Firm Size, Firm Growth and Persistence of Profits:
An Firm-Level Analysis of Turkish Manufacturing
Industry∗
Murat Donduran
†
May 28, 2008
Abstract
In an efficient market economy, profits above or below the average level should
quickly disappear. However, the empirical literature says that there is a persistence
of profits in some industries. Nowadays, the methodology in order to study the
persistence of profits usually focuses on panel unit root tests. Those studies just
conclude if there is a panel unit root also no mean reversion then there will be a
persistency but the detailed analysis about that subject is less. In an historically
empirical literature, there are a lot of methods to analyze the firms’ profits via the
firm level data. In this paper, the firms’ profits are studied by the mean of this
perspective. Firstly, there is a deep analysis about descriptive statistics of firm
level data. Afterthat, step-by-step, the persistence of profits are investigated from
individually estimations to aggregately estimation as panel unit root tests.
∗
The usual disclaimer applies.
[email protected]. Assoc., Prof, Yildiz Teknik Universitesi, Department of Economics, Istanbul.
†
1
1
Introduction
There are two views of competition. One of them belongs to the mainstream economic theory. Other one is about the seeing competition not as a process for allocating a given stock of resources but as a process for transforming these resources
into new products and production techniques. In such a process, creating a new
product causes a monopoly or monopoly power in the industry. Afterthat, other
firms imitate and improve upon the new product, so monopoly will be disappeared.
Therefore, monopoly is the integral part of a dynamically competitive process.
If competition is assumed as an dynamic process, the concept of equilibrium will
not play an important role. In this point, the constellation of prices and allocation
of resources at a particular point in time are not important. Only focused thing
should be their movements over time.
In such a situation, the most notable characteristic of first view of competition’s
models are static and virtually all of the associated empirical work has been crosssectional in character. On the other hand, focusing on innovation, imitation and
adaptation, they are concerned not so much with monopoly as with its persistence.
Both of these alternative line of thought are fundamentally dynamic in character
and it is very appropriate to analyze competition as a process.
On the other hand, in addition to this theoretical motivation, it is very illuminating to analyze a developing country like Turkey. That economy experienced
significant changes in the 80s and 0s associated with a trade and financial liberalization process and with price stabilization that follows nowadays. In principle,
the recent period in Turkey is characterized as more competitive what provides an
interesting setting for this type of investigation.
Another interesting factor for analyzing the profits is emphasized by Mueller
(1977) and Mueller (1990a). These factors those affect the persistence of profits are
sometimes luck or skill for a firm and they provide resources to maintain profits into
the future. Some companies erect entry barriers through increased product differentiation, others via scarce natural resources or land sites. Until now, all the factors
are in an industrial organization perspective. However, in a developing country like
Turkey, some companies obtain legal protection for the positions (e. g. patents,
tariffs, licences) by purchasing the services of scientists and technicians, lawyers or
lobbyists or more directly by contribution to politicians and public officials themselves. The means vary, but in a developing country last words are more important
than others. At the end all conclusions are the same, the preservation of an existing
monopoly rent.
There are two papers which studies the persistence of profits in Turkey. One
of them is Yurtoglu (2004) that is analyzing the manufacturing sector with the
same database with a little bit small time series. It expands the analysis to the
investigation of the parameter about persistence. The other one is Bektas (2007)
that studied the banking system. It rejects the unit root hypothesis for the data and
found that the persistency of profits does not exist in the Turkish Banking System.
Analyzing the persistence of profits is to extend the traditional static, crosssectional empirical models to include market dynamics. In order to that, it is very
useful to begin the study with the Geroski (1998)’s stylised facts: corporate growth
2
rates really are very nearly random and profits are persistent over time.
The descriptive statistics about firm’s performance give the light about the way
of the analysis of persistent profits. Therefore, first of all, the descriptive statistics,
and ANOVA analysis are made for the autoregressive model of the variables such as
firm size, firm’s growth and profits in this paper. After the simple autoregressions,
the panel unit root tests are applied for the persistence of profit analysis on the
behalf of the literature.
2
2.1
Data and Descriptive Statistics
Variables
In this study, we will use three variables about firms size, firm growth and profitability. Accounting profits-Sales ratio is used for the profit measure. Firm size is
the natural logarithm of the sales. For firm growth, the rate of growth of sales is
used. These variables are taken from the ISO database. They are between 1986 and
2005. All the analysis is made on balanced panel data.
Variablesa
Mean Std. Deviation
Profits/sales -6.16e-10
.1663
Profits/assets 1.21e-09
.1610
Profits/capital .3.67e-07
1355
Log of sales
9.796
1.058
Sales’ growth
.0405
.2575
Employment
6.696
.9294
Variance Skewness Kurtosis
.0276
2.574
34.28
.0259
.2779
6.881
1837053
23.59
684.7
1.119
1.050
4.707
.0663
-.1865
7.318
.8639
.5336
3.422
a
For a balanced panel of 95 large Turkish firms over the period 1986-2005 (one year is lost through
differencing for the growth rate).
Table 1: Simple statistical descriptions of company performance
Table (1) reveals that profits/sales ratio peaked and positively skewed. However,
profits/assets ratio (roughly speaking) is accepted as normally distributed. This
means that most firms clustered in the middle of ranking with a few outliers on the
top ant the bottom. The log of sales also normally distributed, meaning that the
level of sales positively skewed. This is also (roughly) true of the other measure of
size in the data, log of employment. In the case of firm size, a skewed distribution
means that most firms are small and only a very few are large. Sales’ growth is
negatively skewed. When the distribution of performance differences between firms
is skewed, the interesting question is who is in the tail (which firm is large?).
2.2
Correlations
According to economists, profitability is a kind of residual which is left after costs
have been deducted from revenues and it is common to think that it is, for this
reason, a natural way to sum up all difference aspects of the performance of any
3
particular firm. The main assumption of mainstream economics about firms is to
maximize the profits. Table (2) shows the correlations between all six measures of
performance. In fact, most of the 15 correlations shown on the table are pretty low
and it is hard to believe that any one of these measures can reliably be taken as a
proxy for any of the others.
Variablesa
Profits/sales (P/S)
Profits/assets (P/A)
Profits/capitalb (P/C)
Log of sales (S)
Sales’ growth (G)
Employment (E)
P/S
P/A P/C
S
G
E
1.0
.7735
1.0
.1004 .1490
1.0
.1700 .1706 .1956
1.0
.0949 .1651 .0192 .1392
1.0
.0102 -.0764 .1687 .6679 .0016 1.0
a
For a balanced panel of 95 large Turkish firms over the period
1986-2005 (one year is lost through differencing for the growth rate).
b
We drop Profits/Capital ratio because there are a lot of outliers
due to the measurement problem on capital.
Table 2: Correlations across different measures of performance
Accounting Profits/Sales ratio is statistically incongruent with many of the other
performance measures. Geroski (1998) states two sources of this incongruence: The
first one is namely the difference in the distribution of performance outcomes across
firms. The second reason is based on the analysis of variance techniques discussed
in details following subsection.
2.3
Analysis of Variance
Using analysis of variance techniques, it is possible to break the total variation
in performance across firms over time into two components: ”between” variation, which reflects differences in firms which prevail on average over a period and
”within” variation which reflects variations in the performance of a typical firm
over time. Data which display a large amount of ”between” variation identify relatively permanent differences between firms. Therefore, it is convenient to apply
a cross-sectional analysis. Data which display large amount of ”within” variation
is explained by various authors as a variations in performance over time are dominant feature of the data. If differences over time insists across firms, then panel
techniques can be explain performance differences. Also, if the within variation is
idiosyncratic, then the only way forward is to analyze the time series of individual
firms one by one.
Table 3 shows that for all variables variations are ”within”. Also, these variables
display no ”between” variation, meaning that year by year differences between firms
do not persist for very long.
4
Variablesa
Profits/sales
Profits/assets
Log of sales
Sales’ growth
Employment
% within variation
99
99
96
90
99
% between variation
1
1
4
10
1
a
For a balanced panel of 95 large Turkish firms over the period 1986-2005 (one
year is lost through differencing for the growth rate).
Table 3: Analysis of variance
2.4
Simple Autoregressions
Much of the empirical literature on corporate growth rates has explored the hypothesis that firm size follows a random walk. There are several ways in which this
hypothesis can be explored, including regressing firm size in period t, Si (t), against
size in period t − 1,
Si (t) = αi + βSi (t − 1) + µi (t)
(1)
and testing whether β = 1 (random walk) or β < 0 (mean reversion). A somewhat more common version of this test involves running the regression,
Gi (t) = αi + γSi (t − 1) + ²i (t)
(2)
where Gi (t) is the rate of growth of firm i in year t and testing whether γ = 0
(random walk) or γ < 0 (mean reversion).
Variablesa
βb
Profits/sales .4617
Profits/assets .3839
Log of sales
.7772
Sales’ growth -.2416
Employment .8058
t-Statisticsc
22.11
17.35
52.19
-10.11
57.01
R2
.4848
.3691
.9410
.0431
.9568
a
For a balanced panel of 95 large Turkish firms over the period
1986-2005 (one year is lost through differencing for the growth rate).
b
These regressions of the form: xit = αi + βxit−1 for each of the
performance variables, xit taken in turn.
c
t-Statistics reports absolute values of the standard test of null that
β = 0.
Table 4: Simple Autoregressions
In table (4) shows an estimate of (1) using the whole panel of firms and allowing
for firm specific fixed effects. β < 0 means that there is a tendency for smaller firms
to ”catch up” with their larger rivals.
The natural logarithm of sales is thought of as a measure of firm size. Its
coefficient in the autoregression model has a positive β with not too close value to
5
one. However, the growth of sales equation is not consistent with the log of sales
equation.
The important observation is that mean reversion is an amazing feature of corporate growth process shown up in the data. However, in order to focus on mean
reversion issue, we have to check that if corporate growth rates are not largely random, then the variance of firm size will not rise over time. But in table (5), the
variance in log sales firstly decreased then rose over the sample period. In fact, the
variability of growth rate and profitability also fall throughout the period.
Variablesa
Profits/sales
Profits/assets
Log of sales
Sales’ growth
Employment
1987
.0427
.0388
1.050
.0696
.9707
1996
.0241
.0259
.9720
.0406
.7731
2005
.0285
.0145
1.367
.0527
.9127
a
For a balanced panel of 95 large Turkish firms over
the period 1986-2005 (one year is lost through differencing for the growth rate).
Table 5: Variance of Variables Over Time
Another analysis can help for the mean reversion. If firm size follows a random
walk, then increments to size will be uncorrelated with each other across firms, while
firm size itself will be highly correlated over time across firms.
Log of Sales
1987
1996
1987 1.0000
1996 -0.0662 1.0000
2005 -0.1139 -0.0844
2005
1.0000
Table 6: Correlation of Log of Sales Over Time
In table (6), the cross section correlations are showed between firm size in 1987,
1996 and 2005. Is is acceptable that there is no correlations over time which is
exactly what one expects. Also, differences in firm size, do not persist at all.
3
3.1
Estimations for Persistence of Profits
Heterogeneities in Performance of Firms
Geroski (1998)’s stylised fact says that heterogeneities in performance between firms
persist into the long run more or less regardless of how performance is measured.
However, performance differences between firms are not constant over time and
many of them widen in recessions.
A largely empirical literature has built up testing the proposition that profit
differences between firms persist even in the long run using a model very like (1),
namely
6
πi (t) = α + βπi (t − 1) + vi (t)
(3)
0
1
Density
2
3
where π is typically one of a number of measures of accounting profitability. The
conventional interpretation of the output of these regressions is that the smaller is
the estimated value of β, the stronger are the forces which induce convergence; if
β = 0, then profit differences between firms do not persist.
For the heterogeneities of firms’ performance, it is easily looking at the first row
of the Table (4). It is obvious that analyzing the data more individually gives us
more detailed capabilities for understanding the persistence. Therefore, we estimate
the equation (3) individually.
−.2
0
.2
.4
persistence1
.6
.8
Density
normal persistence1
kdensity persistence1
Figure 1: Histogram and Kernel Density of Persistence Coefficient in Profit/Sales
Figure (1) shows the histogram and Kernel Density of persistence coefficient also
β for all 95 firms with normal kernel function. Right skewness of the density tells
that β are above 0.5. There are some firms with negative coefficients like in growth
rates’ autoregressions. Skewness-Kurtosis normality test rejects the normality of
the density function.
3.2
Intertemporal Pattern of Profitability
This subsection of the paper is based on the work of Mueller (1990b). Mueller
(1990b)’s model is very simple to generate. Assume that firm i’s return on capital
in year t, πit , is composed potentially of three components: (1) a competitive return
c common to all companies; (2) a permanent rent ri specific to firm i which could
be a premium for risk and (3) a short run rent sit with zero expected value:
πit = c + ri + sit
7
(4)
For a t sufficiently long sit might be assumed to have mean zero and constant
variance over time, and the hypothesis that competition eventually drives all profit
rates to a common normal level could be tested simply by comparing mean profit
rates across firms to see whether they are significantly different from one another,
given their intertemporal variances.
A more reasonable assumption concerning the sit is that they are intertemporally
related but converge o zero. Let sit be defined by
sit = λsit−1 + uit
(5)
where 0 < λ < 1 and uit are distributed N (0, σ 2 ). Assuming equation (5) holds
in every period, it can be used to remove sit from (4) to obtain
πit = (1 − λ)(c + ri ) + λπit−1 + uit
(6)
Let α̂i and λ̂i be the estimates from autoregressive equation then the equation
will be:
πit = α̂i + λ̂i πit−1 + uit
(7)
Afterthat it is easy to derive the estimate of the long run projected profits of
firm i, πip as
πip =
α̂i
1 − λ̂i
(8)
Hypothesis: Competition drives all profit rates to a common competitive level
would be to test whether the πip differ significantly across firms. If there is no significant differences, all long run rents ri are zero.
In order to test the hypothesis, we will use one variable: a company’s return
on capital as its profits net of taxes and gross of interest divided by total assets.
The competitive return c modeled as if it were a constant, it may vary over time as
business cycle factors and long run trends raise and lower the average performance
of firms in the economy. To allow for these common intertemporal patterns, each
firm’s annual profit rate is taken as a deviation from the sample mean for that year.
In effect, it is assumed that the relationship between the competitive return c and
the average return on capital is invariant over time.
Table (7) summarizes the results for the estimations of equation (7) for the 95
firms. The dependent variable is the accounting profits/assets. Starting with the
bottom row, we see that the average fit to the autoregressive equation was not weak,
with a mean R2 of 0.2257. Thus, we can say that the transitory component of a
firm’s profit rate would seem to require more than a year to be eliminated in many
cases. For λ̂, on average 39 percent of any deviation from last year’s sample mean is
expected to reoccur this year. The distribution of λ̂ is obviously skewed, however,
with the mean λ̂ pulled down by negative λ̂’s. Of the latter, all of the 8 λ̂’s that
were negative had a value |λ̂| > 1.72 times its standard error (the critical value for
8
π̂ip
λ̂
R2
Mean St. Error
.0003
.0960
.3905
.2572
.2257
.1801
Min
Max
-.2905 .1994
-.2995 .8356
.0003 .7134
Dependent Variable: Profits/Assets Ratio
For a balanced panel of 95 large Turkish firms over
the period 1986-2005 (one year is lost through differencing for the growth rate).
Table 7: Summary of Results from Autoregressive Profits Equations
0
.5
Density
1
1.5
a two-tailed, 10 percent level test). Thus, the λ̂’s for these firms are consistent with
the hypothesis that short run rents of these firms vary independently over time. Of
the greater that zero, on the other hand, 80 exceeded their standard errors by more
than a factor of 1.72. This figure is more than ten times the 8 one expects under
a 5 percent level, one tailed test if all λ̂’s are zero, but random factors generate
significant coefficients in 5 percent of the equations.
−.5
0
.5
1
Lambda
Density
kdensity Lambda_for_PA
Figure 2: Histogram and Kernel Density of Persistence Coefficient in Profit/Assets (Individual Estimation)
The mean value for λ̂ across sample combined with almost 90 percent of the
λ̂’s being significantly different zero draws out attention to the long run projected
returns for the firms. It is very important that all of the λ̂’s fall between -1 and 1,
implying convergence on the π̂ip . 88 firms exceed their standard errors by a factor of
more than 1.72. Thus, almost all firms in the sample are projected to earn long run
returns significantly from the average firm in the sample. Assuming some positive
rents exist due to market power, the sample mean should exceed c, however.
9
0
2
Density
4
6
In this point, it is very important to answer the question what fraction of firms
have profits rate significantly different from c, the competitive return on capital?
First of all, if all π̂ip equal a common c, all will equal one another. The hypothesis
that all π̂ip converge to a common, competitive c can be tested by seeing whether
restricting all firms to have the same π̂ip results in a significant increase in the sum
of squared residuals from the unconstrained estimates. Also, it does.
If all π̂ip ’s equaled c, then is is convenient to assume a normal distribution around
c. In figure (3) is very easy to see a normal distribution of π̂ip . Then, c and π̂ip are
related because the mean of π̂ip is 0.002. Is is remarkable to conclude that there
exist no firms with nonzero permanent rents. Whereas short run rents appear to
erode quite quickly for most firms, there exists no significant differences in long-run
rents across firms.
−.3
−.2
−.1
0
Long Run Projected Profits
.1
.2
Density
kdensity pi_ip_pa
Figure 3: Histogram and Kernel Density of Long Run Projected Profits (Individual Estimation)
3.3
Panel Unit Roots
Panels with large cross-sectional dimension and long time periods have also been
used by applied economists to examine various area in economics. Recent developments in the econometric testing of unit roots in the context of panel data (see
Levin and Lin (1993) and Pesaran, Im, and Shin (1995)) provide the opportunity for
formal testing of strong form of persistence even with short panels and constitute
therefore a relevant additional tool kit for the profit persistence literature [Resende
(2006)].
10
3.3.1
Methodology
At the beginning of the paper, we said that two important approaches shape the
profit persistence literature. One of them is about entry forces as mainstream and
other is the Schumpeterian competitive process. In this subsection, we emphasize
both of them in order to construct simple theoretical frameworks that provide foundations for empirical analysis of profit persistence. Like in the simple autoregressive
analysis, it is based on the influential example that given by Geroski (1990). The
basic idea can be summarized as follows. Let ρ(t) ≡ π(t)−πp (t) denote firm’s excess
profits at period t and long-run projected profits respectively. In Geroski (1990)’s
model, there are two general classes of factors determining changes in ρ(t). First,
there are systematic factors (like ”entry” E(t)) and set of other factors orthogonal
to the first class that can be generally refereed ”luck” µ(t) that would assumed as
an i.i.d normally distributed process with zero mean and variance σµ2 . Thus, with
a simple expression, following equation it is very convenient to show the relating
changes in excess profitability to the two classes of explanatory factors:
∆ρ(t) ≡ α0 + β0 E(t) + β1 ρ(t − 1) + µ(t)
(9)
where ∆ρ(t) = ρ(t) − ρ(t − 1).
Geroski (1990) refers that there is a difficulty associated with the previous expression. The difficulty is the existence of non-observable components in E(t), for
example, potential entry. The latent variable character of this formulation requires
then a link that expresses E(t) in terms of observable factors, a possibility is given
as follows:
E(t) = φ[ρ(t − 1) − ρ∗ ] + ε(t)
(10)
where ρ∗ denotes the equilibrium value of ρ(t) which does not induce further
entry movements and φ > 0 means for a speed parameter indicating the attractiveness of entry. Even if ρ(t − 1) = p∗ , it is possible to observe an exogenous flow of
entry or exit given by ε(t) which it is assumed to be normally distributed process
with zero mean and variance σε2 . By combining equations (9) and (10) we can write
equation (11) which only involves observable variables,
ρ(t) = α + λρ(t − 1) + v(t)
(11)
where α ≡ α0 − β0 φρ∗ , λ ≡ (β0 φ + β1 + 1) and v(t) ∼ N (0, σv2 = β02 σε2 + σµ2 ).
3.3.2
Panel Data Unit Root Tests
It is well known that traditional unit root tests possess low power against near unit
root alternatives. The development of panel data unit root tests addresses this aspect and additionally allows to consider data sets with a short time dimension.
Levin and Lin (LL) Test
LL test considers unit root testing for different model with different degrees of
heterogeneity across time and units. One of the appropriate model of LL test for a
generic variable y is given by:
11
∆yit = αi + βyit−1 + εit
(12)
with the null hypothesis H0 : αi = βi = 0. The alternative hypothesis is given
by H1 : βi = β < 0 for all i’s. The main limitation of the LL test is to have a
common parameter β across different units. The refereed test can be carried out by
means of the t statistics obtained upon a within group estimator for panel.
Im, Pesaran and Shin (IPS) Test
Pesaran, Im, and Shin (1995) provide a panel data unit root test relaxes the LL
test’s assumption. Considering the model given in expression but with parameter
β varying across units as given below:
∆yit = αi + βi yit−1 + εit
(13)
IPS propose test where H0 : βi = 0 for all i and H1 : βi < 0 for any i. One
therefore relaxes the strong homogeneity assumption of LL tests. The simplest
test proposed by IPS, t-bar statistics is defined as the average of the individual
Dickey-Fuller (DF) or augmented Dickey-Fuller (ADF) say τ1 statistics:
N
1 X
τi
t̄ =
N
(14)
i=1
where
τi =
β̂i
σ̂β̂i
√
where ( N (t̄ − E(τi |βi = 0)/var(τi |βi = 0)1/2 ∼ N (0, 1).
3.3.3
Results
For the variable PA, the 92 sample firms are assigned and there are 20 available
time series observation per firm.
Both the LL and IPS tests are based on the ADF autoregression for the untransformed series. All estimations are made in lag lengths P = {1, 2, 3, 4} without and
with a linear time trend.
The results obtained in two panel data unit testing for two measures of profitability are presented in Table (8) and (9).
In LL Test, first interesting issue is that the coefficient (λ) is very high in panel
unit root test than in averaged individual estimation and panel first lag autoregression model. Augmenting lag is 2 in LL Test without and with trend. However,
including time trend in the test increases the the speed of the statistically insignificance level of the analysis.
When one considers the t-bar statistic in the IPS test at table (9), the results
favour in most cases the existence of unit root for the two excess profitability measures and this is the case whether one includes or not a time trend component. For
the augmenting lag of 3, the evidence favours the rejection of the null hypothesis of
a unit root.
12
Augmenting
Lag
λ
Results for P Aa
t
t∗
p−value
Model without Time Trend
p
p
p
p
=
=
=
=
1
2
3
4
-.5779
-.5686
-.6159
-.6014
-24.45
-20.85
-20.37
-17.54
-11.52
-5.339
-2.976
2.984
.0000
.0000
.0015
.9986
Model with Time Trend
p
p
p
p
=
=
=
=
1
2
3
4
-.7450
-.8123
-.9715
-1.009
λ
-28.87
-25.59
-26.38
-23.16
Results for P S b
t
-11.27
-4.480
-1.059
8.707
.0000
.0000
.1446
1.000
t∗
p−value
Model without Time Trend
p=1
p=2
p=3
-.6104
-.6280
-.7235
-24.57
-21.16
-20.82
-10.46
-4.498
-1.891
.0000
.0000
.0293
p=1
p=2
p=3
-.7724
-.9025
-1.132
-28.88
-26.52
-27.20
-10.31
-4.030
0.2935
.0000
.0000
.6155
Model with Time Trend
a
For a balanced panel of 92 large Turkish firms over the period 1986-2005 (one year is
lost through differencing for the growth rate).
b
For a balanced panel of 95 large Turkish firms over the period 1986-2005 (one year is
lost through differencing for the growth rate).
Table 8: LL Tests Results
13
Augmenting Lag
Model without Time Trend
p=1
p=2
p=3
p=4
p=5
Model with Time Trend
p=1
p=2
p=3
p=4
p=5
Variables
P Aa
P Sb
-10.152
(0.000)
-6.379
(0.000)
-5.045
(0.000)
-2.885
(0.002)
-2.009
(0.022)
-8.828
(0.000)
-5.864
(0.000)
-5.287
(0.000)
-1.755
(0.040)
-0.237
(0.406)
-7.453
(0.000)
-4.220
(0.000)
-4.104
(0.000)
-1.153
(0.125)
-1.615
(0.053)
-5.922
(0.000)
-3.812
(0.000)
-3.881
(0.000)
-0.832
(0.203)
0.033
(0.513)
Note: p-values appear in parenthesis
a
For a balanced panel of 92 large Turkish firms over the period 1986-2005
(one year is lost through differencing for the growth rate).
b
For a balanced panel of 95 large Turkish firms over the period 1986-2005 (one
year is lost through differencing for the growth rate).
Table 9: IPS Tests Results
4
Conclusion
The sample for the present study comprises 95 Turkish large manufacturing firms
with complete annual data for the years 1986 to 2005. Company account data for
the 95 firms are obtained from Istanbul Sanayii Odasi.
Firm size, firm growth and profits are investigated in order to show their relationships with each other and persistence over time. The analysis began with a very
descriptive basis. While the regression for the panel and individual autoregressions
are very simple, the statistics indicate that the regressions are well specified and
have good explanatory power in most cases. The firm size and firm growth are
not persistent at the simple autoregression results. However, profits are, as parallel
persistent profit literature, persist in the favour in all cases.
14
Especially, in the autoregression for profit rate, the mean value of λ̂i is high but
it suggests a rapid speed of adjustment for excess short-run profits. The mean value
of λ̂i is one of the highest sample mean if we compare with the Goddard and Wilson
(1999)’s table.
At the end of the paper, a strong form of non-stationarity referring to the existence of a unit root in the autoregressive process associated with excess profitability
in Turkey is investigated. This simple formulation can be theoretically motivated
and recently developed panel data unit root tests can provide formal testing of
persistence in the context of short panels.
Despite of the trade liberalization in Turkey which one expects as a competitive
pressure in all market in the country, the evidence indicated that an extreme level of
persistence associated with the presence of a unit root in excess profitability cannot
be discarded.
As further research, this paper also provides motivation for dynamic econometric
analysis in the context of IO. Because of the dynamics and evolutionary aspects of
the markets in developing countries are very interesting even with this simple panel
unit root results.
15
References
Bektas, E. (2007): “The Persistence of Profits in the Turkish Banking System,”
Applied Economics Letters, 14, 187–190.
Geroski, P. A. (1990): The Dynamics of Company Profits: An International Comparisonchap. Modelling persisence profitability, pp. 15–34. Cambridge University
Press.
(1998): “An Applied Econometrician’s View of Large Company Performance,” Review of Industrial Organization, 13, 271–293.
Goddard, J., and J. Wilson (1999): “The Persistence of Profit: a New Empirical
Interpretation,” International Journal of Industrial Organization, 17, 663–687.
Levin, A., and C.-F. Lin (1993): “Unit Root Tests in Panel Data: New Results,”
University of California at San Diego, Economics Working Paper Series 93-56,
Department of Economics, UC San Diego.
Mueller, D. C. (1977): “The persistence of profits above norm,” Economica,
44(176), 369–80.
(1990a): The Dynamics of Company Profits: An International Comparison. Cambridge University Press.
(1990b): The Dynamics of Company Profits: An International Comparisonchap. The Persistence of Profits in the United States, pp. 35–60. Cambridge
University Press.
Pesaran, M., K. Im, and Y. Shin (1995): “Testing for Unit Roots in Heterogeneous Panels,” Cambridge Working Papers in Economics 9526, Faculty of
Economics University of Cambridge.
Resende, M. (2006): “Profit persistence in Brazil: a panel data study,” Estud.
Econ. [online], 36(1), 115–126.
Yurtoglu, B. (2004): “Persistence of Firm-LEvel Profitability in Turkey,” Applied
Economics, 36, 615–625.
16
9.4
Average_Log_of_Sales
9.6
9.8
10
10.2
Rate of Growth of the Sales and Average Sales
1985
1990
1995
Year
2000
2005
Average_Growth_Rate
0
.1
.2
Figure 4: Average Sales (1986-2005)
−.1
A.1
Graphs by Yearly Means of Variables
−.2
A
1985
1990
1995
Year
2000
2005
Figure 5: Average Growth Rate of Sales (1987-2005)
17
6.6
Average_Log_of_Employment
6.65
6.7
6.75
6.8
6.85
Average and Total Employment
1985
1990
1995
Year
2000
2005
620
Total_Employment
630
640
650
Figure 6: Average Employment (1986-2005)
610
A.2
1985
1990
1995
Year
2000
Figure 7: Total Employment (1986-2005)
18
2005
14.5
15
Total_Export
15.5
16
16.5
17
Total Exports and Average Profits/Sales
1985
1990
1995
Year
2000
2005
Average_Normalized_Profits
−5.00e−09
0
5.00e−09
Figure 8: Total Exports (1986-2005)
−1.00e−08
A.3
1985
1990
1995
Year
2000
Figure 9: Average Profits/Sales (1986-2005)
19
2005