Gibrat's Law and Diversification

Industrial and Corporate Change, Volume 15, Number 5, pp.
847–875
doi:10.1093/icc/dtl019
Advance Access published August 17, 2006
Downloaded from https://academic.oup.com/icc/article-abstract/15/5/847/733502 by University of Patras, Library & Information Service user on 28 March 2019
Gibrat’s Law and diversification
Giulio Bottazzi and Angelo Secchi
This article presents an analysis of the growth patterns of the worldwide top 200
firms in the pharmaceutical industry. A test of the Gibrat’s Law of Proportionate
Effect is performed and we find, in line with previous literature, a violation con-
cerning the variance of growth. Using disaggregated data on sub-markets, we are
however able to show that this violation can be completely accounted for by a
diversification effect, namely a scale relation between the number of sub-markets
in which a firm is active and its size. To interpret these findings, we propose a
stochastic branching model of firm diversification consistent with a notion of
cumulative corporate competences.
1. Introduction
This work reassesses the stylized fact concerning the relationship between variance in
corporate growth rates and corporate size employing empirical evidence from the
world pharmaceutical industry. Using statistical analyses based on novel disaggre-
gated data, we are able to show that this relation is completely explained by differences
in the diversification structure of firms belonging to different size classes. Moreover,
to interpret these findings, we propose a stochastic branching model of firm diversifi-
cation which tries to account for the existence of “scope economies to diversification.”
The point of departure of the present work is the “Law of Proportionate Effect”
proposed by Robert Gibrat and postulating the independence of the growth rates of a
firm from its size (cf. Gibrat, 1931; Hart and Prais, 1956; Ijiri and Simon, 1977;
Geroski, 2000, and the review by Sutton, 1997). Regarding the validity of this law, a
rich and growing body of empirical analysis seems to suggest two rather robust con-
clusions. On the one hand, at least when young and relatively small firms are
neglected, the Gibrat hypothesis is confirmed by the lack of any relationship between
the (log) size of firms and their average rate of growth (for a summary review of this
literature, see Lotti et al., 2003). On the other hand, it is violated by a clear negative
dependence of the growth variance on size.
It has been suggested that a natural explanation for the reduction of growth vari-
ance with size could be a sort of “portfolio” effect (see, e.g., Hymer and Pashigian,
1962). The basic idea is that a firm can be described as a collection of “atomic” com-
ponents (line of productions, plants, etc.) of roughly the same size whose number
would be proportional to the size of the firm. Under the assumption that the growth
© The Author 2006. Published by Oxford University Press on behalf of Associazione ICC. All rights reserved.
848 G. Bottazzi and A. Secchi
processes for different components are independent, the Central Limit Theorem
would predict a decrease in the variance of the firm growth rates proportional to the
inverse square root of its size. However, since the early studies, the empirical evidence
clearly shows that the observed dependence is milder. The failure of the Central Limit
Theorem has been imputed to the existence of a relation among the firm components
that makes the aggregation of the “atomic” growth not simply additive1 (see, e.g.,
Mansfield, 1962 and Boeri, 1989) or to some complex intra-firm hierarchical struc-
ture supposedly affecting the way in which the dynamics of single components con-
tribute to the aggregate firm growth (Amaral et al., 1997).
Conversely, in this article we show that if one correctly identifies the “atomic”
contributions to growth and the scaling relation between their number and the size
of the firm, the Central Limit Theorem does a good job in explaining the observed
Gibrat violation, without the need of any assumption on the internal structure of
the firm.
The remainder of the article is organized as follows. In section 2, we present a brief
overview of the database used and of the statistical properties of both the firm size dis-
tribution and growth process. Section 3 performs an analysis of the scale relation
between the variance of growth rates and the size of the firm. In section 4, we show
that this relation is totally due to a similar scale relation that exists between the size of
the firm and its diversification structure. In section 5, we propose three stochastic
models of firm diversification: the first based on an idea of pure random accumula-
tion of diversified activities, the second accounting for an “economy of scale” effect on
the ability to enter new markets, and the last assuming the existence of an economy of
scope to diversification. These models are validated against empirical data in section 6
while section 7 concludes.
2. The database
The statistical analyses presented in this article are based on data from the Pharmaceuti-
cal Industry Database (PHID),2 which covers sales figures for the top incumbent firms
in seven western markets (USA, United Kingdom, France, Germany, Spain, Italy, and
Canada) over a time window of eleven years, from 1987 to 1997. This original data-
base contains information about sales of 7654 different drugs. These single product sales
are aggregated in different micro-classes according to the Anatomical Classification
1
However, the sole introduction of cross-correlation in growth rates between the different compo-
nents would not, in general, break the inverse–square relation between the variance of growth and
the size implied by the Central Limit Theorem.
2
This database is developed as part of the EPRIS project for the development of a comprehensive
analysis of the structure and evolution of the European Pharmaceutical Regulation and Innovation
Systems (cf. http://www.epris.it).
Gibrat’s Law and diversification 849
System3 (ACS) developed and maintained by the European Pharmaceutical Marketing
Research Association (EphMRA) together with the Pharmaceutical Business Intelligence
and Research Group (PBIRG) (for details, see EphMRA Guidelines, 2004).
For the purpose of the ACS, a product represents a discrete pharmacological pack
or unit that can be dispensed, prescribed, or purchased. In the ACS, products are clas-
sified in different groups at four different levels:
First level: The first level is based on a letter for the anatomical group (e.g., B for
Blood and Blood-forming organs); there are 16 different main groups.
Second level: The second level is used to regroup several classes together in order to
classify according to indication (e.g., B1 antithrombotics), therapeutic substance
group (e.g., J1 antibiotics), and anatomical system (e.g., S1 ophthalmological).
Third level: The third level describes a specific group of products within the second
level. This specification can be a chemical structure (e.g., J1D cephalosporins) or it can
describe an indication (e.g., N2C antimigraine) or a method of action (e.g., A3F
gastroprokinetics).
Fourth level: The fourth level gives more details of the third level on formulation,
chemical description, and mode of action.
The EphMRA/PBIRG classification has as a primary objective to satisfy marketing
needs of the pharmaceuticals companies. This observation together with the fact that
the substitutability of products belonging to different classes is practically zero sug-
gests that this classification is able to clearly identify different sub-markets.
In what follows we consider a balanced panel comprising the top worldwide 198
firms that are present, for the whole period 1987–1997, in all the countries covered
by the database. In this way, we restrict our analysis only to firms that can be con-
sidered multinational companies operating on a worldwide scale. This choice is
made in order to avoid the obvious problems that would arise from the pooling of
firms based in different countries and operating in different national markets.
Moreover, because this work is focused on the processes of internal growth and
diversification, in order to take into account mergers and acquisitions during the
period of observation, we have constructed “super firms” which correspond to the
end-of-period actual entity (so, e.g., if any two firms merged during the observed
history, we consider them merged from the start). This procedure clearly biases
intertemporal comparisons on actual size distribution, but it helps to highlight
those changes in the distribution which are because of processes of intra-market
competition and inter-market diversification.
3
Note that there exists an alternative, even if very similar, classification system called Anatomical
Therapeutic Chemical (ATC) scheme adapted from the EphRMA scheme by the World Health
Organization for its own needs. A direct comparison between the two schemes implies several diffi-
culties and should be in principle avoided.
Tables 1 and 2 give a general overview of the PHID databank. In the first table,
we report some descriptive statistics of the size (total sales) of firms and of the
number of markets (at the fourth level of the ACS scheme) in which they are active,
that is, markets in which they possess non-zero market shares. As sales are
expressed in US dollar at current prices and current exchange rates, the observed
upward trend in firm size is largely because of a nominal effect. Note that a
remarkable degree of firm heterogeneity emerges both in terms of size (in each
year, larger firms are several times greater than smaller ones) and in terms of
number of markets (some firms are active in very few sub-markets, whereas other
sell products in hundreds of them). This heterogeneity does not seem to decrease
over time. A summary account of the diversification structure of firms in our data-
base is provided in Figure 1, where the empirical probability density of the number
of active sub-markets p(N) is reported. For each value of N, the quantity p(N) is
computed as the fraction of firms, over the entire population, possessing exactly N
active sub-markets.
Table 2 reports descriptive statistics of sub-markets, providing information about
their number, their size, and the number of active firms. These statistics are separately
reported for each level of the ACS.
Table 1 Descriptive statistics of firm size (total sales) and number of sub-markets (four digits)
of the PHID database
Year Firm size Number of markets
Mean Std Min Max Mean Std Min Max
1987 2.36 ´ 105 5.49 ´ 105 2.29 ´ 102 3.08 ´ 106 43.2 46.6 1 220
5 5 2
1988 2.64 ´ 10 6.14 ´ 10 3.06 ´ 10 3.36 ´ 106 43.4 46.2 1 220
1989 2.95 ´ 105 6.91 ´ 105 5.07 ´ 102 3.87 ´ 106 43.6 45.9 1 221
1990 3.31 ´ 105 7.73 ´ 105 7.25 ´ 102 4.20 ´ 106 43.7 45.6 1 220
1991 3.78 ´ 105 8.77 ´ 105 8.18 ´ 102 5.14 ´ 106 44.6 45.9 1 220
1992 4.39 ´ 105 1.03 ´ 106 1.87 ´ 103 6.10 ´ 106 44.0 44.9 1 218
1993 4.57 ´ 105 1.07 ´ 106 1.92 ´ 103 6.60 ´ 106 44.3 44.7 1 221
1994 4.79 ´ 105 1.11 ´ 106 1.88 ´ 103 7.00 ´ 106 44.3 44.4 1 221
1995 5.18 ´ 105 1.19 ´ 106 2.60 ´ 103 7.36 ´ 106 44.4 43.8 2 220
1996 5.63 ´ 105 1.30 ´ 106 2.91 ´ 103 7.63 ´ 106 44.6 43.4 2 218
1997 6.13 ´ 105 1.41 ´ 106 3.44 ´ 103 7.57 ´ 106 44.8 43.3 2 216
Total sales are expressed in thousands of US dollars at current prices and current exchange
rates.
PHID, Pharmaceutical Industry Database.
Table 2 Descriptive statistics of sub-markets size in term of total sales (at current price and
exchange rate) and number of firms
Level Number Size of markets Number of firms

of markets
Mean Std Min Max Mean Std Min Max
First 12 6.73 ´ 106 5.45 ´ 106 1.25 ´ 106 1.80 ´ 107 139.83 27.17 92 176
Second 74 1.11 ´ 106 1.52 ´ 106 8.32 ´ 101 8.29 ´ 106 70.63 18.63 1 154
Third 237 3.59 ´ 105 6.70 ´ 105 6.00 ´ 100 5.75 ´ 106 37.09 28.00 1 128
Fourth 393 2.19 ´ 105 4.64 ´ 105 4.00 ´ 100 3.81 ´ 106 27.64 23.18 1 110
P(N)
0.02
0.01
0
0 50 100 150 200
N
Figure 1 Empirical probability density P(N) of the number of active sub-markets N.

Sub-markets are considered at the fourth level of ACS classification. After checking for
stationarity, observations in different years were pooled together.
3. The variance–size relation

In this section, we explore the relationship between the growth dynamics of a firm and
its size. In particular, we compare the central moments of the growth rates distribu-
tions of firms belonging to different size classes. This is only one of many possible
aspects of the growth dynamics of firms, but, as it will become clear later, its analysis is
particularly relevant for the present discussion. A more complete statistical descrip-
tion of the data under study is provided in the paper by Bottazzi and Secchi (2005),
where the analysis of the evolution of firm size distribution and of the autoregressive
structure of growth process is presented, both at aggregate and disaggregated levels.
Complementary studies on the same type of data are presented by Bottazzi et al.
(2001) and Matia et al. (2004).
Let Si(t) be the size (total sales) of firm i at time t. Because of the presence of a
nominal trend, we introduce a normalized version of the (log) size obtained by sub-
tracting from each observation at time t the average (log) size of all the firms at the
same time,
1 N
si (t ) = log(Si (t )) - å log(Si (t )),
N i =1
(1)
where N stands for the total number of firms in the panel. Bottazzi and Secchi
(2005) show that the firm size distribution concluded that it is characterized by two
noticeable features. First, the density seems to display a remarkable degree of sta-
tionarity in time and, second, it shows a clear bimodal shape confirmed by a statist-
ical test of multimodality. Figure 2 reports the empirical density together with its
kernel estimate.
In order to investigate the relation between the central moments of the growth
rates density and the firm size, we split the (log) sizes si(t) into 50 bins (quantiles). We
then compute the mean, standard deviation, skewness, and kurtosis of the growth
rates gi(t) of the firms in each bin and we fit the linear relation
Y( g | s) = a + bs, (2)
where Y( g | s) is in turn the mean, the log of the standard deviation, the skewness,
and the kurtosis of the growth rates of the firms belonging to the bin whose average
size is s. Results of these fits are reported in Table 3.
First, no clear relation emerges between the average growth rate and the size of the
firm. Second, we observe a strongly significant negative relation between the size and
the standard deviation of growth rates: the estimated slope is b = -0.21 with a stand-
ard error of 0.02. This coefficient is strikingly similar to the one found in the US man-
ufacturing industry by Amaral et al. (1997) and Bottazzi and Secchi (2003).
Moreover, in Figure 3 we report, on a log–log scale, the growth rate standard devi-
ation for each bin versus the average bin size together with the linear fit in (2) and a
non-parametric local regression obtained with a smoothing kernel method.4 As can be
seen, the fit provides a good description of the variance–size relation and the
4
See Pagan and Ullah (1999, Section 3.2), for a description of the method.
0.3
Empirical density
Kernel Est.
0.25
0.2
0.15
0.1
0.05
0
-6 -4 -2 0 2 4 6
s
Figure 2 Empirical density of normalized (log) sales together with a kernel estimate. Data in
different years are pooled together. The kernel used is Gaussian and the bandwidth computed
according to Silverman (1986, Section 3.4).
Table 3 Slopes and standard errors of the linear fit (2) of the
relation between firm size and the first four moments of the
growth rates density (see text for details)
Relation Slope Std error
Size—mean -0.11 0.08

Size—std -0.21 0.02
Size—skewness -0.12 0.09
Size—kurtosis -0.03 0.11
non-parametric regression remains near it for the whole support of s. Finally, we do

not find any evidence that skewness or kurtosis scale with the size of the firm.
In conclusion, the dependence of the growth rates on the size of the firm is com-
pletely described by the variance–size relation
bsi
s( g i ) ~ e . (3)
-0.5
-1
-1.5
σ(gi)
-2
-2.5
-3
-3.5
-3 -2 -1 0 1 2 3 4 5
si
Figure 3 The relation between the standard deviation of growth rates (on a log scale) and the
size of the firm, together with the fit log(s (g | s)) = a + b s and a non-parametric regression of
the relation between log(s (g | s)) and s.
Note that the assumption that the firm’s aggregate sales are the sum of sales of many
independent “atomic” components (plants, business units, or sub-markets), roughly
of the same size, would imply a value for b = 0.5. Conversely, the estimated value for
b is significantly lower, suggesting that this simplistic model cannot be applied.
Hence, in order to understand what is the origin of the observed variance–size relation,
we have to “dissect” the aggregate growth rate in its different pieces and look sepa-
rately at their different contributions. This analysis is performed in the next section.
4. Firm diversification and the violation of Gibrat’s Law

This section is devoted to the detailed analysis of the “variance–size” relation described
in equation (3). We will identify different possible sources of this relation and analyze
them in turn. We will conclude that the only possible explanation of this effect rests in
the diversification structure of the firm. The relevance of our results is twofold: on the
one hand, we are able to identify what part of the business structure of the firm is
responsible for the modification of the firm growth rates variance and, on the other
hand, we are able to show that several possible elements, which are in principle likely
to affect the performances of firms, result indeed irrelevant.
As mentioned in section 2, one of the peculiar features of the PHID database is the
possibility of disaggregating the sales figures of each single firm until the fourth level
of the ACS system. This level of disaggregation allows to identify sub-markets that are
“specific” enough to be considered, roughly speaking, the loci of competition among
firms: the products belonging to a given sub-market possess similar therapeutic char-
acteristics and can then be considered substitutable, whereas products belonging to
different sub-markets are usually targeted to different pathologies. Moreover, it is
widely acknowledged that this disaggregation level is the one at which single research
projects develop (see, for instance, Nightingale and Mahdi, 2006). This evidence sup-
ports the hypothesis that the different sub-markets provide the natural scale at which
firm diversification should be studied. Hence, we use data disaggregated at the fourth
level to study the diversification structure of the firms, looking at the scope of their
activities, at the relative size with which a firm is present in different sub-markets, and
at the possible relation between the performances of a firm in the different sub-mar-
kets in which it operates. To maintain a notation similar to previous sections, let Si,j(t)
be the size (sales) of firm i in sub-market j at time t. We define a “normalized” variable
Si , j
Si , j = N N (4)
å h=1 Sh
to get rid of any nominal effect and the associated sectoral growth rates5
Si , j (t + 1)
Gi , j = - 1. (5)
S (t )
i, j
As a preliminary step, we analyze the cross-correlation in sectoral growth rates. Let

Gi , j be the sample average of the growth rate of firm i in sub-market j in the ten years
covered by our database and si , j the sample standard deviation. If the firm is not act-
ive on the sub-market j for the whole period, we only count the couple of consecutive
years in which it is present, so that a growth rate can be defined. Let Ti , j be the set of
years in which it is possible to define the rate of growth of firm i in sub-market j,
that is
Ti , j = t | t Î{1,…,10}, Si , j (t ) > 0, Si , j (t + 1) > 0

{ } (6)
and let Ti , j be the number of elements of Ti , j . The covariance of the growth rates of
firm i in sub-markets j and j¢ with j ¹ j¢ is defined as
5
This definition is in principle different from the one used in the previous section, indeed it is Gij(t) =
egij(t)-1. Because gij(t) << 1, however, Gij(t) can be safely approximated with gij(t) using a Taylor’s
expansion argument.
1
ci , j , j ¢ = å éGi , j (t ) - Gi , j ùû éëGi , j ¢ (t ) - Gi , j ¢ ùû
Ti , j t ÎT ë
(7)
i, j
and the associated cross-correlation coefficient reads
ci , j , j ¢
ri , j , j ¢ = . (8)
si , j si , j ¢
We ignore sub-markets in which the firm never operates for three consecutive years,
in order to have a meaningful definition of the growth rates sample variance. For the
nature of the data, the variables in equations (7) and (8) are based on estimates of the
central moments obtained with very small samples (at most 10 observations) and are,
consequently, extremely “noisy.” Nevertheless, we are not interested in the cross-
correlation existing between the performances of a firm in two specific sub-markets,
but in the average value, describing the degree of cross-correlation among all the active
sub-markets of a firm. Then we consider the average value of the quantity defined in
equation (8) for a given firm i
1
ri = å
M i (M i - 1) j ÎM
å ri , j , j ¢ , (9)
i j ¢ ÎMi , j ¢¹ j
where Mi denotes the set of sub-markets where firm i operates for more than three
consecutive years and Mi is the number of these sub-markets.
The distribution of the coefficients r on the population of firms has a mean value of
0.030 and a standard deviation of 0.120. Very surprisingly, the coefficients are sharply
centered around zero and the average cross-correlation is absent not only “on aver-
age” but for all the firms in the database. We can conclude that the growth processes
of a firm in different active sub-markets are independent processes.
The previous result will prove extremely useful in the analysis of the relation
between the aggregate size of a firm and the variance of its growth rates. For this pur-
pose, rewrite the aggregate growth rate of firm i at time t, Gi(t), as
Si (t + 1) Si , j (t + 1) 1
Gi (t ) = -1= å -1= åGi , j (t )Di , j (t ), (10)
Si (t ) j S (t )
i N i (t ) j
where Ni(t) is the number of active sub-markets (on which firm i possesses, at time t,
non-zero sales) and where
Si , j (t )
D i , j (t ) = N i (t ) . (11)
S (t )
i
In equation (10) the contribution of each sub-market to the total growth has been
factorized in two terms: Gi,j(t), which represents the actual growth of firm i in sub-
market j, and the coefficient Di,j(t), which describes the relative weight of the sector j
with respect to the aggregate sales of firm i.
Now the previously observed independence of firm growth processes in different
active sub-markets has an important implication: the variance of the aggregate (firm)
growth Gi(t) can be obtained, according to equation (10), adding the variance of the
growth rates in each sub-market
é 1 ù
vari ,t [Gi (t )]= åvari ,t êGi , j (t )D i , j (t ) ú, (12)
j ë N i (t ) û
where vari,t stands for the variance of the distribution computed, consistently with the
previous analysis, using the complete panel (all the firms at all the time steps).
From equation (12) it is clear that the relationship between the variance of aggre-
gate growth rates and the size of the firm can possibly have different origins. Indeed,
the right-hand side contains the product of three different variables that could all be
plausibly related to the size of the firm: a dependence of Gi,j on aggregate size would
imply the existence of increasing returns to scale, whereas an analogous dependence
for Di,j would suggest different diversification structures, plausibly related to different
degrees of “corporate coherence” (in the spirit of Teece et al., 1994), for firms with
different market power. Finally, the “variance–size” relation could be originated by a
dependence of Ni on aggregate size. In the rest of this section we investigate, in turn,
these possible sources.
First, consider the relation between the sectoral growth rates Gi,j(t) and the total
size of firm, si(t), as defined in equation (1). Dividing all the firms in different equi-
populated size classes, Figure 4 shows the relation between the average size of each
class si and the average sub-markets growth rates of all the firms belonging to this
class. It is apparent that the averages of growth rates for different sizes do not differ
significantly. Also the corresponding standard deviations, reported in Figure 5, do not
display any clear relation with firm size.
A second possible origin of the “variance–size” relation is the link between the size
of the firm and its “diversification pattern,” that is, the way in which the total sales of
the firm are distributed on its active markets. One can indeed think of different ways
in which the activities of a company can possibly be distributed on the markets in
which it operates. It can be the case that a given company realizes similar sales in the
different active markets, so that the aggregate company size is the sum of several com-
ponents of a comparable size. Conversely, it is also possible that the diversification
structure of a company is made by a very strong market, where the company realizes
almost all its sales, and by a group of smaller markets, where the company is present
but with relatively low sales. Relatedly, it is interesting to check if the diversification
pattern of a firm bears any relation with the firm size.
10
Mean
-2
-3 -2 -1 0 1 2 3 4
si
Figure 4 The relation between the average of sub-markets (at fourth level) growth rates and
firm size, si. Bars represent two standard errors.
For the present analysis, we need an index that captures the degree of heterogeneity
in the diversification pattern of a firm. Starting from the coefficient in equation (11)
we define a “diversification heterogeneity” index as
N (t ) é ù
2
å j =1i ë D i , j (t ) - 1û (13)
D i (t ) = .
N i (t )[N i (t ) - 1]
It is easy to see that D i (t ) ranges from 0 to 1. In fact, if firm i distributes its sales
evenly among its active markets, it is
Si , j (t ) 1
= "j ,
S (t ) N i (t )
i
hence Di,j = 1 and then D i (t ) = 0 . On the contrary, if the sales of the firm are com-
pletely concentrated in a single market, say j´, whereas it retains negligible positions
on the others, we have
Si , j ¢ (t ) Si , j (t )
~ 1 and ~0 "j ¹ j ¢ ,
Si (t ) Si (t )
so that the index D takes its maximum value of 1.

10
Std
-2
-3 -2 -1 0 1 2 3 4
si
Figure 5 The relation between standard deviation of sub-markets (at fourth level) growth rates
and firm size, Si. Bars represent two standard errors for the sample standard deviation.
In Figure 6 we report a scatter plot of the diversification index D i (t ) versus the

size of the firm si(t) for all the years t in our database. As can be seen, the support
of the D ’s spans almost the whole range [0,1) but the great majority of firms has
values around 0.3. The scatter plot appears like an amorphous blob, suggesting
the lack of any relation between the size of the firm and its diversification pat-
tern. To verify this impression, we fit a linear relation between si and D finding
a slope of -0.025 (with a standard error of 0.0018). The negative slope is signific-
ant and is essentially because of the presence of few large values for D associated with
small- and average-sized firms. Nevertheless, the value of the slope is so small that
we do not dare to give it any economic relevance. Hence, we conclude that in the
pharmaceutical industry we do not observe any peculiar difference between big
and small firms in the way in which they distribute their sales among different
sub-markets.
Finally, consider the contribution from the last term in the left-hand side of equa-
tion (12), namely the relation between the number of sub-markets in which a firm is
active and its size. In Figure 7 we plot the average number of active sub-markets of the
firms belonging to different size bins against the bin average (log) size. A clear positive
exponential relation between the two variables emerges. Then, we fit a linear relation
between the firm (log) aggregated size and the log of the number of its active markets
1
0.9
0.8
0.7
0.6
0.5
Δ
0.4
0.3
0.2
0.1
0
-6 -4 -2 0 2 4 6
si
Figure 6 Scatter plot of the diversification heterogeneity index, D i , versus firm size, Si.
All the firms and all the years are showed.
log(N i (t )) ~ l log(Si (t )) + q. (14)
We obtain a strongly significant slope l = 0.35 with a standard error of 0.01. In the
inset of Figure 7 we report the linear fit in equation (14).
In conclusion, the term Ni(t), the number of active sub-market a firm possesses,
seems the only responsible for the observed dependence of the variance of firm
growth rates on the aggregate size. Note also that the Central Limit Theorem would
predict a relation between the exponent in equation (3) and the slope in the inset of
Figure 7 of the form b = -l /2 , which is in good agreement with our data. This means
that the explanation of the relationship between the variance of growth rates distribu-
tion and the size of firm based on the Central Limit Theorem is valid, as long as one
considers the actual number of sub-markets a firm operates in, instead of assuming
that this number is somehow proportional to the size of the firm.
It remains however to explain why the relation between the number of active sub-
markets, N, and the size of the firm, S, is so well described by a power-like function
N (S)~S l and to try to give some meaning (if any) to the parameter l . This is what
we shall do in the next section where we propose three different models of the firm
diversification process. Although the first two describe the observed diversification
structure as the outcome of successive and uncorrelated events, the last is essentially based
on the assumption of a self-reinforcing effect cumulatively driving the “proliferation”
of sub-markets a firm operates in. We will see that only the latter model is, at least in
200
Lin.
180 Exp.
Log(n i)
160
140 λ=0.35
120
ni
100
si
80
60
40
20
-4 -3 -2 -1 0 1 2 3 4 5 6
si
Figure 7 The (log) number of active sub-markets versus (log) firm size together with a linear
(dotted line) and an exponential (solid line) fit.
part, able to account for our empirical findings, thus providing a simple but sugges-
tive interpretation of Figure 7.
5. Stochastic models of firm diversification

The previous analysis shows that larger firms are present, on average, on a greater
number of sub-markets. For the purposes of this section, it is useful to read this state-
ment in a dynamical fashion: as firms grow, their activities become more and more
diversified because they sell products in an increasing number of sub-markets. In
what follows, we shall try to capture the diversification behavior of firms with a model
describing how the number of active markets, or better, how the probability of having
a given number of active markets changes as the firm grows. Given the nature and
limitations of our data, we consider the firm size as the independent variable and ana-
lyze the changes of diversification patterns in accordance with changes in the former.6
The object of the following analysis thus becomes the probability that a firm which
6
Note that this approach is distinct from considering a dynamical process in time, that is, a model
describing the evolution of diversification patterns “as time goes by.” Even if this kind of model
presents many interesting features, we prefer to study a “cross-sectional” model because the former
turns out to be sensitive to finer specifications of growth dynamics which can be hardly checked
against available data.
possesses n active markets when its (log) size is s will possess m active markets when its
(log) size is s ¢ ³ s, P(n, s; m, s ¢).
In the following we present three different models intended to describe the evolu-
tion of the firm diversification structure discussed in the previous section. They are
characterized by an increasing complexity and are built on different sets of assump-
tions about the nature of the transition probability P(n, s; m, s ¢) . In the first model,
the number of active sectors a firm possesses is described as the outcome of the ran-
dom arrival, during the firm’s history, of independent diversification events. The
arrival dynamics is described through a simple, linear Poisson process. The second
model generalizes this assumption, introducing a relationship between the intensity
with which a firm undertakes diversification events and its size. We account for the
presence of possible economies of scale to diversification using a non-linear Poisson
process with a generic arrival intensity function. Finally, we propose a model where
the arrival of a new diversification event depends on the number of already active sec-
tors. In other terms, the probability that a firm increases the scope of its activity when
its size increases depends on the number of sectors in which it operates. We model
this idea of scope economy to diversification using a Yule arrival process.
Below we will show how different hypotheses about the underlying arrival process
map in different properties of the diversification dynamics. In section 6, these differ-
ent hypotheses will be tested against empirical observations.
5.1 Model I: random diversification

We start with a very simple model which can be considered the benchmark of our
investigation. It is rooted in a relatively small class of models which have attempted to
merge the diversification and the growth aspect of firm’s dynamics. This class of models,
collectively referred as “island models,” originally proposed by Simon (see Ijiri and
Simon, 1977 and the reassessment by Sutton, 1998), does not explicitly address the
diversification dynamics but rather considers that part of a growth process driven by a
successive capture of diverse islands, or business opportunities. In this way, the dis-
tinction between growth and diversification becomes blurred, but nevertheless two
assumptions are generally made concerning the latter. The first one concerns the
nature of the diversification events (i.e., the entry on a previously unexploited market)
and assumes that these events are seen as mutually independent shocks (investment
opportunities, technological discoveries, etc.) that can happen to firms along their his-
tories. The second assumption is that these shocks are uncorrelated in time, that is,
they could happen with the same constant probability at any moment, irrespectively
to the actual firm “state.” If one neglects the possibility of instantaneous multiple
shocks,7 these assumptions provide a complete definition of the transition probability
7
Equivalently: requires the property of orderliness for the associated stochastic process. For a general
discussion on the complete characterization of point processes, see Snyder (1975).
P(n, s; m, s + d) between two sizes differing by an infinitesimal quantity d
ìld + o(d ), m = n + 1,
P(n, s; m, s + d ) = í (15)
îo(d ), m > n + 1,
where l is the constant rate of arrival of diversification events (i.e., the average
number of diversification events per unit time). This form of the transition probabil-
ity defines the well-known Poisson process. If Pn(s) is the probability that a firm of
size s is active in n sub-markets, it must satisfy the (pure-birth) equations
ìï pn0 (s + d ) = pn0 (s)P(n0 , s; n0 , s + d ), n = n0 ,

í (16)
îï pn (s + d ) = pn -1 (s)P(n - 1, s; n, s + d ) + pn (s)P(n, s; n, s + d ), n > n0 ,
where n0 is the initial number of active sub-markets. Using equations (15) and (16)
and taking the limit d ® 0, one obtains a set of differential equations
ìï pn¢ (s) = - l pn (s) + l pn -1 (s), n > n0 ,

í p ¢ (s) = - l p (s), (17)
n > n0
îï n0 n0
with initial conditions
ì1, n = n0
pn ( g 0 ) = í (18)
î0, n ¹ n0 .
A pictorial illustration of this process is shown in Figure 8: the “blobs” stand for the
diversification event “black boxes” that a firm meets along its growth. As far as this
model is concerned, it is irrelevant what really happens inside the black box; what
matters is that after the shock the firm ends up with one more active sub-market.
Solving equation (17) with initial conditions (18) one obtains
n -n
l (s - s0 ) 0 - l(s - s0 ) (19)
pn (s) = e , n ³ n0 .
(n - n0 )!
From this expression, it is straightforward to derive the average number of active mar-
kets a firm possesses, m(s), and its variance s2(s)
+¥
m(s) = å n pn (s) = n0 + l (s - s0 ), (20)
n =0
+¥
s2 (s) = å[n - m(s)]2 pn (s) = l (s - s0 ). (21)
n =0
5.2 Model II: scale economy to diversification
The model based on linear Poisson process presented in the previous section would
predict a linear increase of the average number of active sub-markets with s [see equa-
tion (20)], a property that is clearly in contrast with the empirical results discussed in
the previous sections and illustrated in Figure 7. A straightforward way to account for
the fact that the number of active sub-markets grows more than linearly with the size
of the firm is to consider a non-linear Poisson process, in which the intensity of arrival
of new diversification events increases with the size of the firm. The second model that
we present uses this non-linear Poisson process to introduce an idea of scale economy
to diversification. The increase of the number of sub-markets with size is again
described as the successive undertaking of randomly arriving diversification events.
However, the probability of a firm to experience a diversification event depends now
on its size. According to this hypothesis, the probability in equation (15) is modified
to read
ì L(s)d + o(d ), m = n + 1,
P(n, s; m, s + d ) = í (22)
îo(d ), m > n + 1,
where the function L(s) characterizes the intensity with which a firm of size s engages
in a diversification event. We assume that L(s) > 0 , that is, that an increase in the
size of the firm is never associated with a decrease in the scope of its activities. Of
course, different economic phenomena can be captured assuming different functional
form for L . For instance, the idea of economies of scale and the consequent increased
diversification ability of larger firms would imply L ¢(s) £ 0 . Conversely, the assump-
tion that, at a given scale, the diversification in new markets becomes increasingly difficult
can be captured by assuming that for large enough s it is L ¢(s) £ 0 . Irrespectively of
the precise form of the intensity function, however, the above set of assumptions leads
to the (pure-birth) system of equations
ìï pn¢ (s) = - L(s) pn (s) + L(s) pn -1 (s), n > n0 , (23)

í p ¢ (s) = - L(s) p (s), n = n0 ,
îï n0 n0
with the same initial conditions as in equation (18). The solution of equation (23) reads
q(s; s0 )n - q(s ;s0 ) (24)

pn (s) = e ,
n!
where
s
q(s; s0 ) = ò dx L(x). (25)
s0
s1 s2 s3
Figure 8 Diversification by independent events. The inter-arrival “growths” si - si-1 are

exponentially distributed.
In this case, the average number of active sub-markets and their variance depend explic-
itly on the intensity function L(s) through the integral in equation (25) and read
m(s) = n0 + q(s; s0 ), (26)
s 2 (s) = q(s; s0 ). (27)
Note that because L(s) > 0 , the function q(s; s0) is an increasing function of the firm
size and, consequently, both the number of active markets and the variance of this
number do increase with the size of the firm.
5.3 Model III: scope economy to diversification

Model II discussed above is able to describe, with a proper choice of the intensity
function L(s) , the exponential growth of the number of active sectors with the size of
the firm. Indeed if one sets L(s) = e ls , then equation (26) can account for the
exponential increase of the number of sectors with size reported in Figure 7. This
same result, however, can be obtained also with a different approach. Instead of
assuming the existence of an “economy of scale” for the diversification activity of firms,
one can look for a different model that explicitly considers the existence of an explicit
relationship among the different diversification events. In other terms, one can try to
“open up the black boxes” which represent the diversification events of Figure 8. A
plausible interpretation of the nature of these events draws upon the extensive litera-
ture on the importance of technological capabilities and learning for corporate growth
and diversification (for general arguments, see Dosi, 1988 and Teece et al., 1994; more
specifically, on the pharmaceutical industry, see Mazzucato and Dosi, 2006). Indeed,
s1 s2 s3
Figure 9 Diversification by branching. The inter-arrival “growths” are no longer exponential;

the distance between two events decreases as the number of active markets increases.
the dynamics of the pharmaceutical industry is largely research-driven, so that prod-

uct innovations and imitations constitute essential aspects of the growth of firms and
generate an intense spreading process across different markets (Bottazzi et al., 2001).
Moreover, the development of technological capabilities appears to be an incremental
and cumulative process, involving, in the case of pharmaceuticals, also “horizontal”
innovative competences applicable across several therapeutic targets. Diversification
events are no exception: more often, a firm enters a new market when it has the tech-
nological capability of developing products for that market. It is then quite natural to
expect that these patterns of technological accumulation will shape the growth
dynamics of firms with the introduction of a “scope economy to diversification”: the
capability of a firm to enter a new market increases with its past successful diversifica-
tion events.8
A straightforward way of accounting for the presence of a scope economy to diver-
sification is by supposing that the diversification process proceeds as a branching pro-
cess, where each opened branch (sub-market) becomes eventually the origin of a new
branching (diversification) event. A picture of this process is shown in Figure 9 under
the (minimal) assumption that the branching is uniformly binary. If one neglects the
topology in sub-markets space that emerges from these successive branchings, but is
interested only in the actual number of these sub-markets, the process in Figure 9 can
8
Exceptions to the “technology-driven” diversification events are purely commercial distribution
agreements which, although not uncommon, are not the dominant driver of diversification for the
firms considered here. Nevertheless, their dynamics too can plausibly be characterized by some sort
of “scope economy,” such that their presence does not in principle spoil the present analysis.
be readily described: all the active sub-markets can be sources of a possible diversifica-
tion event à la Poisson, and equation (15) must be modified to read
ìn l d + o(d ), m = n + 1,
P(n, s; m, s + d ) = í (28)
îo(d ), m > n + 1,
neglecting again multiple instantaneous branchings. If pn(s)is the probability that a

firm of size s has n active markets it must satisfy the (pure-birth) set of equations
ìï pn (s + d ) = pn -1 (s)P(n - 1, s; n, s + d ) + pn (s)P(n, s; n, s + d ), n > n0 ,

í p (s + d ) = p (s)P(n , s; n , s + d ), (29)
n = n0 ,
îï n0 n0 0 0
where n0 is the initial number of active markets. Substituting in equation (29) the
definition given in equation (28) and taking the limit d ® 0, one obtains the set of dif-
ferential equations
ïì pn¢ (s) = -nl pn (s) + (n - 1)l pn -1 (s), n > n0 ,

í p ¢ (s) = -n l p (s), (30)
ïî n0 0 n0 n = n0 ,
with the usual initial conditions. This process is known as the Yule process and has
been originally proposed to explain the proliferation in time of animal species (for a
discussion, see Feller, 1968 and the references therein). System (30) can be solved (see
Appendix 1) to obtain the distribution
n - n0
æ n - 1 ö - n0l (s - s0 ) æ - l (s - s0 ) ö
pn (s) = ç e 1 - e , (31)
è n - n0 ÷ø è ø
defined for s ³ s0 and n ³ n0. In this case, the average number of active sub-markets
and their variance (see Appendix 2) are
λ (s - s0 )
m(s) = n0 e , (32)
λ (s - s0 ) é λ (s - s0 ) (33)
s2 (s) = n0 e e - 1ù .
ë û
6. Models validation
Let us turn now to the problem of describing the observed data with the proposed
models. The exponential growth of the number of active sub-markets with the size of
the firm can easily be accounted for using the non-linear Poisson process (Model II)
with an exponential intensity function log L(s)~s [cf. equation (26)]. The same is true
if one considers the expression of the mean obtained from the Yule process and
reported in equation (32). In both cases, the fit of the theoretical model on the data
would produce results analogous to the exponential fit of Figure 7.
Even if they can provide similar predictions for the average number of active sub-
markets, the two models greatly differ when it comes to describe other aspects of the
diversification structure of firms. For each of the three models presented above, we
derived the expression of the mean m and variance s2 of the number of active sub-
markets as a function of the size of the firm s. These relations can be used as a qualita-
tive tools to assess the relative goodness of the three models. Indeed, considering
Model II, from equations (26) and (27) one has
s 2 (s) = m(s) + n0 , (34)
which means that the variance of the number of active sub-markets depends linearly
on its average number. Note that this particular relationship does not depend on the
specification of the intensity function L(s) .9
The case of Model III is different. From equations (32) and (33) one has
1
s 2 (s) = m(s)2 - m(s), (35)
n0
that is a quadratic relationship between the variance of the number of active sub-mar-
kets and its mean.
A comparison between Model II and Model III using equations (34) and (35),
respectively, is shown in Figure 10. Here the mean and variance of the number of act-
ive sub-markets is reported for 25 different size bins, covering all the firms of the data-
base. Together, we plot the two theoretical curves (34) and (35) for different values of
the initial number of sector n0. As can be seen, the linear relation (34) provides a very
bad description of data. In order to obtain a straight line that approximately “crosses”
the graph, we had to choose a very high initial number of active sub-markets. Con-
versely, equation (35) seems to provide a good qualitative description of the empirical
observations for the whole range, with reasonable initial conditions (n0 ~ 3–6).
The previous qualitative analyses rule out both Model I and Model II as possible
candidates for the description of the empirical observations, whereas Model III seem
to be, at least qualitatively, in agreement with them. Its descriptive power, however,
must be judged using a more discriminating object. A good candidate is the diversifi-
cation pattern of the whole industry. To be precise, let P(M0 < M < M1)be the proba-
bility that a firm possesses a number of active sub-markets between M0 and M1. One
9
Incidentally, note that this relation holds also in Model I.
2
σ (s)
Model III Model III Model III
(n0=1) (n0=3) (n0=6)
3
10 Model I-II
(n0=600)
2
10
0 20 40 60 80 100 120 140 160
m(s)
Figure 10 Sample variance of the number of active sectors versus its average for firms in differ-
ent equipopulated size classes. The theoretical prediction of Model I–II and Model III are also
shown for suitably chosen values of n0, the initial number of active sub-markets. Note the log
scale on the y-axis.
way of computing this quantity is through the actual frequency of occurrences in the
database
M1 N T -1
1
P ob (M 0 < M < M1 ) = å åådNi (t ),m , (36)
TN m= M
0 i =1 t =0
where Ni(t) is the number of active markets of firm i at time t and d is the Kronecker
delta.
Given a model pn(s), which predicts the probability, for a firm of size s, to posses n
active sub-markets, one can compute the theoretical counterpart of the previous quantity
M1
+¥
P (M 0 < M < M 1 ) = å òs dsf (s) pm (s), (37)
0
m= M 0
where f(s) is the probability density of firm size and s0 is the lower bound on the latter.
The unknown density f(s) can be approximated using the sizes observed in the data so
that equation (37) becomes
M1 N T -1
1
P th (M 0 < M < M1 ) = å åå pm (si (t )). (38)
TN m= M
0 i =1 t =0
A comparison between the observed Pob and the predicted Pth distributions provides a
way to tune the residual degree of freedom and an indication of the goodness of the
model. As showed in Figure 11, the accordance of Model III with data, considering its
extreme simplicity, is surprisingly good, whereas the “best” non-linear Poisson model
performs badly, especially in the leftmost part of the empirical density.
7. Conclusions
We have shown that the relation between the variance of growth and the size of the
firm, which constitutes the clearest and more often reported violation of the Gibrat
Law, reduces to a diversification effect: bigger firms operate in more sub-markets and
the variance of their growth is consequently reduced. At the same time, our analysis
reveals that, in the industry under study, other effects, which might constitute possible
0.45
observed
Yule
Poisson
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0 20 40 60 80 100 120 140 160 180 200 220
Figure 11 The binned probability density for the number of sub-market computed directly
from the data and theoretically predicted by the Yule process. The theoretical distribution is
characterized by l = 0.35, n0 = 5 and s0 = -12. For comparison, a fit of a Poisson model with
a non-linear intensity function L(s) = s l is also shown.
explanations of the size–variance relation, are absent: first, no correlation has been
identified between the growth rates of the same firm in different sub-markets and,
second, the way in which a given firm distributes its activity among the different sec-
tors [described by the diversification heterogeneity index D, defined in equation (13)]
does not depend, on average, on its size nor on the total number of sectors in which it
operates.
It is interesting to note that the observed relationship between the growth variance
and the size of the firms, being milder than the Central Limit Theorem prediction,
gives evidence against the interpretation of “diversification” as a risk minimization
strategy. If it had been the case, firms would have to be present on much more sub-
markets than they actually are. The scope of diversification, our evidence seems to
suggest, is more limited.
The observed firm diversification structure can be described with good accuracy
using the simple stochastic process proposed in section 5. The model describes a firm
whose ability to effectively enter in new sub-markets progressively increases with the
number of sub-markets already penetrated. One possible interpretation of the result-
ing dynamics stems from the idea that the scope of corporate diversification is shaped
and constrained by “the things a firm already knows how to do.” This interpretation is
well in tune with “Penrosian” competence-based description of corporate diversifica-
tion, in which the ability to diversify the business activities is constrained by the com-
petencies a firm inherits from its past (Penrose, 1958). If one assumes the number of
active sub-markets as a rough estimate of the width of these competencies, our main
results is that competencies are accumulated by a firm at a pace which is less than pro-
portional to its growth rate. The intensity of the dynamics of entry in new markets,
however, does not seem to decrease when the size of the firm increases, or after period
of high entry rate in new sub-markets, so that larger firms seem as capable as small
ones to increase their competencies portfolio and there appears to be no size- or
growth-related “limits to diversification.”
Because of the nature of the industry under study, an explanation based on techno-
logical capabilities and learning seems particularly plausible. The pharmaceutical
industry being naturally fragmented in a number of markets of (almost) not substitut-
able goods, the ability of a firm to enter in new sub-markets is crucial to its growth.
This penetration is possible only through the development of new chemical entities or
effective “me too” drugs (Bottazzi et al., 2001).
The scope of applicability of the present model, however, is by no means limited to
a “science-intensive” industrial dynamics like the one characterizing the pharmaceuti-
cal sector. More generally, one could ground the essential point of the model, that is,
the progressive and increasingly intense proliferation of active sub-markets, on the
evolutionary idea that firms progressively learn how to innovate and also how to
diversify, no matter what the object of this learning might be (cf. Dosi et al., 1995).
Indeed, the penetration of a firm in new sub-markets might well be driven by learning
in other domains, as, for instance, the ability to develop strong vertical relationships
or to exploit quasi-monopolistic positions on domestic markets. Still, our model
would continue to apply in so far two general assumptions hold: first, the existence of
some sort of corporate competencies providing the firm with the ability to diversify its
business and, second, that these competencies (and so the ability to enter a new sub-
market) increase with the number of times they have been effectively used (and so
with the number of active sub-markets).
In the present analysis, we purposely ignored the effect of firms entry and exit.
Because of the nature of the data we used, we lacked any reliable indicator of this
important aspect of industry dynamics. Note, however, that as far as new entrants
(exiters) are characterized by a relation between their initial (final) size and the
number of active sub-markets in line with the one observed on the whole industry, the
entry and exit will have no impact on the proposed model. Even if this assumption is
probably violated in periods of high industrial turbulence, due, for instance, to the
introduction of major technological breakthrough or to an intensive merger and
acquisition process, it is likely to hold for rather mature industries and/or during
period of steady market expansion.
Finally, the proposed model describes the cumulative effect of learning through a
linear function. In general, one may refine the model using more sophisticated rela-
tions, for instance, by dropping the strictly binary nature of the branching or intro-
ducing the possibility of a branch “death.” This finer modeling would require,
however, a stronger phenomenological justification from the data.
Acknowledgements
The authors thank U. Cantner, C. Castaldi, G. Dosi, P. Geroski, S. Klepper, M. Lippi,
and L. Orsenigo for helpful comments and useful discussions. Support from the
Italian Ministry of University and Research (grant A.AMCE.E4002GD), from S. Anna
School of Advanced Studies (grant E6003GB) and from the Merck Foundation is
gratefully acknowledged. Special thanks to F. Pammolli (director of the EPRIS project)
for his constant support and precious help. The usual disclaimers apply.
Address for correspondence

Address for correspondence Giulio Bottazzi, Scuola Superiore Sant’Anna, P.za Martiri
della Liberta’ 33, 56127, Pisa, Italy. e-mail: bottazzi@sssup.it.
References
Amaral, L. A. N., S. V. Buldyrev, S. Havlin, M. A. Salinger, H. E. Stanley and M. H. R. Stanley
(1997), ‘Scaling behavior in economics: the problem of quantifying company growth,’ Physica A,
244, 1–24.
Boeri, T. (1989), ‘Does firm size matter?,’ Giornale degli Economisti e Annali di Economia, 48,
477–495.
Bottazzi, G. and A. Secchi (2003), ‘Common properties and sectoral specificities in the dynam-
ics of U.S. manufacturing companies,’ Review of Industrial Organization, 23, 217–232.
Bottazzi, G. and A. Secchi (2005), ‘Growth and diversification patterns of the worldwide phar-
maceutical industry,’ Review of Industrial Organization, 26, 195–216.
Bottazzi, G., G. Dosi, M. Lippi, F. Pammolli and M. Riccaboni (2001), ‘Innovation and corpor-
ate growth in the evolution of the drug industry,’ International Journal of Industrial Organiza-
tion, 19, 1161–1187.
Dosi, G. (1988), ‘Sources, procedures, and microeconomics effects of innovation,’ Journal of
Economic Literature, 26, 1120–1171.
Dosi, G., O. Marsili, L. Orsenigo and R. Salvatore (1995), ‘Learning, market selection and the
evolution of industrial structures,’ Small Business Economics, 7, 411–436.
EphMRA (2004), ‘Anatomical classification guidelines,’ http://www.ephmra.org/pdfs/
ATCguidelines.pdf.
Feller, W. (1968), An Introduction to Probability Theory and Its Applications. Wiley & Sons:
New York.
Geroski, P. A. (2000), ‘The growth of firms in theory and practice’ in N. Foss and V. Malinke (eds),
New Directions in Economic Strategy Research. Oxford University Press: Oxford, pp. 168–186.
Gibrat, R. (1931), Les inégalités économiques. Librairie du Recueil Sirey:Paris.
Hart, P. E. and S. J. Prais (1956), ‘The analysis of business concentration,’ Journal of the Royal
Statistical Society, 119, 150–191.
Hymer, S. and P. Pashigian (1962), ‘Firm size and rate of growth,’ Journal of Political Economy,
70, 556–569.
Ijiri, Y. and H. A. Simon (1977), Skew Distributions and the Sizes of Business Firms. North Holland:
New York.
Lotti, F., E. Santarelli and M. Vivarelli (2003), ‘Does Gibrat’s Law hold among young, small
firms?’ Journal of Evolutionary Economics, 13, 213–235.
Mansfield, D. E. (1962), ‘Entry, Gibrat’s Law, innovation, and growth of the firms,’ American
Economic Review, 52, 1024–1051.
Matia, K., D. Fu, A. O. Schweiger, S. V. Buldyrev, F. Pammolli, M. Riccaboni, and H. E. Stanley
(2004), ‘Statistical properties of structure and growth of business firms,’ Europhysics Letters,
67, 498–503.
Mazzucato, M. and G. Dosi (2006), Knowledge Accumulation and Industry Evolution.
Cambridge University Press: Cambridge.
Nightingale, P. and S. Mahdi (2006), ‘The evolution of pharmaceutical innovation’, in G. Dosi
and M. Mazzucato (eds), Knowledge Accumulation and Industry Evolution. Cambridge
University Press: Cambridge, pp. 73–111.
Pagan, A. and A. Ullah (1999), Nonparametric Econometrics. Cambridge University Press: Cambridge.
Penrose, E. (1958), The Theory of the Growth of the Firm. Oxford University Press: Oxford.
Silverman, B. W. (1986), Density Estimation for Statistics and Data Analysis. Chapman and Hall:
London.
Snyder, D. L. (1975), Random Point Processes. Wiley & Sons: New York.
Sutton, J. (1997), ‘Gibrat’s legacy,’ Journal of Economic Literature, 35, 40–59.
Sutton, J. (1998), Technology and Market Structure, Theory and History. MIT Press: Cambridge.
Teece, D., R. Rumelt, G. Dosi and S. Winter (1994), ‘Understanding corporate coherence:
theory and evidence,’ Journal of Economic Behavior and Organization, 23, 1–30.
Appendix 1: Solution of the Yule process

To simplify the solution of equation (30), introduce the rescaled size t = l (s - s0 ) and
consider, instead of the probabilities in equation (28), the variables yN(t) defined by
pN (t ) = e - Nt y N (t ). (39)
In these new variables, the set of equation in equation (30) reduces to
ìï y N¢ (t ) = (N - 1)e t y N -1 (t ), N > N 0 ,
í (40)
y ¢ (t ) = 0, N = N0
îï N 0
with initial conditions
ì1, N = N0 ,
y N (0) = í (41)
î0, N ¹ N0 .
From equation (40), iterating over the index N, it is immediate to write the solution as
the multiple integral
t t1
t
N - N 0 +1 t1 +t 2 +…+t N - N
y N (t ) = (N - 1)(N - 2)… N 0 ò dt1 ò dt 2 …ò dt N - N e 0 . (42)
0 0 0 0
Note that, because of the complete symmetry of the integrand, the multiple integral over
the N-N0-dimensional hypercube of side t with the constraints t N - N 0 < < t 2 < t1 < t
reduces to the integral over the whole hypercube divided by all the possible ordering
of the constraints, which are (N-N0)!. One thus obtains
æ N -1 ö t N - N0
y N (t ) = ç ÷ (e - 1) , (43)
è N - N0 ø
which, remembering the factorization in equation (39) and substituting t with its pre-
vious definition, reduces to equation (31) after obvious algebra.
Appendix 2: Moments of the Yule distribution

Let us rewrite the Yule distribution in equation (31) as
h h
æ n0 + h - 1ö n0 h n0 P d -n
pn (s) = ç ÷ø (1 - P ) P = (1 - P ) x 0 (-1)h , (44)
è h h! dx h x =1
- l (s - s0 ) n - n0
where n = n0 + h and P = (1 - e ) . We can now write the moment-generating
function for equation (44)
¥
n0 + h n n
¥
P h d h - n0
M (l , s, n0 ) = ål pn (s) = l 0 (1 - P) 0 å lh h! dx h
x (-11)h , (45)
h =0 h =0 x =1
which, using the Taylor series expansion, reduces to
n0
n n - n0 n0 æ 1- P ö (46)
M (l , s, n0 ) = l 0 (1 - P) 0 (x - l P) =l çè ÷ .
x =1 1- lP ø
It is now straightforward to derive the first two central moments of equation (44)
dM (l , s, n0 ) æ 1 ö
m(s) = = n0 ç , (47)
dl l =1
è 1 - l P ÷ø
whereas
2
d 2 M (l , s, n0 ) dM (l , s, n0 ) æ P ö
s2 (s) = + - m2 (s) = n0 ç . (48)
dl 2 dl l =1
è 1 - l P ÷ø
l =1

Gibrat's Law and Diversification

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gibrat's Law and Diversification

Uploaded by

Copyright:

Available Formats

Industrial and Corporate Change, Volume 15, Number 5, pp.

Year Firm size Number of markets

Mean Std Min Max Mean Std Min Max

Level Number Size of markets Number of firms

Mean Std Min Max Mean Std Min Max

Figure 1 Empirical probability density P(N) of the number of active sub-markets N.

3. The variance–size relation

Relation Slope Std error

Size—mean -0.11 0.08

non-parametric regression remains near it for the whole support of s. Finally, we do

4. Firm diversification and the violation of Gibrat’s Law

As a preliminary step, we analyze the cross-correlation in sectoral growth rates. Let

Ti , j = t | t Î{1,…,10}, Si , j (t ) > 0, Si , j (t + 1) > 0

and the associated cross-correlation coefficient reads

so that the index D takes its maximum value of 1.

In Figure 6 we report a scatter plot of the diversification index D i (t ) versus the

log(N i (t )) ~ l log(Si (t )) + q. (14)

5. Stochastic models of firm diversification

5.1 Model I: random diversification

ìï pn0 (s + d ) = pn0 (s)P(n0 , s; n0 , s + d ), n = n0 ,

ìï pn¢ (s) = - l pn (s) + l pn -1 (s), n > n0 ,

with initial conditions

ìï pn¢ (s) = - L(s) pn (s) + L(s) pn -1 (s), n > n0 , (23)

q(s; s0 )n - q(s ;s0 ) (24)

Figure 8 Diversification by independent events. The inter-arrival “growths” si - si-1 are

m(s) = n0 + q(s; s0 ), (26)

s 2 (s) = q(s; s0 ). (27)

5.3 Model III: scope economy to diversification

Figure 9 Diversification by branching. The inter-arrival “growths” are no longer exponential;

the dynamics of the pharmaceutical industry is largely research-driven, so that prod-

neglecting again multiple instantaneous branchings. If pn(s)is the probability that a

ìï pn (s + d ) = pn -1 (s)P(n - 1, s; n, s + d ) + pn (s)P(n, s; n, s + d ), n > n0 ,

ïì pn¢ (s) = -nl pn (s) + (n - 1)l pn -1 (s), n > n0 ,

s 2 (s) = m(s) + n0 , (34)

Address for correspondence

Appendix 1: Solution of the Yule process

In these new variables, the set of equation in equation (30) reduces to

with initial conditions

Appendix 2: Moments of the Yule distribution

which, using the Taylor series expansion, reduces to

You might also like