# Employment Effects of Innovation at the Firm Level

**ABSTRACT** This paper analyzes empirically the effects of innovation on employment at the firm level using a uniquely long panel dataset of German manufacturing firms. The overall effect of innovations on employment often remains unclear in theoretical contributions due to reverse effects. We distinguish between product and process innovations and additionally introduce different innovation categories. We find clearly positive effects for product and process innovations on employment growth with the effects for process innovations being slighthly higher. For product innovations that involved patent applications we can identify an additional positive effect on employment.

**1**Bookmark

**·**

**81**Views

- Small Business Economics 12/2014; · 1.55 Impact Factor
- SourceAvailable from: Michaela Fuchs[Show abstract] [Hide abstract]

**ABSTRACT:**Empirical research on agglomeration and regional economic growth puts high emphasis on the impact of specialization, diversity, and competition on regional employment dynamics (Glaeser et al. 1992, Henderson et al. 1995, Blien et al. 2006, Fuchs 2009). However, Beugelsdijk (2006) and Raspe/van Oort (2008) argue that this relationship should most profoundly hold at the micro or firm level. This paper centres on the labour demand of individual plants and assesses the influence of regional features in direct contrast to plant-specific characteristics as well as conventional labour-demand determinants. Hence, it contributes to the sparse literature on the importance of regional character-istics for firm performance and additionally integrates research from industrial as well as labour economics. The analysis is based on the IAB Establishment Panel, a comprehensive data set on German plants. For the years from 2004 to 2008 it encompasses observations on roughly 8,000 plants. The regional variables are added on the NUTS3-level. First econometric results confirm the basic hypotheses derived from labour-demand theory: wages exert a significantly negative and output a positive influence on the number of employees. Among the plant-specific characteristics, it is mainly plant size, exporting behaviour and R&D / innovation activities that foster employment. There are also distinctive differences regarding the single sectors. Last but not least, the regional environment plays a decisive role for plant-level labour demand. The size of the region the plant is located in, the degree of sectoral concentration as well as of competition within a sector have a positive and highly significant impact. By contrast, accessibility to highways, specialization, and diversity seem to be of minor relevance01/2011; - SourceAvailable from: Jens Horbach[Show abstract] [Hide abstract]

**ABSTRACT:**The employment effects of environmental technologies are in the focus of politicians but there are only few studies analyzing these effects for different environmental innovation fields. We use the 2009 wave of the German part of the Community Innovation Panel (CIS) allowing for such an analysis at the firm level. The main focus of the paper lies on the analysis of the adaptation behavior of firms with respect to the relationship of employment and (environmental) innovation. We use an endogenous switching regression approach to take the simultaneous haracter of innovation activities and employment demand into consideration. Our econometric analysis shows that innovative firms in general are characterized by a significantly more dynamic employment development. Especially the realization of environmental process innovations leads to a higher employment within the firm. The theoretical background of this finding is that process innovation induced cost savings improve the competitiveness of firms. This has a positive effect on demand and thus also increases employment. A more detailed analysis by different environmental innovation fields shows that material and energy savings are positively correlated to employment because they especially help to increase the profitability and competitiveness of the firm. On the other side, air and water process innovations that are still dominated by end-of-pipe technologies have a negative impact on the employment development. --Journal of Cleaner Production 01/2012; · 3.59 Impact Factor

Page 1

Employment Effects of Innovation at the Firm Level

Stefan Lachenmaier

Horst Rottmann

Ifo Working Paper No. 27

April 2006

An electronic version of the paper may be downloaded from the Ifo website www.ifo.de.

Page 2

Ifo Working Paper No. 27

Employment Effects of Innovation at the Firm Level

Abstract

This paper analyzes empirically the effects of innovation on employment at the firm level

using a uniquely long panel dataset of German manufacturing firms. The overall effect of

innovations on employment often remains unclear in theoretical contributions due to reverse

effects. We distinguish between product and process innovations and introduce in addition

different innovation categories. We find clearly positive effects for product and process

innovations on employment growth with the effects for process innovations being slightly

higher. The effects are stronger in small firms and differ between firms in former West and

East Germany.

JEL Codes: J23, O30, L60

Keywords: innovation, labour demand, employment, firm size, panel data

Stefan Lachenmaier

Ifo Institute for Economic Research

at the University of Munich

Poschingerstr. 5

81679 Munich, Germany

Phone: (+49) 89-9224 1696

E-mail: lachenmaier@ifo.de

Horst Rottmann

Research Professor at the Ifo Institute for

Economic Research at the University of Munich

and

University of Applied Sciences Amberg-Weiden

Hetzenrichter Weg 15

92637 Weiden, Germany

Phone: (+49) 961-382 179

Fax: (+49) 961-382 110

E-mail: h.rottmann@fh-amberg-weiden.de

Page 3

Page 4

3

1. Introduction

This paper delivers empirical evidence for the effects of innovations on employment. It

contributes to the existing research by using a uniquely rich dataset of German

manufacturing firms. The dataset combines annually surveys over the last 22 years and

thus delivers a panel dataset, that allows analyses over a long time horizon. The

theoretical literature stresses the importance of the distinction between product and

process innovations. But for both types the overall effects on employment remain

unclear, with the effect depending mainly on the demand elasticity of the affected

products. Thus, pure theoretical analyses are not able to deliver clear predictions for the

effects of innovations on employment, which raises the need for empirical evidence.

With our data set we can analyze the German manufacturing sector for two decades

with the possibility to distinguish between product and process innovations. In addition,

we introduce different categories of innovation representing different importance levels

of the respective innovations. We also address the questions of whether the effects

differ between small and large firms or differ between firms which are located in former

West and East Germany. In this paper we concentrate on longer periods and do not try

to model the year-to-year employment adjustment processes. This becomes especially

difficult for small firms, which are also part of our dataset.1

The paper is structured as follows. Section 2 gives a short overview about the

existing theoretical and empirical literature in this research field. Section 3 presents our

identification strategy. Section 4 describes the data base and presents the descriptive

statistics. The results are presented in section 5; section 6 concludes the paper.

2. The Literature on Innovation and Employment

2.1 Theory

In theoretical contributions on the impact of innovation on employment, the direction

of the effect of technological progress often remains unclear. Researchers have been

analyzing this task for a long time, and their analyses differ mainly in the methodology

and the data available. An historical overview about the evolution of this field of

research is given in Petit (1995).

1 See e. g. Hamermesh / Pfann (1996)

Page 5

4

In the theoretical literature the distinction between product and process innovations

has been proved important (Stoneman 1984, Hamermesh 1993, Katsoulacos 1986).

Whereas for product innovations it is meanwhile generally assumed that they enhance

employment via a higher demand created by the introduction of new products or an

improved quality of existing products, it is especially the effects of process innovations

that leave open questions.

But for both types of innovation there are effects on employment that go in opposite

directions. The introduction of new or improved products creates a new demand for

these products. This increasing demand leads to an increase in employment in the

innovating firm. But the innovation can also lead to a (temporary) monopoly of the firm

or at least to a very high market share of the firm. If the firm takes advantage of this

situation and increases the product price to maximize its profits the employment level

may suffer from this reduction in the amount of output. Also for process innovations the

overall effect is not clear in theory. As a process innovation improves the labour

productivity, the direct effect of a process innovation might reduce the number of

workers since the same output can be achieved by fewer workers. But, if this advantage

of a cheaper production process is passed on to the prices, this might increase the

demand for the product. This increase in demand might then ─ depending on the

demand elasticity ─ lead to an increase in employment.

To sum up the theoretical contributions, a clear statement is not possible on the

direction of the effect of innovations on employment at the firm level. The effects can

differ significantly depending on the size of the contrary direct and indirect effects,

which depend on the prevailing market structure and on the price elasticity of product

demand.

2.2 Empirical evidence

The empirical literature on technological progress and its impact on different

economic measures is extensive. What we will concentrate on in this paper is the

microeconometric analysis of the effects of innovation on employment.2 This strand of

literature started mainly in the 1990s with the increasing availability of micro data on

firms’ innovation behaviour. An excellent overview of microeconometric analyses in

2 Topics not to be covered in this paper include the effects on wages and skill-biased technological

change.

Page 6

5

this field of research is given in Chennells / Van Reenen (1999). As suggested by

theoretical contributions, the empirical analysis usually distinguishes between product

and process innovation. In almost all analyses a positive effect of product innovations is

found; for process innovations there is also a tendency for a positive effect but the

analyses are not that clear.

The methods used are widespread as are the countries covered and the employed

variables. These include the innovation variables (or proxy variables for innovation) as

well as control variables. In terms of econometric models one can divide the existing

literature mainly in three parts: cross-sectional analyses, analyses of the growth rates

with data of two different points in time and panel data analysis.

Early contributions are mainly based on cross-sectional data due to the limited data

availability. Contributions in this line are Zimmermann (1991), Entorf and Pohlmeier

(1990) and König et al. (1995). Zimmermann (1991) and Entorf and Pohlmeier (1990)

also use data of the Ifo Insitute, but from a different survey, in which the innovation

data is not as detailed as in the innovation survey. Zimmermann (1991) concludes that

technological progress played an important role in the decrease of employment in 1980.

Entorf and Pohlmeier (1990), however, show a positive effect of product innovations on

employment while process innovations showed no significant effect. König et al. (1995)

also use German data, stemming from the “Mannheimer Innovationspanel” in 1993 and

also found a positive effect of product innovations on labour demand.

Newer analyses combine two surveys of different points in time and therefore are

able to explain the growth rate of employment between these two points in time.

Brouwer et al. (1993) are in this line of literature with their analysis of Netherlands data

of 1984 and 1989. The authors show a negative effect of total R&D investment on

employment growth, but a positive effect for those R&D expenses related to creating

new products. Blanchflower and Burgess (1998) find a positive relation between

process innovations and employment growth in the UK in 1990 and in Australia in

1989/1990. Doms et al. (1995) also show a positive relation between the use of modern

technology and employment growth between 1987 and 1991 using firm data of the U.S.

manufacturing sector together with data from a technology survey in 1988. Klette und

Forre (1998) have matched different data sets for Norway. Census data was combined

with several surveys between 1982 and 1989. Their, mainly descriptive, analysis did not

show a clear positive relation between innovations (measured as firms conducting R&D

Page 7

6

vs. firms not conducting R&D) and employment. Using German data from the

Community Innovation Survey (CIS3), Peters (2004) analyzes employment growth

between 1998 and 2000. Product innovations show a significantly positive effect on

employment growth whereas process innovations showed a negative effect for German

manufacturing firms. Also using CIS data, Blechinger et al. (1998) find positive effects

of product as well as process innovation on employment growth for the Netherlands

between 1988 and 1992 and for Germany between 1992 and 1994.

The third type ─ panel studies ─ are the rarest ones. A first step in this direction is

Greenan and Guellec (2000), who use firm panel data, but they match it with a cross-

sectional innovation survey. Their results show that innovating firms (and innovative

sectors) have created more new jobs than non-innovating firms (less innovative sectors).

Their results suggest that, on the firm level, process innovations play the more

important role whereas on the sector level it is the product innovations that are more

important. Real panel analyses over a longer time horizon are the contributions of

Smolny (1998), Flaig and Rottmann (1999), van Reenen (1997) and

Rottmann/Ruschinski (1997). Smolny (1998) analyzes data of German firms from the

Ifo Business Survey and the Ifo Investment Survey from 1980 to 1992. Using pooled

OLS regressions, he shows a positive effect of product innovations as well as process

innovations. Van Reenen (1997) matches firm data of firms listed at the London Stock

Exchange with the English innovation database of the SPRU (Social Policy Research

Unit). With this data set for 1976-1982 he estimates panel models, which allows him to

control for fixed effects, dynamics and endogeneity. But still he finds positive effects of

innovation on employment. Rottmann and Ruschinski (1997) carried out analyses with

data from the Ifo Institute. The authors show, in their analysis of the effects of

technological change on employment growth, the importance of controlling for

unobserved firm heterogeneity and adjustment processes. Controlling for these effects

the authors find significantly positive effects of product innovations and significantly

negative effects of process innovations on employment growth. An additional important

variable in their models is the expected demand growth, which shows a positive effect

on employment. Building on these results the authors also use a dynamic panel method,

the Anderson-Hsiao framework (Rottmann and Ruschinski 1998). The positive effect of

product innovations was also found in this analysis, but process innovations showed no

significant impact. Flaig and Rottmann (1999) control for unobserved firm

Page 8

7

heterogeneity and estimate a recursive equation model with output, output expectations

and employment as endogenous variables. They also find positive effects for product

and process innovations. All these studies, even the panel studies, are restricted to a

relatively limited time horizon. In addition, these studies do not include any quality

measures of the innovation outputs.

3. The Estimation Strategy

Our identification and estimation strategy combines different elements of the literature

mentioned above. We extend the existing literature on innovation and employment not

only in terms of a broader variety of innovation variables but also on applying a

different estimation strategy.

We assume that labour demand can be described by the following equation in levels,

,

(1)

where L is labour demand, T is a measure for the technology used in the production

process, Q is a measure for the quality of the product and X denotes other control

variables, which we specify in more detail in equation (3). In our analysis we

concentrate on the growth rates and thus transform the function: First, we take log

values (denoted by lower case letters) and second, we first difference the equation

(denoted by the difference operator Δ). This procedure basically is a first-difference

panel approach, by which we also already account for the possible unobserved firm

heterogeneity. Otherwise a spurious relationship between innovation and employment

could be generated due to unmeasured factors that are reasonably stable over time like

quality or risk tolerance of management. If such effects were present in the level

equation, these time-constant firm specific effects drop out by taking first differences:

(2)

For the estimation of equation (2) we need a measure for the progress in the applied

technology and for the improvement in the product quality. These changes can be

approximated by our innovation variables. The implementation of a process innovation

),(

XQTfL =

xqtl

3210

ββββ

+Δ+Δ+=Δ

Page 9

8

can be interpreted as the change in the production technology, and the introduction of a

product innovation can be interpreted as a change in the product quality. Substituting Δt

and Δq with our innovation variables and introducing additional control variables on the

sector level we get the following estimation equation:

(3)

Ipc denotes the process innovations and Ipd denotes the product innovations. Δw and

Δg are additional control variables at the sector level (NACE two-digit classification).

Δw denotes the growth rate of the real hourly wage rate, which of course may influence

the employment demand of a firm. Since the wage rate of the individual firms are not

observed, the average sectoral real hourly wage rate is used here as the best proxy

available. Δg denotes the growth rate of the Gross Value Added in the sector and is

included as a control variable for the demand situation in the respective sector.

Since the unobserved firm effects are already differenced out, we can – following the

first difference panel approach – estimate this differenced equation by least squares

regressions. Equation (3) is a static version of a labor demand equation. Adjustment

costs for employment and expectation formation will induce dynamics to equation (3).

Modeling these adjustment processes is a very complex topic (Hamermesh and Pfann

1996), especially within small firms. Furthermore, innovations do not only have

employment effects in the year of their introduction; they are likely to influence

employment growth in the following years, too. Little is known about the delayed

effects of innovation. Therefore, we use an estimation strategy employed in labor

market analyses, where one does not expect instant (yearly) effects of different

institutional arrangements on unemployment (e. g. Nickell 1997, 2003 and Blanchard

and Wolfers 2000). In this kind of analyses averages for longer time periods are

calculated, usually for 5-year-periods, to smooth out the year-on-year noise and detect

long-term effects of institutions on the labour market. Assuming that innovations do not

show their effects on employment growth in a short time horizon, i.e. from year to year,

we apply these estimation technique and calculate averages over four and five year

it it it

it

Pd

it

Pc

it

ugwIIl

+Δ+Δ+++=Δ

43210

βββββ

Page 10

9

periods.3 We then use these periods as time units in our panel estimations. That means

the time index t in our estimation equation does not denote a single year anymore but a

whole time period. The values of the variables are the calculated averages per period.

So Δlit stands for the average yearly employment growth rate per firm within one

period. Ipc and Ipd are the average number of years per period in which a firm gave a

positive answer to the questions whether any process or product innovation was

introduced. Δw and Δg are averages of the yearly growth rate per period, but on a

sectoral level. Additionally we introduce the variable eit, which denotes the log of the

employment start level of a firm in the respective four and five year period.

(4)

eit controls for the possible differences of the growth rate in small and large firms.

Or, in other words, it is a test for Gibrat’s Law, which states that the growth rate of a

firm is independent of the size of a firm (Gibrat 1931). Many studies have dealt with the

empirical test of Gibrats’s Law, especially in manufacturing firms. The underlying

result of these studies is that Gibrat’s Law does (often) not hold in the manufacturing

sector, especially for small firms (e.g. Sutton 1997, for Germany: Wagner 1992,

Harhoff et al. 1998 and Almus / Nerlinger 1999). There is a strong tendency that

initially smaller firms tend to grow at a faster rate than initially large firms. Only for

special samples, large manufacturing firms (Hall 1987, Evans 1987) or for service firms

(Audretsch et al. 2004) are there empirical results that lead to the assumption that

Gibrat’s Law is valid in these cases.

Our estimation strategy might raise some concern about estimating causal effects.

The reason for that is the problem of endogeneity of the innovation variables. They

might be correlated with the error term of the labour demand function. But, following

this argument, one has to keep in mind that the unobserved individual effects cannot be

responsible for such a correlation since they dropped out as we took first differences of

3 Due to our sample of 22 years, we calculate averages for three 4-year periods and two 5-year

periods. These are the periods from 1982-1986, 1987-1990, 1991-1995, 1996-1999 and 2000-2003. By

setting a border between 1990 and 1991 we also account for the problem that arises in data due to

German reunification. All data up to 1990 refer to former West Germany; all data since 1991 refer to

Germany. We also tested several other lengths of periods; details are described in chapter 5.2.

ititit it

it

Pd

it

Pc

it

uegwIIl

++Δ+Δ+++=Δ

543210

ββββββ

Page 11

10

our estimation equation. If there is no autocorrelation in the error terms, the only factor

leading to an endogeneity problem might be a contemporaneous correlation of the

innovation variables with the error term uit, resulting from a shock simultaneously

affecting employment and innovation. In case that such a shock occurs, a possible

solution of this problem in our estimation strategy would be an instrumental variable

strategy. The questionnaire contains two questions that might offer useful instruments.

First, firms are asked which innovation impulses led the firm to start the innovation

process. Second, thay are asked for their innovation expenses. But the construction of

these instruments leads to additional problems. Beside the question of whether these

instruments are uncorrelated with the error term, the construction of the survey

questionnaire raises some concerns: The information on innovation impulses and

innovation expenses is only available for those firms that introduced any innovation.

Therefore we have to make questionable assumptions for those firms that did not

introduce any innovations: For all those firms we have to replace the missing

information in innovation impulses and innovation expenses by the value zero as a best

approximation. But, using this strategy, our results did not show robust results. Either

the instruments used showed a low explanatory power of the innovation variables or the

exogeneity assumption was rejected by Sargan tests.4

4. Database and Descriptive Statistics

4.1 The Ifo Innovation Survey

The data source used in this analysis is the Ifo Innovation Survey. The Ifo Innovation

Survey is conducted yearly by the Ifo Institute for Economic Research at the University

of Munich. It was started in 1982, since that time the Ifo Institute has collected the

answers of, on average, 1500 respondents every year, including eastern German firms

since 1991. The latest data, used in this analysis, stem from the questionnaire in 2004,

which describes the innovation behavior of the year 2003. The observation unit of this

survey is not necessarily always a whole firm. For firms, that produce more than one

product, the questionnaire refers only to a certain product range, i.e. for multi-product

4 Results are not presented but are available from the authors on request.

Page 12

11

firms the survey delivers even more detailed data than firm level data. For reasons of

clarity, in the following we use the expression “firm” as the cross-sectional unit, even if

it might not be correct in the case that there are different product ranges from one firm

in the sample. This survey gives us a total sample of 33,159 observations from 7,023

different firms over 22 years from 1982 to 2003.

The questionnaire offers different innovation measures. The first one is the simple

information of whether the firm has introduced any innovation during the last year. This

information is available for product as well as for process innovations as required by the

theoretical models (see section 2.1). One can argue that a potential drawback of the

simple innovation variable is the lack of detailed information about the importance of

the innovation. But, as the discussion for a “correct” measurement of innovation is still

ongoing in the literature, we of course do not claim to have a perfect measure for

innovation here. Other innovation variables like R&D or patents also have advantages

and disadvantages. A comparison of the Ifo innovation measure with other popular

measures is given in Lachenmaier/Wößmann (2004). In addition to the simple

innovation dummy variable we also try to increase the explanatory power of this

innovation variable by introducing different categories of innovations. We use different

questions relating to the “importance” of an innovation. These questions give

information on whether R&D was necessary for the implementation of a new

innovation and if any patent applications were filed during the innovation process.

4.2 Descriptive statistics

The dataset consists of an unbalanced panel with 33,159 observations, collected from

7,023 firms over the 22 years 1982-2003. The survey is conducted among German

manufacturing firms. But as described in our estimation strategy in section 3, we do not

use yearly data but the averages over four or five year periods. Therefore we will

present the descriptive statistics according to the observation units in our regressions,

which are the averaged values per period. If a firm has not answered in all years during

a period, we calculate the averages of the available observations as the best estimation

for the whole period. Due to the estimation strategy of calculating growth rates, we need

for each firm at least two observations within one period to be able to calculate a growth

rate. This leads to an unbalanced panel data set of 9,142 observations, which stem from

4,567 different firms. Table 1 shows the descriptive statistics.

Page 13

12

Table 1: Descriptive Statistics

Employment growth (Δlog)

Innovation

Product innovation

Process innovation

Employment start level (log)

Sectoral GVA growth

Sectoral real wage growth

n=9142, N=4567, Avg.T=2.002

Mean

-0.016

0.497

0.406

0.317

4.682

0.005

0.018

Std. Dev.

0.261

0.412

0.410

0.365

1.506

0.046

0.026

Min

-2.708

0

0

0

0

-0.265

-0.231

Max

2.996

1

1

1

11.513

0.283

0.428

The mean of the dependent variable – the average yearly employment growth rate per

period – shows a negative sign. That means, on average, the employment level in the

firms of our sample is slightly declining within a period. This growth rate is measured

as the difference in log values divided by the respective length of the period

((log LT – log L1)/T).5 The innovation variable is the average of how often a firm

responded with “yes” to the question of whether an innovation was introduced during a

four or five year period. Thus, a firm that has innovated in all years has an innovation

value of one, a firm that has not introduced any innovations during a period has an

innovation value of zero and a firm that has reported an innovation in half of the years

has an innovation value of 0.5. The sample mean of this variable is 0.497. But it is also

important to know that in 2,964 cases (out of the 9,142 observations) firms have not

innovated at all during a period (i.e. their average for the period equals zero) and in

2,903 cases, the innovation value is one, i.e. the firm has innovated in all observations

during a period. This gives us 5,867 of 9,142 cases (equals 64%) where no change in

the innovation variable is observed within one period. With our dataset we are able to

split this variable into product and process innovations – which are not mutually

exclusive, i.e. a firm can either tick no innovation, one of the innovation types or both

types. The dataset shows that product innovations were implemented more often than

process innovations. The employment start level, which is the number of employees in

the first year of a period is, on average, about 108 employees (or 4.682 in log values).

The next two variables of table 1 are calculated as the average yearly growth rates

Page 14

13

within the corresponding period. The growth in the gross value added is added on the

industry sector level to accounts for economic development of the corresponding sector.

The mean value is slightly positive. Also as a control variable we include the sectoral

real wage rate growth, which is also positive in our sample.

5. Results

In this section we present the results of several specifications of estimating equation (4).

In section 5.1 we only distinguish between product and process innovations, in section

5.2 we present results for different firm sizes and different regional locations of the

firm. In section 5.3 we introduce different categories for both types of innovation.

5.1 Product and process innovations

Table 2 presents the specifications in which the innovation is split into product and

process innovations, which are not mutually exclusive (see section 4.2). The innovation

variables are, as described in section 3, the average per period of how many times the

firms responded with “yes” to the yearly questions of whether any product (or process)

innovations were introduced. So the regression coefficient has to be interpreted as the

difference between a firm that has innovated each year during the period and a firm that

had no innovation during the period.

5 Log LT denotes the log of the employment level in the last year observed during a period, log L1

denotes the log of the level of employment in the first year observed during a period and T denotes the

time between the first and the last year observed during a period.

Page 15

14

Table 2: Product and process innovations

Dependent variable: average yearly employment growth

(1) (2) (3)

Estimated

Coefficients

OLS standard

errors

Heteroskedastictiy

robust standard

errors

(0.003)***

(0.162)***

(0.102)**

(0.009)***

(0.009)***

(0.024)***

Covariance

robust standard

errors

(0.003)***

(0.161)***

(0.102)**

(0.009)***

(0.009)***

(0.026)***

Employment start level

Real wage growth

Real GVA growth

Product innovation

Process innovation

Year

Sector

States

Constant

Observations

-0.034

-0.437

0.257

0.033

0.057

incl.

incl.

incl.

0.112

9142

(0.002)***

(0.132)***

(0.081)***

(0.008)***

(0.009)***

(0.026)***

Adj. R-squared

Regression coefficients are * significant at 10%; ** significant at 5%; *** significant at 1%

0.039

Table 2 shows different specifications in terms of heteroskedasticity and of

correlation between error terms, but as can be easily seen the difference in the standard

errors is very small. Specification (1) shows standard OLS standard errors, specification

(2) corrects for possible heteroskedasticity and specification (3) additionally relaxes the

assumption of independency within the observations of the same firm in different time

periods. The very small change in the size of the standard errors can be taken as a sign

for a robust specification. In the following we will only present results which allow for

heteroskedasticity and dependence within firms, as in specification (3).

The control variables show the expected signs. The employment start level shows a

negative sign and is significantly different from zero at the 1% level. This gives strong

evidence for the hypothesis that large firms grow more slowly than smaller firms. The

sectoral gross value added growth rate shows a positive sign and is significant at the 5%

level. This is no surprise since it shows that a single firm benefits from the sectoral

development. The wage growth has a negative effect on the employment level. The

coefficient can be interpreted as the wage elasticity. A one percent higher real hourly

wage rate in the sector leads to a 0.4% smaller yearly employment growth rate in the

firm. This result is clearly in line with theory that high wages hinder employment