LEM

WORKING PAPER SERIES

Aggregate Productivity Growth in the Presence of

(Persistently) Heterogeneous Firms

Giovanni Dosi a
Marco Grazzi b
Le Li c
Luigi Marengo d
Simona Settepanella e

a
Institute of Economics and EMbeDS Department, Scuola Superiore Sant’Anna, Pisa, Italy.
b
Department of Economic Policy, Università Cattolica del Sacro Cuore, Milan, Italy.
c
Institute of Business Research, Chuo University, Tokyo, Japan.
d
Department of Business and Management, LUISS University, Rome, Italy.
e
Department of Mathematics, Hokkaido University, Sapporo, Japan.

2019/26 August 2019

ISSN(ONLINE) 2284-0400
 AGGREGATE PRODUCTIVITY GROWTH IN
THE PRESENCE OF (PERSISTENTLY)
HETEROGENEOUS FIRMS∗
Giovanni Dosi† , Marco Grazzi◦ , Le Li‡ , Luigi Marengo¶ , and
Simona Settepanella
†
Institute of Economics, Scuola Superiore Sant’Anna, Italy
◦ Department of Economic Policy, Università Cattolica del Sacro Cuore, Italy
‡ Institute of Business Research, Chuo University, Japan
¶ Department of Business and Management, LUISS University, Italy
 Department of Mathematics, Hokkaido University, Japan

Abstract
In this article we propose a new methodology for computing the aggregate
productivity of an industry, its variations and decompositions of the latter into
changes of individual productivities (within effect) and changes in industry compo-
sition (between effect). Current aggregate measures rely on some weighted average
of individual productivities, and decompositions distinguish between the effect of
productivities and weights on variations of the average. However such aggregate
measure is incoherent with the disaggregate one (the two are computed with dif-
ferent methodologies), is subject to aggregation biases, arbitrariness in the choice
of weights, and information loss. Such problems are particularly serious when het-
erogeneity among firms is high. We propose instead a geometric approach where
aggregate productivity can be measured directly on industry data, but nevertheless
its variations can be decomposed into between and within effects plus an hetero-
geneity effect. We show that our measure does not incur in many of the problems
of the weighted average and we also present an empirical application to European
data.

JEL codes: D24; C67; C81; O30

Keywords: Productivity measurement; Decomposition of aggregate productivity
growth; Firm heterogeneity
∗
We benefited from participants’ comments on this work (including its early version) at Comparative
Analysis of Enterprise Data (CAED) Conference 2017, International Schumpeter Society (ISS) Confer-
ence 2018, Productivity Research Network (PRN) Conference 2018, Japan Association for Evolutionary
Economics (JAFEE) conference 2019 and seminars at Jorge Tadeo Lozano University, Meijing University,
the University of Tokyo and Chuo University. Le Li is a JSPS International Research Fellow (P18759)
supported by the Japan Society for the Promotion of Science. Le Li also gratefully acknowledges the
research support by the IBIMET-CNR (grant CrisisLab-ProCoPe) in the early stage of this article. This
work has been partly supported by the European Commission under the H2020, GROWINPRO, Grant
Agreement 822781.

1
1 Introduction
The considerable increase of the availability and reliability of firm-level data in recent
years has given a much more solid empirical basis on at least two very important research
lines in industrial economics focusing, respectively, on productivity and heterogeneity.
The former has produced a progressive refinement of establishment- firm- industry- and
country-level productivity and productivity change measures which have also fed the
debate on whether we have entered a time of slow productivity growth in spite of the
apparent wave of technological change (3rd and 4th industrial revolutions). This is not
the place to discuss whether any slowdown has happened and, if it has, its causes. Rather,
let us just notice that persistent heterogeneity is a very robust stylized fact, irrespectively
of the level of statistical observation, in terms of e.g. productivity, profitability and other
measures of firm characteristics and performance.
In turn, heterogeneity makes it more problematic to produce good aggregate measures
of such characteristics and performance indicators and, most notably, of productivity. If
firms had very similar individual productivities, computing aggregate values would be
straightforward. Or, at least, if competition forces were pushing individual productivities
towards those of the most efficient firms, homogeneity would be increasing and aggrega-
tion would become less problematic in the long run.
However this is not what happens in the real world. Firms operating in the same
industry display large asymmetries in productivity, relative input intensities (also with
the same relative prices) and high intertemporal persistence of such asymmetries (e.g.
Baily et al., 1992; Baldwin and Rafiquzzaman, 1995; Bartelsman and Doms, 2000; Dis-
ney et al., 2003; Dosi, 2007; Syverson, 2011). Finally, such heterogeneity is persistent
also when defining industries more and more narrowly, as summarized by Griliches and
Mairesse (1999) “ We [...] thought that one could reduce heterogeneity by going down
from general mixtures as “total manufacturing” to something more coherent, such as
“petroleum refining” or “the manufacture of cement”. But something like Mandelbrot’s
fractal phenomenon seems to be at work here also: the observed variability-heterogeneity
does not really decline as we cut our data finer and finer. There is a sense in which
different bakeries are just as much different from each others as the steel industry is from
the machinery industry.”
It is fundamental to note that firms are not heterogeneous in terms of a scalar which
multiplies a commonly shared production function. In a deeper way, firms, and even
plants of the same firm, differ substantially in terms of relative combinations of inputs
and input/output relations.
Current measures of aggregate productivity (AP henceforth) for an industry are typ-
ically averages of the productivities of individual firms, weighted by some indicator of
the shares of individual firms in the total size of the industry, i.e. either input (typi-
cally labour) or output (total sales) shares. However, weighted average measures have
clear drawbacks. The choice of weights is somewhat arbitrary and different weights can
produce very different results. Moreover there is a kind of incoherence when measuring
individual productivities with one method (e.g. using a production function) and the
aggregate productivity with a different method (a weighted average). This incoherence
makes aggregate measure sensitive also to the level of observation of individual units: for
instance computing AP as weighted average over productivities of establishments or pro-
ductivities of firms yields, in general, very different results, though the latter are simply
aggregations of the former. It seems therefore more natural to compute AP directly on

2
aggregate data. Industry productivity could be simply computed using industry input
output data and treating the industry as a single entity.
On the other hand, weighted average measures have apparently a very important ad-
vantage over such a “direct” measure. Usually we are not so much interested in measuring
productivity per se but its variations, that is the “aggregate productivity growth” (APG
henceforth), and, especially, to understand what caused such variations: changes of in-
dividual productivities or changes in the composition of the industry? Weighted average
measures provide a straightforward answer to these question, since it is in principle easy
to decompose the variation of the weighted average into a component due to the variation
of firm productivities (the so-called “within” effect) and another component due to the
variation of weights/shares (the “between” effect). Moreover, also the contribution of
firms entering and exiting the market can be easily accounted for. On the contrary, it is
usually believed that similar decompositions into within, between, entry and exit effects
are not possible with direct measures.
In this article we challenge this view and propose a “direct” (i.e. using aggregate in-
put and output data rather than averages of firm data) measure of AP which allows for a
rigorous decomposition of its variations into within, between, and net entry effects. More-
over, we can further decompose the between effect into two parts, one which is a proper
“shares” effect which captures the changes of the individual contributions to the size of
the industry, and the other which instead captures the variations of the heterogeneity
of the combinations of input used by productive units. Thus, our methodology is par-
ticularly appropriate when heterogeneity is significant and the standard decompositions
would deliver a spurious measure of the between effect.
Our measure builds upon a representation of an industry empirical production activi-
ties set first proposed by Hildenbrand (1981) and further developed by Dosi et al. (2016)
and it is based on some geometric properties of such a set. In addition to overcoming
most of the problems encountered by weighted average measures, our methodology has
also an additional important advantage: it is based solely upon the empirical production
set and does not require to fit a production function in order to estimate total factor
productivity (TFP henceforth).
Finally, we show that our methodology can indeed be applied to real data and we
provide as an illustration and empirical application to some selected industries in France,
Italy and the UK. A comparison between the results of our methodology and a standard
one on the same empirical data set highlights the differences.
The rest of the article is organized as follows. Section 2 briefly reviews the current
methods that compute APG decompositions and discusses the problems they face. Sec-
tion 3 develops our own suggested measure and compares it to the current one by means of
some artificial toy examples. Some mathematical details are left to appendix A, whereas
the extension to the case of multiple outputs is discussed in appendix B. Section 4 presents
instead a test of our methodology against a standard one on real empirical data of some
selected industries in France, Italy and the UK. Finally, section 5 concludes.

3
2 Decomposition of APG: A brief discussion of the
state of the art
Many empirical studies1 try and measure the effect of changes of firm-level productivities
and changes of industry composition on APG.
In this section we do not want to review this large and growing body of literature,
but simply to sketch the methodology they rely upon (leaving aside many variations
and refinements) and discuss its limitations. In order to do so, we will mainly refer to
three methodologies introduced, respectively, by Baily et al. (1992), Grilliches and Regev
(1995), and Haltiwanger (1997), and to the critical review given by Foster et al. (2001).
In subsection 2.1, we summarize three most influential decomposition methods, and
in section 2.2 we discuss the potential issues with these methods.

2.1 APG and its decomposition

Foster et al. (2001) compare the major APG decomposition methods. AP at time t,
denoted by P t , is defined as the weighted average of the productivities of individual
firms: X
P t := wit pti (1)
i∈I t

where I t is the set of all firms in the industry at time t, wit is the weight/share (e.g.
output share) of firm i in this industry, and pti is the productivity of firm i. APG is then
defined as the difference of the aggregate productivities between two (consecutive) time
periods.
A decomposition method for APG was introduced by Baily et al. (1992):
X X X X X X
wit pti − wit−1 pit−1 = wit−1 ∆pti + pti ∆wit + wit pti − wit−1 pt−1
i (2)
i∈I t i∈I t−1 i∈C i∈C i∈N i∈X

where C denotes continuing firms, i.e. those active in both periods, N and X denote the
set of firms that have, respectively entered and exited the industry, and the operator ∆
represents change from time t − 1 to time t, i.e. ∆pti = pti − pt−1
i and ∆wit = wit − wit−1 .
The first term, originally called “fixed shares”, represents a within-firm component given
by each firm’s productivity change, weighted by the initial shares in the industry. The
second term, originally called “share effect”, represents a between-firm component that
reflects changing shares, weighted by the productivities of the final period. The last two
terms represent the contribution of the firms that, respectively, entered and exited the
industry (see Melitz and Polanec (2015) for a detailed discussion of the entry and exit
effects).
A second decomposition method was introduced by Grilliches and Regev (1995). It
only differs from the previous one in the weights used to compute the “within” and the
“between” effects. Rather than using either the initial or the final weights, this method
employs their averages:
X X X X X X
wit pti − wit−1 pt−1
i = w̄i ∆pti + p̄i ∆wit + wit pti − wit−1 pt−1
i (3)
i∈I t i∈I t−1 i∈C i∈C i∈N i∈X

1
Among the others, Baily et al. (1992); Olley and Pakes (1996); Bartelsman and Dhrymes (1998);
Dwyer (1998); Haltiwanger (1997) use data from US, Aw et al. (1997) from Taiwan, Liu and Tybout
(1996) from Chile and Colombia, Grilliches and Regev (1995) from Israel, and Dosi et al. (2015) from
France, Germany, and the UK.

4
 wt +wt−1 pt +pt−1
where the bar over a variable indicates the average w̄i = i 2 i and p̄i = i 2 i . In this
decomposition, the first term can be interpreted as a within effect which is measured by
the sum of productivities weighted by the average (across time) shares. The second term
represents a between effect where the changes in the shares are indexed by the average
firm-level productivities. The last two terms represent the contribution of the firm which
entered and exited the industry.2
Finally, Haltiwanger (1997) suggests a refinement of decomposition (2) where 1) the
“share effect” actually captures between effect and covariance terms, and 2) the between
and the net entry effects are weighted by the deviations of firm productivities from the
initial industry index:
X X X X X
wit pti − wit−1 pit−1 = wit−1 ∆pti + pt−1 t−1 t
∆pti ∆wit


i − P ∆w i +
i∈I t i∈I t−1 i∈C i∈C i∈C
X X (4)
wit pti t−1
wit−1 pt−1 − P t−1



+ −P − i
i∈N i∈X

where a continuing firm whose output increases and a new entry will contribute positively
to the index only if their productivities are higher than the aggregate, whereas an exiting
firm will contribute positively only if its productivity was lower than the aggregate.
Without such a deviation term the between effect may be different from zero even when
all individual productivities remain constant if the share of entering and exiting firms are
different.

2.2 Some problems with the current decomposition methods

Defining AP as a weighted average of the productivities of individual firms leaves open
the question of which weights should be used. Different weights will, in general, deliver
different aggregate values and different decompositions.
As an illustration, consider the extreme case of a hypothetical industry composed of
the two highly heterogeneous firms described in Table 1.

Table 1: A hypothetical industry with two firms and one input

Firm Labour Output Labour Productivity

A 1 100 100
B 100 1 0.01
Aggregate (Industry) 101 101 1

To compute the AP of this industry we can use either input or output shares as
weights. With the former we obtain an AP of 1, whereas with the latter we obtain a
value very close to 100, i.e. the productivity of firm A.
Alternatively, instead of computing AP as a weighted average, we could compute it
directly, by considering the industry as one large firm producing the industry total output
with the industry total input. This measure is 1 in our example, and it is equal to the
value obtained with inputs as weights because we are considering a case with only one
input. With multiple inputs this equality does not generally hold.
2
To be more precise, Grilliches and Regev (1995) treat all entering and exiting firms as one firm.

5
 In fact, when we have multiple inputs we normally aggregate them into a synthetic
measure called total factor productivity (TFP). But the calculation of TFP requires
the assumption of a specific production function and the empirical estimation of its
parameters. Consider for instance a hypothetical industry composed of three firms using
capital and labour as in Table 2.

Table 2: A hypothetical industry with three firms and two inputs

L-based K-based Q-based

L K Q TFP Weights Weights Weights

Firm 1 30 10 20 0.877 0.882 0.625 0.357

Firm 2 1 3 6 4.559 0.029 0.188 0.107
Firm 3 3 3 30 10.000 0.088 0.188 0.536
L K Q TFP w.a. T F Pi w.a. T F Pi w.a. T F Pi
Industry 34 16 56 1.989 1.786 3.285 6.161
*
w.a. stands for weighted average.
The first three columns of Table 2 report input-output data for each firm, the fourth
column gives the corresponding TFP values under the assumption that the production
function is a Cobb-Douglas with constant returns to scale:
Qi
T F Pi = 0.75 0.25
(5)
L i Ki
and the three rightmost columns give the weights computed, respectively, on labour,
capital and output. Finally the bottom row gives the aggregate values for L, K and
Q, the aggregate TFP directly computed on such values (treating the industry as one
production unit with the same production function) and the last three columns give
instead the TFP values computed as weighted averages using, respectively, L, K and Q
shares as weights.
Also in this case APs greatly vary with the chosen weights and they also vary from
AP/TFP measure directly computed on the aggregate input-output data rather than as
a weighted average of individual TFPs.
This “direct” aggregate measure presents some important advantages over weighted
averages. The first and most important feature is that productivity is computed exactly in
the same way for the individual firm and for the aggregate, whereas standard measures are
defined in radically different ways at firm (TFP) and aggregate (weighted average) levels.
As a consequence, the direct measure is invariant to changes of the level of aggregation
at which individual units are observed. Suppose for instance that some production units
merge into a single entity (for instance two firms merge legally, if our unit of analysis
are firms) or simply we change our level of observation from individual establishment to
multi-establishment firms, keeping exactly the same input-output structure. Weighted
average measures would change3 although the input-output structure of the industry has
3
To be more precise, AP will be different if computed using outputs as weights and using TFP with
multiple inputs. It would not change in the case of a single input.

6
remained exactly the same. The productivity computed directly on the aggregate would
instead remain unchanged. For instance, imagine that firms 2 and 3 of Table 2 merge
or, equivalently, that 2 and 3 are establishments of the same firm and establishment
data are no longer available forcing us to rely on firm-level data only. It is easy to show
that all APs computed as weighted averages would change and, in particular, AP would
decrease by 3% when computed with labour shares, increase by 10% with capital shares
and decrease by 10% with output share, whereas TFP directly computed on industry
input-output data would remain obviously unchanged, correctly signalling that nothing
as changed in the industry.
More in general, no matter the method adopted, when we try to summarize multiple
pieces of information, e.g. quantities of inputs and outputs, into an aggregate value,
e.g. the productivity level, we inevitably lose information. Thus, since every time we
compute a productivity value we lose information, in principle it is preferable to minimize
the number of such computations. Now, if we compute AP as a weighted average of the
productivities of n firms we will lose information n + 1 times, whereas if we compute
productivity directly on the aggregate level, we only do it once. More information is
preserved when moving from firm to industry level if we aggregate firms’ production
activities rather than their productivities.
Finally, when we compute APG, we have to make an additional arbitrary choice,
that is the weights of either the initial, or the final, or some intermediate time. Also in
this case the choice will matter and different weights may produce significantly different
decompositions.
To summarize, it seems more reasonable to define productivity for an industry exactly
as we define it for a firm, rather than computing it as an average. But then the next
question is whether without using the weighted-average we can still perform a decom-
position of AP variations into within, between and net entry components. Contrary to
what is usually believed, in the next session we show that the answer to this question is
yes, provided we use an appropriate methodology.

3 Productivity growth in firms and industries: a uni-

fied framework
In this section we propose our decomposition method of APG. We start with an em-
pirical representation of an industry as a set of heterogeneous firms first introduced by
Hildenbrand (1981) and later developed by Dosi et al. (2016). Such a representation does
not assume the existence of a production function, but nevertheless allows to compute
rigorous aggregate measures of productivity and, as we show below, decompositions of
their variation which preserve coherence when passing from individual firms to industry
aggregates. We first introduce some notation and definitions in subsection 3.1. Then, in
subsection 3.2, we propose our measure of productivity both for individual firms and for
an industry and show how the latter can be decomposed into a “within” and a “between”
effect. In section 3.3, we further discuss the role of heterogeneity among firms in APG. In
section 3.4, we extend our proposed decomposition method by taking into account firm
entry and exit and, finally, in section 3.5 we illustrate our methodology with some toy
examples, before presenting, in section 4, an empirical application to real data.

7
3.1 Notation and definitions
Following Koopmans (1977) and Hildenbrand (1981) we represent the actual technique
of a production unit (e.g. firm) i by means of the vector of its production activity:

ati = αi,1
t t t
∈ Rl+ ,


, · · · , αi,(l−1) , αi,l (6)
 
t t t
where αi,l is the output in period t and αi,1 , · · · , αi,(l−1) is the vector of inputs. In
order to keep our formalism simpler, in this section we confine our discussion to activities
producing only one output. However, in appendix B we will show that our methodology
can be easily extended and deal with production activities with more than one output.
If I t denotes the set of all production units within one industry at time t, the aggregate
(industry) production activity can be defined as:
!
X X X
t
dt = β1t , · · · , βl−1 , βlt = t t t
∈ Rl+ ,


αi,1 ,··· , αi,(l−1) , αi,l (7)
i∈I t i∈I t i∈I t

i.e. the sum of all individual firm production activities in the industry.
The productivity pti of the production unit i at time t can be measured as the tangent
of the angle θ(ati ) that the vector ati forms with the space of inputs (Dosi et al., 2016).
To give the intuition behind this measure of productivity, let us consider the case of only
one input. Clearly, the larger the angle that the vector representing a production activity
forms with the input axis, and therefore the smaller the angle it forms with the output
axis, the more productive is the activity. By extension to the case of multiple inputs we
obtain the following productivity indicator:
t
αi,l
pti := tg θ(ati ) =


(8)
||pr (ati ) ||
where the map

pr : Rl → Rl−1
(x1 , · · · , xl−1 , xl ) 7→ (x1 , · · · , xl−1 )

is the projection map on the space of inputs4 .

Similarly, we define the AP of the industry at time t, denoted by P t , as

βlt
P t := tg θ(dt ) =


. (9)
||pr (dt ) ||
t t
P
Notice that whereas β l = i∈I t αi,l is the total output of the industry, in general
t t
P
||pr (d ) || =
6 i∈I t ||pr (ai ) ||, the equality holding either in the case of a unique input or,
in the case of multiple inputs, only when all the vectors pr(ati ) lie on the same line and
therefore production activities are perfectly homogeneous and differ only in their scale.
If instead techniques are heterogeneous and firms use different combinations of inputs,
t t
P
the inequality ||pr (d ) || = 6 i∈I t ||pr (ai ) || holds. This heterogeneity component is an
important feature of our model and we will further exploit it later in the article.
4
This can be easily generalised to multi-output case simply considering a different projection map
(see Dosi et al. (2016)).

8
3.2 Decomposing APG into within and between Effects
AP, defined in equation (9), can be further written as a weighted average of individual
productivities pti , since
t t
P
t βlt i∈I t αi,l
X ||pr (at ) ||
i
αi,l
P = = =
||pr (dt ) || ||pr (dt ) || t
||pr (dt ) || ||pr (ati ) ||
i∈I

from which, by (8), we get the decomposition

X
Pt = wit pti (10)
i∈I t

where the weights

||pr (ati ) ||
wit := (11)
||pr (dt ) ||
represent the input-based weights defined as the relative length of individual input vectors
||pr (ati ) || over industry input vector ||pr (dt ) ||. As already mentioned, the length of the
industry input vector ||pr (dt ) || is not necessarily equal to the sum of the lengths of the
t t
P
individual input vectors ||pr (ai ) ||, thus i∈I t wi is not necessarily equal to 1.
Equality (10) indicates that AP can be written as a weighted average of individual
productivities pti . However, it is important to stress a fundamental methodological differ-
ence between how we obtain this weighted average and the standard approaches that we
briefly surveyed in section 2 above. In the latter, AP is defined and hence computed as
a weighted average of individual productivities, no matter how the latter are measured.
On the contrary, in our framework AP is defined and computed in exactly the same way
as we compute individual productivities, as it clearly emerges by comparing definitions
(8) and (9), not as a weighted average. Equation (10) shows that our AP has the prop-
erty of being equal to a weighted average, but it is neither defined nor computed in that
way. Moreover, the measure of AP we propose is also a straightforward generalization of
the one-input-one-output case. When there is only one input, the industry input vector
degenerates to one number and the tangent of the angle we use to measure AP, according
to (9), becomes the quotient of the sum of all outputs divided by the sum of all inputs.
αt
Input based weights in (11), wit = P i,1 t are nothing else than input shares with sum
i∈I t αi,1
equal to 1.
Now we show that our measure of AP can be decomposed and the standard effects
outlined by the traditional literature can indeed be easily computed also in our framework.
For the sake of simplicity we first introduce our decomposition method only for the set C
of continuing firms, i.e. all those that are active both in both periods t − 1 and t. Entry
and exit will be introduced later in section 3.4.
Such continuing firms are described at time t − 1 by the vector set {at−1 i }i∈C ∈ R+
l

and at time t by the vector set {ati }i∈C ∈ Rl+ . Let dt−1 and dt , computed according to
(7), represent the corresponding aggregates at time t − 1 and t respectively. Given all the
production activity vectors, aggregate and individual productivities at t − 1 and t can be
easily computed according to (9) and (8) respectively.
We can now decompose APG, defined as the difference of AP between two consecutive

9
years, into within-firm and between-firm components5 :
X X
∆P t = w̄i ∆pti + p̄i ∆wit , (12)
i∈C i∈C
| {z } | {z }
W ithin Between

where ∆ represents the variation from year t − 1 to year t, whereas w̄i and p̄i are averages
of weights and productivities respectively.
In the above decomposition, the within term represents the contribution given to
APG by the variations of the individual productivities and it is therefore similar to the
“within” effect in the current literature. The between term presents instead an important
difference when compared to the “between” effect in the standard literature. In the
latter, the weights wit are defined either as input or output shares and, in both cases,
t
P
i∈C wi = 1. This is not the case in our decomposition (12) where, since the sum of the
lengths of individual input vectors is not necessarily equal to the length of the sum of
individual input vectors, i.e. the length of the industry input vector, we have, in general,
t
P
i∈C wi 6= 1. We will discuss this point in details in the next subsection and show that,
actually, our between effect can be further decomposed.

3.3 Further Decomposing the Between Effect

In this section, we show that we can further decompose our Between term into two parts,
an input weights component and a heterogeneity component.
Given the individual ati and aggregate dt production activities, consider their projec-
tions on the input space and call them pr (ati ) and pr (dt ). Figure 1 provides a graphical
representation for the case with two inputs, capital K and labour L, and one output.
Notice that if all individual production activities used the same proportion of inputs and
differed only in scale and/or productivity, all the projection vectors pr (dt ) and pr (ati )
would overlap. On the other hand, the further away pr (ati ) is from pr (dt ), the more
the combination of inputs used by firm i differs from the industry average combination.
To measure this difference, we introduce ϕti which is the angle formed by the individual
projection vectors pr (ati ) and the industry projection vector pr (dt ). Notice that given
vectors dt and ati , and thus pr (dt ) and pr (ati ), it is easy to compute cos ϕti for each firm
i at time t.
In the input space, we denote by bti the projection of pr (ati ) onto the industry pro-
jection vector pr (dt ) (see Figure 1 for a graphical illustration). Notice that the length
||bti || of the vector bti can be regarded as the contribution of pr (ati ) to the length of the
industry input vector pr (dt ). Hence, from now on, we will refer to the length ||pr (ati ) ||
of the firm input vector pr (ati ) as the actual input size of firm i and to the length
||bti || of the vector bti as the contributing input size of firm i. It is easy to see that
||pr (dt ) || = i∈I t ||bti ||.
P
Since
||bti ||
||pr ati || =


,
cos ϕti
we can decompose wit as
||pr (ati ) || ||bti || 1 ||bti || 1
wit = = t
=
t
||pr (d ) || t
cos ϕi ||pr (d ) || ||pr (d ) || cos ϕti
t

5
All the mathematical details which lead to this decomposition can be found in appendix A.

10
 Figure 1: A graphical explanation for the decomposition of the between effect

VA
dt

pr(dt )
ati
bti

ϕti
pr(ati )

that is as the product:

wit = sti · hti (13)
of what we could call the “input weights”:

||bti ||
sti = (14)
||pr (dt ) ||

and a “heterogeneity coefficient”:

1
hti = . (15)
cos ϕti

Equation (13) shows that the input based weights wit in (11) combine together two
different effects. The first one, sti , represents the contribution of individual firms to
the length of the industry input vector, i.e. our equivalent
P to the input weights in an
t
multiple input case. Notice that in this case we have i∈C si = 1 and that, in the
t
case of only one input si is the standard input share weight. The second one, that
we named “heterogeneity coefficient” hti , measures to which degree the individual input
combinations diverge from the industry average combination. The larger P this divergence,
the bigger the angle ϕi and therefore the coefficient hi . Thus, the sum i∈C hti can be
t t

regarded
P as ant index of the heterogeneity of input combinations among productive units,
and i∈C ∆hi measures the variations of such heterogeneity.

11
 Given these two effects, we can further decompose the Between term in (12) into two
parts and thus refine our decomposition of APG as follows:6
X X X
∆P t = w̄i ∆pti + p̄i h̄i ∆sti + p̄i s̄i ∆hti (16)
i∈C i∈C i∈C
| {z } | {z } | {z }
W ithin InputShares Heterogeneity

where the Within term is the same as the one in (12) and the sum of InputShares plus
Heterogeneity is equal to the Between term in (12). The term InputShares measures
the changes of contributions of individual firms to the industry total input size, whereas
the term Heterogeneity measures the changes of heterogeneity among the combinations
of inputs used by different firms. It is easy to see that when there is only one input, hti is
always equal to 1, Heterogeneity is 0 and the two decompositions (16) and (12) coincide.
As an illustration, consider the hypothetical example of Table 3 which describes an
industry made of three firms. In year 1 such firms are identical, but in year 2 one
of them has switched to a more capital intensive technique, moving from production
activity vector (1.814, 0.842, 1) to (0.564, 1.919, 1) where output has remained unchanged
and equal to 1 and the length of the input vector has also remained equal to 2.
Table 3 reports data for each firm and for the industry (last row). In each panel
the last two columns report the productivity measures computed, respectively, with our
methodology and with the standard one. The latter consists of TFPs for the individual
firms (assuming production function (5)) and of the three weighted averages for the
industry, where weights are given, respectively, by labour, capital and quantity shares.

Table 3: Example: Between, Input Shares and Heterogeneity Effects.

Year 1 Year 2
Li Ki Qi tg(·) T F Pi Li Ki Qi tg(·) T F Pi
Firm 1 1.814 0.842 1.000 0.500 0.668 1.814 0.842 1.000 0.500 0.668
Firm 2 1.814 0.842 1.000 0.500 0.668 1.814 0.842 1.000 0.500 0.668
Firm 3 1.814 0.842 1.000 0.500 0.668 0.564 1.919 1.000 0.500 1.305
P P P P P P
Li Ki Qi tg(·) w.a. T F Pi Li Ki Qi tg(·) w.a. T F Pi
0.668 0.754
Industry 5.442 2.526 3.000 0.500 0.668 4.192 3.603 3.000 0.543 1.007
0.668 0.880
*
w.a. stands for weighted average.

If we use our methodology we find that all individual productivities remain constant,
whereas AP increases from 0.5 to 0.543 and this variation can be decomposed as follows:

0.043
| {z } = |{z}
0 |−0.001
{z } + |0.044
{z } (17)
AP G W ithin InputShares Heterogeneity

These values correctly reflect that individual productivities have remained unchanged
(firm 3 has switched to a new technique which delivers the same output with a vector
input of the same length and then has grown with constant returns to scale), so the
W ithin effect is null, whereas Heterogeneity has considerably increased.
6
Appendix A contains the mathematical details of the derivation of equation 16.

12
 On the other hand, when we measure productivities with TFPs, we observe that
the productivity of firms 3 has almost doubled (because of the different exponents of L
and K in (5)) and therefore also AP undergoes an increment entirely due to the Within
effect. For instance, if we use output shares as weights, APG is equal to 0.212 and can
be decomposed as follows:
| {z } = |0.212
0.212 {z } + |{z}
0 (18)
AP G W ithin Between

It may be worth explaining into greater detail the origin of such a substantial differ-
ence between the two decompositions. In the standard method a production function is
imposed on data and, in our example, such a function reads the change of (L, K) combi-
nation of firm 3 as an increase of its productivity which, in turn, translates into a positive
within effect in the aggregate. Our methodology instead interprets such a change as an
increase of heterogeneity because firms 3 was using at time 1 the same input combina-
tion as the other two firms, whereas at time 2 its combination is substantially different.
But since we do not impose a production function on data we cannot conclude that
(1.814, 0.842, 1) is less productive than (0.564, 1.919, 1) as the two techniques have equal
output and the same input size (i.e. the length of the vector of inputs). Nevertheless
the change of technique of firm 3 determines a movement of the industry diagonal which
results in an aggregate productivity increase. The Input Shares effect is negative because
the diagonal moves towards firm 3 and therefore the projections of firms 1 and 2 on it
become shorter.

3.4 Decomposing APG with Entry and Exit

In this section we expand our decomposition of APG by accounting for the contributions
given by production units which enter or exit the industry during the period under con-
sideration. Let C be the set of all continuing firms, i.e. those that are active both in t − 1
and in t, N the set of entering firms which are active in t but not in t − 1, and X the set
of exiting firms which are active in t − 1 but not in t. Let vector sets {at−1i }i∈{C∪X} ∈ R+
l

and {ati }i∈{C∪N } ∈ Rl+ represent all firms active in the industry in t − 1 and t respectively.
According to equation (10) we have
X X
∆P t = wit pti − wit−1 pt−1
i
i∈{C∪N } i∈{C∪X}
X X X
= (wit pti − wit−1 pt−1
i )+( wit pti − wit−1 pt−1
i )
i∈C i∈N i∈X

where for all the continuing firms, the term i∈C (wit pti − wit−1 pt−1
P
i ) can be further decom-
posed as in equation (16) and finally we have
X X X
∆P t = w̄i ∆pti + p̄i h̄i ∆sti + p̄i s̄i ∆hti
i∈C i∈C i∈C
| {z } | {z } | {z }
W ithin InputShares Heterogeneity
X X (19)
+ wit pti − wit−1 pt−1
i .
i∈N i∈X
| {z }
N etEntry

13
3.5 An Illustrative Example
Table 4 provides an illustrative example of a hypothetical industry composed of 5 pro-
duction units (firms) producing a unique output with two inputs, L and K. The first
three columns report the input-output data in year 1. The fourth and fifth columns
report the length ||pr (a1i ) || of the input vectors and the length ||a1i || of the production
activity vectors respectively. In the sixth column, we compute productivities according
to equation(8). Columns 7th to 12th report the same data referred to year 2. The only
change taking place between the two years is an increase of the heterogeneity of the in-
put combinations among firms, as visualized in Figure 2.7 The industry becomes more
productive, as productivity increases from 0.5001 to 0.5294. Applying decomposition
method in equation (16) this 0.0293 increase of productivity can be decomposed in the
following way:
| {z } = |{z}
0.0293 0 −0.0006
| {z } + 0.0299
| {z } .
AP G W ithin InputShares Heterogeneity

which confirms that the within effect is null and the between effect is basically due to
the increase of heterogeneity among firms, which is indeed the only phenomenon taking
place between the two years.8
Table 4: Toy Example with Five Firms and Increasing Heterogeneity

Year 1 Year 2
1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th 12th
length of length of length of length of
L K Output input vector vector
tg(·) L K Output input vector vector
tg(·)

Firm 1 1.414 1.414 1.000 2.000 2.236 0.500 1.414 1.414 1.000 2.000 2.236 0.500
Firm 2 1.464 1.362 1.000 2.000 2.236 0.500 1.764 0.942 1.000 2.000 2.236 0.500
Firm 3 1.424 1.404 1.000 2.000 2.236 0.500 1.864 0.724 1.000 2.000 2.236 0.500
Firm 4 1.374 1.453 1.000 2.000 2.236 0.500 1.044 1.706 1.000 2.000 2.236 0.500
Firm 5 1.394 1.434 1.000 2.000 2.236 0.500 0.564 1.919 1.000 2.000 2.236 0.500
Industry 7.071 7.068 5.000 9.998 11.178 0.5001 6.651 6.705 5.000 9.444 10.686 0.529

Let us now assume that in year 3 everything remains unchanged from year 2 except
that firm 3 doubles its output with the same inputs. The first six columns of Table 5
report these hypothetical data for year 3. Because of the increase of firm 3’s productivity,
AP increases from 0.5294 to 0.6353 and this increase can be decomposed as:

| {z } = |0.1059
0.1059 {z } + 0
|{z} + 0
|{z}
AP G W ithin InputShares Heterogeneity

which indicates that APG is completely driven by the technical change operated by Firm
3, and therefore only the W ithin term is different from zero. Finally, let us suppose that
between year 3 and year 4 all firms hold their productivities and heterogeneity coefficients
hti constant and only the input shares weights sti change (the last six columns in Table 5).
APG is now totally imputed to variations of the input weights, i.e. to the InputShares
7
Vectors in Figure 2 are the projections pr(ai ) on the space of inputs of the vectors ai . So, for
example, the projection of the three dimensional vector a11 = (1.414, 1.414, 1.000) of firm 1 in year 1 is
simply given by the two dimensional vector pr(a11 ) = (1.414, 1.414).
8
Consistently, the heterogeneity measure introduced by Dosi et al. (2016) increases from 2.09025e-06
to 0.00728504.

14
 Figure 2: Toy Example with Five Firms and Increasing Heterogeneity

6 6
Capital

Capital
4 4

2 2

0 0

0 2 4 6 0 2 4 6
Labor Labor
Year 1 Year 2

effect:
−0.0060 0 −0.0060
| {z } = |{z} | {z } + 0
|{z} .
AP G W ithin InputShares Heterogeneity

These hypothetical examples show that our measure correctly captures the phenomena
driving APG. The next section provides an empirical application to real data.
Table 5: Toy Example with Five Firms - Dynamic from Year 3 to Year 4

Year 3 Year 4
L K Output ||pr(·)|| tg(·) sti hti L K Output ||pr(·)|| tg(·) sti hti
Firm 1 1.414 1.414 1.000 2.000 0.500 0.212 1.000 1.494 1.494 1.057 2.113 0.500 0.224 1.000
Firm 2 1.764 0.942 1.000 2.000 0.500 0.202 1.046 1.766 0.943 1.001 2.002 0.500 0.203 1.046
Firm 3 1.864 0.724 2.000 2.000 1.000 0.193 1.095 1.777 0.690 1.906 1.906 1.000 0.184 1.095
Firm 4 1.044 1.706 1.000 2.000 0.500 0.206 1.028 1.081 1.766 1.035 2.071 0.500 0.213 1.028
Firm 5 0.564 1.919 1.000 2.000 0.500 0.186 1.137 0.533 1.811 0.944 1.888 0.500 0.176 1.137
Industry 6.651 6.705 6.000 9.444 0.635 1.000 1.000 6.651 6.705 5.943 9.444 0.629 1.000 1.000

4 An Empirical Application
4.1 Data and methodology
In order to show that our methodology is also empirically relevant, in this section we
apply our APG decomposition to real firm-level data and compare it to the results that
standard methods would produce on the same data. The database we use is the October
2015 release of AMADEUS, a commercial database provided by Bureau van Dijk and
containing balance sheets and income statements for over 21 million European firms over

15
the period 2004-2013. We focus on firms from several industries at the 4-digit NACE
classification for three European countries, namely France, Italy and UK. These industries
have been randomly selected from those with at least 20 firms (including continuous,
exiting, and entering firms) during the time period under investigation for all the three
countries. For reasons of space we report here only the results for seven selected industries,
listed in Table 6. Results for other industries are available upon request.
Number of employees and fixed assets are chosen as proxies for the two inputs, labour
and capital, and turnover as a proxy for output. In some cases we add a third input,
proxied by material costs. All these values, except the number of employees, are measured
in thousands Euros and expressed in 2010 prices using the appropriate deflator for the
4-digit industry and the country under consideration9 .
Table 6: List of Selected Industries

NACE Name of Industry

2014 Manufacture of other organic basic chemicals
2120 Manufacture of pharmaceutical preparations
2593 Manufacture of wire products, chain and springs
2630 Manufacture of communication equipment
2712 Manufacture of electricity distribution and control apparatus
2813 Manufacture of other pumps and compressors
2920 Manufacture of bodies (coachwork) for motor vehicles; manufacture of trailers and semi-trailers

We compute APG and its decomposition for two time periods: between 2004 and
2007 and between 2010 and 2013. We compute four-years rather than yearly variations
because productivity, however measured, tends to be very sticky over time, hence a longer
period is more suited to investigate its variation. Moreover, we omit from consideration
2008 and 2009, the two years when the economic crisis hit more harshly the countries
under consideration. The analysis of APG in these years is indeed interesting per se, but
we want to make sure that our results are not driven by very abrupt variations in those
years mostly due to plunging output and high exit rates.
Table 7 shows APG and its decomposition, all computed following our methodology10
for the period 2004-07 in the left panel and 2010-13 in the right panel. In the left panel,
column 1 shows the value of APG, columns 2-4 the contributions to APG given by,
respectively, entering, exiting and continuing firms. The latter is further decomposed
in a W ithin and Between effect in column 5. Then the Between effect of column 5 is
further decomposed in a InputShares and a Heterogenenity effects computed as in (16)
and reported in column 6. Finally, in column 7 we report an heterogeneity coefficient
coherent with our methodology and introduced by Dosi et al. (2016).11 Columns 8 to 14
report the same results for APG between 2010 and 2013.
The table contains some interesting results. For instance, for what concerns sector
“2014 ” from 2004 to 2007, we observe positive W ithin and negative Between effects in
9
Deflators for 4-digit industries are provided by Eurostat (https://ec.europa.eu/eurostat/data/
database). In a few cases the 4-digit deflators for a specific industry are not available, hence more
aggregate deflators, e.g. 3-digit or 2-digit deflators for that country, have been used.
10
The software developed for computing AP, APG and its decompositions following the methodology
developed in this article has been written in R and is available upon request.
11
The coefficient is a Gini volume coefficient and is given by the ratio between the volume of the
zonotope formed by the actual production activities and the volume of the zonotope of an industry with
the same size but maximum heterogeneity. For more details see Dosi et al. (2016) p. 885.

16
all three countries. The former points to an increase in productivity at the firm level,
whereas the latter signals that the reallocation of input shares had a negative effect on
industry productivity. An opposite dynamics is observed for instance in industry “2120 ”
in UK and Italy. As far as the between term is further decomposed, one can notice that
although the variation in input shares tends to dominate over the heterogeneity term, the
latter is far from being irrelevant and in a few instances even outweighs the other. Other
things being equal, a variation in the adopted techniques by firms has an effect on AP.

4.2 Comparison with Standard Methods on Empirical Data

Finally, we compare our methodology with the more traditional one on the same set of
data.
As a standard benchmark we adopt the method proposed by Grilliches and Regev
(1995) and summarized in equation (3) above. In order to estimate TFPs of individual
firms, which are then averaged to determine AP, we follow Levinsohn and Petrin (2003)
and use material costs as the additional proxy variable necessary in their methodology.12
Notice that by doing this, we lose all the firms from UK and some from France and Italy
for which there is no indication of material costs in our original dataset.
Using employment shares as weights, we compute APGs from 2004 to 2007 for each
country/industry and report them in column 1 of Table 8. Decomposition results, com-
puted according to (3), can be found in columns 2 (Entry and Exit) and 3 (Within and
Between). Similar APGs and their decompositions but using output instead of labour
shares as weights can be found in column 4 to 6. In column 7 we report APG computed
with our methodology, and in columns 8-10 its decomposition into the various compo-
nents: column 8 gives the Entry and Exit effects, column 9 the W ithin and Between
and column 10 splits the latter into InputShares and Heterogeneity effects. Finally,
column 11 gives our measure of variation of industry heterogeneity, proxied by the Gini
coefficient for the industry zonotope.
The right panel of Table 8 (columns 12-22) replicates the same indicators for the
productivity variations between 2010 and 2013.
In commenting the results in Table 8, we start by reminding that the absolute mea-
sures of APG provided by our proposed method and by the more traditional one are the
outcome of two very different approaches, hence they cannot be directly compared. The
related rates of change (reported in brackets below APG) provide a better benchmark
for comparison and reveal that both in terms of signs and order of magnitude of the
variation, the two methods point to a common direction, of course with some notewor-
thy exceptions. Note however that on average - and as expected - the differences across
the methods remain large. When we focus on the decomposition terms across the two
methods, similar comments apply.

12
We use Stata command levpet (Petrin et al., 2004) to estimate TFP with number of employees
as free variable, fixed assets as capital variable, and material cost as proxy variable respectively. The
dependent variable is total revenue instead value added.

17
 Table 7: APG Decompositions for Selected Industries in France, Italy and UK, 2004-07 and 2010-13 - Our methodology

From 2004 to 2007 From 2010 to 2013

W ithin Gini W ithin Gini

NACE Ctry. APG Entry Exit Continue Growth APG Entry -Exit Continue Growth
InputShares InputShares
Btw. (%) Btw. (%)
Heterogeneity Heterogeneity
1 2 3 4 5 6 7 8 9 10 11 12 13 14

2.57 -0.26
2014 FR 1.60 0.31 -0.34 1.63 42.24 -0.61 0.67 -0.20 -1.08 -40.92
-3.27 -0.81
-0.94 -0.81
2.34 0
5.46 1.28
2014 UK 0.31 0.36 -0.00 -0.05 37.99 0.30 0.32 -0.17 0.15 -10.47
-10.58 -1.13
-5.51 -1.12
5.07 0.01
0.49 -0.46
2014 IT -0.00 0.40 -0.01 -0.40 98.76 1.97 0.49 -0.14 1.63 -3.70
-0.89 2.09
-0.89 2.09
0 0
0.01 -0.04
2120 FR 0.89 0.25 -0.47 1.11 -7.88 -0.32 0.41 -0.47 -0.25 -7.69
1.25 -0.09
1.1 -0.21
-0.14 -0.12
-0.23 0.19
2120 UK 0.47 0.36 -0.01 0.12 2.98 -1.15 0.90 -0.19 -1.86 75.09
0.36 -2.1
0.35 -2.05
-0.01 0.05
-0.49 -0.48
2120 IT -0.12 0.17 -0.24 -0.04 3.37 -0.27 0.81 -0.37 -0.72 -10.22
0.44 -0.2
0.44 -0.24
0 -0.04
0.54 0.17
2593 FR 1.25 0.71 -1.11 1.65 -8.46 4.54 4.89 -2.74 2.39 -6.48
1.09 2.23
1.11 2.22
0.02 -0.01
0.69 2.15
2593 UK 0.45 1.13 -0.44 -0.24 36.36 -0.51 0.52 -0.15 -0.87 -17.64
-0.96 -4.64
-0.93 -3.02
0.02 1.62
0.14 -0.04
2593 IT 0.41 0.86 -0.07 -0.38 21.33 -0.00 0.57 -0.18 -0.39 10.44
-0.52 -0.36
-0.52 -0.36
0 0
-1.93 -0.49
2630 FR 1.96 2.64 -3.30 2.62 165.30 -8.18 2.90 -6.47 -4.61 -56.58
4.64 -4.13
4.56 -4.12
-0.08 0.01
0.66 -0.61
2630 UK -1.07 1.40 -0.25 -2.21 -15.48 -0.47 0.21 -0.26 -0.42 -30.54
-7.54 0.11
-2.87 0.18
4.67 0.07
-2.03 -33.76
2630 IT -2.26 0.88 -0.47 -2.67 53.77 -1.71 0.17 -0.76 -1.12 -29.79
1.31 38.31
-0.64 32.65
-1.94 -5.66
-0.01 -0.03
2712 FR 0.10 3.00 -1.09 -1.81 -23.71 -1.87 1.74 -1.70 -1.91 102.80
-1.78 -1.88
-1.79 -1.88
-0.01 0
-3.63 -1.54
2712 UK 6.44 5.63 -0.19 1.00 279.46 -5.40 0.20 -5.16 -0.44 -76.85
5.84 1.09
4.63 1.1
-1.21 0.01
0.42 -1.52
2712 IT 0.58 0.17 -0.01 0.42 -11.67 -0.09 0.41 -0.04 -0.45 -25.45
0 1.08
0 1.07
0 -0.01
-0.01 -0.44
2813 FR -0.70 0.37 -2.00 0.93 34.75 0.59 1.81 -2.28 1.05 -28.72
0.95 1.49
0.95 1.49
0 -0.01
0.09 -0.15
2813 UK -2.67 0.47 -0.03 -3.12 109.51 0.20 0.06 -0.03 0.17 -24.87
-3.26 0.32
-3.21 0.31
0.05 0
0.52 -1.17
2813 IT 0.57 1.35 -0.14 -0.63 56.87 -0.45 0.76 -0.55 -0.66 1.08
-1.16 1.32
-1.16 0.51
0 -0.8
1.04 -0.04
2920 FR 1.00 2.12 -2.56 1.44 6.02 -0.17 2.30 -4.67 2.20 -6.80
0.87 2.24
0.4 2.23
-0.46 -0.01
0.47 1.26
2920 UK 0.16 0.95 -0.26 -0.52 14.39 1.42 0.40 -0.27 1.29 -2.67
-0.99 0.03
-0.99 0.02
0 -0.01
1.49 0.46
2920 IT 1.00 1.29 -1.02 0.72 7.49 0.11 0.85 -0.10 -0.64 -44.96
-1 -1.08
-0.77 -1.09
0.23 -0.01

18
 Table 8: APG Decompositions for Selected Industries in France and Italy 2004-07 and 2010-13 - Our methodology vs. Grilliches and Regev (1995)

From 2004 to 2007 From 2010 to 2013

Decomposition Method Decomposition (3) Proposed Method (19) Decomposition (3) Proposed Method (19)

Entry W ithin Entry W ithin Entry W ithin Gini Entry W ithin Entry W ithin Entry W ithin Gini
APG APG APG APG APG APG
NACE Ctry. Growth Growth
(% Growth) (% Growth) (% Growth) InpSh (% Growth) (% Growth) (% Growth) InpSh
-Exit Btw -Exit Btw -Exit Btw (%) -Exit Btw -Exit Btw -Exit Btw (%)
Heter Heter
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

1.05 1.8 1.05 4.59 0.13 0.43 4.37 -1.37 2.56 -5.07 0.36 -0.1
1.77 2.39 0.04 -3.35 -12.05 -0.18
2014 FR -9.82 -58.54
(12.12) (14.28) (2.53) -0.4 (-20.44) (-40.55) (-11.59) -0.32
-2.13 1.05 -2.73 -0.53 -0.27 -0.26 -1.5 -4.85 -0.69 -8.84 -0.1 -0.34
0.14 -0.02
1.1 0.43 1.04 0.34 0.32 0.17 1.75 -0.49 0.91 -0.02 0.22 0.02
0.67 0.6 0.05 -0.25 0.02 0.39
2014 IT 291.02 -25.29
(25.82) (23.94) (4) -0.5 (-5.89) (0.81) (35.02) 0.44
-0.03 -0.84 -0.03 -0.76 -0.01 -0.44 -1.22 -0.28 -0.7 -0.15 -0.11 0.27
0.06 -0.17
3.08 0.61 1.41 0.87 0.17 -0.16 3.9 0.53 3.55 1.05 0.22 -0.02
0.59 3.25 0.27 2.01 -3.42 -0.07
2120 FR -10.34 -22.48
(2.95) (11.74) (13.21) 0.63 (10.15) (-11.77) (-3.84) -0.03
-3.38 0.27 -3.29 4.27 -0.35 0.61 -3.13 0.7 -8.53 0.52 -0.27 -0.01
-0.02 0.03
1.34 0.47 1.02 0.96 0.1 0.07 7.74 -0.05 6.59 0.31 0.58 -0.08
0.52 1.12 0.02 -0.24 0.17 0.05
2120 IT 15.61 -10.96
(2.82) (5.89) (1.35) 0.1 (-1.13) (0.8) (2.78) -0.26
-2.84 1.54 -2.16 1.3 -0.21 0.08 -0.87 -7.06 -2.8 -3.92 -0.21 -0.24
-0.03 0.01
2.53 -0.13 1.89 0.22 0.21 0.02 8.75 -0.26 9.46 -0.1 0.97 -0.02
-0.23 -0.09 -0.07 -1.21 -0.73 -0.15
2593 FR -26.31 -35.31
(-1.36) (-0.49) (-3.94) 0.17 (-7.19) (-4.19) (-8.16) 0.2
-3.99 1.36 -4.05 1.86 -0.4 0.11 -12.06 2.36 -12.19 2.1 -1.28 0.19
-0.06 -0.01
0.36 -0.05 0.32 -0.04 0.45 -0.08 0.54 -0.04 0.38 -0.03 0.4 -0.02
19

0.05 0.06 -0.01 0.25 0.16 0.06

2593 IT 10.30 47.68
(5.27) (7.56) (-0.36) -0.34 (24.73) (17.93) (4.02) -0.21
-0.03 -0.23 -0.03 -0.19 -0.04 -0.33 -0.07 -0.18 -0.06 -0.12 -0.12 -0.2
0 0.01
3.81 0.06 4.59 1.02 0.53 -0.25 16.92 -0.12 21.54 -0.54 1.9 -0.04
1.65 3.51 -0.36 5.33 5.68 -0.23
2630 FR 519.39 -73.20
(8.9) (18.65) (-13.9) 0.7 (28.04) (26.75) (-8.6) -0.67
-8.2 5.98 -7.79 5.69 -1.26 0.61 -11.04 -0.44 -10.19 -5.14 -1.53 -0.57
-0.09 0.11
9.62 -0.17 15.99 1.58 0.5 0.04 3.96 -44.3 5.09 -39.54 0.17 -28.14
-1.25 5.84 0 3.11 1.05 -0.38
2630 IT 104.81 -14.13
(-3.02) (11.11) (-0.09) -0.24 (6.93) (1.94) (-19.77) 33.91
-5.66 -5.05 -5.95 -5.78 -0.3 -0.25 -9.13 52.57 -14.08 49.57 -0.42 28.02
-0.01 -5.89
5.99 0.18 7.13 0.26 1.04 -0.06 5.15 0.11 4.13 0.12 0.62 0.01
0.4 0.49 -0.07 0.25 0.07 0.1
2712 FR -45.69 22.45
(3.22) (3.93) (-3.67) -0.67 (1.88) (0.53) (5.21) -0.09
-3.02 -2.76 -2.68 -4.22 -0.4 -0.66 -3.68 -1.33 -3.15 -1.03 -0.44 -0.09
0.02 0
2.41 0.43 1.36 0.37 0.14 0.2 3.93 -1.59 3.25 -2.43 0.32 -0.14
0.48 0.53 0.35 -1.08 -3.92 -0.09
2712 IT 24.67 -55.92
(3.12) (3.5) (33.72) 0.05 (-6.67) (-21.16) (-6.01) -0.18
-0.18 -2.19 -0.14 -1.06 -0.01 0.02 -0.62 -2.8 -0.34 -4.4 -0.03 -0.24
-0.03 -0.07
0.05 1.34 0.03 1.97 0.13 -0.14 0.15 0.65 0.14 1.17 0.88 -0.04
1.52 1.6 -0.19 0.66 1.88 0.13
2813 FR 39.92 -46.77
(204.58) (172.92) (-8.58) 0.51 (46.69) (108.15) (6.3) 0.47
-0.24 0.37 -0.23 -0.17 -0.69 0.52 -0.48 0.34 -0.3 0.87 -1.21 0.49
0.01 0.02
7.68 0.3 9.13 0.68 0.51 -0.01 7.79 -0.86 7.97 -1.54 0.42 -0.12
1.38 2.17 0.02 0.08 -1.89 0
2813 IT 92.80 11.53
(5.52) (7.98) (1.07) -0.41 (0.3) (-6.36) (0.17) -0.06
-0.74 -5.86 -0.89 -6.76 -0.06 -0.42 -3.18 -3.67 -5.81 -2.5 -0.27 -0.03
-0.01 0.03
11.93 7.29 11.2 9.64 0.34 0 12.52 1.53 14.53 1.5 0.46 0.01
10.9 14.08 0.02 3.73 -2.04 0.02
2920 FR 24.57 -2.34
(17.17) (19.93) (1.25) 0.14 (6.51) (-3.05) (1.09) 0.47
-12.61 4.29 -14.68 7.91 -0.44 0.13 -24.6 14.28 -34.12 16.04 -0.92 0.46
-0.02 0
14.62 -2.03 14.82 -14.98 0.36 -0.05 14.95 1.16 13.28 -0.64 0.41 0.04
-4.83 -25.87 0.04 6.2 -3.02 0.08
2920 IT 19.42 -45.42
(-7.28) (-26.43) (2.75) 0.07 (14.99) (-5.07) (5.44) -0.32
-11.4 -6.02 -22.53 -3.17 -0.33 0.07 -10.94 1.03 -1.27 -14.39 -0.05 -0.32
0 0
5 Conclusions
Thanks to the increasing availability of longitudinal establishment- and firm- level data, a
rapidly growing body of empirical literature has shown a highly significant and persistent
degree of heterogeneity among firms and establishments in the input combinations and in
their productivities even in the presence of the same relative input prices and in narrowly
defined industries, thus with relatively homogeneous types of output. Moreover, the
analyses addressed the issue of the relative importance between firm-level increase in
productivity and the reallocation of market share to the APG, i.e. so-called “within”
and “between” effects, across individual producers within narrowly defined sectors. The
message, overall, is that “within” effects, i.e. learning, dominate upon “between” effects,
i.e. selection.
However, such widespread and persistent heterogeneity poses serious challenges to the
use of standard aggregate production functions and AP one can derive from them. In
addition, it challenges also the innocence of aggregations based on sheer weights given by
input or output shares. In this article, building upon a geometric representation of the
empirical production possibility set first suggest by Hildenbrand (1981) and developed in
Dosi et al. (2016), we introduce a new decomposition method for APG which 1) computes
individual and aggregate productivities in the same way, instead of computing the latter
as some arbitrary weighted average of the individual indicators; 2) reduces the loss of
information implied by standard decomposition methods; 3) allows for a precise measure
of the contribution given by variations in heterogeneity.
Our methodology can be applied to empirical data and the preliminary application
we present in this article, on some selected industries in France, Italy and the UK, show
that indeed the contributions to APG that can be attributes to changes of firm-level
heterogeneity are far from negligible.
To put it differently, given the overwhelming evidence on heterogeneity at all levels
of observation, aggregation of “micro production functions” is not appropriate. However,
it is also partly misleading to use simple “evolutionary decompositions” just relying on
learning vs. selection components, as revealed by the often “perverse” sign of their inter-
action in empirical estimates (firms with higher productivity growth shrinking in size).
Rather, it seems more informative to make explicit a heterogeneity component which
is orthogonal to either, somehow similar to the drift component found in evolutionary
biology.

20
References
Aw, B.-Y., S. Chen, and M. Roberts (1997): “Firm-level Evidence on Productivity
Differentials, Turnover and Exports in Taiwanese Manufacturing,” NBER Working
Paper 6235.

Baily, M. N., C. Hulten, and D. Campbell (1992): “Productivity dynamics

in manufacturing plants,” Brookings Papers on Economic Activity (Microeconomics),
1992, 187–249.

Baldwin, J. R. and M. Rafiquzzaman (1995): “Selection versus evolutionary adap-

tation: Learning and post-entry performance,” International Journal of Industrial Or-
ganization, 13, 501–522.

Bartelsman, E. J. and P. J. Dhrymes (1998): “Productivity dynamics: US manu-

facturing plants, 1972–1986,” Journal of Productivity Analysis, 9, 5–34.

Bartelsman, E. J. and M. Doms (2000): “Understanding productivity: Lessons from

longitudinal microdata,” Journal of Economic literature, 38, 569–594.

Disney, R., J. Haskel, and Y. Heden (2003): “Entry, exit and establishment survival
in UK manufacturing,” The Journal of Industrial Economics, 51, 91–112.

Dosi, G. (2007): “Statistical regularities in the evolution of industries. a guide through

some evidence and challenges for the theory,” in Perspectives on Innovation, ed. by
F. Malerba and S. Brusoni, Cambridge University Press, 153–186.

Dosi, G., M. Grazzi, L. Marengo, and S. Settepanella (2016): “Production

Theory: Accounting for Firm Heterogeneity and Technical Change,” The Journal of
Industrial Economics, 64, 875–907.

Dosi, G., D. Moschella, E. Pugliese, and F. Tamagni (2015): “Productivity,

market selection, and corporate growth: comparative evidence across US and Europe,”
Small Business Economics, 45, 643–672.

Dwyer, D. W. (1998): “Technology locks, creative destruction, and nonconvergence in

productivity levels,” Review of Economic Dynamics, 1, 430–473.

Foster, L., J. C. Haltiwanger, and C. J. Krizan (2001): “Aggregate productivity

growth: lessons from microeconomic evidence,” in New Developments in Productivity
Analysis, ed. by C. R. Hulten, E. R. Dean, and M. J. Harper, University of Chicago
Press, 303–372.

Griliches, Z. and J. Mairesse (1999): “Production functions: the search for identi-
fication,” in Econometrics and Economic Theory in the Twentieth Century: The Rag-
ner Frisch Centennial Symposium, ed. by S. Strøm, Cambridge: Cambridge University
Press, 169–203.

Grilliches, Z. and H. Regev (1995): “Productivity and firm turnover in Israeli

industry,” Journal of Econometrics, 65, 175–203.

21
Haltiwanger, J. (1997): “Measuring and analyzing aggregate fluctuations: the im-
portance of building from microeconomic evidence,” Federal Reserve Bank of St. Louis
Review, 79, 55–78.

Hildenbrand, W. (1981): “Short-Run Production Functions Based on Microdata,”

Econometrica, 49, 1095–1125.

Koopmans, T. C. (1977): “5 Examples of Production Relations,” The Microeconomic

Foundations of Macroeconomics, 144.

Levinsohn, J. and A. Petrin (2003): “Estimating production functions using inputs

to control for unobservables,” The Review of Economic Studies, 70, 317–341.

Liu, L. and J. R. Tybout (1996): “Productivity growth in Chile and Colombia: the
role of entry, exit, and learning,” in Industrial Evolution in Developing Countries:
Micro Patterns of Turnover, Productivity and Market Structure, ed. by M. J. Roberts
and J. R. Tybout, Oxford: Oxford University Press, 73–103.

Melitz, M. J. and S. Polanec (2015): “Dynamic Olley–Pakes productivity decom-

position with entry and exit,” The RAND Journal of Economics, 46, 362–375.

Olley, G. S. and A. Pakes (1996): “The dynamics of productivity in the telecom-

munications equipment industry,” Econometrica, 64, 1263–1297.

Petrin, A., B. P. Poi, and J. Levinsohn (2004): “Production function estimation

in Stata using inputs to control for unobservables,” Stata Journal, 4, 113–123.

Syverson, C. (2011): “What determines productivity?” Journal of Economic literature,

49, 326–365.

22
A Derivation of APG decomposition
In this Appendix we spell out the mathematical details behind the derivation of equations
(12) and (16).

Decomposition in equation (12) By the decomposition of productivity as P t =

t t
P
i∈C wi pi described in equation (10) we get that
X X
∆P t = P t − P t−1 = wit pti − wit−1 pt−1
i .
i∈C i∈C

The above equality holds if we add and subtract same quantity from its right side, that
is
X X X X X X
∆P t = wit pti − wit−1 pt−1
i + wit pt−1
i − wit pt−1
i + wit−1 pti − wit−1 pti
i∈C i∈C i∈C i∈C i∈C i∈C

which we can collect as

X wt−1 + wt X pt−1 + pt
i i i i
∆P t = (pti − pt−1
i ) + (wit − wit−1 )
i∈C
2 i∈C
2
which becomes equation (12)
X X
∆P t = w̄i ∆pti + p̄i ∆wit
i∈C i∈C
| {z } | {z }
W ithin Between
wit−1 +wit pt−1 +pti
using notation w̄i = 2
, p̄i = i
2
and ∆pti = pti − pt−1 t t t−1
i , ∆wi = wi − wi .

Decomposition in equation (16) Following decomposition

X X
∆P t = w̄i ∆pti + p̄i ∆wit
i∈C i∈C

in equation (12), we further decompose the coefficient ∆wit = wit −wit−1 as follows. because
wit = sti hti (see equation (13)), we get equality
∆wit = sti hti − st−1
i hi
t−1

which can be modified adding and subtracting same quantity in the right side as follows
∆wit = sti hti − sit−1 ht−1
i + (sti ht−1
i − sti ht−1 t−1 t t−1 t
i ) + (si hi − si hi ) .

Terms in the second part of the equality can be collected as

hit−1 + hti t st−1
i + sti t
∆wit = (si − st−1
i ) + (hi − ht−1
i ) .
2 2
ht−1 +ht st−1 +st
If we denote, as usual, by h̄i = i 2 i and s̄i = i 2 i the average sums, and by
∆sti = sti − sit−1 and ∆hti = hti − ht−1
i the variations, then ∆wit becomes
∆wit = h̄i ∆sti + s̄i ∆hti
which, replaced in equation (12), gives
X X X
∆P t = w̄i ∆pti + p̄i h̄i ∆sti + p̄i s̄i ∆hti
i∈C i∈C i∈C
| {z } | {z } | {z }
W ithin InputShares Heterogeneity

that is equation (16).

23
B APG decomposition with multiple outputs
In this appendix we give a quick description of the AP and its decomposition when firms
produce more than one output.

Notations and definitions During period t the production unit i, which is described
by the vector
ati = αi,1 t t t t
∈ Rm+n


, · · · , αi,m , αi,m+1 , · · · αi,m+n + , (20)


produces ati,out = αi,m+1 t t
, · · · αi,m+n units of n kinds of outputs by means of ati,in =
t t


αi,1 , · · · , αi,m units of m kinds of inputs. Denote by I t the set of all firms within one
industry at time t. Then the set of vectors {ati }i∈I t ∈ Rm+n + represents the production
activities of firms within the same industry at time t. Thus the aggregate (industry)
production activity dt is the sum of individual firm production activity and can be written
as

dt = β1t , · · · , βmt t t


, βm+1 , · · · , βm+n
!
X X X X
t t t t
= αi,1 ,··· , αi,m , αi,m+1 ··· , αi,m+n ∈ Rm+n
+ . (21)
i∈I t i∈I t i∈I t i∈I t

If we denote by

prout : Rm+n → Rn
ati 7→ ati,out

and by

prin : Rm+n → Rn
ati 7→ ati,in

the projections of the production activities ati (analogously of the industry vector dt ) on
the spaces of outputs and inputs respectively, then the formula to compute the industry
and firms productivities become

||prout (dt ) ||
P t := tg θ(dt ) = (22)
||prin (dt ) ||

and

||prout (ati ) ||
pti := tg θ(ati ) = (23)
||prin (ati ) ||
respectively, where θ(.) denotes the angle of vectors dt and ati with the space of inputs.
Notice that, because
P in this case output ati,out is a multidimensional vector then, in general,
t t
||prout (d ) || 6= i∈I t ||prout (ai ) ||, unless all output vectors are proportional or there is
only one output. If we denote by ϕti the angle formed by the vectors prin (ati ) and prin (dt )
and by σit the angle formed by the vectors prout (ati ) and prout (dt ), we get that
X
||prout dt || = ||prout ati || cos σit




(24)
i∈I t

24
which replaced into equation (22) gives
t t
P t t
i∈I t (||prout (ai ) || cos σi )
X 
t t ||prin (ai ) || ||prout (ai ) ||
P = = cos σi ,
||prin (dt ) || t
||prin (dt ) || ||prin (ati ) ||
i∈I

that is X
Pt = uti pti
i∈I t

where the “weight” coefficient

uti := kit · wit
is defined as the product of the output-based-homogeneity measure
kit := cos σit
and the input-based-weight
||prin (ati ) ||
wit := .
||prin (dt ) ||
Notice that kit is a decreasing function of σit when σit ∈ 0, π2 . That is smaller σit ’s corre-
 

spond to bigger kit ’s and indicate that the vector prout (ati ) is closer to the vector prout (dt ),
i.e. less output-based-heterogeneity. The fact that more output-based-homogeneity co-
incides with bigger kit explains why we name kit as output-based-homogeneity measure.

Decomposing the aggregate industry growth Let’s denote by C, N , and X the

sets of continues, entering, and exiting firms respectively. APG from time t − 1 to time
t is given by
X X X X
∆P t = ūi ∆pti + p̄i ∆uti + uti pti − ut−1
i pi
t−1
. (25)
i∈C i∈C i∈N i∈X

where for any variable xt ∈ R at time t, operator ∆ represents its change from t − 1 to
t t−1
t, i.e. ∆xt ≡ xt − xt−1 , and x̄ ≡ x +x2 . We can further decompose ∆uti as
kit−1 + kit t t−1 wit−1 + wit t
∆uti = (wi − wi ) + (ki − kit−1 )
2 2
= k̄i ∆wit + w̄i ∆kit
= k̄i h̄i ∆sti + s̄i ∆hti + w̄i ∆kit


(26)
where the last step substitutes the decomposition of ∆wit = wit − wit−1 as indicated in
appendix A. Finally, by substituting (26) into (25) we have
X X X X X X
∆P t = ūi ∆pti + p̄i k̄i h̄i ∆sti + p̄i k̄i s̄i ∆hti + p̄i w̄i ∆kit + uti pti − ut−1
i pi
t−1
.
i∈C i∈C i∈C i∈C i∈N i∈X
| {z } | {z } | {z } | {z } | {z }
W ithin InputShares Heterogeneity Homoout N etEntry
(27)
Notice that when n = 1, i.e. there is only one output, the angle between the individual σit
output vector and the aggregate output vector degenerates to 0. Thus for all firm i over
all time t, we have
kit = cos σit = 1
and thus k̄i = 1, ∆kit = 0 and uti = wit . As a result, the decomposition (27) degenerates
to (19).

25