Embed Size (px)
Citation preview
Determinants of Firm Profitability - TheEffect of Productivity and its Persistence †
Andreas StierwaldMelbourne Institute of Applied Economic and Social Research
The University of MelbourneJune 2009
Melbourne Institute for Applied Economic and Social ResearchThe University of Melbourne
Parkville Victoria 3010 AustraliaTelephone: +61 (0) 3 8344 2138.
Fax: +61 (0) 3 8344 2111.www.melbourneinstitute.com
† Preliminary version. Please do not quote from this paper. Comments are welcome and can be addressed
Determinants of Firm Profitability - TheEffect of Productivity and its Persistence
Abstract
The study investigates the determinants of firm profitability. Using data for 961large Australian firms for the period 1995-2005, the paper applies random and fixedeffect regression and corrects for dynamic panel bias. The profit model includes,among other variables, a time-variant, firm-level measure for total factor productivityobtained from an auxiliary cost function estimation. The analysis reveals that firm-level variables, such as lagged profit, productivity level and size, have a positive andlarge impact on firm profitability. Sector effects are present but play a minor role.
Keywords: Total Factor Productivity, Firm Performance, Determinants of Profit, DynamicPanel Bias.
JEL Classification: C23, D24, L25.
1 Introduction
The primary objective of this paper is to answer the question which factors are relevant and
important in determining firm profitability. The analysis reveals that firm-level variables, such
as lagged profit, productivity level and firm size, have a positive and large impact on firm
profitability. Sector effects are present but play a minor role.
Firm profitability and its determinants are a well addressed research topic in the field of
industrial organization. Modern literature provides two schools of competing models of firm
profitability. The structure-conduct-performance (SCP) model postulates that the degree of
concentration in an industry determines firm behavior and profitability. A higher concentra-
tion enables collusion between firms which can lead to higher profits. Firm effect models
argue that differences in firm-level characteristics, such as efficiency level, organizational
structure or quality of management, exist, persist and cause differences in profitability.
The fundamental assumption in firm effect models is that firms are heterogenous within an
industry. Specifically, the superior firm hypothesis, introduced by Demsetz (1973), states that
firms can be distinguished with respect to their level of efficiency.1 More productive firms
have a competitive advantage over their less productive rivals which is likely to be reflected
in profitability. Firms with higher levels of total factor productivity earn higher profits.
The superior firm hypothesis establishes a positive relationship between productivity and
profitability at the firm level. Taking these arguments further, Jovanovic (1982) postulates
1Throughout this chapter, the term firm effect models refers to studies that emphasize differencesbetween firms and their effect on profitability, for example Demsetz (1973), Rumelt (1991) and Hawawiniet al. (2003). Firm effect models are also referred to as revisionist, heterogeneity or resource basedmodels. The term efficiency refers to the level of cost-efficiency in the production process and will beused interchangeably with the terms total factor productivity (TFP) and productivity.
1
that only efficient firms stay in the market, and that less productive firms will eventually exit
the market.
There is an abundance of publications devoted to the analysis of firm profitability. Overall, the
evidence suggests that the SCP and firm effect models are plausible. This implies that firm
and industry characteristics are important to determine profitability. The role of total factor
productivity has, presumably due to measurement difficulties, received very little attention
in applied work. The present analysis advances the limited literature in this area.
The question of whether firm or industry effects determine firm profitability is important and
has implications for welfare analysis and, ultimately, for the design of competition policy. In
firm effect models, markets function competitively, and high firm profitability coincides with
industry concentration but is not caused by it. Demsetz (1973) warns of adverse effects
on overall efficiency levels through anti-trust policy. He argues that if concentration is high
because of high firm efficiency, anti-trust policy would eradicate incentives for efficiency
increases, see also Peltzman (1977).
Using a dynamic firm-level panel data set, this study identifies the determinants of firm
profitability and assesses the relative importance of firm and industry effects. The profit
model includes a firm-specific and time-varying estimate for total factor productivity which
is derived from an auxiliary cost function estimation. To the best of our knowledge, this is
the first study that employs firm-level measures for TFP and its persistence to estimate a
profit model.2
2In a broader context, one exception might be studies that use frontier efficiency scores and relatethem to other performance measures, such as stock market returns. Recent applications can be foundin, for example, Edirisinghe and Zhang (2008).
2
The remainder of the paper is organized as follows. Section 2 discusses related literature.
An dynamic panel data model for firm profitability is developed in Section 3, and Section
4 produces descriptive statistics for the sample. Section 5 documents results from random
and fixed effects regressions and from methods that correct for dynamic panel bias. Section
6 concludes the paper.
2 Related Literature
Models of firm profitability can be classified into two major groups, structure-conduct-
performance (SCP) and firm effect models. In the SCP model the market structure de-
termines firm behavior and profitability. In firm effect models, market structure is the result
of the distribution of firms and firm profits.
The SCP model is embedded in neoclassical theory and asserts that firms in concentrated
industries are more profitable than firms in perfectly competitive markets, see Bain (1951).
A reason for that can be that high industry concentration facilitates the exertion of market
power, for example in the form of monopoly pricing. Colluding firms impose a higher mark-
up on those goods with lower elasticity of demand without suffering the loss of demand to
competitive rivals. The increased price allows firms to earn profits that exceed competitive
rates. Due to the restricted quantity of supply, industry concentration and high profits are
associated with sub-optimal welfare levels.
The fundamental assumption in firm effect models is that firms are heterogeneous. According
to the superior firm hypothesis, introduced by Demsetz (1973), firms can be distinguished
with respect to their level of cost- or production efficiency. Efficient firms have a competitive
3
advantage over their non-efficient rivals. Higher levels of cost-efficiency can be caused by
lower costs of production, economies of scale or higher quality of products.
In the Demsetz model, superior performance can exist for some period of time. Potential
reasons for that can be the firm’s reputation, complex organizational structures, resource
heterogeneity, factor immobility or uncertainty of investments. Jovanovic (1982) argues that
only efficient firms survive, stay in the market, grow larger and obtain a higher market share.3
At the same time, efficient firms are more profitable than non-efficient ones.
Peltzman (1977) asserts that high market concentration, in the form of high market shares,
and high firm profitability occur simultaneously and are the result of the same cause, differ-
ences in productivity levels. Markets function competitively, and no collusion between firms
takes place that restricts supply or enables firms to raise their price above marginal costs.
For this reason, high firm profitability is not necessarily associated with welfare losses in firm
effect models.
There has been a substantial amount of empirical research undertaken in the area of profits,
market structure and firm-level effects.4 Taken together, the evidence suggests that both SCP
and firm effect models are plausible. This implies that industry effects, such as concentration
and entry barriers, and firm effects, such as productivity differences or strategic management,
are empirically important. Depending on the study, firm-level or industry-specific effects are
found to be the dominant factor on firm profitability.
3Non-efficient firms shrink, their market share declines and, eventually, they exit the market. Theflow of entry and exit into the industry prevents domination of few very large firms.
4Schmalensee (1989) and McGahan and Porter (2002) provide comprehensive overviews of appliedwork.
4
3 Empirical Specification
Following the literature, the specification below allows for lag-dependency in profitability and
the contribution of other firm characteristics in explaining firm profits.5 In a reduced form,
the basic model is given as
πij,t = f(πij,t−1, Xij,t−1, Dj, t), (1)
where πij,t and πij,t−1 represent current and lagged profitability for firm i in sector j respec-
tively, Dj are sector dummy variables and t a time trend.6 The lagged dependent variable
accounts for a dynamic component in profitability. The term Xij,t−1 contains a set of lagged
firm characteristics such as firm size and leverage ratio and age and financial risk. It also com-
prises a lagged estimate for total factor productivity and firm-level measures for productivity
persistence.
A linear dynamic model of firm profitability takes the form
πij,t = α+ βπij,t−1 + δXij,t−1 + dDj + εij,t, (2)
where α, β, δ and d are the parameters to be estimated. The model is based on the state
dependence model in Anderson and Hsiao (1982). The dependent variable πij,t is the current
profit rate of firm i at time t. Consider an error structure of the form εij,t = eij,t + νi, with
eij,t ∼ i.i.d.N(0, σ2e) and νi ∼ i.i.d.N(0, σ2
ν). The term eij,t is an idiosyncratic error that
accounts for the proportion of firm profit that correlates neither across time nor across firms.
5See, for example, Mueller (1977) for lag-dependency and Slade (2004) for a discussion of explanatoryfactors in profit models.
6The firms in the present sample are classified into sectors and sub-sectors according to the GlobalIndustry Classification System (GICS), which was developed by Standard and Poors. For this reason thispaper refers to sector effects what is traditionally called industry effects. Continuous sector variables areunobserved. However, the empirical specifications include, wherever possible, sector dummy variablesto account for sector-wide observed and unobserved factors.
5
The term νi captures unobserved heterogeneity in firm profitability. It can be interpreted as
a collection of factors that are specific to firm i but unobserved.
Nickell (1981) points out that using fixed effects to estimate (2) yields inconsistent parameter
estimates. A solution to this is to correct the coefficients for dynamic panel bias. Kiviet
(1995) derives a formula for the bias and corrects the original estimator. Bruno (2005)
provides an extension to unbalanced panels.7
Let ψ contain the fixed effect parameter estimates from the model in (2), ψfe = [β, δ]. Then
the bias correction is implemented through
ψbc = ψfe − bias, (3)
where the estimate for the bias approximation is subtracted from the original fixed effect
estimator.8 In particular, the bias approximation is a function of the unbiased coefficient ψm
and its variance σ2e,m, bias = f(ψm, σ
2e,m).
The subscript m indicates the method chosen to initialize the bias correction. The present
analysis uses Anderson and Hsiao (1982) instrument variable, Arellano and Bond (1991)
difference GMM and Blundell and Bond (1998) system GMM methods. Bun and Kiviet
(2001) argue that results are not sensitive to the choice of the initial estimator. Section 5
presents only results from difference GMM.
Commonly, all three approaches estimate a first-difference model. First differencing (2)
7In general, the larger the sample size N and time dimension T , the smaller the dynamic panelbias. Bun and Carree (2005) provide evidence that a bias-corrected fixed effect estimator may be moreappropriate than alternatives. Judson and Owen (1999) compare the performance of various dynamicpanel data estimators and find that the corrected fixed effect estimator can outperform instrumentvariable (IV) and generalized-methods-of-moments (GMM) estimators. Kiviet (1995) argues that abias corrected fixed effect estimator is much more efficient than IV or GMM estimators.
8Appendix A-2 presents the correction algorithm algebraically.
6
removes the fixed effect νi and, thus, eliminates a potential source of the dynamic panel
bias. The underlying assumptions are that the same relationships apply to the first-difference
and to the untransformed model and that the coefficients are identical. On the downside,
time-invariant firm characteristics disappear in first-difference models.9
4 Data and Descriptive Statistics
The sample comprises 961 large Australian firms for the period 1995-2005 and is, because
of missing observations, moderately unbalanced. The main source of the data is IRESS, a
financial information system. The information stems from balance sheet items, and from
profit and loss and cash flow statements. The database is supplemented with information
from the Australian Bureau of Statistics (ABS) and the Australian Securities and Investments
Commission (ASIC).10
Firm profits are computed as the ratio of profit level to the value of total assets. The level of
profit is defined as the difference between sales revenue and total operating expenses. Total
costs include labour, material and opportunity costs of capital.11
Accounting profit rates are not necessarily an unbiased measure for firm profitability. Sources
of distortions can be differences in depreciation practices, the treatment of R&D investments,
9It seems noteworthy that estimating (2) using system GMM and through the unofficial Stata com-mand -xtabond2- developed by Roodman (2006) includes sector dummies and time-invariant firm char-acteristics but does not change the direction of results.
10All firms in the sample are listed at the Australian Stock Exchange (ASX). In comparison to thetotal population of Australian firms, the average firm size in the sample is greater. The sample maycontain a self-selection bias because seeking listing and fulfilling the listing requirements is initially thefirm’s decision.
11Total costs as in the balance sheet item Total Operating Expenses do not include opportunitycosts of capital. Opportunity costs of capital are defined as the product of Total Fixed Assets andthe long-term money market interest rate. To obtain total costs, they are added to Total OperatingExpenses.
7
advertisement outlays and cost of human capital as current expenses instead of capitalized
stock. Furthermore, neither the profit level in the numerator nor total assets in the denomi-
nator of the profit measure defined above account for intangible assets.
However, despite the criticism accounting profit levels and rates are a useful measure to
approximate for firm profitability. Firms use accounting data themselves in internal decision
making. The rules and regulations to document financial information, in particular for firms
listed on the stock market, are stringent and enforced. Furthermore, the stock market is
highly sensitive to numbers published in accounting and financial reports.12
Table 1 presents observed profit rates by sector. On average over sectors and years, firms
in the sample earned an accounting profit of 0.5%. The finding of almost zero profit is
not pervasive across all sectors. Substantial differences exists between the sectors. The
large magnitude of the standard deviation illustrates a wide horizontal profitability dispersion
across firms.13
The evidence from Table 1 suggests that substantial heterogeneity in terms of profitability
exists, within the sample and within each sector. This finding can be interpreted as prima
facie evidence for the importance of sector-specific and firm-level effects on profitability.
Table 2 presents an overview of the variables used in the empirical estimation of the profit
model in (2).
12See Schmalensee (1989). Fisher and McGowan (1985) and Mueller (1990) discuss strengths andweaknesses associated with the use accounting profit rates. Lindenberg and Ross (1981) argue in favorof using Tobin’s q as an unbiased measure of firm performance. A correlation matrix in Appendix A-1indicates that alternative profitability measures are significantly positive correlated.
13Alternative profitability measures, such as return on assets and return on equity, indicate a similarpattern. The profit rates in Table 1 state unweighted averages within sectors and across time. Comput-ing size-weighted sector average profitability instead reveals that all sectors, except Financials, reporta sector average profitability that is positive.
8
Table 1: Firm Profitability by Sector, 1995-2005.Number Number Profit Rateof firms of obs. Mean S.D.
SectorDiscretionary 130 668 0.098 0.189Energy 77 304 -0.071 0.335Financials 99 339 0.014 0.254Health 100 359 -0.133 0.366Inform. Technology 94 369 -0.040 0.326Industrials 143 749 0.086 0.176Materials 239 860 -0.050 0.277Staples 39 218 0.058 0.146Telecommunication 25 102 -0.004 0.332Utilities 15 59 -0.011 0.155
All Sectors 961 4,027 0.005 0.271S.D. - Standard deviation. Table A-2 in the appendix produces descriptives statistics for the explana-tory variables.
Table 2: Variables in the Profit Model.Dependent Variableπij,t Current profit rate
Explanatory Variablesπij,t−1 Lagged profit rate
ln Aij,t−1 Lagged productivity estimate
di(Aij,t−1) Productivity persistence dummyΨij,t−1 Interaction of lagged productivity level
and productivity persistenceemplij,t−1 Lagged firm size (No. of employees)levij,t−1 Lagged leverage ratioageij,t Age of firmriskij Financial riskDj Sector dummy variable
Firm effect models, such as Demsetz (1973) and Jovanovic (1982), state that highly produc-
tive firms are more profitable than their less productive rivals and that this effect strengthens
with increasing persistence in high productivity levels. In terms of research design, this implies
the inclusion of variables that directly account for the the level and persistence of productiv-
9
ity in the profit model.14 The profit model in (2) contains an estimated for firm-level total
factor productivity obtained from an auxiliary cost function estimation.
Total factor productivity refers to the level of cost-efficiency in the production process and
is defined as the log-difference between predicted and empirical cost:
ln Aijt = ln C(Yijt,Wjt)− lnCijt(Yijt,Wjt), (4)
where ln C are the common, time-invariant costs and lnCijt the empirical costs of firm i in
sector j at time t.15 The resulting value Aijt is an index value, not specified in any units
and bound in the interval [−1, 1]. Larger values of Aijt imply higher levels of efficiency and
identify higher productivity firms.
In addition to the productivity estimate, the profit model in (2) employs a dummy variable
for productivity persistence and an interaction term. Both terms are based on firm-level
measures for productivity persistence. There is no unique approach to capture productivity
dynamics at the firm level. For example, the intertemporal autocorrelation (IAC) is defined
as
θ1i = Σt
(λi · σ2
ω
1− λ2i
)/Ti, (5)
where λi is the AR(1) parameter and σ2ω the variance in the specification aij,t = ci + λi ·
aij,t−1 + ωij,t with aij,t = ln Aij,t, Greene (1993). The term θ1i states the within-firm
14Traditionally, profitability studies approximate firm effects with the variable market share. Theunderlying assumption is that a firm’s relative position in the market is the result of its combinedcharacteristics. However, there are reasons to question the exogeneity and adequacy of the variablemarket share. Market share is a complex variable itself and potentially influenced by a number ofobserved and unobserved factors, see Shepherd (1972).
15Sector input prices and firm output are denoted with W and Y , respectively. Specifically, a three-input translog cost function is estimated together with two cost share equations using Iterative Seem-ingly Unrelated Regression (ITSUR). Coelli et al. (2005) summarize the cost function approach andprovide a classification of methods measuring TFP.
10
average covariance between past and current productivity outcomes. Larger values indicate
more persistent patterns of productivity.16
Using θ1i, firms are classified into persistent and non-persistent firms. Persistent firms are
those with values of θ1i exceeding the sample-wide 75th percentile.17 A persistence dummy
variable di takes the value unity when a firm is classified as persistent, and di = 0 when the
productivity level shows no persistent pattern. The latter group is a combined category and
includes firms that show very little or a moderate degree of persistence.
An interaction term is constructed as the product of the productivity level itself and the
persistence dummy variable, Ψij,t = di · aijt. In the case of a persistent firm, it takes
the value of aijt and is zero otherwise. Together, the persistence dummy variable di and
the interaction term Ψij,t account for the effect of high productivity and persistently high
productivity levels on firm profitability.
The regression analysis includes the firm-level control variables firm age, size, leverage ratio
and financial risk. Firm age serves as an approximation for intangible capital, such as market
experience. Firm size has a positive impact on profitability if larger firms benefit from
economies of scope, exploit scale economies or access capital at lower costs than smaller
firms.
Following the capital assets pricing mode, a firm with higher risks should compensate its
stakeholders with higher profits. Firm financial risk is defined as the variance of weekly stock
16Results presented in Section 5 are robust to a number of alternative persistence measures that allfocus on the volatility and dispersion of productivity within a firm. Examples are the mean relativedeviation from the within-firm average productivity θ2i = Σt
∣∣ aijt−¯ai¯ai
∣∣/Ti and the coefficient of variationθ3i = σai/
¯ai with σai and ¯ai as the within-firm standard deviation and average productivity, respectively,and Ti the number of periods observed.
17The results are not sensitive to alternative cut-off points, such as mean plus one standard deviation.
11
market returns. The larger the variation in stock market returns the higher the individual
risk. Similarly, it can be argued that a firm with more borrowed capital represents greater
financial risks to equity holders than a firm with relatively less debt. The variable leverage
ratio, given as the ratio of total liabilities to total assets, captures the effect of the capital
structure on profitability.
5 Results
The first column in Table 3 produces results for random effects. In order to compare the
findings to fixed effect and bias corrected fixed effect estimates in columns (III) and (IV),
the model is re-estimated in column (II) with a reduced set of explanatory variables.
All four models unveil a similar finding. Controlling for firm size and capital structure, the
major determinants of firm profitability in all four models are lagged profit rates, lagged
productivity level and persistence of high productivity. The more profitable and productive
firms were in the past, the higher their current profit.
Furthermore, the estimation results prove the presences of sector effects. In the random
effects models (I) and (II), a Wald test rejects the null hypothesis that all sector dummy
variables are jointly equal to zero. However, about the components of the sector effects can
only be speculated. The reason for this lies in the nature of dummy variables which capture
a collection of effects, and any significance cannot be attributed to a particular component.
Further information on sector attributes is not available due to the GICS classification in the
IRESS database.
12
Table 3: Determinants of Firm Profitability.Augmented Corrected
Variable random effects Random effects Fixed effects fixed effectsI II III IV
Lagged profit rate 0.360∗∗∗ 0.375∗∗∗ 0.130∗∗∗ 0.271∗∗∗
(0.015) (0.015) (0.018) (0.021)
Lagged productivity 0.104∗∗∗ 0.196∗∗∗ 0.142∗∗ 0.123∗∗
(0.047) (0.028) (0.038) (0.045)
Lagged productivity - 0.122∗∗∗ – – –persistence interacted (0.053) – – –
Productivity -0.014∗∗∗ – – –persistence dummy (0.013) – – –
Lagged no. of employees 0.030∗∗∗ 0.033∗∗∗ 0.018∗∗∗ 0.015∗∗
(0.003) (0.003) (0.005) (0.006)
Lagged leverage ratio 0.066∗∗∗ 0.061∗∗∗ 0.038∗ 0.039(0.020) (0.019) (0.023) (0.028)
Age -0.001 – – –(0.001) – – –
Financial risk -0.000∗∗∗ – – –(0.000) – – –
Constant -0.157∗∗∗ -0.189∗∗∗ -0.102∗∗∗ –(0.045) (0.045) (0.024) –
Time trend yes yes yes noSector dummy yes yes no no
R2 (overall) 0.511 0.510 0.464 –Correlation of residuals 0.132 0.131 0.241 –Wald test (χ2) 41.89 39.02 – –
No. of observations 3,926 4,027 4.027 2,874No. of firms 939 961 961 782
Statistical significance: *** at 1%, ** at 5%, * at 10%. For (IV), bootstrap standard errors from 1,000repetitions in parentheses. In comparison to (II) and (III), the number of observation in (I) is reducedbecause of missing values in the additional explanatory variables. In (IV), only dGMM is presented.The number of observation is further reduced because the first-difference model requires at least threeconsecutive observed periods with non-missing values for each firm. First-differencing also eliminates theconstant. R2 is not reported because the residual sum of squares is not constrained to be smaller thantotal sum of squares. Regressions implemented using the -xtreg- and -xtlsdvc- commands in Stata 9.2.
13
The coefficient for lagged profit rate is positive and significant. Firm profitability is moder-
ately serially correlated. A potential explanation is that high earnings in the past provide the
opportunity to realize high profits in the future. The larger the value, the more successful
the firm has been in maintaining its competitive position. Firms can benefit from previous
profits if, for instance, retained earnings are re-invested into research and development, and
successful product and process innovation increases future profits.18
The sign of the coefficient for lagged productivity is significantly positive. More productive
firms are more profitable. This finding lends support to the superior firm hypothesis in
Demsetz (1973). High levels of total factor productivity cause high firm profitability. A
potential explanation is that high productivity, manifested in, for example, in low average
costs of production, higher product quality or higher output quantities produces with fewer
inputs, leads to higher profits.
The coefficient for the interaction term between lagged productivity and productivity persis-
tence is positive and significant. This implies that the higher the level of productivity and
the more persistent the high productivity, the more profitable the firm. Both factors together
have a significant and positive impact on firm profitability.
Interestingly, in column (I) the sum of the coefficients of lagged productivity and the produc-
tivity - persistence interaction term exceeds the value of the lagged productivity in columns
(II), (III) and (IV). Evidently, high productivity levels and persistence of high levels together
are more important than high productivity alone. The productivity persistence dummy vari-
able is negative and significant. This can be explained by the fact that the dummy variable
18The finding of lag dependency of profits is in line with the literature. For example, Geroski andJaquemin (1998) find parameter values of 0.410 to 0.488 and Waring (1986) of, on average, 0.36 to 0.55.
14
indicates whether a firm shows a persistent pattern of productivity, independent at which
level.
The size of the firm significantly enhances firm performance. The positive and significant
parameter estimate for firm size illustrates that, in comparison to smaller firms, larger firms
are more profitable. This finding can be an indicator that larger firms exploit scale economies
and benefit from economies of scope. An alternative interpretation is that larger firms can
access capital at lower costs than smaller firms.
The coefficient for leverage ratio is significantly above zero. The higher the extent to which
debts were used as the source of financing, the higher the profits. An explanation can be that
profitable firms have had easier access to debt financing and do not need to rely exclusively
on equity capital. Alternatively, higher leveraged firms bear greater risks of bankruptcy and
need to compensate stakeholders with higher profits.
A potential explanation for the absence of an age effect in Table 3.6 is that the benefits from
intangible capital are already incorporated in high levels of cost-efficiency. Since productivity
is, in this application, defined in terms of cost-efficiency, it seems possible that the effect of
age is accounted for in the coefficient for lagged productivity.
The coefficient for financial risk is small but significantly negative. The result is not consis-
tent with the positive risk-return relationship predicted by the capital asset pricing model.
Literature in strategic management aims to explain the negative influence of risk on firm per-
formance, see, for example, Bromiley et al. (2001). Hurdle (1974) argues that the positive
coefficient for the variable leverage already accounts for firm’s risk.
15
Comparing the results from Models I and II to III, it can be concluded that firm fixed effects
are present. The magnitude of the lag-dependency decreases substantially in the fixed effects
model. Column (IV) produces results for coefficients that are corrected for dynamic panel bias
using difference GMM. In comparison to uncorrected coefficient values, the same structure
of correlations is unveiled, and the results do not change substantially.
Table 4 produces marginal effects and quantifies the relative importance of the determinants
of firm profitability. The numbers in the table illustrate the variation in the dependent variable
due to the transition in one of the explanatory variables from one standard deviation below
its mean to one standard deviation above its mean, leaving all other variables equal. For
sector dummy variables the change is from zero to one.
Table 4: Marginal Effects of Determinants of Firm Profitability.Marginal Effects
augmented bias correctedrandom effects random effects fixed effects
Variable (I) (II) (IV)Lagged profit rate 0.1946 0.2014 0.1494Lagged productivity level 0.0314 0.0595 0.0372Lagged productivity -
persistence interacted 0.0270 – –Productivity persistence dummy -0.0120 – –Lagged no. of employees 0.1452 0.1640 0.0754Lagged leverage ratio 0.0301 0.0281 0.0179Age -0.0131 – –Financial risk -0.0531 – –Sector dummy variables 0.0800 0.0771 –Number of firms 939 961 782Number of observations 3,926 4,027 2,874
Note: Results are derived from earlier regressions in Table 3. Augmented random effects from column(I), random effects from column (II) and bias corrected estimation from column (IV). For sectordummy variables, only the largest contribution is shown.
Table 4 unveils that past profits and firm size are the principal determinants of firm prof-
itability. Other firm characteristics, such as productivity level and its persistence, leverage
16
ratio and financial risk, are relevant but have a much smaller relative importance. Sector
effects are present, but play a minor role (at most one half of the size effect). Results from
bias corrected fixed effects illustrate similar findings.
6 Conclusion
Economic literature suggests competing models of firm profitability. The structure-conduct-
performance model postulates that the degree of industry concentration determines firm
behavior and profit. The higher the concentration in an industry the higher the profit of
firms in that industry. Firm effect models consider heterogeneity within industries. The
distribution of profits depends on firm characteristics. Empirical work has supported both
types of models.
The primary objective of this study is to verify the predictions of firm effect models using panel
data for 961 large Australian for the period 1995-2005. Results from random and fixed effect
regressions and from procedures to correct for dynamic panel bias unveil a similar pattern.
The sample is characterized by a large amount of heterogeneity in terms of profitability.
The determinants of firm profitability are lagged profit rate, lagged productivity level, its
persistence, firm size and sector effects. In terms of relative importance, lagged profits, size
and productivity level have the largest impact on current profits. Sector effects are present
but play a minor role.
The analysis verifies the predictions of firm effect models that firm-level effects determine
differences in profitability and that sector-wide effects have little impact. This has implica-
tions for welfare analysis because in firm effect models high firm profitability is the result of
17
competitive processes and not market failure. The design of competition policy should be
mindful of heterogeneity among firms.
An imminent extension to the analysis in this chapter could be to investigate how long firms
can maintain the competitive advantage in the form of relatively high productivity levels,
and whether its performance enhancing effect eventually disappears or not. Further pursuing
these questions requires an extended panel dataset with more observed periods.
18
Bibliography
Anderson, T. W. and C. Hsiao (1982). Formulation and Estimation of Dynamic ModelsUsing Panel Data. Journal of Econometrics 18, 47–82.
Arellano, M. and S. Bond (1991). Some Tests of Specification for Panel Data: Monte CarloEvidence and an Application to Employment Equations. Review of Economic Studies 58,277–297.
Bain, J. S. (1951). Relation of Profit Rate to Industry Concentration: American Manufac-turing, 1936-1940. Quarterly Journal of Economics 65(3), 293–324.
Blundell, R. and S. Bond (1998). Initial Conditions and Moment Restrictions in DynamicPanel Data Models. Journal of Econometrics 87, 114–143.
Bromiley, P., K. D. Miller, and D. Rau (2001). Risk in Strategic Managment Research. InM. A. Hitt, R. E. Freeman, and J. S. Harrison (Eds.), The Blackwell Handbook of StrategicManagement, pp. 259–288. Oxford: Blackwell.
Bruno, G. S. F. (2005). Approximating the Bias of the LSDV Estimator for Dynamic Unbal-anced Panel Data Models. Economic Letters 87, 361–336.
Bun, M. J. G. and M. A. Carree (2005). Bias-Corrected Estimation in Dynamic Panel DataModels. Journal of Business and Statistics 23(2), 200–210.
Bun, M. J. G. and J. F. Kiviet (2001). The Accuracy of Inference in Small Samples ofDynamic Panel Data Models. Tinbergen Institute Discussion Paper TI 006/4, 1–78.
Bun, M. J. G. and J. F. Kiviet (2003). On the Dimishing Returns of Higher-Order Terms inAsymptotic Expansions of Bias. Economic Letters 79, 145–152.
Coelli, T. J., D. Rao, C. J. O’Donnell, and G. E. Battese (2005). An Introduction to Efficiencyand Productivity Analysis (Second ed.). New York: Springer.
Demsetz, H. (1973). Industry Structure, Market Rivalry, and Public Policy. Journal of Lawand Economics 16(1), 1–9.
Edirisinghe, N. C. P. and X. Zhang (2008). Portfolio Selection Under DEA-Based RelativeFinancial Strength Indicators: Case of US Industries. Journal of Operational ResearchSociety 59, 842–856.
Fisher, F. M. and J. J. McGowan (1985). On the Misuse of Accounting Rates of Return toInfer Monopoly Profits. American Economic Review 73(1), 82–97.
Geroski, P. A. and A. Jaquemin (1998). The Persistence of Profits: A European Comparison.Economic Journal 98(391), 375–389.
Greene, W. H. (1993). The Econometric Approach to Efficiency Analysis. In H. O. Fried,L. A. K. Lovell, and S. Schmidt (Eds.), The Measurement of Productive Efficiency: Tech-niques and Applications, pp. 68–119. New York: Oxford University Press.
19
Hurdle, G. J. (1974). Leverage, Risk, Market Structure and Profitability. Review of Economicsand Statistics 56, 478–485.
Jovanovic, B. (1982). Selection and the Evolution of Industry. Econometrica 50(3), 649–670.
Judson, R. A. and A. L. Owen (1999). Estimating Dynamic Panel Data Models: A Guidefor Macroeconomists. Economic Letters 65, 9–15.
Kiviet, J. F. (1995). On Bias, Inconsistency, and Efficiency of Various Eestimators in DynamicPanel Data Models. Journal of Econometrics 68, 53–78.
Lindenberg, E. B. and S. A. Ross (1981). Tobin’s q Ratio and Industrial Organization. TheJournal of Business 54(1), 1–32.
McGahan, A. M. and M. E. Porter (2002). What do we Know About Variance in AccountingProfitability. Managment Science 48(7), 834–851.
Mueller, D. C. (1977). The Persistence of Profits Above the Norm. Economica 44, 369–380.
Mueller, D. C. (1990). The Dynamics of Company Profits. Cambrigde: Cambridge UniversityPress.
Nickell, S. (1981). Biases in Dynamic Models with Fixed Effects. Econometrica 49(6),1417–1426.
Peltzman, S. (1977). The Gains and Losses from Industrial Concentration. Journal of Lawand Economics 20(2), 229–263.
Roodman, D. (2006). How to do xtabond2: An Introduction to Difference and System GMMin Stata. Working Paper, Centre for Global Development, Washington 103, 1–51.
Schmalensee, R. (1989). Inter-Industry Studies of Structure and Performance. InR. Schmalensee and R. D. Willlig (Eds.), Handbook of Industrial Organization, Vol. II, pp.952–1009. Amsterdam, New York, Oxford, Tokyo: North-Holland.
Shepherd, W. G. (1972). The Elements of Market Structure. Review of Economics andStatistics 54(1), 25–37.
Slade, M. (2004). Competing Models of Firm Profitability. International Journal of IndustrialOrganization 22, 289–308.
Waring, G. F. (1986). Industry Differences in the Persistence of Firm-Specific Returns.American Economic Review 86(5), 1253–1265.
20
Appendix
A-1 Alternative Profit Measures.
Table A-1: Correlation Matrix for Alternative Profit Measures.
AdROE AdROA ROE ROA EBITM NPBTM Tobin’s q
AdOE 1AdROA 0.448∗ 1ROE 0.839∗ 0.610∗ 1ROA 0.467∗ 0.962∗ 0.679∗ 1EBITM 0.340∗ 0.703∗ 0.550∗ 0.759∗ 1NPBTM 0.335∗ 0.693∗ 0.569∗ 0.752∗ 0.941∗ 1Tobin’s q 0.497∗ 0.251∗ 0.418∗ 0.274∗ 0.273∗ 0.266∗ 1
Number of observations: 4,027. Number of firms: 961. * - Correlation statistically significant on the1% level. ROA: accounting return on assets. ROE: accounting return on equity. AdROE: adjustedreturn on equity ([EBITDA - (fixed assets · 10year interest rate)] / equity). AdROA: adjusted returnon assets ([EBITDA - (fixed assets · 10year interest rate)] / total assets). EBITM: EBIT margin(EBIT / sales revenue). NPBTM: NPBT margin (NPBT / sales revenue). Tobin’s q: market valueof assets / book value of assets.
Table A-2: Descriptive Statistics, 1995-2005.
Variable Mean Median S.D. Min Max
Profit rate 0.005 0.073 0.271 -1.648 0.719Productivity 0.032 0.018 0.152 -0.970 0.907No. of employees 1,849 148 7,239 1 213,000Leverage ratio 0.415 0.439 0.230 0.001 1.000Age 12.4 10.0 10.3 0.000 43Risk 0.022 0.013 0.040 0.001 0.833
Number of observations 4,027Number of firms 961
Note: Number of persistent firms: 135 (θ1i) and 255 (θ2i). S.D. - standard deviation.
21
A-2 Correction for Dynamic Panel Bias
The dynamic panel bias approximation can be simplified into eight steps. An extensivederivation can be found in Kiviet (1995) and Bruno (2005).
(1). Consider the dynamic panel data model from (2)
πij,t = α+ βπij,t−1 + δXij,t−1 + εij,t, (A-1)
with εij,t = eij,t + νi, eij,t ∼ i.i.d.N(0, σ2e) and νi ∼ i.i.d.N(0, σ2
ν).
(2). Collecting data over time t and firms i gives the matrix form of (A-1) as
π = Wψ +Dν + e, (A-2)
where π and e are (NT × 1) vectors of the dependent variable and disturbances,
W = [π−1...X] a (NT × k) matrix of stacked observation, ν a (n × 1) vector of
fixed effects, D = (IN⊗
ιT ) a (NT × N) matrix of firm-specific dummy variables,ψ = [β, δ] a (k × 1) vector of parameters and N , T , k the number of observations,periods and parameters, respectively.
(3). Extending to unbalanced panels gives
Sπ = SWψ + SDν + e,
where S is a (NT × NT ) block-diagonal matrix with the dynamic selection rule sij,t
on the diagonal. Define rij,t = 1 if (πij,t, Xij,t) is observed and 0 otherwise. Similarly,sij,t = 1 if (rij,t, rij,t−1) = (1, 1) and 0 otherwise. The dynamic selection rule sijt
ensures that the unbalanced panel contains only pairs of observations for which currentand one-period lagged values are not missing.
(4). The fixed effect estimator for model (A-2) is given as
ψfe = (W ′MsW )−1W ′Msπ, (A-3)
with Ms as a symmetric and idempotent (NT × NT ) matrix that eliminates fixedeffects and selects suitable observations, Ms = S(I −D(D′SD)−1D′)S.
(5). The bias of the estimator in (A-3) can be expressed as
E[ψfe] = E[(W ′MsW )−1W ′Mse]. (A-4)
(6). Kiviet (1995), Bun and Kiviet (2003) and Bruno (2005) show how after replacing Ms
and some algebraic transformation an expression for the bias approximation can beobtained that is in essence a function of the consistent parameter estimate ψm andthe variance σ2
e,m,
biasm(ψfe) = f(ψm, σ2e,m). (A-5)
22
Given the consistent estimator ψm, an estimate for the variance is obtained from
σ2e,m =
e′mMsem
(N − k − T ), (A-6)
where em = π −Wψm.
The subscript m indicates the method chosen to initialize the bias correction. Thepresent analysis uses Anderson and Hsiao (1982) instrument variable (IV-AH), Arellanoand Bond (1991) difference GMM (dGMM) and Blundell and Bond (1998) systemGMM (sGMM). The bias approximation is of the order O(N−1T−2).
(7). The true parameter value and variance are, of course, unobserved and not feasiblefor bias correction. Instead, IV-AH, dGMM and sGMM methods are used to obtainconsistent estimators for ψm and σ2
e,m.
(8). Lastly, insert ψm and σ2e,m into (A-5) to correct the original estimator with the bias
approximation to obtain the bias corrected estimator
ψbc = ψfe − biasm(ψm, ˆσ2e,m). (A-7)
23
