You are on page 1of 45

A Framework for Conducting

Numerical Experiments on Cost System Design*†

Vic Anand
University of Illinois at Urbana-Champaign
vanand@illinois.edu

Ramji Balakrishnan
University of Iowa
Ramji-Balakrishnan@uiowa.edu

Eva Labro
University of North Carolina Chapel Hill
Eva_Labro@kenan-flagler.unc.edu

November 21, 2017

*
The appendix and computer code that accompany this paper are available for downloading at http://vicanand.wee-
bly.com/abl_jmar_code.html. We request that anybody who downloads the code and amends it for their research
paper purpose, acknowledges their use of this code and references this working paper in an acknowledgement sec-
tion.

We thank Karen Sedatole (editor), two anonymous referees, Anna Rohlfing-Bastian (discussant), and participants
at the Accounting Research Workshop at the Universtiy of Basel for comments.
A Framework for Conducting
Numerical Experiments on Cost System Design1

This paper aims to advance the use of numerical experiments to investigate issues that surround the design
of cost systems. As with laboratory and field experiments, researchers must decide on the independent
variables and their levels, the experimental design, and the dependent variables. Options for dependent
and independent variables are ample, as are the ways in which we can model the relations among these
variables. We provide a modular framework that provides structure to these variables, their definitions,
and the modeling of the connections among them. Further, we offer some insights into the design and
layout of output data files, which will allow for easier data analysis. We also present tips on how to report
the results from such numerical experiments effectively. Finally, we furnish online the source code in C# for
many of these modules. We hope that the framework and guidance provided in this paper will help spur
and focus further meaningful work in this important area of management accounting.

INTRODUCTION
Product- and capacity-planning are among the more important problems that firms confront as

these decisions shape revenue potential and cost structure. Ideally, these problems should be solved

jointly because they are inter-dependent (Balakrishnan and Sivaramakrishnan 2001). However, doing so

is a daunting task from both a computational and an informational perspective. Balachandran et al. (1997)

show that a joint formulation is computationally intensive even if the firm has complete information about

input and output markets, production technology, and product costs. Moreover, as Hwang et al. (1993)

point out, firms usually operate with limited information. Thus, there is demand for mechanisms that

decompose the joint problem into smaller pieces that are both consistent with available information and

are computationally feasible (Balakrishnan and Sivaramakrishnan 2002). We use the term “cost system”

to describe this decomposition process. 2

1
The appendix and computer code that accompany this paper are available for downloading at http://vicanand.wee-
bly.com/abl_jmar_code.html. We request that anybody who downloads the code and amends it for their research
paper purpose, acknowledges their use of this code and references this working paper in an acknowledgement sec-
tion.
2
All but the smallest of organizations employ such systems to make product- and capacity-planning decisions. There
is a vast literature on activity-based costing (Cooper and Kaplan 1988). Time-driven activity based costing (Kaplan
and Anderson 2007) is the latest iteration in the debate of how to construct a cost system, taking information and
computational constraints into account.

2
A rich literature has sought to determine the economic loss from such disaggreagation, after en-

dowing the firm with full information. Here, the loss results from using the disggregated data to make

individual decisions relative to using the full information to make a joint decision. Among others, Kaplan

and Thompson (1971) argue that the disaggregated estimates or fully-allocated product costs serve as

good approximations of long-run marginal cost. Banker and Hughes (1994) establish conditions under

which such a disaggregation is without economic loss. These works provide the theoretical justification

for the observed practice of firms employing the reported full cost of a product as an estimate of its mar-

ginal cost when making decisions (Shim and Sudit 1995). 3 However, these works focus on the computa-

tional complexity aspect of the problem because they endow the firm with full information about output

and input markets, and production technology.

Over the past two decades, an emerging literature has examined the impact of limited infor-

mation on cost system design. This literature rests on the premise that firms do not know the true con-

sumption of resources by products and must employ cost systems to estimate such resource consumption.

That is, firms design their cost systems with incomplete and imprecise information about their production

technology. The incompleteness might occur because of the dispersal of relevant information within an

organization or because of a lack of knowledge about the production process. The imprecision stems from

the cost-benefit tradeoff inherent in measuring resource consumption. The firm must therefore decide

how much to invest in gathering information about resource consumption, and hence a cost system is

always the outcome of a design choice.

While there is a recent rise of time-driven cost systems, empirically, the vast majority of cost sys-

tems are still two-stage systems (Drury and Tayles 2005), and this is where the majority of the numerical

3
Noreen and Burgstahler (1997) show from an analytic perspective, and Coller and Collini (2015) show via numerical
experiments that pricing strategies based on full costs may lead to the firm achieving sub-optimal profitability. This
loss due to informational limitations differs from formulations wherein there is uncertainty regarding the nature of
the problem itself.

3
experiment literature has focused.4 With two-stage systems, cost system designers must make two sets

of design choices. In stage 1, firms decide how to group resource costs from the financial accounting sys-

tems into activity cost pools, with direct tracing where possible and with allocations otherwise. In stage 2,

firms decide how to allocate the costs in each activity cost pool among cost objects. Thus, the key design

choices in constructing a cost system are: which costs to allocate, how many cost pools to form, how to

group resource costs into activity cost pools, and which drivers to employ in both stages of the cost system.

These choices influence the extent of the error in costs reported by the cost system relative to a (unob-

servable in practice) full-information based benchmark. Accordingly, this research stream has focused on

quantifying the impact of these choices on the error in reported costs (Datar and Gupta 1994) and the

economic loss that results because of decisions that rely on erroneous estimates.

It is difficult, if not impossible, to address analytically questions about the effect of cost system

design choices on error and on decision quality. 5 Hence, this literature has largely relied on numerical

experiments (“simulations”) to examine these research questions, 6 as this method allows one to examine

complex tradeoffs and yet generate generalizable insights. For instance, related to the accuracy of cost

systems, numerical experiments have allowed researchers to address questions such as the impact of

errors in design chocies on the accuracy of reported costs, where costing system designers can get the

most “accuracy bang for their buck” when improving their systems, and whether the rules of thumb that

are advocated in management accounting textbooks and practitioner publications improve accuracy. Re-

lated to the effect of cost system accuracy on decision quality, numerical experiments have provided in-

sights into questions on the efficiency of cost-based decision rules in determining the expected cost of

4
Hoozee and Hansen (forthcoming) is a notable exception.
5
Noreen (1991) shows analytically the linearity and separability conditions under which Activity-Based costs (ABC)
provide relevant costs. These conditions can be broken in many ways. Christensen and Demski (1997) numerically
explore the ability of different accounting procedures (e.g. ABC) to capture marginal costs, and show that it depends
on the underlying production technology that the cost system is trying to capture.
6
Some papers have used research methods such as surveys or field research (e.g. Anderson and Sedatole 2013). Our
focus is on the use of numerical experiments.

4
under- and over-stocking of capacity and how the degree of competition affects the economic conse-

quences of a firm’s choice of allocation bases. Research also has examined whether some production

environments are more (or less) sensitive to costing accuracy and decision quality outcomes, relating the

demand for accuracy to core characteristics of the firms.

While the insights on costing systems that can be gained are substantial, there is limited consen-

sus on how to best structure these numerical experiments. There are numerous choices about how to

model the dependent and independent variables, and the relationships among them. Moreover, a lack of

computer code has posed a significant barrier to entry. In this paper, we describe a framework for struc-

turing numerical experiments that addresses the design of cost systems and provides modular building

blocks. 7 We also provide guidance on the selection of building blocks, structuring of output files, analysis

of the generated data, and reporting of results. Online, we also provide suitable code (in C#) for several

modules in the overall experiment. The appendix to the paper (available online) describes that code and

its use in detail. A researcher can use these modules, with or without modifications, to ease the task of

generating a suitable sample for analysis. We hope that the framework and guidance provided in this

paper will help encourage and direct further work on the important topic of cost system design.

We organize the remainder of this paper as follows. In the next section, we describe the key com-

ponents of a numerical experiment that addresses cost system design, focusing on the full-information

benchmark. We then describe the modular blocks to model the limited information setting. We next ad-

dress choices encountered when setting up the numerical experiment. We offer pointers on how best to

organize and analyze the resulting data. We conclude with some conjectures about open research ques-

tions.

7
Hocke et al. (2015) likewise stress the importance of experimental design when conducting numerical experiments
in management accounting.

5
COST SYSTEM DESIGN: COMPONENTS
It is evident that researchers should design numerical experiments in the same manner, and with

the same care, as laboratory and field experiments. Thus, researchers must decide which independent

and dependent variables to include and how to model the relationships between them. The set of poten-

tial independent and dependent variables is large and there are numerous choices for how to model the

relations among them. In this and the next section, we provide a structure to think about these options

and present modular building blocks.

Figure 1 provides an overview of our structure. In this figure, Panel A represents the “real” eco-

nomic operations of the firm. These choices, which represent physical flows, define the scope of the prob-

lem. In panel B, we depict choices made when constructing the cost model. Items in bold are usually

modeled in the literature and are included in our computer code, while items in regular font have received

little attention in the literature to date.

Insert Figure 1 about here

The firm
The firm is one of two key data structures in our framework (the other being the cost system).

The literature on cost system design typically models the firm as a single decision maker that owns a

production technology that converts inputs (resources) into outputs (products and services) and maxim-

izes an objective like profit. 8 The firm attempts to solve a capacity acquisition and planning problem. Given

forecasts of demand, the firm must decide how much capacity to acquire and sets prices for its outputs.

The chosen prices affect demand, the input to the capacity acquisition problem. Collectively, the capacity

8
Note that some analytical accounting literature has modeled capacity acquisition, its allocation and pricing optimi-
zation as a decision with multiple decision makers in the firm. For example, in Banker and Hughes (1994), the mar-
keting department makes pricing decisions while the production department makes capacity decisions. The role of
full product cost is to decompose the problem in two separable programs where communication between the two
departments can be limited to full product costs, rather than all available information (which is prohibitively costly
to communicate). The focus on a single-decision maker obviates important issues such as decentralization and trans-
fer pricing, however. Extant research also has suppressed the impact of taxes, or variation in taxes across jurisdic-
tions.

6
acquisition and pricing problems are known as the Grand Program (Balakrishnan et al. 2011). Even with

full information about products’ costs, the Grand Program becomes complex once we consider character-

istics (e.g., uncertainty) of input and output markets, and production processes. 9 A vast operations man-

agement literature has modeled these complexities and focused on finding an effective solution in an

efficient fashion (see Van Mieghem (2003) for an overview). This literature also considers computational

complexity, motivating attention to implementable (as opposed to optimal) solutions, and the speed of

such heuristic approaches.

Limited information about products’ true costs complicates the above problem substantively.

Limited information constrains the design of the cost system and hence the ability of the firm to compute

error-free costs. This view naturally leads to two features: first, the full-information optimal solution pro-

vides a natural benchmark, the first-best solution, for assessing alternate cost systems and heuristics for

making decisions. This first-best solution specifies the capacity and price choice that maximizes profit un-

der the Grand Program. Second, there are two sources for the loss relative to the first-best solution. The

first is due to limitations on available information and the second arises from employing heuristical solu-

tion techniques that do not lead to the optimal solution with limited information. As in Anand et al. (2017),

we term the loss arising from information limitations as the second-best solution. The solution from heu-

ristics is then the third-best. Different papers focus on different aspects of this overall problem, and make

specific assumptions about resource, product and process characteristics. Variations in the availability of

information are particularly interesting to accounting researchers.

9
See Balakrishnan and Sivaramakrishnan (2002, p. 13) for a mathematical formulation of the Grand Program. A
monopolistic firm facing uncertain demand must choose prices for each of its N products. The firm uses M resources
to make these products. The firm must choose purchase quantities (capacities) of the M resources before observing
demand. After observing demand, it can purchase additional capacity at a premium. In Balakrishnan and Sivarama-
krishnan’s formulation, variable cost per unit is known, as is consumption of capacity resources by product. In this
paper, we assume both must be estimated (with error) by the cost system, further complicating the Grand Program.

7
In the remainder of this section, we describe an ideal situation in which a firm has full information

about its production technology. In such a case, a cost system is not needed since the firm knows its

products’ costs without error. We argue that the full information setting provides a natural benchmark

against which researchers can evaluate cost system designs. We then discuss the role of limited infor-

mation in numerical experiments on cost system design, and argue that limited information is the main

friction in such experiments.

Full information setting


Resources or inputs
The first major component of a model of a firm is a listing of all of the resources that serve as

inputs to the firm, as shown in the left-most box of panel A in Figure 1. The number of resources to con-

sider depends in part on the number of cost objects (discussed later). Second, the researcher must specify

the nature of the markets for these inputs. Much of the numerical experiment research to date has as-

sumed a competitive market for input resources. That is, the firm can buy any desired quantity at a con-

stant marginal cost. A third choice deals with the useful life for the resource. Most numerical experiments

to date assume that the firm can acquire capacity on an as needed basis. 10

Allowing for long-lived resources, we model resources and resource consumption as follows. For-

mally, let 𝑖𝑖 be the index for resources, and 𝑹𝑹𝑹𝑹𝑹𝑹 be the total number of resources. 11 Then, let 𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 be

the capacity available for resource 𝑖𝑖 at the start of period 𝑡𝑡,and let 𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 be the additional capacity

(if any) purchased in period 𝑡𝑡. From a timing perspective, the term 𝑃𝑃𝑃𝑃𝑃𝑃 indicates that the firm might ac-

quire resources ahead of learning demand realizations. However, some (e.g., raw materials) resources

may be acquired after learning demand; for other resources, it might be possible to augment capacity in

10
Balachandran et al. (1997) is an exception. In contrast, analytical research (Banker and Hughes 1994) generally
allows for long-lived capacity. In the context of a numerical experiment, a researcher who chooses to model long-
lived capacity also needs to specify how available capacity diminishes over time, and the periodic expensing of the
cost for the diminution in resource capacity.
11
See panel A of Table 1 for a listing of the notation. As in Table 1, boldface text indicates parameters that are
included in our computer code. Notation relating to the firm is termed “exogenous” because they are a given from
the perspective of designing a cost system.

8
the spot-market (𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 ). 12 Next, as Balakrishnan et al. (2004) argue, users have differing de-

grees of control over how much of available capacity can be consumed. Let 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 ≤

(𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 + 𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 + 𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 ) be the capacity used up. Notice that resource characteristics

influence 𝐶𝐶𝐶𝐶𝐶𝐶𝑆𝑆_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 . For example, it is not possible to carry over some resources such as labor from

one period to another. Then, 𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖+1 = 𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 + 𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 + 𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 − 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 .Note

that 𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 , 𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 and 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 are all decisions made by the firm within the con-

straints posed by the production technology. We defer the discussion of the determinants of these deci-

sions until after we have discussed the other two components – production process and outputs -- in the

value chain.

Insert Table 1 about here

Next, turning to costs, let pre-demand purchases be valued at 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑇𝑇𝑖𝑖𝑖𝑖 (𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 ) and let

𝜽𝜽𝒊𝒊𝒊𝒊 (𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 ) be the cost of spot purchases of capacity. 13 The sum of these values determines the

cash outflow in period 𝑡𝑡 due to resource 𝑖𝑖. However, the “cost” from an accounting perspective might

differ due to the capitalization of expenses. Thus, let 𝐴𝐴𝐴𝐴𝐴𝐴_𝐶𝐶𝐶𝐶𝐶𝐶𝑇𝑇𝑖𝑖𝑖𝑖 be the accumulated cost in the books

of resource 𝑖𝑖 at the start of period 𝑡𝑡. Let the diminution in value (due to use) be 𝛾𝛾𝑖𝑖𝑖𝑖 (𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 ). Then,

just as with physical capacity, we can write an inventory equation for costs: 𝐴𝐴𝐴𝐴𝐴𝐴_𝐶𝐶𝐶𝐶𝐶𝐶𝑇𝑇𝑖𝑖𝑖𝑖+1 =

𝐴𝐴𝐴𝐴𝐴𝐴_𝐶𝐶𝐶𝐶𝐶𝐶𝑇𝑇𝑖𝑖𝑖𝑖 + 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑇𝑇𝑖𝑖𝑖𝑖 (𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 ) + 𝜽𝜽𝒊𝒊𝒊𝒊 (𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 ) − 𝛾𝛾𝑖𝑖𝑖𝑖 (𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 ).

The above framework is much too general to solve in closed form. Thus, few analytical papers

distinguish between the economic and the accounting definitions of “cost” in a multi-period framework. 14

Most papers assume a constant marginal cost of capacity purchased in advance, as well as an additional

12
Some companies outsource their capacity investments. For example, Porsche outsources production of some of
its vehicles to a Finnish company (Dougherty 2009). It is able to order as many or as few cars as it wants in a period.
Thus, it is able to augment its capacity by purchasing additional capacity as needed on the spot market.
13
The literatrure uses the terms hard and soft capacity constraints to denote the feasibility of such purchases. See
Göx (2002).
14
For an exception, see Dutta and Reichelstein (2002) who identify a class of depreciation rules and a capital charge
rate that aligns the investment incentives of a better informed divisional manager with headquarters.

9
cost for capacity purchased on the spot market, rushed in once demand is observed. That is, these papers

specify 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑇𝑇𝑖𝑖𝑖𝑖 (𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 ) = 𝑐𝑐𝑖𝑖 ⋅ 𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 and 𝜽𝜽𝒊𝒊𝒊𝒊 (𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 ) = 𝜽𝜽𝒊𝒊 ⋅ 𝑐𝑐𝑖𝑖 ⋅ 𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 ,

where 𝜽𝜽𝒊𝒊 ≥ 1. Finally, several papers assume that capacity lasts for one period only, which considerably

simplifies the problem. The simplification is appealing if we assume that the firm acquires capacity on an

as-needed basis (i.e. all spot purchases) and eliminate the distinction between pre- and post-demand pur-

chases of capacity. Then, 𝜽𝜽𝒊𝒊 = 1, 𝑐𝑐𝑖𝑖 = 1, 𝑃𝑃𝑃𝑃𝑃𝑃_𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑖𝑖𝑖𝑖 = 0, 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 = 𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷_𝑨𝑨𝑨𝑨𝑨𝑨𝑷𝑷𝒊𝒊𝒊𝒊 , and 𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 =

0 for all 𝑖𝑖 and 𝑡𝑡 (these are the assumptions implemented in our current computer code). While the above

discussion pertains to a single resource 𝑖𝑖, firms usually acquire several resources, with differing procure-

ment, life and usage characteristics. We are not aware of papers that explicitly model these differences

and consider how they influence decisions regarding how much of what resource to acquire. Rather, the

literature imposes the same underlying conditions on all of the resources for a firm.

����������⃗𝒕𝒕 , for the 𝑖𝑖 resources in period 𝑡𝑡. Each ele-


Next, we conceive of a vector of resource costs, 𝑹𝑹𝑹𝑹𝑪𝑪

ment of this resource corresponds to the cost for resource 𝑖𝑖 in period 𝑡𝑡. We note that this information is

usually available to the firm from its financial accounting system. Finally, even in simple settings, the re-

searcher has to consider the dispersion in resource costs. In a numerical experiment, the parameter

𝑹𝑹𝑹𝑹_𝑫𝑫𝑫𝑫𝑫𝑫𝑫𝑫 (resource cost dispersion) determines the variation in the costs of individual resources, the

elements of ���������⃗
𝑹𝑹𝑹𝑹𝑹𝑹 . 15 Low values of 𝑹𝑹𝑹𝑹_𝑫𝑫𝑫𝑫𝑫𝑫𝑫𝑫 correspond to environments with many resources with

roughly equal monetary importance (i.e., low dispersion in resource costs) such as might be found in a

firm producing a wide and varied product line. High values of 𝑹𝑹𝑹𝑹_𝑫𝑫𝑫𝑫𝑫𝑫𝑫𝑫 indicate an environment with

many small resource cost pools and a few large cost pools such as might be found in a refinery or a law

firm where machine and human resources account for a majority of costs, respectively (Balakrishnan et

15
We implement resource cost dispersion (RC_DISP) using two parameters: the first 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷1 resources account for
𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷2 percent of total resources.

10
al. 2011). As we discuss later, this value has the potential to affect the magnitude of the aggregation error

in reported costs, and thereby the quality of decisions made with limited information.

The appendix provides computer code for generating sample data that implements the model

described in the preceding paragraphs. The code allows the researcher to specify the number of resources

and the bounds for the cost per resource unit. In particular, with spot-purchase of capacity (the usual

���������⃗�. The program


simplifying assumption), the program generates a vector of resource costs per unit �𝑹𝑹𝑹𝑹𝑹𝑹

generates the elements of this vector, the random cost per resource unit, within the bounds specified by

the researcher and assuming a uniform distribution. This module also computes a measure of resource

cost dispersion (𝑹𝑹𝑹𝑹_𝑫𝑫𝑫𝑫𝑫𝑫𝑫𝑫) as the percent total cost accounted for by the top x% of all resources as used

in Balakrishnan et al. (2011). To illustrate the diversity in modeling approaches in the literature, note that

Labro and Vanhoucke (2008) measured resource dispersion by the variance in the distribution of total

resources in a setting where total resources is a constant.

Cost objects or outputs


We next define cost objects, which specify “where the resource cost is used”. This item is the right

most box in panel A of Figure 1, with the notation in panel B of Table 1. The researcher needs to specify a

list of cost objects. While we use the term “products” for convenience, the methodology can apply to any

item that we wish to cost, like services, customers, or distribution channels. We conceive of a product

����⃗𝒕𝒕 be the product prices chosen by the firm in


market structure with 𝑪𝑪𝑪𝑪 products, indexed by 𝑗𝑗. Letting 𝑷𝑷

���⃗𝑡𝑡 , 𝜀𝜀���⃗�
period 𝑡𝑡, let 𝑅𝑅𝑅𝑅𝑅𝑅_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑃𝑃 𝑡𝑡 be the associated realized demand for product 𝑗𝑗 (we ignore more com-

plex pricing strategies such as allowing for volume discounts). The vector notation on the input reflects

possible interdependencies in demand, which is assumed to be stochastic. The term ε���⃗𝑡𝑡 represents prod-

uct-level demand shocks. At a general level, the correlations in demand shocks can be over time and

across products within a period. The demand function might vary over time to reflect concepts such as

product life cycles. Of course, because the firm does not have to fill (satisfy) all realized demand, let

11
����⃗𝒕𝒕 , 𝜀𝜀���⃗�
𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑷𝑷 ����⃗ ���⃗�.
𝑡𝑡 ≤ 𝑅𝑅𝑅𝑅𝑅𝑅_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑷𝑷𝒕𝒕 , 𝜀𝜀𝑡𝑡 Next, letting 𝐼𝐼𝐼𝐼𝑉𝑉𝑗𝑗𝑗𝑗 be the inventory of product 𝑗𝑗 at the start

of period 𝑡𝑡. The firm has to decide on the amount to produce, 𝑃𝑃𝑃𝑃𝑃𝑃𝐷𝐷𝑗𝑗𝑗𝑗 . Notice that the quantity produced

is a function of current and future demand as well as current and anticipated decisions regarding resource

����⃗𝒕𝒕 , 𝜀𝜀���⃗�.
capacity. The inventory equation then yields: 𝐼𝐼𝐼𝐼𝑉𝑉𝑗𝑗𝑗𝑗+1 = 𝐼𝐼𝐼𝐼𝑉𝑉𝑗𝑗𝑗𝑗 + 𝑃𝑃𝑃𝑃𝑃𝑃𝐷𝐷𝑗𝑗𝑗𝑗 − 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑷𝑷 𝑡𝑡 Each el-

�����������⃗𝒕𝒕 is given by 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑷𝑷


ement in the associated revenue vector, 𝑹𝑹𝑹𝑹𝑽𝑽 ����⃗𝒕𝒕 , 𝜀𝜀���⃗�
𝑡𝑡 ⋅ 𝑷𝑷𝒋𝒋𝒋𝒋 .

Researchers must make choices about the markets for the firm’s outputs, and these choices can

introduce significant complexity to the model. Thus, as with resources, researchers typically impose many

simplifying assumptions. Most papers ignore the possibility of accumulating inventory, meaning that

𝐼𝐼𝐼𝐼𝑉𝑉𝑗𝑗𝑗𝑗 = 0 for all 𝑗𝑗 and 𝑡𝑡. Next, researchers have typically used a linear demand function with an additive

demand shock 16 to specify a market structure (Banker and Hughes 1994). Most papers have modeled an

additive demand shock with mean zero. 17 Some papers (e.g. Hwang et al. 1993) allow for a more general

demand function but suppress a demand shock. The choice of demand function allows the researcher to

manipulate price elasticity, which determines the relation between price and demand. It seems reasona-

ble to model a linear demand function and to choose price elasticity randomly from a pre-specified distri-

bution. The choice of demand shock captures the idea that realizations of demand are uncertain and may

differ from expectations. Another important choice is whether the firm can or must fill realized demand

���⃗𝑡𝑡 , 𝜀𝜀���⃗�
(that is, whether to set 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑃𝑃 ���⃗ ���⃗�).
𝑡𝑡 = 𝑅𝑅𝑅𝑅𝑅𝑅_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑃𝑃𝑡𝑡 , 𝜀𝜀𝑡𝑡 The assumption that all demand must

be filled imposes restrictions on the values for the cost of acquiring resources. Because in this case we

need resources to make the products and fill realized demand, spot-market purchases cannot be prohib-

itively expensive when available capacity is insufficient. In sum, it is evident that allowing for an output

16
For example, demand for a product is a linear function of price, plus a demand shock, e.g. 𝑄𝑄𝑗𝑗 = 𝛼𝛼𝑗𝑗 + 𝛽𝛽𝑗𝑗 𝑃𝑃𝑗𝑗 + 𝜀𝜀𝑗𝑗 ,
where for product 𝑗𝑗, 𝑄𝑄 is quantity demanded, 𝑃𝑃 is price chosen by the firm, 𝜀𝜀 is the demand shock, and 𝛼𝛼 and 𝛽𝛽 are
constants.
17
Some papers also model a multiplicative demand shock (e.g. Reichelstein and Rohlfing-Bastian 2015).

12
market introduces considerable complexity to the problem. Thus, several papers even suppress the out-

put market altogether, specifying ����⃗


𝑸𝑸𝒕𝒕 as the exogenous realized product demand in period 𝑡𝑡 (Balakrishnan

et al. 2011; Labro and Vanhoucke 2007). Some papers (Anand et al. 2017) bridge the gap, allowing the

firm to choose the product portfolio but not to otherwise set demand quantities.

While much of the above discussion pertains to a single product or service, even with the most

restrictive set of assumptions (e.g., the firm fills exogenous demand in its entirety each period), the re-

searcher needs to take into account demand patterns among products. In particular, we need to consider

variations in quantities demanded. Such variations are useful for testing whether refined cost systems

have more value in settings with a few high volume products and many low volume products. That is,

dispersion in product demand is likely an important factor to consider in evaluating the role for infor-

mation. Further, we can model other characteristics of the output market such as high or low true profit

margins and the variance in these margins (Anand et al. 2017). Arguably, the same error in computing

costs can alter decisions and have an impact when profit margins are slim, relative to when profit margins

are more likely to be either highly positive or highly negative. The appendix describes features of the code

that the researcher can use to induce desired demand patterns (e.g., a few high volume and many low

volume products) by modifying the code suitably.

Production technology
We conceive of the production technology (the middle box in panel A of Figure 1; see also panel

C of Table 1) as a function that relates inputs and outputs. At a general level, this technology is a mapping

from the amounts of resources used, 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝐶𝐶𝐶𝐶𝑃𝑃𝑖𝑖𝑖𝑖 to production quantities, 𝑃𝑃𝑃𝑃𝑃𝑃𝐷𝐷𝑗𝑗𝑗𝑗 . This mapping has

to satisfy feasibility constraints such as having the ability to supply needed resources (via prior or current

period purchases). In addition, it is possible that this function is stochastic if the researcher wishes to

model variations in the production process. Economies of scale and scope introduce non-linearity into the

relation as well. While these factors are realistic, virtually all numerical experiment papers to date have

13
employed a deterministic Leontief technology. The implicit assumptions are that each product uses re-

sources in fixed proportions and any kind of scale economy or substitutability among resources is absent.

These assumptions imply that we can express a product’s full cost of production as a linear function. This

assumption corresponds well with the design of cost systems, most of which also make a similar assump-

tion about the production function. From a modeling perspective, Leontief technology implies that we

can express resource consumption patterns in the form of a resource-consumption-pattern matrix,

𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷. The rows of this matrix represent resources and the columns represent products. Each

element 𝑎𝑎𝑖𝑖𝑖𝑖 represents the production coefficients: the quantity of input resource i required for one unit

of output j.

Manipulating the pattern of cell entries in 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷 allows the researcher to model al-

ternate production environments, as defined by three attributes. First, the parameter 𝑫𝑫𝑫𝑫𝑫𝑫 (density of

consumption matrix) captures the extent of resource traceability (or, its counterpart, resource sharing) as

measured by the number of zeros in the resource consumption matrix, 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷. When 𝑫𝑫𝑫𝑫𝑫𝑫 is

low, the resource consumption matrix is sparse, meaning that only a few products consume any given

resource (that is, we have many zeros in the consumption matrix). As might occur in a job shop, there is

high traceability of costs to products. For instance, we can directly trace much of a lawyer’s time to indi-

vidual cases. In contrast, a dense matrix implies a setting with many common costs and low traceability.

A bottler is a good example because all products (e.g., many kinds of drinks) go through the same line.

Second, the researcher can vary the extent of diversity in resource consumption by changing the correla-

tion patterns in resources needed. That is, the researcher can manipulate the correlation between any

two rows of the resource consumption pattern matrix, 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷, to capture the extent of varia-

tions in resource use across products. 18 A large positive correlation (𝑪𝑪𝑪𝑪𝑪𝑪) induces similarity between the

18
Other papers have taken different approaches to model diversity in resource consumption. Labro and Vanhoucke
(2008) use two parameters to model diversity in resource consumption across products: the number of cost driver

14
consumption patterns of resources across products. A negative value for 𝑪𝑪𝑪𝑪𝑪𝑪 implies significant disparity

between the consumption patterns of resources across products. Balakrishnan et al. (2011) use this prop-

erty to distinguish between resources whose consumption is proportional to the number of units made

and to the number of batches made, reflecting a cost hierarchy.

The appendix describes our heuristics for constructing a consumption matrix. In the appendix, we

discuss our methods for ensuring that the resulting matrix meets the researcher’s desired values of re-

source dispersion and density. Figure 2 shows sample resource consumption pattern matrices at different

values of the correlation and density parameters.

Insert Figure 2 about here.

Computing Product Costs Under Full Information


Once the production technology has been defined, it is possible to compute true, error-free prod-

uct costs as ��������⃗


𝑷𝑷𝑪𝑪𝑩𝑩 = ���������⃗
𝑹𝑹𝑹𝑹𝑹𝑹 ⋅ 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷, where the B superscript denotes benchmark costs. Note

that, under a Leontief technology, true costs are a linear combination of the rows of the resource con-

sumption pattern matrix, with each row weighted by the corresponding unit resource cost. We compute

true (error-free) profit margins on individual products as the ratio of the product’s selling price to its true

cost. Thus, (un)profitable products have a margin (less) greater than 1.0, and breakeven products have a

margin of exactly 1.0. In our notation, the profit margin for product 𝑗𝑗 in period 𝑡𝑡 is 𝑀𝑀𝑀𝑀𝑅𝑅𝑗𝑗𝑗𝑗 = 𝑃𝑃𝑗𝑗𝑗𝑗 ⁄𝑃𝑃𝐶𝐶𝑗𝑗𝑗𝑗𝐵𝐵 .

Linking choices
We end this section by emphasizing the importance of picking a setting that is tractable while

capturing the core tensions in the research question. For example, Labro and Vanhoucke (2007)’s research

question relates to how different types of errors (aggregation and measurement errors) interact, and pos-

sibly trade-off in determining the accuracy of product costs. Hence, they actively manipulate many of the

parameters that we will discuss in the next section on limited information, while keeping the one-period

links and the variance of the distribution of these links over the cost pools. They also separately model diversity in
proportional resource usage by products at individual cost pools.

15
production technology very simple by randomizing it. On the other hand, Labro and Vanhoucke (2008)’s

research question relates to how different aspects of diversity in resource consumption patterns affect

the accuracy of reported product costs. Hence, they actively manipulate production technology parame-

���������⃗ (such as RC_DISP). Moreover, we note that the


ters of 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷 (such as DNS) and of 𝑹𝑹𝑹𝑹𝑹𝑹

choices for the various aspects of the firm’s model are linked. Modeling long-lived capacity immediately

brings into play issues such as the rate of dimunition in capacity. The timing for the purchase of capacity

has implications for whether it is possible (or profitable) to fill all realized demand. While allowing for

inventory accumulation adds realism, it also complicates the problem substantively by linking the physical

(or “real”) problem across periods. In general, we advise caution in adding features, and to suppress any

that do not directly speak to the question at hand.

Limited information
Full-information based analyses allow us to examine interactions among problem features. How-

ever, this assumption is tenous even when we impose simplifying assumptions about market structures

and production functions. Firms typically do not have full information about the consumption of resources

by individual products, even if they have one centralized accounting system that collects all available in-

formation. The problem is excacerbated in settings with decentralized decisions. Production managers,

for example, might know some (but not all) details about resource consumption and costs, whereas mar-

keting has insight into demand distributions (Banker and Hughes 1994). As another example, depreciation

methods start playing a role to optimize decisions in decentralized settings (Dutta and Reichelstein 2002).

In sum, we argue that firms must necessarily resort to some simplification to overcome the informational

barriers that prevent them from analyzing decisions in the context of full information.

Cost Systems Arise When Available Information is Limited


When dealing with limited information, the first task is to specify the nature of the limitations.

���������⃗, and the demand-price relation in


The typical assumption is that the vector of total resource costs, 𝑹𝑹𝑹𝑹𝑹𝑹

the output market are known. The limitations in information concern the firm’s production function, which

16
is where the demand for a cost system arises. As noted earlier, research on cost system design has pri-

marily considered a Leontief technology, which means we can represent the product to resource mapping

by the resource consumption matrix 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷 . Models that consider limited information assume

that the firm has information about only some rows of 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷. That is, the firm knows resource

consumption patterns, by product, only for select resources. That is, while the firm knows the total ex-

penditure on each resource, it does not know which products consumed non-select resources. This limi-

tation constrains the design of a cost system.

Design of two-stage cost systems


The bulk of the research on the design of cost systems has focused on two-stage systems. 19 In

their first stage, the systems re-group resource costs into costs contained in activity cost pools. In the

second stage, systems allocate to cost objects the costs in the activity cost pools using cost drivers (e.g.,

labor hours). Formally, we first compress the number of rows in matrix 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷 to yield a

smaller matrix of activity consumption patterns, 𝑨𝑨𝑨𝑨𝑨𝑨_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷. Each row in 𝑨𝑨𝑨𝑨𝑨𝑨_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷 is a lin-

ear combination of the rows in 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷, with the restriction that the sum of the weights (across

rows of 𝑨𝑨𝑨𝑨𝑨𝑨_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷) is 100% for each row of 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷. Applying the same linear combi-

nation to a vector of resource costs means that we aggregate resource costs into activity cost pools (the

totals are the same due to the restriction on the weights). In the second stage, we distribute the cost in

each activity cost pool to the 𝑗𝑗 cost objects. 20 We represent this structure in panel B of Figure 1, with the

attendant notation in panel D of Table 1. Here, the link between the first two boxes, resouces and activi-

ties, is the first stage and the link between activities and cost objects is the second stage. Note that the

19
Exceptions include Hoozee et al. (2012) who analytically examine the effect of driver choices on error in time-
driven activity based costing. Hoozee and Hansen (forthcoming) use numerical experiments to compare and contrast
traditional and time-driven systems. Time-driven cost systems seemingly skip the intermediate stage of allocating
resource costs to activity cost pools, and use time equations to immediately allocate resource costs to products.
TDABC requires some aggreagation of resource costs, however.
20
Some work (Labro and Vanhoucke 2007; Labro and Vanhoucke 2008) has more explicitly modeled the two-stage
nature of the allocation to allow for differences in both production function and cost allocation in the first and second
stage.

17
parameter 𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷 (for choice relating to the grouping of resources into activities) deals with stage 1, while

the parameter 𝑷𝑷𝑷𝑷𝑷𝑷 (for choosing cost drivers) deals with the allocation in the second stage.

Cost pools
Executing the first stage requires that we specify the number of activity cost pools (𝑨𝑨𝑨𝑨𝑨𝑨) and the

linear combinations of resource costs into activity costs. Typically, the number of cost pools is a design

variable, manipulated by the researcher (Balakrishnan et al. 2011). The code provided online allows the

researcher to specify a fixed set of alternatives for the number of cost pools. Obviously, the number of

resources provides an upper bound for the number of cost pools.

Grouping resources into cost pools


There are many possibilities for mapping resources to activities. The first major choice is whether

a given resource maps to one or many activity cost pools. Labro and Vanhoucke (2007) and Labro and

Vanhoucke (2008) are primary examples of works that consider a one-to-many mapping. Because of their

interest in aggregation error (among other items), they systematically vary the extent of this mapping.

Later works adopt the one-to-one mapping between resources and activity cost pools. The next major

choice is the “rule” for allocating resource costs to activity cost pools. For example, with 𝑖𝑖 resources and

𝐴𝐴 activities, we could assign the largest 𝐴𝐴 resources (as measured by resource cost) to the activities to

simulate a size-based assignment. We can then randomly assign the remaining 𝑖𝑖 − 𝐴𝐴 resources among the

𝑨𝑨𝑨𝑨𝑨𝑨 activity cost pools or pool the low-cost resources into a single “miscellaneous” cost pool. To simulate

a correlation-based assignment, we could pick “like” resources with a correlation in consumption patterns

with a base resource when assigning them to the activity cost pools. Additional refinements are possible.

For example, Balakrishnan et al. (2011) split resources into “volume” and “batch” level resources based

on the correlation in resource consumption patterns. They then examined the effect of comingling these

resources in the grouping process versus keeping the costs separate in activity cost pools. This analysis

18
helps to shed light on the gains obtained from using “batch” level drivers relative to only employing “vol-

ume” level drivers. 21

The appendix provides detail on the options available in the code. As noted earlier, the input pa-

rameter 𝑷𝑷𝑷𝑷𝑷𝑷𝑷𝑷 allows the researcher to specify the method to be considered. We note that all of these

options require that the cost of each resource be assigned in its entirety to a given activity cost pool.

Choice of cost drivers


In stage 2 of a two-stage system, we allocate costs from activity cost pools to cost objects. The

key limitation here is that each cost pool can have only one cost driver. That is, the number of activity cost

pools equals the number of drivers. In addition, given that there is no benefit from having two or more

pools with the same driver, the number of drivers on which usage information is available constrains the

number of cost pools. Babad and Balachandran (1993) analytically model the tradeoff between the infor-

mation loss from using few drivers to the costs of adding more drivers. Homburg (2001) builds on this

model to allow the use of composite drivers. Despite the insights provided by these models, numerical

experiments tend to rely on simple rules to select drivers.

The driver determines the ratios that correspond to the split of the total cost among cost objects.

For example, consider a 2-product firm. If product A consumes (in total) 3 units of a driver resource, and

product B consumes 2 units of that driver resource, then 60% (40%) of the cost from the pool will be

allocated to product A (B). There are several methods for selecting drivers. The “Big Pool” method first

identifies the resource that contributes the most cost to the given pool. This method is consistent with

limited information in that we only need information concerning the consumption patterns for this driver

resource. It also conforms to the idea that a firm will have greater information about its more expensive

resources. An alternative is to employ an “indexed driver.” We construct this synthetic driver as the

21
We are not aware of papers that have formally considered the choice of drivers among alternates in this first stage.
The results in Labro and vanhoucke (2007) suggest that this choice is not likely to be critical in determining the
accuracy of reported costs.

19
straight or weighted average of the consumption patterns for two or more resources contained in the cost

pool. For example, a firm might use a combination of labor hours and machine hours to develop an in-

dexed driver, and hence improve the specification of the cost driver used. The researcher has to specify

both the number of resources combined and their weighting, keeping in mind the associated implications

for available information. 22 Research to date has modeled indexed drivers as a straight average of the

resource cost of a subset of the resources pooled in a cost pool. A key issue to consider is whether to use

“higher-order” drivers because Homburg (2001, 2004) demonstrates that drivers that capture heteroge-

neity in resource consumption are of value.

The online code provides the ability to choose between the big pool and indexed driver methods.

As noted earlier, the parameter 𝑷𝑷𝑷𝑷𝑷𝑷 governs this choice. We note that prior work has, for the most part,

focused on the Big Pool method, as there are few economically reliable gains from using indexed cost

drivers (Balakrishnan et al. 2011).

Comparing alternate designs – dependent variables


As with laboratory experiments, designers of numerical experiments must specify one or more de-

pendent variables. In this section, we argue that researchers must first consider whether to model the

firm’s use of cost information in a managerial decision, before deciding which dependent variables to

calculate. If, for example, the researcher’s goal is to evaluate the accuracy of different cost system de-

signs, it is not necessary to model a cost-based decision. The researcher could simply vary features of

the cost system and compare the accuracy of reported costs to a full-information benchmark, as in

Balakrishnan et al. (2011). Alternatively, the researcher may wish to explicitly incorporate a cost-based

managerial decision, i.e. a decision-context, into the numerical experiment. We discuss both scenarios

below.

22
It is easy to verify that we recover the full-information solution if we use all of the resources and weight them by
the total resource cost.

20
One period models: No decision context
The researcher also has to specify the criteria for comparing the outputs of competing cost system

designs to use as dependent variables. As there is no decision, profit is not a useful measure. We therefore

accomplish this task by comparing each system based on limited information (the second-best solution)

to a benchmark first-best solution obtained with full information. The setting in Labro and Vanhoucke

(2007) is a useful starting point. Here, the authors fix the total cost, the number of cost objects, the asso-

ciated demand, and the production function. The lack of uncertainty allows them to examine how cost

system design choices, such as number of cost pools, affect errors in reported costs as categorized by

Datar and Gupta (1994). Labro and Vanhoucke (2007) compute several error metrics. Following Babad

and Balachandran (1993) and Homburg (2001), they calculate the Euclidian distance as 𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬 =

𝐽𝐽 ��������⃗
�∑𝑗𝑗=1(𝑃𝑃𝑃𝑃𝑗𝑗𝐵𝐵 − 𝑃𝑃𝑃𝑃𝑗𝑗𝑅𝑅 )2 where 𝑗𝑗 indexes products, 𝑷𝑷𝑪𝑪 𝑩𝑩 is the vector of benchmark (full-information) prod-

��������⃗
uct costs, and 𝑷𝑷𝑪𝑪 𝑹𝑹 is the vector of costs reported by the cost system. (The superscript R indicates a re-

ported cost.) Following Christensen and Demski (1997) and Datar and Gupta (1994), they also calculate

1 𝐽𝐽 �𝑃𝑃𝑃𝑃𝑗𝑗𝑏𝑏 −𝑃𝑃𝑃𝑃𝑗𝑗𝑅𝑅 �
Mean Absolute Percentage Error, 𝑴𝑴𝑴𝑴𝑴𝑴𝑴𝑴 = 𝐽𝐽 ∑𝑗𝑗=1 . Furthermore, they calculate a “materiality”
𝑃𝑃𝑃𝑃𝑗𝑗𝑏𝑏

measure, %ACC, as the percent of products whose costs are reported without substantial error (say, 5%):

1
%𝐴𝐴𝐴𝐴𝐴𝐴 = ∑𝐽𝐽𝑗𝑗=1�1|0.95 × 𝑃𝑃𝑃𝑃𝑗𝑗𝑏𝑏 < 𝑃𝑃𝑃𝑃𝑗𝑗𝑅𝑅 < 1.05 × 𝑃𝑃𝑃𝑃𝑗𝑗𝑏𝑏 ; 0 𝑜𝑜𝑜𝑜ℎ𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒� (see panel F of Table 1). Fortunately,
𝐽𝐽

the rankings provided by alternate measures exhibit strong positive correlation, reducing the criticality of

this choice. However, we note that these error metrics are subject to the criticism that the reported costs

are not independent of each other. After all, the total cost allocated equals the total in the activity cost

pools and is the sum of resource costs.

One period models: Decision context


Hwang et al. (1993) consider a single period decision context where they allow the firm to make

pricing choices in markets with an inverse demand function. Thus, they naturally focus on profit as the

basis for a comparison. However, this choice requires that they sacrifice on the dimension of cost system

21
design. In particular, they fix the number of cost pools and the resources in each pool. Moreover, they

implicitly allow capacity to be acquired on an as needed basis. 23 The paper then considers choices for

alternate drivers. While these assumptions limit some insights, they allow the researchers to characterize

analytically the loss in profit that results if the firm employs reported costs to set prices. The paper then

employs numerical methods for comparing alternate heuristics for selecting cost drivers in individual

pools.

The discussion of Labro and Vanhoucke (2007) and Hwang et al. (1993) highlights a conundrum

that arises when we study cost system design in a single period setting. Ideally, we should compare sys-

tems based on the expected profit they generate, allow for pricing decisions, and permit capacity to be

acquired ahead of demand. The problem arises in that capacity decisions depend on the expected demand

for the products which need to be manufactured with these resources, which are a function of pricing

decisions. However, pricing decisions depend on cost information, which is derived from the total cost of

resources, an output of the capacity planning decision. Single period models break this loop by fixing total

capacity (Labro and Vanhoucke 2007) or by assuming that capacity is acquired based on demand (Hwang

et al. 1993). Relaxing these conditions requires that we consider a multi-period framework.

Multiple periods and decision context


The literature has made limited inroads into considering the effects of cost system design on

decision quality in a multi-period setting. Anand et al. (2017), among the first papers to consider this issue,

highlights a subtle issue that arises when the firm’s decisions rely on limited information. Specifically, a

firm’s cost accounting system predicts the expected costs of implementing a particular decision (say, pro-

duce a given mix). However, implementing the decisions is a “real” action. Thus, the patterns in the un-

known matrix 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑨𝑨𝑨𝑨 determine the quantity of each resource required to produce any given

product mix, regardless of what the cost system reports to be the costs of individual products. Further, a

23
Specifically, the model computes rates based on expected total costs and volumes.

22
firm’s financial accounting system records the aggregate cost of resources used. That is, the financial ac-

counting system provides a second (aggregate) read on the total cost to produce a given mix. If we con-

sider rational decision makers, the two values must agree. Else, the decision maker knows that some un-

derlying assumption is incorrect and will seek to correct the cost system.

The above requirement presents a substantive obstacle for the researcher. The difficulty is that a

random cost system is unlikely to be informationally consistent (in the sense defined above) with the real

costs of producing the optimal mix defined by reported data. Fortunately, Anand et al. (2017) show that

simple heuristics are surprisingly efficient and effective at finding such consistent systems, and that the

third-best profit they generate is very close to the second-best profit, the best that is achievable under

limited information. 24

Links among choices


As with choices pertaining to “real” activities, the researcher needs to be deliberate when mak-

ing choices regarding cost systems. For instance, it is realistic that costs from a single resource cost pool

will flow to many activity cost pools, as is modeled in Labro and Vanhoucke (2007) and Labro and

Vanhoucke (2008). However, more recent reseach (e.g. Balakrishnan et al. 2011) suppresses this feature

in order to focus on generating other insights. More importantly, the researcher must be extra careful to

consider the information available to the decision maker. For instance, constructing an indexed driver

requires considerably more knowledge about the production process than is required for calculating the

big pool driver.

SETTING UP A NUMERICAL EXPERIMENT


In the previous two sections, we developed a set of modeling components that researchers should

consider including in a numerical experiment on cost system design. While numerical experiments allow

for inclusion of more variables than analytical methods, there are limits to the number of variables that

24
Profit efficiency of the third-best solution is best compared against the profit second-best solution, but can also
be compared against the first-best solution.

23
can be modeled and there is a tension between simplicity and elaboration (Harrison et al. 2007). Research-

ers must therefore make three choices: which variables to include, whether to manipulate or randomize

each variable, and where to set their levels. We discuss each of these choices in turn. Table 2 provides a

summary of our commentary in this and the next section.

Insert Table 2 about here.

Choices to be Made
The first choice is which of these modeling components to include. Clearly, the research questions

drive these choices. For example, Anand et al. (2017) include a product mix decision and the dynamic link

between ex post and ex ante information because their core research question is to explore the efficacy

of heuristics in obtaining informational consistency when decisions use error-prone costing data. In con-

trast, Balakrishnan et al. (2011) do not include these components because their interest is in the impact

of cost system design heuristics on the accuracy of the reported cost numbers. While the temptation to

add factors is strong, it also accentuates the problem of seeing the forest for the trees, where the re-

searcher has trouble understanding the many effects they are documenting and the reader is unable to

register the key takeaways of the paper (Labro 2015).

The second choice is which of the included components to manipulate as core variables of interest

and which components to either randomize as control variables or fix at a certain level. As before, research

questions should guide which variables are manipulated. The choice between whether to randomize the

other variables or fix them at a certain level is less clear-cut particularly because some components always

have to be included. For example, the researcher has to define a way for grouping resources into pools

and choosing a cost driver at each pool, even if such cost pool aggregation and driver selection choices

are not the focus. If prior research suggests that the impact of these choices on the outcome variable of

interest (e.g. the impact of aggregating resources into cost pools based on correlation in resource con-

sumption patterns versus size on costing error) is limited, it is sufficient to fix the variable at one level (e.g.

24
all aggregation into pools is based on size). However, prior research may document that some parameters

influence the outcome variable of interest strongly. Alternatively, researchers may have a prior on the

potential impact of the variable. In either case, it becomes important to randomize the variable so that

generalizability (or not) of the results can be addressed. For example, Labro and Vanhoucke (2008) and

Balakrishnan et al. (2011) manipulate the extent of resource sharing in the production technology as a

parameter of interest, and find that this variable affects costing error as errors are more likely to cancel

out when there is more sharing of resources by products. Building on this insight, Anand et al. (2017)

randomize the variable and measure its value to assess the robustness of their results to this production

technology characteristic. Indeed, they find that greater resource sharing leads to a greater probability of

existence of a fixed-point solution.

The appropriateness of the decisions to include modeling components and to either manipulate,

randomize, or fix these variables should receive considerable thought during the design phase of numer-

ical experiments. Writing out hypotheses is very useful in this phase. What are the expectations research-

ers have on the impact of independent variables on the dependent variable of interest, and why? If they

have expectations founded on intuition or prior literature, this may be an indication that it may be worth

manipulating the independent variables in a controlled design. If they do not, it may be sufficient to con-

trol for the variables. In the current state-of-the art literature, we see few, if any, papers that contain such

a hypotheses section. We believe that the relative youth of this literature that has neither seen a lot of

theoretical development nor established a vast amount of generally accepted results inhibits the formu-

lation of hypotheses. It is our hope that such guiding hypotheses will increase the internal validity of the

design of the numerical experiments and accelerate the maturing of this literature.

A third experimental design choice contributes to the external validity of the numerical experi-

ment: what are sensible levels at which to manipulate the variables of interest and what are reasonable

distributions (and bounds) for control variables? Ideally, the researcher should calibrate the numerical

25
experiment with values and distributions that are reflective of what occurs in practice. When existing

survey and case-based evidence is available about the likely values certain variables may take in practice

(e.g. firms form between 1 and 20 cost pools), we can impute these values in the experimental design.

Results from the numerical experiment will then speak to settings that are likely to obtain in practice. This

is the approach taken in, for example, Balakrishnan et al. (2011). Unfortunately, only a limited number of

surveys were published in managerial accounting during the last two decades. With respect to document-

ing what firms do in terms of cost system design in practice, we still refer to dated surveys (e.g. Shields

1995; Shim and Sudit 1995; Drury and Tayles 2005). Sometimes, we are even worse off in that no evidence

whatsoever is available on the range of values parameters are likely to take in practice. In this case, the

researchers’ only option may be to simulate the variables of interest and control variables over their entire

theoretical range (Labro 2015). This choice allows for managers who may have more information on

where their firms are situated on the entire theoretical spectrum to assess how the results apply to their

situation. For example, this is the calibration choice made in Labro and Vanhoucke (2007) and Labro and

Vanhoucke (2008).

Experimental Design
With the above three choices, researchers are now ready to put the components together in an

overarching experimental design. A nested design is typical here. The typical first step is to generate many

“true” or “benchmark” firms, under perfect information conditions. For each of those benchmark firms,

we manipulate the limited information that supports cost system design choices and generate several

“false approximations.” For example, Balakrishnan et al. (2011) manipulated a number of features of the

production technology of the firm, whereas Anand et al. (2017) manipulated both production technology

features and parameters of the market in which the firm operates. From a programming perspective, the

reader can think about such experimental designs as a series of nested “FOR”-loops, as depicted in Figure

26
3. We present the overarching programming code for the nested experimental design of Anand et al.

(2017) in Section 2.5 of the appendix.

Insert Figure 3 about here.

The programming code needs to write out the output that the researchers subsequently want to

analyze. In our experience, it is useful to think about the analyses that need to be run (see below) when

designing the output files. Section 2 developed on both independent variables and dependent variables,

but so far, we have only discussed how to consider choices for independent variables in the design of the

experiment. The choice of the dependent variable(s) of interest is, however, also very important and both

the independent and dependent variables should be written in the main output file that will serve as the

core dataset for the analysis. Alternate dependent variables that the researchers plan to use in robustness

checks should also be included in the main output file. With a nested design, it is helpful to organize the

output files as a relational database (Figure 4 provides sample output). From an error checking and data

analysis perspective, it is useful to assign IDs to each simulated benchmark firm (FirmID) and each simu-

lated cost system approximation (CostSysID). We can then use composite key (FirmID and CostSysID) as

the primary key in the database to facilitate identification of each false system with each true system.

It is worthwhile to write several ancillary output files that are likely to never make it into print.

These files collect data such as intermediate calculations that are not intended to be part of the main

analyses. These ancillary output files play a major role in assessing the internal validity of the program and

its output. These intermediate calculations are vital for the researcher to verify each step in the calcula-

tions for a set of examples drawn from various settings that cover the ranges of the parameters simulated.

Working out small examples by hand and comparing those to the ancillary output files generated when

parameters are set at these specific levels also helps in assessing the internal validity. 25 We recommend

25
We also recommend using an Integrated Development Environment (IDE) such as Microsoft Visual Studio to step
through the code, line-by-line, and observe the values of the variables. Verify these by hand to ensure that the code
is working correctly.

27
that at least two persons in the research team independently work on this validation. In addition, setting

the parameters at extreme values (e.g. simulating a case with only one cost pool) is likely to uncover

coding errors, even if the researchers opt to present results only for more practically relevant parameter

choices.

ANALYZING DATA AND REPORTING RESULTS


If the design of the modeling components covered in sections 2 and 3 and the choice related to

the variation to simulate for these modeling components covered in section 4 are both well thought

through and executed, the execution of the analyses should be straightforward at this point. We presume

the reader has knowledge of standard statistical packages such as SAS and SPSS and is familiar with ANOVA

and regression analysis. We therefore focus on those aspects of the analyses related to numerical exper-

iment methods. One of the core strengths of numerical experiments is that sample size is not constrained;

the researchers’ ability to simulate data points is only limited by computing power and the capacity of

statistical packages to read in those data points. Large datasets, however, also make the analysis of sim-

ulated data different. With large sample sizes, any difference can become statistically significant even

though economic significance is lacking. Thus, we believe that researchers should not stress statistical

significance in their analyses too heavily, as it conflates effect size and sample size. In addition, they should

not “over-interpret” high order interactions that look significant but for which they may not have any

intuition or theoretical rationale.

Alternate analysis methods, which may not be very common in the analysis of smaller scale data

in accounting, play a more important role in numerical experiments. Measures of effect size are more

informative when running an ANOVA on a large sample. Two popular measures of effect size are eta

squared and partial eta squared. Eta squared is interpreted as the proportion of total variability in the

dependent variable that is accounted for by variation in the independent variable, or in mathematical

𝑆𝑆𝑆𝑆𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒
notation, 𝜂𝜂 2 = 𝑆𝑆𝑆𝑆𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡
, whereby SS denotes the sums of squares. Eta squared presents a measure of the

28
relative effect size of the different independent variables in the study. We compute partial eta squared

𝑆𝑆𝑆𝑆𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒
as: 𝜂𝜂𝑃𝑃2 = 𝑆𝑆𝑆𝑆 . Here, the denominator is not the total variation in the dependent variable, but
𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 +𝑆𝑆𝑆𝑆𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒

the unexplained variation in the dependent variable plus the variation explained by the independent var-

iable of interest, which removes the effect of variation explained by other dependent variables. The ad-

vantage of partial eta squared over eta squared lies in its comparability of effect size across different

studies that do not contain the same covariates. Other authors have stressed the importance of just look-

ing at graphs rather than statistical tests with large samples (Runkel 2012).

Reporting Results
After analyzing the data, researchers turn to reporting results. The main hindrance in effective

reporting relates to the problem of seeing the forest for the trees, discussed briefly earlier (see Labro

(2015) for a longer discussion). When a paper tries to discuss every variation of every result, the reader

has difficulty remembering the key takeaways of the paper. We recommend sticking to a handful of main

results, with some development of more minor results as an option, as long as we can present these minor

results as related to one of the main results reported. If hypotheses drive the design of the numerical

experiment, the selection of the main results to report is easy. However, exploratory research is more

common because of the difficulty in developing such hypotheses (see section 4). In such case, it is useful

to have at least a “theme” worked out at the design phase so that it can be the “umbrella” in the write-

up.

The following considerations can be helpful in guiding which results to select for presentation in

the paper. First, researchers should select those results for which they can gain reasonable intuition, as

presenting some intuition substantially improves the understanding of the reader, and the likelihood of

this insight making it to long-term memory. However, unlike in analytical work, where model design helps

develop such intuition, numerical experiments keep such insights hidden in a black box from which the

researchers may have a very hard time extracting insight in the inner workings. Researchers can work

29
through (unreported) smaller examples with heterogeneous parameter values to obtain such intuition.

Second, researchers should not focus on the extreme values in the data, especially not when they selected

to manipulate a parameter over its entire theoretical range. Strange results may obtain in the extremes,

and while those results may be useful for the researcher in checking the internal validity of the data, they

are typically not very interesting to the reader, as those extremes are unlikely to obtain in practice. Third,

given the difficulty of calibrating parameter values, researchers should not stress magnitudes of particular

effects as main takeaways. Numerical experiments are not the appropriate method to address questions

on magnitude; other research methods such as large-scale empirical data analysis or structural modeling

are more suitable to such conclusions. Fourth, consistent with the advice given in the prior section on data

analysis, try to present takeaways as graphs, as pictures tend to “stick” better with the reader.

Because the average accounting scholar is unlikely to be familiar with numerical experiments and

lack “hands-on” experience, reporting of the results of a numerical experiment should done carefully to

address a lack of confidence the reader may have in results generated by this method. A more than aver-

age level of elaboration and explanation on experimental design and methodological choices is necessary

to overcome that hurdle. We hope the framework presented in this paper will be useful in helping re-

searchers make methodological choices that will be easier to communicate to the reader convincingly.

DISCUSSION AND CONCLUSION


Future issues to be Addressed
We believe that the literatue has made limited inroads into understanding the tensions and

tradeoffs that influence the design of cost systems. While extant works have paved the path, these models

make several simplifying assumptions. For instance, Anand et al. (2017) are the first to explicitly incorpo-

rate a decision context, after Hwang et al. (1993) made an initial foray. This decision making avenue will

allow for a substantially stronger connection to the large body of existing analytical work on the use of

cost information in decision making. However, cost systems exist in a dynamic world, with multiple

30
sources of information. As noted earlier, while the financial accounting system provides totals by re-

sources, a cost accounting system provides totals by cost objects. Moreover, a rational decision maker

will expect the two to match. Integrating this view of information consistency is likely an important step

in the evolution toward multi-period models.

Incorporating decisions in the context of capacity planning means that the researcher must ad-

dress the timing differences between when a firm purchases capacity and when the firm consumes it

(Reichelstein and Rohlfing-Bastian 2015). Current work either assumes fixed capacity (absent a decision

context) or assumes that we can acquire capacity on an as-needed basis. However, we know that cost

allocation problems arise because firms commit to expenditures on capacity resources in anticipation of

demand. Anand et al. (2017)’s analysis assumes acquisition of capacity occurs on an as-needed basis. The

next major challenge on this topic is to define consistency and examine how we might identify such infor-

mationally consistent systems when firms buy capacity ahead of demand. Indeed, dealing with the re-

sulting imbalance between the supply and the demand for capacity is the source of tension in short-term

planning decisions (Balakrishnan et al. 2012). Dealing with this timing difference therefore requires that

the researcher specify the solution to the short-run problem. In settings with excess capacity, how should

a firm identify and deal with the cost of unused capacity? In addition, in settings with excess demand, how

should the firm deal with the shortfall? Here, one can exploit a spot market for resources, with the pre-

mium determining the desirability of obtaining additional capacity. An extreme is to assume that the spot

market premium is so low that it is always worthwhile to fill realized demand; the other extreme is to

assume that the premium is so high that it is never worthwhile to augment capacity in the short-run.

The issue of capacity life arises when we consider issues in capacity planning. In order to connect

numerical experiments to analytical modeling work (e.g. Banker and Hughes 1994) , researchers will need

to specify how available capacity diminishes over time, and the periodic expensing of the cost for the

31
diminution in resource capacity. On the one hand, we can visualize this as a problem that deals with re-

peated demand draws and thereby lending meaning to expected value computations. Such an approach

becomes particularly attractive if we can also decouple the periods. Such decoupling can occur by, for

example, assuming that spot market premiums are such that it is worthwhile to fill all realized demand.

On the other hand, we know that many capacity resources impose “hard” constraints. How this feature

impinges on the optimal capacity to acquire and the impact of errors on the planning decision is an inter-

esting area of enquiry. Insights from the Theory of Constraints might come into play in such settings.

While the above discussion pertains to a single resource i, firms usually acquire several resources, with

differing procurement, life and usage characteristics. We are not aware of papers that explicitly model

these differences either and consider how they influence decisions regarding how much of what resource

to acquire. Furthermore, future research might consider both scale and scope economies in acquisition,

which result in decreased marginal costs.

Once we move to a setting with excess capacity, a natural question arises: Are some costing sys-

tems more appropriate to measure unused capacity? The bulk of existing numerical experiment research

has focused on two-stage costing systems. 26 Propoents of time-driven costing (e.g. Kaplan and Anderson

2004) make the argument that this one-stage method is able to identify unused capacity more accurately

than other methods. However, Cardinaels and Labro (2008) report laboratory experiment evidence that

puts this claim into question. The answer to the research question is likely found when couching the cost

system choice within a decision-making setting with long-lived capacity.

Along with capacity, products too have limited lives. We can model product life by assuming that

a certain percentage of the current product mix is not available in next period’s opportunity set, and to

be replaced by a new set of products, with different production characteristics. The interaction between

26
Hoozee and Hansen (forthcoming) which compares the two-stage Activity-Based Cost system with the one-stage
Time-Driven Cost system is an exception.

32
product and capacity life is of particular interest. Full cost pricing may be optimal when product lives

match or exceed the anticipated lives of capacity resources. However, marginal cost based or tactical pric-

ing might be optimal when capacity resources last longer than products. Understanding this tension might

lead to testable empirical propositions about pricing practices in differing industries.

Extant research has suppressed the impact of taxes, or variation in taxes across jurisdictions. How-

ever, the production- and capacity planning of multinational firms as well as their cost system design

choices will be strongly influenced by taxes they pay for their inputs, allowable tax depreciation schedules

for their acquired assets, and the income taxes of the countries where they sell their products. Given the

interdependencies of these decisions, it would be interesting to model the impact of taxes on each of

building blocks we described. 27 This interaction emphasizes the tradeoff between the gains from strategic

allocation of costs to reduce taxes versus the gains from more accurate allocations that improve decisions

but do not minimize taxes owed.

Most of the capacity acquisition and allocation literature (as relevant to the issue of limited infor-

mation) has not included strategic game playing behavior or incentive provision in an agency context, bar

some notable exceptions. Christensen and Demski (2003) show how strategic influences can distort the

demand for accurate costs. Hemmer (1996) shows how, in the presence of capacity costs in a joint pro-

duction setting, the optimal cost allocation method to deal with incentives is based on the joint products’

estimated net realizable values. Baldenius et al. (2016) develop accrual accounting-based performance

measures when investment decisions are delegated to a manager with unknown time preferences. Since

the role of information in agency settings is different from that in decision making settings, it would be

interesting to see how results on cost system design change in this setting.

Thus far, we have considered cost systems that are static and unchanging. However, we know that

firms operate in environments that are constantly changing. Examining the robustness of a cost system to

27
We thank an anonymous referee for this suggestion.

33
exogenous change seems an interesting avenue to explore (Anand et al. 2014). A more subtle issue stems

from the nature of the design influencing the extent of change in the environment. Such endogeneity

arises when, for example, the superior insight provided by an activity-based cost system (relative to a one-

pool system) also leads to greater changes in the production technology.

Concluding Comments
Our goal for this paper is to assist researchers interested in employing numerical methods to ex-

plore issues in the design of cost systems. Thus, we begin with a listing of the elements of the problem

and the necessary choices on each front. We next provide thoughts on structuring the experiment. While

it is easy to overlook, we emphasize how one might best organize and analyze the mountains of data

usually generated.

We end by offering some caveats and lessons learned the hard way. It is easy to use the provided

code to generate data around a loosely structured research question. It is even easier to spend time ana-

lyzing vast amounts of data. However, the bulk of the work needs to happen before writing code. Successful

research in this stream, like the work in all other branches, requires clarity of thought about the research

question. The researcher then needs to think carefully about the consistency of information available to

the decision maker. This step is particularly important if we are evaluating competing methods for com-

puting costs. Else, we can give one method an edge by inadvertently endowing the implied decision maker

with more information. Likewise, we foresee research that embeds a rational decision maker acting on

the information provided by the cost system. In such instances, we need to be sure that the decision

maker optimally employs all available information and that the information sets are consistent with each

other. This forethought is required so that we can design the elements with just the right amount of com-

plexity to answer the question, without dragging the project into uncharted and unwanted territory

(Balakrishnan and Penno 2014; Labro 2015).

34
FIGURES AND TABLES
Figure 1: Overview of Problem

Bolded items represent choices available in the online code. Unbolded items are of interest but have not
been considered extensively in the literature.

Panel A: Structural (“real”) factors

Production
Input Resources Outputs
Technology
• Number • Leontief technology • Number
• Cost per unit • Resource requirements • Unit price
• Amount Purchased • Density of use • Quantity
• Amount used • Correlation in consump- • True costs
tion
• Useful life • Market structure
• Starting capacity • Economies of scale • Demand uncertainty
• Premium for spot pur- • Substitutable resources • Correlation in demand
chase • Inventories
• Market conditions

Panel B: Choices made when constructing a two-stage cost system

Resource Cost Stage 1 Stage 2


Activity Cost Pools Cost Objects
Pools
• Number of resources • Number (ACP) • Reported cost (PC)
(RCP) • Methods for grouping • Error in reported costs
resources (PACP) (EUCD, MAPE, etc.)
• Methods for choosing
drivers (PDR)

35
Figure 2: Sample Resource Consumption Pattern (RES_CONS_PAT) Matrices Illustrating Variations in
Correlation and Density

In our code, the resource consumption pattern matrix (𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷) models the firm’s production
technology. We assume a Leontief technology, i.e. a fixed usage of each resource by each product, no
substutability of resources among products, and no economies of scale. Our code allows the researcher
to manipulate two attributes of the resource consumption pattern matrix, correlation and density. Panel
A shows a simplified process for generating the 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷 matrix. Panels B – D show sample ma-
trices for different values of the correlation and density parameters

Panel A: Simplified Process for Generating Resource Consumption Pattern Matrix

Following is a simplified process for generating the matrix 𝑹𝑹𝑹𝑹𝑹𝑹_𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪_𝑷𝑷𝑷𝑷𝑷𝑷. See the appendix for the
full process.

1. Generate a random, standard normally distributed vector of baseline resource 𝑋𝑋⃗ of dimension
1 × 𝑪𝑪𝑪𝑪, where 𝑪𝑪𝑪𝑪 is the number of cost objects.
2. Generate (1 − 𝑹𝑹𝑹𝑹𝑹𝑹) additional random, standard normally distributed vectors ���⃗ 𝑌𝑌2 through 𝑌𝑌 ��������⃗
𝑅𝑅𝑅𝑅𝑅𝑅 ,
where 𝑹𝑹𝑹𝑹𝑹𝑹 is the number of resources.
3. Compute vectors 𝑍𝑍 ����⃗2 through 𝑍𝑍 ���������⃗ ���⃗ ⃗ 𝟐𝟐 ��⃗
𝑅𝑅𝑅𝑅𝑅𝑅 as: 𝑍𝑍𝚤𝚤 = 𝑪𝑪𝑪𝑪𝑪𝑪 ⋅ 𝑋𝑋 + √1 − 𝑪𝑪𝑪𝑪𝑹𝑹 ⋅ 𝑌𝑌𝚤𝚤 , where 𝑪𝑪𝑪𝑪𝑪𝑪 is the user-
specified correlation.
4. For the vectors 𝑋𝑋⃗ and ����⃗𝑍𝑍2 through 𝑍𝑍 ���������⃗
𝑅𝑅𝑅𝑅𝑅𝑅 ,
a. Take the absolute value of all elements.
b. Multiply all elements by 10.
c. Replace all elements with the nearest integer that is greater than or equal to the ele-
ment (the Ceiling function).
5. For each vector ����⃗
𝑍𝑍2 through 𝑍𝑍 ���������⃗
𝑅𝑅𝑅𝑅𝑅𝑅 :
a. Loop over the elements.
b. For each element, generate a random number 𝑟𝑟𝑟𝑟𝑟𝑟 ∼ 𝑈𝑈[0.0,1.0]. If 𝑟𝑟𝑟𝑟𝑟𝑟 < 𝑫𝑫𝑫𝑫𝑫𝑫, where
𝑫𝑫𝑫𝑫𝑫𝑫 is the user-specified density parameter, keep the element as is. Otherwise, replace
the element with 0.0.
6. Form the matrix 𝑅𝑅𝑅𝑅𝑅𝑅_𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝑃𝑃𝑃𝑃𝑃𝑃 as the concatenation of the row vectors �𝑋𝑋⃗; ���⃗ ��������⃗
𝑌𝑌2 ; … ; 𝑌𝑌𝑅𝑅𝑅𝑅𝑅𝑅 �
7. Check that all elements of the first row are nonzero. Check that there are no rows with all zeros,
and no columns with all zeros. If any of these conditions are met, discard the matrix and repeat
the process.

Panel B: Sample Resource Consumption Pattern Matrices

The following 𝑅𝑅𝑅𝑅𝑅𝑅_𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝑃𝑃𝑃𝑃𝑃𝑃 matrices with 5 resources and 3 products were generated using the pro-
cedure above. Each element specifies the number of units of resource i consumed by product j. For ex-
ample, in each matrix, product 1 uses 19 units of resource 1.
19 12 7 19 12 7 19 12 7
⎡18 15 3⎤ ⎡ 2 11 13⎤ ⎡18 15 0⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢20 12 9⎥ ⎢5 2 9⎥ ⎢ 0 12 0⎥
⎢20 13 7⎥ ⎢6 7 4⎥ ⎢0 0 7⎥
⎣21 9 5⎦ ⎣9 8 5⎦ ⎣0 9 5⎦

36
The first matrix uses a correlation parameter of 0.95, indicating very high positive correlation between
the rows of the matrix. The second matrix uses a correlation of 0.05, indicating almost no correlation
between the rows. Both have a density of 1, meaning that every resource is used by every product. On
the other hand, the last matrix has a density parameter of 0.4. Here, 6 of the 15 elements (40%) are
zero, indicating that these resources are not used by these products.

37
Figure 3: Example of a nested experimental design
The example covers a numerical experiment on the impact of two cost system design choices on costing
error that is generalizable to firms with different production technologies as characterized by two tech-
nology parameters.

EXAMPLE

For a set of values of production parameter 1 drawn from a distribution with specific bounds or simu-
lated at different levels do:
For a set of values of production parameter 2 drawn from a distribution with specific bounds or
simulated at different levels do:
Generate a true benchmark firm
For a set of values of costing design parameter 1 simulated at different levels do:
For a set of values of cost system design parameter 2 simulated at different
levels do:
- Generate a false approximation based on limited information
- Calculate the error between the true benchmark and the false approx-
imation
- Report output

38
Figure 4: Sample Output Organized as a Relational Database
This figure shows sample output from a numerical experiment, organized as a relational database. Panel
A shows sample output on the sample of firms. Each firm is assigned a unique ID (FirmID). The sample
output shows the parameters for each firm. These parameters may be randomly chosen or manipulated
at levels pre-selected by the researcher. Panel B shows sample output on the sample of cost systems.
Each cost system is identified by 2 values, FirmID and CostSysID. Thus, the researcher can easily identify
which cost system is associated with each firm. Further, the sample output illustrates how many numeri-
cal experiments loop over pre-selected levels for the independent variables. In this case, each firm has 3
cost systems that differ only in the number of activity cost pools.

Panel A: Sample Ouput for the Sample of Firms Generated in a Numerical Experiment
FirmID COR DNS RCP CO RCU1 RCU2 …
1 0.28 0.87 50 20 14.94 19.17 …
2 0.93 0.34 50 20 24.87 22.47 …
3 0.47 0.76 50 20 20.59 10.08 …
… … … … … … … …

Panel B: Sample Ouput for the Sample of Cost Systems Generated in a Numerical Experiment
FirmID CostSysID ACP PACP PDR …
1 1 1 1 0 …
1 2 4 1 0 …
1 3 10 1 0 …
2 1 1 1 0 …
2 2 4 1 0 …
2 3 10 1 0 …
3 1 1 1 0 …
3 2 4 1 0 …
3 3 10 1 0 …
… … … … … …

39
Table 1: List of Notation Used in the Paper
This table lists all variables used in the paper, and their notation. Environmental parameters are exogenously specified by the modeler and characterize the firm.
Variables with accounting value correspond to quantities in accounting systems. Choice variables are classified as exogenous (specified by the modeler, even
though a firm might make this decision in a richer context), endogenous (a choice variable; the modeler must specify a rule for making this choice), or computed
(intermediate quantities computed during the numerical experiment). Finally, cost system parameters are exogenous parameters that define a cost system.
Items in bold are included in the accompanying computer code, while those in normal font are not.

Environmental Accounting Choice variable: Exoge- Cost sys-


parameter value nous (as a design choice tem pa-
that simplifies the ex- rameter
periment), endogenous
(decision), or com-
puted?

Panel A: Resources
RCP Number of resources 
CAPit Capacity available for resource i Computed
in period t
PRE_ACAPit Capacity acquired prior to de- Endogenous
mand realization for resource i in
period t
POST_ACAPit Capacity acquired after demand Endogenous
realization for resource i in pe-
riod t
CONS_CAPit Capacity consumed for resource i Endogenous
in period t
RCOSTit(PRE_ACAPit) Cost of capacity acquired prior to
demand realization for resource i 
in period t
θit Penalty parameter for spot pur-

chases of resource i in period t
ACC_COSTit Accumulated cost on books of re-

source i at start of period t

40
Environmental Accounting Choice variable: Exoge- Cost sys-
parameter value nous (as a design choice tem pa-
that simplifies the ex- rameter
periment), endogenous
(decision), or com-
puted?
γit(CONS_CAPit) Diminution in value for resource i

in period t
RC_DISP Dispersion in resource cost Exogenous
���������⃗
𝑹𝑹𝑹𝑹𝑹𝑹 Vector of unit resource costs 

Panel B: Cost objects


CO Number of cost objects 
����⃗
𝑷𝑷𝒕𝒕 Prices for products in period t Exogenous (ideally, en-
dogenous)
��������⃗
𝑷𝑷𝑪𝑪𝑩𝑩 Vector of benchmark, or true, Computed
product costs
MARjt Margin for product j in period t  (only relevant
if price if fixed)
����⃗𝒕𝒕 , 𝜀𝜀���⃗�
𝑅𝑅𝑅𝑅𝑅𝑅_𝐷𝐷𝐷𝐷𝑀𝑀𝑗𝑗𝑗𝑗 �𝑷𝑷 𝑡𝑡
Realized demand Computed / exogenous
FILL_DEM Filled demand, or total sales vol- Endogenous / computed
ume. This is the amount of
REL_DEM satisfied by the firm.
εjt Demand shock in period t for

product j
����⃗
𝑸𝑸𝒕𝒕 Exogenous realized product de-

mand in period t.

Panel C: Production technology


RES_CONS_PAT Resource consumption pattern 
DNS Density of consumption matrix 
COR Correlation in resource consump-

tion

41
Environmental Accounting Choice variable: Exoge- Cost sys-
parameter value nous (as a design choice tem pa-
that simplifies the ex- rameter
periment), endogenous
(decision), or com-
puted?

Panel D: Cost system


PACP Method for assigning resources
 
to activity cost pools
PDR Method for choosing driver  
ACP Number of activity cost pools  
��������⃗
𝑷𝑷𝑪𝑪 𝑹𝑹 Vector of costs reported by the Computed
cost system

Panel E: Computed Intermediate Variables


����������⃗𝒕𝒕
𝑹𝑹𝑹𝑹𝑪𝑪 Vector of resource costs. Element
i is the total cost of consuming 
resource i in period t.
ACT_CONS_PAT Each row is a linear combination Computed
of rows of RES_CONS_PAT
REVt Revenue realized in period t. 
𝑰𝑰𝑰𝑰𝑽𝑽𝒋𝒋𝒋𝒋 Inventory of product j at start of Computed

period t
𝑷𝑷𝑷𝑷𝑷𝑷𝑫𝑫𝒋𝒋𝒋𝒋 Chosen production quantity of Computed

product j in period t

Panel F: Dependent Variables (measures of error)


EUCD Euclidian distance Computed
MAPE Mean absolute percent error Computed

42
Table 2: Summary of advice on researchers’ decisions in each project stage

Research project stage Researchers’ decisions Advice


Modeling components Which to include? Driven by research questions and hy-
potheses; Don’t add too many compo-
nents.
Manipulate as variables of in- Driven by research questions and hy-
terest or randomize as con- potheses; Guided by prior research on
trols? effects of components.
Parameter choices? Reflective of range of practice observa-
tions; If such information unavailable,
use theoretical range.
Experimental design How to set up experiment Nested design generating many true
structure? benchmark systems, and for each of
those “false” costing approximations.
How to organize output files? Main dependent and independent varia-
bles in main output file; Organize as re-
lational database;
Internal validity check of How to ensure code accuracy? Save ancillary and intermediate outputs
code and data for checking of code; Independent
checking by researchers; Work small ex-
amples with extreme values.
Data analysis How to deal with large sample Don’t stress statistical significance; Use
size? measures of effect size.
Reporting results How to see the forest for the Report a handful of main results; Work
trees? in a “theme”; Select results for which re-
searchers can present intuition; Don’t
focus one extreme values; Use graphs.
Should researchers focus on No – numerical experiments are inap-
magnitudes of effects? propriate to draw conclusions on magni-
tudes.
How to increase accessibility Given many readers are unfamiliar with
for readers? numerical experiments, educate in your
write-up.

43
VII. REFERENCES
Anand, V., R. Balakrishnan, and E. Labro. 2014. Updating Cost Systems. In Working Paper.
Anand, V., R. Balakrishnan, and E. Labro. 2017. Obtaining Informationally Consistent Decisions When
Computing Costs with Limited Information. Production and Operations Management 26 (2):211-
230.
Anderson, S. W., and K. L. Sedatole. 2013. Evidence on the cost hierarchy: The association between
resource consumption and production activities. Journal of Management Accounting Research
25 (1):119-141.
Babad, Y. M., and B. V. Balachandran. 1993. Cost driver optimization in activity-based costing. The
Accounting Review 68 (3):563-575.
Balachandran, B. V., R. Balakrishnan, and K. Sivaramakrishnan. 1997. On the efficiency of cost-based
decision rules for capacity planning. Accounting Review 72 (4):599-619.
Balakrishnan, R., S. Hansen, and E. Labro. 2011. Evaluating Heuristics Used When Designing Product
Costing Systems. Management Science 57 (3):520-541.
Balakrishnan, R., and M. Penno. 2014. Causality in the context of analytical models and numerical
experiments. Accounting, Organizations and Society 39 (7):531-534.
Balakrishnan, R., and K. Sivaramakrishnan. 2001. Sequential solutions to capacity planning and pricing
decisions. Contemporary Accounting Research 18 (1):1-25.
Balakrishnan, R., and K. Sivaramakrishnan. 2002. A critical overview of the use of full-cost data for
planning and pricing. Journal of Management Accounting Research 14 (1):3-31.
Balakrishnan, R., K. Sivaramakrishnan, and G. Sprinkle. 2012. Managerial accounting. Textbook.
Hoboken, NJ: John Wiley and Sons, Inc.
Balakrishnan, R., K. Sivaramakrishnan, and S. Sunder. 2004. A resource granularity framework for
estimating opportunity costs. Accounting Horizons 18 (3):197-206.
Baldenius, T., A. A. Nezlobin, and I. Vaysman. 2016. Managerial Performance Evaluation and Real
Options. The Accounting Review 91 (3):741-766.
Banker, R. D., and J. S. Hughes. 1994. Product costing and pricing. Accounting Review 69 (3):479-494.
Cardinaels, E., and E. Labro. 2008. On the Determinants of Measurement Error in Time-Driven Costing.
The Accounting Review 83 (3):735-756.
Christensen, J., and J. S. Demski. 1997. Product Costing in the Presence of Endogenous Subcost
Functions. Review of Accounting Studies (2):65-87.
———. 2003. Factor Choice Distortion Under Cost Based Reimbursement. Journal of Management
Accounting Research 15:145-160.
Coller, and Collini. 2015. The optimality of full-cost pricing: a simulation analysis of the price-adjustment
dynamics. Journal of Management Control.
Cooper, R., and R. S. Kaplan. 1988. Measure Costs Right - Make the Right Decisions. Harvard Business
Review 66 (5):96-103.
Datar, S. M., and M. Gupta. 1994. Aggregation, specification, and measurement errors in product
costing. The Accounting Review 69 (4):567-591.
Dougherty, C. 2009. Porsche Finds Fortune From Unlikely Outsourcing. The New York Times, April 3,
2009.
Drury, C., and M. Tayles. 2005. Explicating the design of overhead absorption procedures in UK
organizations. The British Accounting Review 37 (1):47-84.
Dutta, S., and S. Reichelstein. 2002. Controlling investment decisions: Depreciation-and capital charges.
Review of Accounting Studies 7 (2-3):253-281.
Göx, R. F. 2002. Capacity Planning and Pricing under Uncertainty. Journal of Management Accounting
Research 14 (1):59-78.

44
Harrison, J. R., Z. Lin, G. R. Carroll, and K. M. Carley. 2007. Simulation modeling in organizational and
management research. Academy of Management Review 32 (4):1229-1245.
Hemmer, T. 1996. Allocations of Sunk Capacity Costs and Joint Costs in a Linear Principal-Agent Model.
The Accounting Review 71 (3):419-432.
Hocke, S., M. Meyer, and I. Lorscheid. 2015. Improving simulation model analysis and communication
via design of experiment principles: An example from the simulation-based design of cost
accounting systems. Journal of Management Control 26 (2):131-155.
Homburg, C. 2001. A Note on Optimal Cost Driver Selection in ABC. Management Accounting Research 1
(12):197-205.
———. 2004. Improving Activity-Based Costing Heuristics by Higher-Level Cost Drivers. European Journal
of Operational Research 157:332-343.
Hoozee, S. M., and S. Hansen. forthcoming. A Comparison of Activity-based Costing and Time-driven
Activity-based Costing. Journal of Management Accounting Research.
Hoozee, S. M., L. Vermeire, and W. Bruggeman. 2012. The impact of refinement on the accuracy of time-
driven ABC. Accounting Horizons 48 (4):439-472.
Hwang, Y., J. H. Evans III, and V. G. Hegde. 1993. Product cost bias and selection of an allocation base.
Journal of Management Accounting Research 5:213-242.
Kaplan, R. S., and R. S. Anderson. 2004. Time-driven activity-based costing. Harvard Business Review 82
(11):131–138.
Kaplan, R. S., and S. R. Anderson. 2007. Time-driven activity-based costing : a simpler and more powerful
path to higher profits. Boston: Harvard Business School Press.
Kaplan, R. S., and G. L. Thompson. 1971. Overhead allocation via mathematical programming models.
The Accounting Review 46 (2):352-364.
Labro, E. 2015. Using Simulation Methods in Accounting Research. Journal of Management Control 26
(2-3):99-104.
Labro, E., and M. Vanhoucke. 2007. A simulation analysis of interactions among errors in costing
systems. Accounting Review 82 (4):939-962.
Labro, E., and M. Vanhoucke. 2008. Diversity in Resource Consumption Patterns and Robustness of
Costing Systems to Errors. Management Science 54 (10):1715-1730.
Noreen, E. 1991. Conditions under which Activity-Based Cost Systems Provide Accurate Costs. Journal of
Management Accounting Research 91 (3):159-168.
Noreen, E., and D. Burgstahler. 1997. Full Cost Pricing and the Illusion of Satisficing. Journal of
Management Accounting Research 9:239-263.
Reichelstein, S., and A. Rohlfing-Bastian. 2015. Levelized Product Cost: Concept and Decision Relevance.
The Accounting Review 90 (4):1653-1682.
Runkel, P. 2012. Large Samples: Too Much of a Good Thing? In The Minitab Blog.
Shields, M. D. 1995. An Empirical Analysis of Firms' Implementation Experiences with Activity-Based
Costing. Journal of Management Accounting Research 7:148-166.
Shim, E., and E. F. Sudit. 1995. How Manufacturers Price Products. Management Accounting 76 (8):37-
39.
Van Mieghem, J. A. 2003. Commissioned Paper: Capacity Management, Investment, and Hedging:
Review and Recent Developments. Manufacturing & Service Operations Management 5 (4):269-
302.

45

You might also like