Rdad 005
Rdad 005
1093/restud/rdad005
© The Author(s) 2023. Published by Oxford University Press on behalf of The Review of Economic Studies Limited.
Advance access publication 25 January 2023
First version received May 2016; Editorial decision November 2022; Accepted January 2023 (Eds.)
Multiproduct firms constitute a considerable share of firms and account for an even greater share of
production. Nevertheless, the vast majority of production function estimates are based on the assumption
that firms are single-product manufacturers. This assumption is due to a lack of data on how firms allocate
their inputs across their various product lines. I provide a strategy for estimating product-specific input
allocations and production functions of multiproduct firms involved in monopolistic competition. The
strategy is based on using firms’ product prices and output demand in solving for product-level input
allocations.
Key words: Multiproduct firm, Production function, Productivity
1. INTRODUCTION
Firms’ production functions are estimated in order to identify important values like marginal
costs, returns to scale, and productivity differences between firms. These attributes are essential
in evaluating how changes in, for example, competition and regulation affect market outcomes.
Production function estimates are typically based on the often implicit assumption that the firm
produces all of its output with a firm-level production technology that is independent of the
firm’s product set and that the firm is equally productive in manufacturing all of its products.
If a firm produces a range of products, however, it is possible that the firm uses multiple pro-
duction technologies, in addition to facing distinct demand for these products. Recognising this
is important because a substantial share of firms are multiproduct firms, and an even greater
share of products are provided by these firms. For example, in the US manufacturing sector
from 1987 to 1997, 39% of firms manufactured more than one product, and these multiprod-
uct firms accounted for 87% of the sector’s output (Bernard et al., 2010). In international trade,
multiproduct firms are even more prominent: they accounted for more than 99% of US exports
in 2000 (Bernard et al., 2007). Moreover, empirical findings in the international trade literature
suggest that multiproduct firms have productivity differences across product lines.1 The reason
for assuming a firm-level production technology that is independent of the product set is simply
1. Reductions in barriers to international trade are shown to raise firms’ productivity and reduce their product
scopes, i.e. the number of products produced, presumably because firms drop the least productive products from their
selections of exported products (e.g. Baldwin et al., 2005; Bernard et al., 2011; Mayer et al., 2014).
3315
3316 REVIEW OF ECONOMIC STUDIES
pragmatic: in a typical data set of a cross-section of firms, the input allocation across the firm’s
product lines is unobservable.
In this paper, I present a strategy for estimating product-level production functions of firms
that are multiproduct producers and face monopolistic competition in the output market. Pro-
duction functions may vary across products, and there may be productivity differences across
products within firms. As usual, inputs are observed only at the firm or establishment level. The
challenge is to solve for the product-level inputs, which are functions of, among other factors, the
unobservable productivity levels. The key to the estimation strategy is that for firms that are not
2. Other data sets reporting outputs and revenues at the product-firm or product-plant levels are available from,
for example, Belgium (Statistics on the Production of Manufactured Goods by Statistics Belgium), Colombia (Annual
Manufacturers Survey by the National Administrative Department of Statistics of Colombia), India (Prowess Data by
the Centre for Monitoring the Indian Economy), and the United States (Longitudinal Research Database by the U.S.
Census Bureau).
3. In this paper, I estimate Cobb–Douglas production functions but, as shown in Supplementary Material,
Appendix A, the identification strategy is also compatible with the translog production function.
4. The demand function estimated in this paper is isoelastic, but also other functional forms can be used, as
shown in Supplementary Material, Appendix A.
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3317
The first assumption I make in order to solve for the product-level input allocation concerns
economies of scope and flexible manufacturing. Economies of scope are obtained when simulta-
neous manufacturing of multiple products results in lower fixed or variable costs of production.
They are a potential reason for firms to produce multiple products (e.g. Panzar, 1989). An alter-
native supply-side explanation for the existence of multiproduct firms is that firms can add new
products to their product assortments without making large investments in production technol-
ogy, even if the product-specific variable costs of the new products are higher. This is referred
to as flexible manufacturing, and it is assumed in several theoretical studies concerning multi-
5. For example, Eckel and Neary (2010) and Mayer et al. (2014) assume economies of scope in the form of lower
fixed costs, whereas Qiu and Zhou (2013) and Flach and Irlacher (2018) do not.
6. “Core competency” has an alternative definition in the multiproduct firm literature: it is the most profitable
activity of a firm (Bernard et al., 2011).
7. There are empirical findings consistent with firms having core competencies in producing some of their
products (Arkolakis et al., 2021; Dhyne et al., 2021).
8. However, the estimation strategy may be adjusted to allow for the number of products to have an effect on the
unobservable product-specific productivity level, as in De Loecker et al. (2016).
9. In other words, the firm’s production correspondence is additively separable in the production functions for
each product.
10. An additional problem arises when firm-level revenue production functions of multiproduct firms are esti-
mated. As the composite output (i.e. the product-level outputs aggregated to the firm level) depends on the products’
relative prices, so do the elasticity estimates of the firm-level function, even when controlled for the endogeneity between
product prices and input choices (Mundlak, 1963).
3318 REVIEW OF ECONOMIC STUDIES
output quantities have to be observable and, for any given product, comparable across manufac-
turers. This data requirement is a matter of how heterogeneous the products in the industry are
and, eventually, how detailed the product classification is.
Another data requirement relates to dealing with product selection for which the production
function estimation strategy does not account. It is well known in the literature that firm produc-
tivity and capital are correlated among the observations some distance above the market entry
threshold, which may give rise to selection bias (e.g. Wedervang, 1965). This kind of selection
may take place also at the product level. To avoid selection bias, one needs a data set with a low
using the estimated input allocations. Orr finds that plants in machinery manufacturing have
sizeable within-plant variation in efficiency.
The estimation strategy of the present study differs from the strategies discussed above
because it enables estimation of non-joint and product-specific production functions with pro-
ductivity differences across products, while the unobservable product-level inputs are solved
accordingly. This requires, in contrast to the methodologies discussed above, simultaneous
estimation of product-level input allocations and production functions.
The model and the estimation strategy are presented in Sections 2 and 3. In Section 4, I intro-
2. THE MODEL
The model consists of product-specific production and demand functions and of assumptions
about the timing of production decisions. Firm-level production functions are typically estimated
without considering demand for the products. Estimation of product-specific production func-
tions, however, requires solving for the unobservable product-level input allocations, which in
turn requires the econometrician to specify the output demand functions.11 In demonstrating the
estimation strategy, I assume Cobb–Douglas production and isoelastic demand functions but,
as shown in Supplementary Material, Appendix A, the estimation strategy is compatible also
with other functional forms, including demand functions that allow for substitution and com-
plementarity patterns between other products and output of other producers. The identifying
assumptions are discussed in more detail in Section 6.1.
2.1. Production
In an industry, a group of firms produces N similar or related products i, i ∈ {1, . . . , N }. Firm
j produces at least one of these products, Oi jt = 1 denotes that firm j produces title i at time
t and Oi jt = 0 otherwise, and firm j does not make products in other industries. The physical
output of product i that firm j produces at time t is denoted by Q i jt . The firm uses three inputs
in making product i: materials Mi jt , labour L i jt , and capital K i jt .12 Labour and capital are sub-
stitutable across the product lines of the firm.13 All the factors of production are continuously
divisible and exclusive across product lines. This means that they can be flexibly allocated across
the different lines and that any given share of a firm-level input is used in only one product line.
Outputs or by-products of other product lines are not used as factors of production. All the fac-
tors other than Mi jt , L i jt , and K i jt that affect the firm’s production volume of product i at time
t are comprised in productivity, ωi jt . Productivity may represent (product line-specific) factors
11. De Loecker (2011) estimates a demand system together with production functions to control for unobserved
product prices and demand shocks. In a single-product setting, Doraszelski and Jaumandreu (2013) allow for imperfect
competition in the output market, but they need to specify only the own-price elasticity of demand.
12. Materials Mi jt may be an aggregate of a variety of materials. Quantities of different types of materials add up
to Mi jt with weights. The econometrician does not need to observe these weights provided that the model assumptions
hold and that the quantities of and expenditures on the different material types are observable, as discussed in Section 4.1
13. As discussed in Section 2.3, firm-level L jt and K jt are chosen at t − 1 and they are predetermined at the time
of production. However, if the econometrician rather assumes L jt to be flexible, i.e. to be chosen at t, and the require-
ments for making this assumption are fulfilled as discussed in Section 6.1, the assumption of labour substitutability
across product lines can be relaxed.
3320 REVIEW OF ECONOMIC STUDIES
such as management and organisation of production, downtime due to, for example, maintenance
work, and defect rates in the manufacturing process (Ackerberg et al., 2015). Allowing for pro-
ductivity differences across product lines within firms is in line with the concept of flexible
manufacturing, as discussed in the Introduction. The production technology is product-specific
and independent of production of other products; that is, the firm’s production correspondence
is additively separable in the production functions for each product. This implies that there are
no economies of scope in the form of lower variable costs.14 The production function takes the
form of Cobb–Douglas, which is a special case of the translog production function considered
Productivity ωi jt follows an exogenous first-order Markov process. It can be divided into a con-
ditional expectation of productivity, which is a function of the productivity attained at t − 1, and
a deviation from the conditional expectation of productivity; that is, a productivity shock ξi jt :15
ωi jt = E ωi jt | ωi jt−1 + ξi jt . (2)
The productivity shock ξi jt is mean zero and independent across products, firms, and time.
2.2. Demand
Competition in the output market is monopolistic, that is, price changes of other products or by
other producers have only a negligible effect on the demand for product i of firm j. Firm j faces
a downward-sloping and isoelastic demand curve for product i at time t. The demand for and
price of product i of firm j at time t are denoted by Q i jt and Pi jt , respectively. The demand
function is written as
η
Q i jt = exp αi j Pi jti exp(εi jt ). (3)
The demand for product i of firm j depends on factors unobservable to the econometrician and
denoted by αi j . These factors vary across products and firms but are constant over time. Any
shocks to the product- and firm-specific demand levels are captured by εi jt . These shocks can be
caused by changes in buyers’ preferences or income or by the number of buyers in the market,
for example. The demand shock εi jt is mean zero and independent of any information at time
t − 1. The price elasticity of demand ηi is product-specific. As firm j faces its own demand
curve for product i, it operates like a monopolist in the elastic part of the demand function.
Hence, to ensure positive input allocations to product line i, ηi is assumed to be lower than −1.
As competition is monopolistic, Pi jt is determined as a function of Q i jt , ηi , αi jt , and εi jt , as
discussed in Section 2.3.
14. However, it is possible to allow for economies of scope in the form of lower variable costs by modelling
productivity ωi jt as a function of the producer’s product scope, as De Loecker et al. (2016) do.
15. Defining the structure of the productivity process is essential for structural methods of production function
estimation, such as Olley and Pakes (1996).
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3321
expects from manufacturing any product i ∈ {1, . . . , N } at time t. Product choices are not cor-
related with the productivity shocks ξi jt or the demand shocks εi jt that take place only at time t.
The three types of inputs Mi jt , L i jt , and K i jt differ in how they are determined. The firm-level
labour16 L jt and capital stock K jt are predetermined at the time of production. L jt is chosen
in the previous period, at t − 1, and changing labour from L jt−1 to L jt may involve adjustment
costs. In addition, K jt is chosen at t − 1. It is determined in a dynamic process as a function
of the previous period’s capital stock K jt−1 and investment I jt−1 , as K jt = f (K jt−1 , I jt−1 ).
Both L jt and K jt depend on the (expected) prices of labour, capital, and materials. L jt and
N
max jt = Pi jt Q i jt − P jtM Mi jt
Mi jt ,L i jt ,K i jt
i=1|Oi jt =1
N
N
s.t. L i jt ≤ L jt and K i jt ≤ K jt . (4)
i=1|Oi jt =1 i=1|Oi jt =1
16. L jt is often assumed to be a flexible input. I assume L jt to be predetermined because it is more realistic
for the Finnish labour market, as discussed in Section 6.1. However, the empirical model can be adjusted to allow for
flexible labour input if the wage (level) is observable and firms are price-takers in the labour market.
17. Firm-level labour and capital are predetermined so their costs are not included in the static profit function.
3322 REVIEW OF ECONOMIC STUDIES
Substituting in the inverse demand, Pi jt = (Q i jt (exp(αi jt + εi jt ))−1 )1/ηi jt , and the production
function, the static profit maximisation problem becomes
N −(1/ηi )
max jt = exp(αi j + εi jt )
Mi jt ,L i jt ,K i jt
i=1|Oi jt =1
1/ηi +1
β β β
× exp (β0i ) Mi jtMi L i jtLi K i jtK i exp ωi jt − P jtM Mi jt
The optimisation problem yields a Lagrangian equation with two constraints. The constraints
account for not exceeding L jt and K jt in setting L i jt and K i jt for each active product line.
maximises profit, L jt and K
More precisely, given that the firm jt are always fully utilised, and
the constraints are binding as i=1,...,N ;Oi jt =1 L i jt = L jt and i=1,...,N ;Oi jt =1 K i jt = K jt . The
Lagrangian is
N
Lagr = exp(αi j + εi jt )−(1/ηi )
i=1|Oi jt =1
1/ηi +1
β β β
× exp (β0i ) Mi jtMi L i jtLi K i jtK i exp ωi jt − P jtM Mi jt
⎛ ⎞ ⎛ ⎞
N N
+ λ L jt ⎝ L jt − L i jt ⎠ + λ K jt ⎝ K jt − K i jt ⎠ . (6)
i=1|Oi jt =1 i=1|Oi jt =1
The first-order conditions of static profit maximisation for firm j, ∀ j ∈ {1, . . . , J } at time t,
∀t ∈ {1, . . . , T } are
∂ Lagr 1 −(1/ηi )
= + 1 exp(αi j + εi jt )
∂ Mi jt ηi
1/ηi +1 β Mi
β β β
× exp (β0i ) Mi jtMi L i jtLi K i jtK i exp ωi jt − P jtM
Mi jt
= 0 ∀i ∈ {1, . . . , N } | Oi jt = 1, (7)
∂ Lagr 1 −(1/ηi )
= + 1 exp(αi j + εi jt )
∂ L i jt ηi
1/ηi +1 β Li
β β β
× exp (β0i ) Mi jtMi L i jtLi K i jtK i exp ωi jt − λ L jt
L i jt
= 0 ∀i ∈ {1, . . . , N } | Oi jt = 1, (8)
∂ Lagr 1 −(1/ηi )
= + 1 exp(αi j + εi jt )
∂ K i jt ηi
1/ηi +1 βK i
β β β
× exp (β0i ) Mi jtMi L i jtLi K i jtK i exp ωi jt − λ K jt
K i jt
=0 ∀i ∈ {1, . . . , N } | Oi jt = 1, (9)
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3323
∂ Lagr N
= L jt − L i jt = 0, (10)
∂λ L jt i=1|O =1 i jt
∂ Lagr N
= K jt − K i jt = 0. (11)
∂λ K jt i=1|O =1 i jt
18. I provide solutions for the product-level input allocations for alternative functional forms of production and
demand in Supplementary Material, Appendix A.
19. The identification strategy does not require firm-level flexible inputs to be observable. But as data sets like
the one used in this paper often include firm-level expenditures on flexible inputs, they can, together with the firm-level
price of materials, be used in overidentifying moment conditions such as this one.
3324 REVIEW OF ECONOMIC STUDIES
is based on the firm’s input demand function, as opposed to the production function. Finally, in
Section 3.3, I discuss how the own-price elasticities of demand are identified.20 They are the only
parameters of the demand function that need to be estimated. The model can be estimated in one
step, as in the empirical application of this paper. Alternatively, the own-price elasticities may be
estimated first and then be used in estimating the product-level inputs and production functions.
1
ηi
+ 1 Pi jt Q i jt β K i K jt
K i jt = where i ∈ {1, . . . , N } | Oi jt = 1. (14)
i=1,...,N ;Oi jt =1 ηi + 1 Pi jt Q i jt β K i
1
These solutions22 for product-level inputs are expressed only in terms of the observable data
(Pi jt , Q i jt , P jtM , L jt , K jt ) and parameters to be estimated (ηi , β Mi , β Li , β K i ).
20. I discuss identification of the own-price elasticity of demand of alternative demand functions in Supplemen-
tary Material, Appendix A.
21. Doraszelski and Jaumandreu (2013) also define the first-order conditions for static profit maximisation. They
solve for the productivity term by parametrically inverting the intermediate input demand function thus obtained.
22. The demand function, as specified in equation (3), does not allow for substitution and complementar-
ity between different products and products of different producers. However, as shown in Supplementary Material,
Appendix A, equations (12)–(14) are also the solutions for the product-level inputs when the demand function is modified
to allow for substitution and complementarity.
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3325
After solving for the product-level inputs, the productivity shock ξi jt can be written, given
the productivity process as defined in equation (2), as
ξi jt = ωi jt − E ωi jt |ωi jt−1 , (15)
where ωi jt is inverted from the production function in equation (1); that is, ωi jt =
β β β
log(Q i jt /exp(β0i )Mi jtMi L i jtLi K i jtK i ), Mi jt , L i jt , and K i jt are as defined in equations (12)–(14), and
the lagged ωi jt−1 , Mi jt−1 , L i jt−1 , and K i jt−1 are defined accordingly. In this way, ξi jt is written
23. For a discussion on identification of gross output production functions, and especially on avoiding collinearity
problems, see Gandhi et al. (2020).
24. The discussion on identification of β Li and β K i is based on the assumption that the probability of change in
the firm’s product assortment is low and therefore product selection does not require particular attention. The potential
firm or product selection bias is discussed in Section 6.1.
25. In addition, Mi jt−1 , L i jt−1 , and K i jt−1 are solved for because they are needed for solving ωi jt−1 , which
is the explanatory variable in E[ωi jt | ωi jt−1 ].
26. Lagged output prices have previously been used as instruments by Doraszelski and Jaumandreu (2013).
27. Potential functional dependence problems in production function estimation are discussed in Doraszelski and
Jaumandreu (2013), Ackerberg et al. (2015), and Gandhi et al. (2020). Gandhi et al. (2020) use a support condition on
the regressors adapted from Newey et al. (1999) to avoid the functional dependence problem.
28. As P jtM is observed after K jt and L jt have been set, it is a source of variation for Mi jt that does not determine
K jt and L jt .
3326 REVIEW OF ECONOMIC STUDIES
the productivity shocks are defined at the product-firm level. The moment conditions29 are
E ξi jt | P jtM = E ξi jt | L jt = E ξi jt | K jt = E ξi jt | P jt−1
M
= E ξi jt | L jt−1 = E ξi jt | K jt−1 = E ξi jt | Pi jt−1 = 0
where i ∈ {1, . . . , N }. (16)
3.2. Overidentifying restriction from firm-level demand for the flexible input
In addition to estimating the production functions, I estimate the firm-level demand function
for the flexible input M jt . The firm-level M jt comprises the unobservable product-level materi-
als choices Mi jt in equation (12). Observing the firm-level M jt is not necessary for identifying
the production functions, but if M jt is observable,30 it can be used to write an overidentifying
restriction for the output elasticity of the flexible input, β Mi . In practice, firm-level expenditures
on materials are often observable in financial statement data. Given the firm-level price of mate-
rials, P jtM , a firm-level measure of materials input quantity M jt can be obtained, as discussed in
Section 4.1.
If the econometrician is confident that M jt is measured with zero measurement error, the
restriction of M jt = i=1,...,N ;Oi jt =1 Mi jt can be imposed. I require the firm-level M jt to be
measured with a multiplicative mean zero measurement error M jt . As discussed in Section 4.1,
the source of measurement error in M jt is not P jtM but the variable of firm-level expenditures on
materials in the financial statement data. The firm-level demand for the flexible input is written
as
M jt = Mi jt 1 + M jt , (17)
i=1,...,N ;Oi jt =1
where Mi jt is defined as in equation (12). The firm-level input demand function can be iden-
tified using the product-level sales Pi jt Q i jt as the instruments. Pi jt Q i jt is correlated with the
underlying input choice Mi jt for two reasons. First, the higher the product-firm-specific demand
level αi j + εi jt and hence Pi jt , the higher the profit maximising Mi jt . Second, the higher the
productivity and the resultant output Q i jt , the higher the input choice Mi jt . Since Pi jt Q i jt is
uncorrelated with the measurement error M jt , Pi jt Q i jt is a valid instrument for identifying β Mi .
One more moment condition for identifying β Mi is therefore
E M jt | Pi jt Q i jt = 0 where i ∈ {1, . . . , N }. (18)
29. These moment conditions do not identify the price elasticity of demand, ηi . The moment conditions that
identify ηi are discussed and presented in Section 3.3.
30. Observing the firm-level materials input M jt does not eliminate the need for the observable materials
price P jtM .
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3327
As the above moment condition is an overidentifying one, it can also be left out of estima-
tion. But if M jt is observable, allowing for non-zero measurement error and including the
overidentifying assumption in estimation uses the available information to its fullest extent.
31. For a discussion of the instruments used in demand estimation, see, for example, Ackerberg et al. (2007).
32. The discussion of the identification of ηi is based on the assumption that the probability of change in the
firm’s product assortment is low, and therefore product selection does not require particular attention. Firm and product
selection are discussed in Section 6.1.
33. The output sold to the paper and paper products industry and the energy sector are by-products of wood
manufacturing. The data used in this paper have seven products that can be considered by-products: 16.10.23.03 Conif-
erous wood in chips or particles; 16.10.23.05 Non-coniferous wood in chips or particles; 16.10.41.00.10 Sawdust;
16.10.41.00.20 Woodchips; 16.10.41.00.40 Lathes, borders, etc.; 16.10.41.00.60 Bark; 16.10.41.00.80 Other wood waste
3328 REVIEW OF ECONOMIC STUDIES
year, I observe the output measured in a physical unit, Q i jt , as well as the sales revenue. These
two factors yield the average price of the product in a given year, Pi jt .
Similarly for the intermediate products and materials, I observe physical quantities and
expenditures by the PRODCOM titles. These data are used to obtain the “price” of materials,
P jtM , computed as the Elteto–Koves–Szulc (EKS) multilateral price index (e.g. Hill, 2004; Neary,
2004). For firm j, it can be expressed as follows (suppressing subscripts t for time):
1/J
J
PF q j , q j , p j , p j
where q j and p j are the quantity and price vectors of firm j , and PF (q j , q j , p j , p j ) is the
bilateral Fisher price index between firm j and firm j , j = 1, . . . , J (J is the number of firms),
which is given by
1/2
qj ∗ pj qj ∗ pj
PF q j , q j , p j , p j = × ,
q ∗p
j j q j ∗ p j
j j
where q j ∗ p j = n=1,...,N qi pi (N is the number of intermediate product and material titles);
similarly for PF (q j , qb , p j , pb ), where b stands for the base firm chosen. The EKS multilateral
index satisfies the circularity (transitivity) requirement, which implies that the same index is
obtained irrespective of whether firms are compared with each other directly or through their
relationships with other firms (Hill, 2004; Neary, 2004). The EKS multilateral index is thus well
suited for my purpose of comparing firms when no representative firm exists and bundles of
products differ between firms.
In addition to the physical quantities of and expenditures on the intermediate products and
materials by PRODCOM title, I observe plant-level intermediate product and material purchase
values in the LDPM, which I denote by M jtL D P M . To obtain a plant-level measure of intermediate
product and material use that does not comprise input price differences across plants, M jt , I
deflate M jtL D P M by the price index for intermediate products and materials, P jtM . While P jtM is
assumed to be measured without error, M jtL D P M and hence M jt are allowed to be measured with
error, as discussed in Section 3.2.
The labour input is measured in labour costs, which comprise salary and social payments.
The capital stock is estimated using the perpetual inventory method, K jt = (1 − δ)K jt−1 +
I jt−1 , where δ = 0.1 and I jt is investment.34
The estimation strategy poses certain requirements on the estimation sample. First, obser-
vations with missing variables cannot be used in estimation. Second, each product has to be
observed in at least four pairs of observations, with each observation pair being from two
consecutive years in a given plant.35 The four observation pairs may be from, for example, five
(excluding sawdust, woodchips, bark, lathes, borders, pellets, briquettes, etc.). The data show that about two thirds of the
wood manufacturing plants have either 0% or 100% sales in these by-products, which suggests that shared inputs in the
production of “main products” and by-products, which are assumed away in the estimation strategy, are not important.
34. I have set δ = 0.1 following the examples of, among others, Levinsohn and Petrin (2003), who estimate
production functions of Chilean manufacturing firms, and Doraszelski and Jaumandreu (2013), who estimate production
functions of Spanish manufacturing firms.
35. The requirement of at least four observation pairs can cut down the data set considerably if there are many
products that only very few plants or firms produce. For most industries, however, the aforementioned requirement is
not likely to affect the sample considerably.
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3329
TABLE 1
Product titles
PRODCOM Title
16.10.10.33 Coniferous wood; sawn or chipped lengthwise, sliced or peeled, of a thickness >6 mm,
end-jointed, sanded or planed
16.10.10.33.10 Spruce wood (Picea abies Karst.), sanded or planed, end-jointed, sawn or chipped
lengthwise, sliced or peeled, of a thickness >6 mm
16.10.10.33.20 Pine wood (Pinus sylvestris L.), sanded or planed, end-jointed, sawn or chipped lengthwise,
sliced or peeled, of a thickness >6 mm
consecutive years in a single plant, or two consecutive years in four plants. This is because,
for each product, there are four parameters to be estimated by solving the non-linear optimisa-
tion problem,36 and because estimating the first-order Markov process of productivity evolution
requires sequences of at least two observations. To be included in the estimation sample, a plant-
level observation has to contain at least one product that the plant also produced in the previous
period or in the following period. Observations that do not fulfil the aforementioned criteria are
dropped from the sample.
Recall that measurement error in output is assumed to be zero. Unfortunately, there is no
other output variable that could be used to verify the accuracy of the product-level sales rev-
enue variables. The only other output variable available is the plant-level gross output value
reported in the LDPM. Gross output value is defined as the sum of sales revenue, deliveries
to other plants of the firm, changes in inventories, production for own use, and other business
revenue, deducting capital gains and acquisition of merchandise. Not surprisingly, gross out-
put value is not equal to the sum of product-level sales revenues from production in all of the
plants. There are several possible explanations for this. Plants may produce output that is not
included in the sales revenue from production (due deliveries to other plants of the firm, positive
changes in inventories, production for own use), or the sales revenue data may include output
produced in some previous year (due to negative changes in inventories). Moreover, because
capital gains and acquisition of merchandise are deducted from gross output value, it is not pos-
sible to make strong inferences about potential measurement error in output. Unfortunately, the
various components of gross output value are not reported in the LDPM, and hence I cannot
identify why gross output value may differ from sales revenue. However, to reduce the likelihood
of using observations with major measurement error in output, I use only those observations for
which the ratio of the sum of sales revenue to gross output value is at least 0.6 but not more
than 1.4.
In practice, the aforementioned requirements on data restrict the estimation sample as fol-
lows. I start with 3660 product-plant-year observations (exactly 1200 plant-year observations)
and sixty-six products once the observations with any missing variables have been excluded. I
first drop products with fewer than four observations pairs (step 1). I then drop (1) those plant-
year observations for which at least one product has been dropped in step 1, as well as (2)
plant-year observations for which the ratio of the sum of sales revenue (reported in the Indus-
trial Output data) to gross output value (reported in the LDPM) is less than 0.6 or more than 1.4
(step 2). I repeat these two steps sequentially until no observations drop out. Repeating steps 1
36. The rest of the parameters are concentrated out of the non-linear optimisation problem, as discussed in
Section 4.3.
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3331
TABLE 2
Descriptive statistics
Percentiles
and 2 is required because executing step 2 may result in some of the remaining products having
less than four observation pairs. Similarly, dropping products brings about plant-year observa-
tions for which step 2 has to be carried out again. After repeating these two steps, there are 2936
product-plant-year observations (904 plant-year observations) and forty-two products left. This
estimation sample comes from 190 plants during an eight-year time period.
In this estimation sample, 673 out of 904 plant-year observations have a preceding observa-
tion. These 673 plant-year observations have altogether 2280 product-plant-year observations,
of which 2053 are of products that the plant also produced in the previous year. This means that
the product entry rate, excluding plants entering the data, is 10%. The product exit rate, exclud-
ing plants exiting the data, is 7%. Completely changing the product selection is very uncommon:
there is only one plant-year observation with a totally different product selection in year t than
in t − 1, with only one product produced in each year. Plants’ product assortments range from
one to 17 products, with 659 out of 904 plant-year observations containing at least two product-
level observations. Product assortments vary across plants such that there are no typical product
combinations. The correlation between making two products is low for most product pairs: the
absolute value of the correlation coefficient is lower than 0.05 (0.1) [0.2] for 63% (81%) [91%]
of all product pairs.
The estimation sample is summarised in Table 2. The descriptive statistics show that the
plants vary in size, as measured by their sales and by their input use (capital, labour, and materi-
als, reported in hundreds of thousands of euros in Table 2). For example, the fifth percentile plant
has sales of 240,000 euros in a year, while the 95th percentile plant has sales of 8 million euros.
The price of materials, which is measured by the EKS multilateral price index, also varies across
plants. For the fifth percentile plant, the materials price index is 0.42, while for the 95th per-
centile plant the index is 2.38. A median plant produces three products, and the 95th percentile
plant produces seven products. In Table 2, the product-specific output prices are characterised as
relative prices, where the relative price is the ratio of the price to the product-year-specific mean.
The relative output prices also have a rather wide distribution, with the fifth percentile relative
price being 0.21 and the 95th percentile being 1.73.
Output prices vary not only between plants but also to a large extent within plants. Table 3
summarises the mean differences between product prices at the plant level. In a two-product
plant, the mean difference between the relative prices is 0.69. In a three-product plant, the mean
difference is 0.35 between the highest and the second highest and 0.27 between the second
3332 REVIEW OF ECONOMIC STUDIES
TABLE 3
Within-plant price differences
Mean differences between relative prices, ordered from highest to lowest
(Standard deviation in brackets)
Number of products 1st and 2nd 2nd and 3rd 3rd and 4th 4th and 5th Other
2 0.69
(1.07)
3 0.35 0.25
highest and the lowest price. Plants with larger scopes also have within-plant price differences,
as shown in Table 3. This suggests that demand varies across the different products that a plant
produces.
37. The objective function appears smoother and the estimates are less responsive to the starting values when the
optimal instruments are used instead of the traditional instruments. This is because the functional forms are exploited to
a fuller extent.
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3333
the product-firm-specific constant αi j is estimated by OLS given the output and price data and
some value for ηi . Finally, I compute the moments.
As multiple parameters are estimated by solving the non-linear optimisation problem, I
need to make sure that the estimation routine reaches the global minimum among the local
minima of the GMM objective function. I experiment with various minimisation algorithms;
the Gauss–Newton algorithm turns out to be the least sensitive to starting values. I also run the
estimation routine with a large set of alternative starting values.40 The GMM estimator uses the
efficient weighting matrix. Since the parameter estimates are obtained in one step, the asymptotic
5. RESULTS
The estimation results are presented in Table 4. The production and demand functions are esti-
mated for the products in “Sawmilling and planing of wood” (PRODCOM 161) and “Products of
wood, cork, straw, and plaiting materials” (PRODCOM 162). All the output elasticities of inputs
and price elasticities of demand are statistically significant.41 In addition, the estimates of the
output elasticities are statistically different for the technologies in the two product groups. The
output elasticity of materials β Mi is considerably higher for PRODCOM 162 than for PROD-
COM 161 (0.73 and 0.37, respectively). The output elasticity of labour β Li is again considerably
lower for PRODCOM 162 than for PRODCOM 161 (0.13 and 0.36, respectively). The order of
these estimates is not what we may expect as the output titles in PRODCOM 162 are more
processed than the output titles in PRODCOM 161. However, a plausible explanation for these
estimates relates to the composition of the material inputs used for PRODCOM 161 and PROD-
COM 162. The materials used in manufacturing PRODOM 161 are raw in the sense that they
have barely been processed before being transported to the plant. At least some of the materi-
als and intermediate products used in manufacturing PRODCOM 162, on the other hand, are
products of other industries. In other words, some labour and capital inputs have already been
used in manufacturing the material inputs for PRODCOM 162, raising its value. As a con-
sequence, β Mi for PRODCOM 162 is also rather high. The output elasticity of capital is of
the same magnitude for PRODCOM 161 and PRODCOM 162 (β K i is 0.20 and 0.18, respec-
tively). Returns to scale are different across the products in PRODCOM 161 and PRODCOM
162: the technology for the products in PRODCOM 161 is subject to decreasing returns to
scale (β Mi + β Li + β K i = 0.93 < 1), while the technology for the titles in PRODCOM 162 has
increasing returns to scale (β Mi + β Li + β K i = 1.04 > 1). In short, the various products in the
product groups “Sawmilling and planing of wood” and “Products of wood, cork, straw, and plait-
ing materials”, which many multiproduct firms produce simultaneously, are not manufactured
with a single production technology.
The demand for the titles in PRODCOM 161 is more price elastic than the demand for the
titles in PRODCOM 162, as ηi for the titles within PRODCOM 161 is −1.29, while ηi for the
titles within PRODCOM 162 is −1.13. This is intuitive because products of wood, cork, straw,
and plaiting materials are likely to be more differentiated than the output of sawmilling and
planing of wood. Hansen’s J-test does not reject the null hypothesis of valid overidentification
restrictions (Prob[Chi-sq.(26442 )>J] is 0.4632).
40. The starting values for β Mi , β Li , and β K i range between 0.15 and 0.5, and the starting values for ηi between
−8 and −1.5.
41. The forty-two constants β0i and the parameters governing the productivity process g(ωi jt−1 ) are not
reported.
42. The number of moment conditions is 272.
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3335
TABLE 4
Parameter estimates
Parameter estimate (standard error)
6. DISCUSSION
In Section 6.1, I discuss the assumptions and data requirements underlying the identification
and compare them to the assumptions and requirements of other structural production func-
tion papers. In Section 6.2, I discuss the extent to which the empirical strategy accommodates
economies of scope and differentiated products.
6.1. Identification
Most of the assumptions underlying the identification strategy are familiar from the struc-
tural production function literature (e.g. Olley and Pakes, 1996; Levinsohn and Petrin, 2003;
Doraszelski and Jaumandreu, 2013; Ackerberg et al., 2015; Gandhi et al., 2020). I also make
some novel assumptions. One of the key assumptions in all the estimation strategies concerns
the timing of the input choices with respect to the evolution of the productivity process. As in
other structural production function models, at least one flexible (product-level) input is required
to solve the firm’s static profit maximisation problem. In this paper, firm-level labour is consid-
ered to be a predetermined input. The reason for this is the environment in which the data have
been generated: employment protection legislation plays a significant role in Finland.43 How-
ever, the empirical model may be adjusted to allow for flexible labour input if the wage (level)
is observable and the assumptions concerning materials input and the exogeneous price of mate-
rials are also plausible for labour and wage. In that case, instead of L jt and L jt−1 , wage and
lagged wage are valid instruments for identifying β Li .
I further specify that, while the firm-level L jt and K jt are predetermined, the product-level
allocations of labour and capital, L i jt and K i jt , are flexible. This assumption not only facilitates
solving for the product-level L i jt and K i jt but also allows firms to reallocate labour and capital in
response to demand and productivity shocks. Whether the assumption of zero adjustment costs
43. The OECD indicators of employment protection (OECD, 2013) measure the strictness of legislation on indi-
vidual and collective dismissals and the strictness of hiring employees on temporary contracts. The measures are based
on information about statutory and case law, collective bargaining agreements, and advice from officials from OECD
member countries and country experts. According to these indicators, the Finnish labour market was at the OECD
average in terms of the strictness of employment protection during the 2004–2011 period. Based on this measure,
predetermined labour input is a realistic assumption.
3336 REVIEW OF ECONOMIC STUDIES
holds depends on the set of technologies the firm uses and whether the firm is a single-plant firm
or observed at the plant level. While most multiproduct firms manufacture products in a single
industry, a considerable share of capital and labour is likely to be applicable to several product
lines, even if predetermined at the plant level. On the other hand, the adjustments producers
are willing to make in response to demand and productivity shocks are likely to be moderate in
magnitude. Hence, the assumption of costless reallocation is likely to be realistic to the extent
that single-plant firms, or multiplant firms observed at the plant-level, active in a single industry
are willing to make reallocations due to demand and productivity shocks.
44. In the wood products industry considered in the present study, input prices vary across plants because of
transportation costs that constitute a considerable share of the cost of wood purchases.
45. This assumption is plausible, at least, in the wood products industry. First, the most important material for
wood product plants or firms is raw wood. The majority of raw wood is purchased by the paper and paper manufacturing
industry. Thus, demand shocks for wood products are not likely to have an important effect on wood prices. Second, the
Natural Resources Institute Finland (Luke) publishes monthly statistics on wood prices. These statistics are based on
more than 90% of the volume purchased by the forest industry. This suggests that buyers of raw wood are well informed
about the current price levels of various raw wood types, leaving little room for negotiation.
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3337
markups for a given product. Identification of the isoelastic demand function of the empirical
application is based on two assumptions. First, the level of demand for a given product of a
given firm is constant over time, except for mean independent demand shocks. If appropriate,
potential industry-level changes in demand can be accounted for by estimating product-time-
specific fixed effects, as shown in Supplementary Material, Appendix A. Second, changes in
input prices and predetermined input stocks shift the supply curve, while the demand curve,
including the demand shock εi jt , is not affected. Using material prices and predetermined input
stocks as instruments is a standard practice.
46. According to the two data sets I use, the annual exit rate of plants in the wood products industry is 5.6%.
Unfortunately, the data set is not informative about whether the exits from the data are real plant closures or whether the
plants have not replied to the survey. Most of the plants are large, though, which suggests that the probability of plant
exit is low.
47. I use data on 904 plant-year observations in which 659 plants produce at least two products; the average
scope of all plants is 3.25 products.
48. De Loecker et al. (2016) estimate product-specific production functions using a panel of single-product firms,
some of which may be multiproduct firms in other periods. They estimate the production functions correcting for the
selection to become a multiproduct firm using the insights of Olley and Pakes (1996) and compare the results to those
estimated without correcting for such selection. They conclude that “the use of the unbalanced panel of single-product
firms (which includes firms that are always single-product and firms that ultimately transition to a multi-product status)
likely alleviates most of the concerns about the selection bias” (De Loecker et al., 2016, p. 483).
49. I use data where product entry rate, excluding plants entering the data, is 10%, and product exit rate, excluding
plants exiting the data, is 7%.
3338 REVIEW OF ECONOMIC STUDIES
of change in the firm’s product assortment is low. Alternatively, the model may be extended to
control for, first, firm selection and, second, selection into various product lines by computing
propensity scores for entry, as in Olley and Pakes (1996).
The production function parameters are estimated by solving a non-linear optimisation prob-
lem. This implies that the parameters may not be estimated as specific to the most detailed
product classification available, as discussed in Section 4.4. In the empirical example in the
present study, the production functions are estimated at the eight- or 10-digit level of the PROD-
COM classification, but the output elasticities of inputs are estimated as specific to the three-digit
6.2.1. Economies of scope and flexible manufacturing. Economies of scope may arise
due to lower average fixed or average variable costs of production. Average fixed costs decrease
when the producer’s fixed costs are spread over a greater number of products. For example, open-
ing a new product line does not necessarily require larger premises and thus higher rental costs.
The estimation strategy of the present study assumes away fixed costs, as usual in production
function estimation.
Economies of scope due to lower variable costs arise if, for example, the production process
of a product yields by-products that can be used as inputs for other products. Alternatively, there
VALMARI ESTIMATING PRODUCTION FUNCTIONS OF MULTIPRODUCT FIRMS 3339
can be some units of inputs that are joint between product lines, leading to a decrease in average
variable costs. These kinds of inputs cannot be solved for when only firm-level inputs are observ-
able. The estimation strategy therefore assumes away the aforementioned types of economies of
scope and is not suitable for industries where such economies of scope are plausible. However,
not allowing for economies of scope in the form of lower variable costs is in line with the theory
of flexible manufacturing (e.g. Eckel and Neary, 2010). Instead, the estimation strategy allows
firms to have productivity differences between product lines and hence core competencies in
making some products, which is an essential feature of flexible manufacturing, as discussed in
6.2.2. Product heterogeneity. The estimation strategy is suitable for industries where com-
petition is imperfect, where product differentiation is likely. Hence, the product classification
has to be detailed enough for the physical output quantities of a given product to be compara-
ble across manufacturers.50 At the same time, the data have to include a sufficient number of
observations of the product for identification to be possible. If the physical quantities of a prod-
uct cannot be compared due to unobservable quality variation or different units of output, the
estimation strategy is not suitable for that industry.
A large share of manufacturing industries make products that are comparable in physical
units even across manufacturers. Examples include food products, textiles, paper and paper
products, printing and reproduction of recorded media, chemicals and chemical products, basic
pharmaceutical products and pharmaceutical preparations, rubber and plastic products, other
non-metallic mineral products, basic metals, and fabricated metal products. Industries that do
not make products comparable in physical units are likely to manufacture products that involve
advanced technology. Certain transport equipment such as ships and boats, or air and spacecraft,
are good examples. If an econometrician wished to estimate production functions for cars, for
example, (s)he should consider whether subcompact cars such as the Citroë n C1 and Renault
Twingo, or entry-level luxury cars such as the Audi A4 and Mercedes-Benz C-Class, are com-
parable in physical units, and if not, whether the product classes can be defined more narrowly
without making the number of observations in product classes too small.
In short, product differentiation is not a challenge to identification as long as quality differ-
ences are observable, or if differentiation is horizontal and does not require different production
technologies. Ultimately, this requirement is a matter of whether the output data are sufficient
regarding the output characteristics. Many national statistics authorities provide data sets that are
likely to be sufficiently rich for several manufacturing industries. There are also several manu-
facturing industries, such as automakers, whose output is not reported in sufficient detail in these
data sets and would have to be complemented with additional output data such as car type and
characteristics.
50. In the empirical example used in the present study, unobservable quality differences or different units of
output are not admissible for a product defined at the eight- or 10-digit level of the PRODCOM classification.
3340 REVIEW OF ECONOMIC STUDIES
7. CONCLUSION
This paper contributes to the substantial empirical literature on production function estimation,
which underlies an even larger body of applied economic research. The standard assumption
made in production function estimation is that firms produce all of their output with a firm-level
production technology that is independent of the firm’s product set and that the firm is equally
productive in manufacturing all of its products. However, an empirical fact is that a remarkable
share of firms are multiproduct firms, and it is possible that they use several, product-specific
Supplementary Data
Supplementary data are available at Review of Economic Studies online.
(Statistics Finland, 2012b). These data sets are provided by the Research Services department of Statistics Finland.
Statistical legislation and data protection and confidentiality practices specified in legislation are applied in releasing
the data. Release of data is subject to a user licence. Instructions for applying for a user licence are provided here:
https://www.tilastokeskus.fi/tup/mikroaineistot/hakumenettely en.html. The replication package can be accessed here:
http://doi.org/10.5281/zenodo.7474464.
REFERENCES
ACKERBERG, D., BENKARD, C. L., BERRY, S. and PAKES, A. (2007), “Chapter 63 Econometric Tools for Analyzing