that are

special, constrained cases of the agency model. The standard, effort-aversion agency model

appears quite limiting as a model of managerial behavior, and expanding to multiple actions

allows one to focus on effort allocation in addition to effort supply. The reader should recognize

that there is nothing intrinsic in the setting of these models that makes them qualitatively

different from those of the former section. The assumptions of negative exponential utility and

normal outcome distributions can be viewed as special cases of a Holmström model. However,

the restriction of contracts to linear functions of the observable variables constrains the principal

to an inefficient set of choices.

        Lambert has done a good job of describing the motives behind these special cases, and I

agree that tractability is probably the most common rationale. To be more specific, we are

interested in accounting issues like performance measurement, and the LEN model's assumptions

allow greater analyses of alternative performance statistics with multiple actions. The insights

this model provides on window-dressing, valuation versus incentive compensation, and so on,

are quite nice. Lambert's discussion of the use of stock price information in contracting (section

3.3.5) is very insightful. As noted above, one of the significant disciplines of the agency model

is that one must be very careful about the action set available to the agent, and the LEN analysis

highlights the fact that discouraging unwanted actions is just as important as motivating the

actions that are wanted. In a multi-measure environment, the purpose of some measures is to

alert the principal to an agent’s attempts to look good without doing good.

       These results are reasonable enough that they are probably not driven by the inefficient

linear contracts, but the reader should recognize that the whole setup is quite carefully

orchestrated to yield interpretable results and that there are limits on the insights that might be

obtained. For instance, it is not possible to assess the values of different performance measures,

because those measures are not being used in an efficient fashion. And, the linear contracts do

not allow any insights about contract shapes, budget-like incentives such as stock options, and

similar issues. For instance, a LEN model conclusion that it would be beneficial to add a second

performance measure could disappear if the principal were allowed to use an option-like

incentive based on a single performance measure.

       As Lambert points out in the third principal section, extending the model to include

private information is essential to examining many management control issues. Management

control systems are designed to influence delegated decision making, and the delegation of

decisions is almost always motivated by a desire to take advantage of the agent’s private

information and to limit the information rents of the agent. Even when an agent’s action is

observable, such as in the quantity of a resource used, we cannot achieve first-best solutions

because the principal doesn’t know what action is optimal conditional on the agent's information.

Questions regarding budget participation, transfer pricing, cost allocation and capital budgeting

are most naturally examined when the agent has such information. Lambert speculates that

private information has something to do with the frequent use of aggregated performance

measures like revenues or expenses or profits, and I concur. As the agent knows more (relative

to the principal) about the opportunities to increase revenues and reduce costs, it seems likely

that the principal will base incentives on a simple aggregation of these outcome measures – the

principal’s payoff outcome. To do otherwise could cause the agent to, say, forego an opportunity

to decrease expenses by $1 in order to increase revenues by 80¢ because the incentive weight on

revenues exceeded that on expenses. Having a wide array of action choices, with the resulting

arbitrage opportunities, would also make it difficult to place differential weights on components

of income.

       Finally, Lambert provides a short section on multi-period models. If researchers are to

say anything about investment decisions or capitalization and depreciation or similar issues, we

need a model of multiple periods. Given the discussion of accounting aggregation above, it

would be convenient if the model had a good reason to combine information into a composite

performance statistic. Such a model would allow us to consider the efficiency of different

treatments of research and development costs, for instance.

       While Lambert calls for more work on multi-period issues, we need to recognize that

there are multiple ways in which multiple periods might be incorporated – either through the

Acceptable Utility (AU) Constraint or the Incentive Compatibility (IC) Constraint. That is, do

we have to worry about the agent's consumption smoothing or that the agent might learn

something in one time period and then leave before the end of the model (or that the principal

might fire the agent before the end of the model)? Or, do we have to worry that the agent might

learn something in the course of the model and alter future decisions? Or, do we have to worry

about both? Then, of course we need to consider how the private information arrives through

these periods. Is all of the agent's private information revealed in the first period or does it come

to light more slowly? What sorts of commitment are available (or simply assumed)? What

limitations on communication are in place? All of these issues arise in structuring a multiperiod

model, making it difficult to achieve comparability across studies

       Various researchers have worked on aspects of this setting, but it doesn’t take much for

the set of constraints to increase to the point where it becomes difficult to achieve insightful

results. As Lambert notes, it is necessary to make some modeling choices to develop results and,

given the limits of our technology, the assumptions must be tailored to the researcher's question.

It seems to me that multi-period accounting models should let the "information action" occur

over multiple periods. That is, a multi-period model in which everyone learns everything they're

ever going to know in the first period doesn't capture the tension of having to make an accrual

accounting decision when there is substantive uncertainty to be resolved in future periods.

       Such assumptions should probably be thought of in much the same way as using up

degrees of freedom for an empirical researcher, because it is possible to generate almost any

model result, if one has sufficient latitude in the choice of assumptions. Lambert's analysis of

the Economic Value Added/discount rate choice is an excellent example. Is there a setting in

which it is optimal to establish incentives using Economic Value Added calculated at the

principal’s discount rate? The answer is “yes,” but Lambert shows that the result appears rather

fragile: a slight change in assumptions yields a different result. Should models be evaluated

based on the realism of their assumptions or the usefulness (accuracy) of their predictions? In

general, we would choose the latter, but suppose the researcher has observed behavior and then

built a model to predict this behavior. Assessing the predictive usefulness of such a model

becomes difficult, unless it makes additional predictions that may be confirmed (or refuted) by

further observation.

         In their ten-year retrospective on positive research in accounting, Watts and Zimmerman

[1990] stated that theory should play a larger role in informing empirical tests and that we should

focus on the simultaneous choice of efficient contracts and efficient accounting procedures.

While recognizing that managerial accounting is a difficult area for empirical testing, I would

say that the agency theory model has provided mixed success on this program. However, agency

theory is by no means alone in falling short of this rather lofty goal. Most models take as given

the set of contractible information and do not maximize over the contract and its arguments.

         Questions of incentive function characteristics and methods for dealing with private

information have been addressed with some success. However, one of agency theory's insights

is the importance of the agent's action set, and this set is usually not observable to the empirical

researcher. The same is true of utility functions and the probability distributions of outcomes

conditional on agent actions. Moreover, measurement questions like cost allocation,

aggregation, and depreciation have been very difficult to tackle when we assume that the

information "primitives" must be used in an optimal fashion. Even when linear contracts are

imposed, unweighted aggregation of financial outcomes is rarely found in the solution. We

could view this as an indication that accounting aggregation is inefficient and that incentives

should be based on a large set of such primitives. Or, we have omitted some key elements in the


         As a field, we have tended to focus on the decision impact of information that derives

from a "user orientation," and we commonly assume perfect knowledge of action sets, perfect

memory (or costless history), costless information production and costless contracting. As a

result, many models make "corner" predictions about the amount of information desired – either

use every piece of information or suppress every piece of information. But in general, it is not a

simple thing to produce and process information and to optimize such incentives, even in an

example in which we've been given a utility function and outcome probability functions.

Incorporating some kind of cost for information and/or contract complexity might be an

interesting avenue to pursue. There has been some research in this area (e.g., Dye [1985]), but it

has not been applied to agency theory information problems.

       The late mathematician Paul Erdös had a practice of posing problems to the mathematics

community at large and assigning to each problem a monetary value commensurate with its

difficulty. He posed $10 problems, $100 problems, and $1000 problems. If you solved the

problem, he would write you a check for that amount. In his review, Lambert provides a very

nice overview of the agency theory/managerial accounting literature, with appropriate attention

to the economic forces underlying the principal insights. He notes some remaining puzzles,

suggests some areas for future research and, in the concluding section, lists what might be

considered "$1000 problems" because they are central to furthering our understanding of

accounting practices.2

       Let me close my discussion by noting two things about this list of issues. First, I perceive

a common denominator in these questions – our uncertainty about the role of accounting

measurement when the economics of information delivery are changing rapidly. What are the

forces behind aggregation, accrual practices, and reporting discretion? If it is costly information

processing, then the trend in computing power could endanger these practices. Accounting

reports could become unaggregated reports of financial and nonfinancial items that the user

brings together into some model for valuation or contracting or some other purpose. On the

other hand, if cognitive limitations or contracting costs are behind these accounting practices,

then they may endure. We often use these models to derive cross-sectional hypotheses about

accounting practices and the use of accounting information, but this approach might also be used

to predict time-series changes in accounting systems.

           Second, I believe we should put a premium on developing models of these practices that

are "robust" to the details of the model. For instance, I could write an LEN model with a set of

revenue generation and cost reduction opportunities that have identical returns to effort (though

of opposite signs for revenues and expenses) and independent, identically distributed errors.

Accounting aggregation would seem efficient in that model – that is, simply adding revenues and

subtracting expenses would produce an optimal contract. However, this result is not very robust

to minor changes in specifications (e.g., if it were harder for the agent to increase revenues than

to decrease costs), so I don’t think it provides much insight about a practice that appears

widespread. Lambert’s results on the cost of capital used in Economic Value Added incentives

raise the same types of concerns. Of all the possible parameterizations of these models, the set

that yields the predicted behavior seems quite small.

           In every model, predicted behavior may depend on variables that are observable to the

researcher and on variables that are not observable to the researcher. I don’t mean to imply that

we cannot learn from models in which results are sensitive to parameters and structures that we,

as researchers, would never be able to observe. But results that depend more on variables that

can be observed by a researcher, and subsequently confirmed by empirical verification, would

provide additional comfort that the model has captured the essentials of the accounting problem

at hand.

    I should add that no one at the conference volunteered to write any checks!


Banker, R., S. Datar. 1989. Sensitivity, precision and linear aggregation of signals for

       performance evaluation. Journal of Accounting Research 27(1), 21-39.

Blackwell, D. 1953. Equivalent comparisons of experiments. Annals of Mathematical Statistics

       24, 267-272.

Demski, J. 1972. Optimal performance measurement. Journal of Accounting Research 10(2),


Demski, J. 1981. Cost allocation games, in: S.Moriarity, ed., Joint cost allocations, (Center of

       Economic and Management Research, University of Oklahoma) 142-173.

Dye, R. 1985. Costly contract contingencies. International Economic Review 26(1), 223-250.

Holmström, B. 1979. Moral hazard and observability. Bell Journal of Economics 10(1), 74-91.

Magee, R. 1988. Variable cost allocation in a principal-agent setting. The Accounting Review

       63(1), 42-54.

Watts, R., J. Zimmerman. 1990. Positive accounting theory: A ten year perspective. The

       Accounting Review 65(1), 131-156.


To top