THE INDUSTRIAL REVOLUTION IN THE N ETHERLANDS :

THE INDUSTRIAL REVOLUTION IN THE N ^ETHERLANDS :

A CHANGE IN PERSISTENCE ?

Krista Hoekstra

THE INDUSTRIAL REVOLUTION IN THE N ^ETHERLANDS :

A CHANGE IN PERSISTENCE ?

Krista Hoekstra Turfsingel 52E

9711 VV Groningen

University of Groningen Faculty of Economics

Master thesis June 2007

Supervisor: dr. J.P.H. Smits

ABSTRACT

This thesis investigates the possibility of the occurrence of a shift in persistence in Dutch output during the nineteenth century. It has often been argued by both economists and historians that the industrial revolution constituted the beginning of modern economic growth.

The Netherlands clearly followed a different growth path than Great Britain – the first industrialised nation. Several historians have argued that a break occurred in Dutch economic development around 1870. This is the moment when the Dutch economy experienced a transition towards a modern economy.

Using advanced econometric techniques, I attempt to detect a change in persistence in Dutch output, labour productivity, and total factor productivity data of the nineteenth century. Two different types of test procedures are applied. First, the methodology of Banerjee et al. (1992) will be followed where rolling and recursive augmented Dickey-Fuller statistics show the evolution of the unit root test statistic throughout the time sample. Second, the approach of Harvey et al. (2006) is followed where Kim’s original ratio statistic and a modified version is used to find a change in persistence.

It turns out that a change in persistence from an I (0) to an I (1) process can be detected in the services sector for both value added and labour productivity. Furthermore, there is some evidence for a change in persistence in total output and total factor productivity. However, total factor productivity in the manufacturing sector turns out to have experienced a shift from an I (1) to an I (0) process.

It can be concluded that the services sector was of great importance in the development of the

Netherlands. The introduction of new technologies, most important steam engines and the

building of railways, caused a huge increase in productivity for the Dutch trade and transport

sector.

CONTENTS

Abstract 2

1 Introduction 5

2 Theoretical framework 8

2.1 Endogenous growth theory 8

2.2 Modern economic growth 11

2.3 Literature overview 12

2.3.1 Industrial revolution in Great Britain 12

2.3.2 Development of the Netherlands 15

2.3.3 Industrial revolution in the Netherlands 17

3 Methodology 20

3.1 Time series analysis 20

3.2 Unit root analysis 21

3.3 Structural breaks 22

3.4 Rolling and recursive unit root tests 23

3.5 Tests for a change in persistence 24

3.5.1 Dickey-Fuller test for a change in persistence 24 3.5.2 Ratio-based test for a change in persistence 26 3.5.3 Comparison of change in persistence test approaches 29

4 Data 30

5 Results 32

5.1 Augmented Dickey-Fuller test for a change in persistence 32 5.2 Ratio-based test for a change in persistence 36

6 Conclusions 38

References 41

Appendices 46

List of figures

5.1 Recursive ADF total value added 33

5.2 Recursive ADF services value added 33

5.3 Recursive ADF services value added / worker 34

5.4 Rolling ADF total factor productivity 34

5.5 Recursive ADF total factor productivity 34

List of tables

5.1 Rolling and recursive ADF analysis 35

5.2 Results test for a change in persistence 37

Appendices

1 Graphs of the data series

2 Rolling ADF statistic - value added 3 Recursive ADF statistic - value added 4 Rolling ADF statistic - labour productivity 5 Recursive ADF statistic - labour productivity 6 Rolling ADF statistic - total factor productivity 7 Recursive ADF statistic - total factor productivity 8 Test for a change in persistence

9 Critical value tables

1. INTRODUCTION

This thesis focuses on economic growth in the Netherlands in the nineteenth century, the era of the Dutch industrial revolution. Given earlier research about economic development in the Netherlands, I want to find out whether the transition from a pre-modern economy towards a modern economy, as is observed by historians, can be distinguished by evaluating the statistical properties of the historical time series using advanced econometric techniques.

Mokyr has suggested that the industrial revolution has caused a change in the nature of economic growth. According to Mokyr, the real breakthrough of the industrial revolution was not the invention of new technologies itself, but the observation that after the industrial revolution, growth did not slow down as has always been the case before. Mokyr states that the industrial revolution changed economic growth from a system dominated by negative feedback to one dominated by positive feedback, where growth initiated growth. After the industrial revolution, growth ‘fed on itself’ (Mokyr, 1999, pp. 2 – 3).

The industrial revolution started in Great Britain in the eighteenth century and can be separated in two phases: the first industrial revolution (1770-1870) and the second industrial revolution (1870-1913). Aided by revolutions in agriculture, transportation, communications and technology, the United Kingdom was able to become the ‘first industrial nation’. At the point where the first industrial revolution was completed for Great Britain, the process of industrialisation just started in the Netherlands. There are differences in timing and location of new technological inventions, resulting in differences in overall growth potential and structural change. The Netherlands industrialised almost a century later than Great Britain and followed a different growth path.

It is interesting to study the Netherlands, because of its special history of the flourishing

‘golden age’ in the seventeenth century, but its relatively late industrialisation in the late nineteenth and early twentieth century. Furthermore, the nineteenth century data estimates of the Netherlands are very reliable so that they can be used for a thorough econometric analysis.

It is useful to get insight in the causes of economic growth during the industrial revolution, in order to obtain new explanations about why growth accelerated in the Netherlands.

At the end of the sixteenth century, Dutch economic performance became very

successful. Continuous economic growth during the sixteenth and seventeenth century

featured the Netherlands, while other European economies were stagnating. There is a break in the second half of the seventeenth century; the Dutch economy stagnated and the international competitive power of the Netherlands decreased. This continued into the eighteenth and nineteenth century.

I will test whether the industrial revolution caused a shift from a deterministic, exogenous growth process to a stochastic, endogenous growth process. This shift would imply that the Dutch economy changed from a pre-modern towards a modern economy. In other words, if this shift indeed can be distinguished, the industrial revolution changed the nature of economic growth in a fundamental way.

Many historical time series exhibit a trend. It is important to know whether this trend is deterministic (stationary) or stochastic (non-stationary). According to Jacobs and Smits are there not many studies in which modern times series techniques are applied to historical time series (Jacobs and Smits, 2006, pp. 3). Jacobs and Smits postulate that one of the most important properties of the transition of a pre-modern, agricultural society to a modern society is that the time series evolve from a stationary into a non-stationary process (Jacobs and Smits, 2006, pp. 8). However, based on the work of Van Eijkel and Romp (2005), they must conclude that in contrast to the expectations of historians, there is no evidence of a structural break in the second half of the nineteenth century. However, they argue that “it may be that the focus at a macro level obscures important underlying causes in the economy. We therefore argue that HTSA methodologies should also be applied to series at a meso level, which can be found in many databases on historical national accounting.”

When analysing historical time series, it is necessary to make a clear link between the results of econometric analysis and the (historical) implications. It is very important to question what the econometric analysis tells you about the nature of economic growth. This is emphasised by both Jacobs and Smits who state that an econometrician should know about the historical context of the data and that the historian should be familiar with modern techniques (Jacobs and Smits, 2006, pp. 3 – 4) and Jacobs et al. who argue that data users and data constructors should cooperate. Feedback between data users (econometricians) and data constructors (historians) thereby leads to positive externalities (Jacobs, Sturm, and Groote, 1999, pp. 20).

1 Jacobs and Smits (2006), “Historical time series analysis”, pp. 12.

In this thesis I will take into account these two above mentioned points. First, the historical time series analysis will not only be applied to gross domestic product, but also to the broad sectors that comprise GDP, which are agriculture, industry and services. In addition, the analysis will also be applied to labour productivity and total factor productivity data. Second, the econometric analysis will be done with the objective of finding out about a change in the nature of economic growth. The results of the empirical analysis should therefore be used to further investigate the causes of the structural shift.

The structure of the thesis will be as follows. Section two provides a description of the concepts of ‘endogenous growth’ and ‘modern economic growth’, and gives an overview of existing literature about the industrial revolution in both Great Britain and the Netherlands.

Section three describes the methodology of time series analysis and unit root analysis, and

rolling and recursive unit root analysis in particular. Furthermore, two types of test procedures

are described that can test for the occurrence of a change in persistence. Section four gives

detailed information about the data that is used to investigate the persistence of output and

productivity. Section five gives the results of the empirical investigation. Section six

summarizes and discusses the conclusions of the analysis.

2. THEORETICAL FRAMEWORK

This section will discuss the theories on which the analysis of the thesis is based. First, the concepts of endogenous growth and modern economic growth will be discussed. Then an overview of existing literature will be given about the industrial revolution in Great Britain, the development of the Netherlands during the nineteenth century and the industrial revolution in the Netherlands.

2.1 E

According to traditional neoclassical growth theory, capital accumulation is subject to diminishing returns. Long-run growth thus requires exogenous improvements in productivity.

This means that long-run growth depends on factors which cannot be explained by neoclassical economic models. A weakness of the exogenous growth models is that they do not give a convincing explanation about the cause of economic growth. Some part of economic growth remains unexplained, which is called the residual. Historical experience shows that economic performance was continuingly growing. Endogenous growth models are able to explain continuing economic growth without relying on exogenous, unexplained, technological changes.

Endogenous growth models assume that the factors of production are not subject to diminishing returns to scale. Growth can be caused by the accumulation of human and physical capital. The endogenous approach to economic growth explains output growth from a historical perspective. These models predict continuing economic growth without exogenous technological progress. In the endogenous model it is assumed that individuals maximize utility and thereby choose their savings rate, which is no longer exogenously given.

Romer (1994) evaluates the origins of endogenous growth models. Before the 1980s, the Ramsey model was mostly used by economists for analysing economic growth. This model, however, assumes that capital and output are a decreasing function of the capital stock.

Countries are thus expected to converge. Empirical analysis, however, showed that

convergence is not taking place. Growth has been increasing over time, instead of decreasing

as was expected. This was one of the main reasons for Romer to develop an endogenous

growth model.

Romer (1986) developed a model in which endogenous technological change is driven by accumulation of knowledge by forward-looking individuals. “Diminishing returns in the production of knowledge are required to ensure that consumption and utility do not grow too fast. But the key feature in the reversal of the standard results about growth is the assumption of increasing rather than decreasing marginal productivity of the intangible capital good knowledge.”

Romer’s model is an alternative for earlier views on long-run growth. Per capita output can grow at an increasing rate, furthermore the rate of return on capital and investment may be increasing for higher levels of the capital stock. As a result, output in different countries does not have to converge.

Lucas (1988) further investigates the problem of accounting for the observed levels of income across countries and across time. He develops a model in which only the population growth rate is given, there are no other exogenous forces. The initial conditions play an important role in convergence as they determine the point where the country will converge to. Poor countries can have the same long-run growth rate as rich countries, and thus stay poor relatively. Lucas’ model explains both sustained growth and sustained differences in income levels between countries.

Rebelo (1991) developed an endogenous growth model which exhibits constant returns to scale (contrasting to the increasing returns to scale in Romer’s model). He tries to explain differences in growth performance from a government policy perspective. Rebelo makes a distinction between two types of production factors. The first type is reproducible and can be accumulated over time. This type includes both physical and human capital. The second type is non reproducible and is available in the same quantity every period. This type of production factor includes land. According to Rebelo, the assumptions about increasing returns to scale and externalities made by Romer are not necessary for endogenous growth to occur. He argues that as long as there are capital goods whose production involves the use of non- reproducible production factors, endogenous growth is compatible with constant returns to scale production technologies.

Greasley and Oxley (1997) provide a useful definition of exogenous and endogenous growth models. “The exogenous model implies industrial output follows a predetermined trend in

2 Rebelo, “Increasing returns”, pp. 1004

which the long-run growth path is predictable, and any deviations around the trend are transient.”

In contrast, the endogenous model is defined as follows. “The endogenous model implies output follows a stochastic process in which particular episodes, for example shifts in human capital formation, changes in economic policy, or wars, influence industrial growth.

Future industrial growth cannot be predicted if output follows a stochastic process since the effects of historical events accumulate in uncertain fashion to shape the trend.”

One way of analysing endogenous growth theory within the context of the industrial revolution, is investigating the statistical properties of industrial production time series. The statistical properties of a time series are important because different modelling strategies (exogenous or endogenous) have different implications for the persistence of output fluctuations. An endogenous growth model implies that fluctuations around trend growth can have a persistent effect on output. In an exogenous growth model, however, there are diminishing returns to scale which implies that shocks are not persistent.

If fluctuations are persistent, the time series are characterised by a stochastic process.

The behaviour of the time series is then also called non-deterministic. Contrasting, a deterministic process is a process without randomness in the development of future states.

Deterministic models thus produce the same output repeatedly for a given initial condition.

Concluding, the exogenous model implies that (industrial) output follows a predetermined trend with transient fluctuations and a predictable long-run growth path.

Contrasting, the endogenous model implies that (industrial) output follows a stochastic process in which future economic growth can not be predicted.

Thus, at first sight, endogenous growth models seem to be better able to explain historical experience. However, as Greasley and Oxley (1997) state, “historical experience does post important challenges for the new growth theorists, given their desire to construct models with empirical credibility. The question of whether the Industrial Revolution can be understood within an endogenous growth modeling strategy has relevance for theorists and historians. An affirmative answer would help to associate theory with the empirics of economic growth and offer valuable perspectives for historians of the Industrial Revolution.”

3 Greasley and Oxley, “Endogenous growth”, pp. 937

4 Greasley and Oxley, “Endogenous growth”, pp. 938

5 Greasley and Oxley, “Endogenous growth”, pp. 936

Several authors have investigated the industrial revolution in Great Britain. Crafts (1995) and Greasley and Oxley (1997), among others, have investigated whether the economic growth process in Great Britain during the industrial revolution was endogenous or exogenous.

2.2 M

The notion of modern economic growth was first developed by Kuznets (1966). Modern economic growth is defined as sustained growth of per capita national income. The industrial revolution is often seen as the point of departure for modern economic growth. “Economic history as a discipline has long been virtually unanimous in its answer: modern economic growth (that is, sustained, longterm growth of per capita output) is the achievement of the industrial revolution, an event that began in a definite geographical location (northern England) at a definite period of time (the reign of George III), and that spread gradually to embrace the societies that now can be called advanced industrial economies. This event stands as the gatekeeper between two types of economic life; only an economy that has been transformed by an industrial revolution can pass on to experience modern economic growth.”

Some authors have argued, however, that in the case of the Netherlands, the industrial revolution in the nineteenth century was not the beginning of modern economic growth. For example, De Vries states that “this country never passed through the eye of the needle that was the British Industrial Revolution, and which long has defined the nature of a modern economy for the entire world.”

Another often used classification for Dutch pre-industrial growth, advocated by Van Zanden (1991), is the concept of merchant capitalism which is based on the idea that international forces drive economic growth. De Vries and Van der Woude (1997) state that the invocation of the concept of merchant capitalism is “conceptually flawed in its overemphasis of the discontinuity in economic history represented by the British Industrial Revolution.”

The authors argue that it is misleading to compare pre-industrialised economies with the success of Great Britain as ‘first industrialised nation’.

6 De Vries (2000), “Dutch economic growth”, pp. 444.

7 De Vries (2000), “Dutch economic growth”, pp. 463.

8 De Vries and Van der Woude (1997), “The first modern economy”, pp. 693.

De Vries and Van der Woude (1997) give an alternative definition of a modern economy that is not related to the industrial revolution. “A ‘modern economy’ need not be one with the outward attributes of a twentieth-century industrial economy; rather it should incorporate the generic features that make those outward signs possible. Foremost among those features are:

- markets, for both commodities and the factors of production (land, labor, and capital), that are reasonably free and pervasive;

- agricultural productivity adequate to support a complex and social and occupational structure that makes possible a far-reaching division of labor;

- a state which in its policy making and enforcement is attentive to property rights, to freedom of movement and contract, and at the same time is not indifferent to the material conditions of life of most inhabitants; and

- a level of technology and organization capable of sustained development and of supporting a material culture of sufficient variety to sustain market-oriented consumer behaviour.”

2.3 L

In this section, different views on the industrial revolution in Great Britain will be discussed, and thereafter, the economic development and interpretations on the industrial revolution in the Netherlands will be evaluated.

It is important to discuss Great Britain for two reasons. First, it is useful to be able to compare the development of the Netherlands with Great Britain’s. The evolution of Great Britain often serves as a model for defining the industrial revolution. Second and most important, the type of analysis that will be performed in this thesis, i.e. to distinguish different growth regimes before and after the industrial revolution, has only been applied to Great Britain before. More precisely, there is a large volume of literature about determining whether growth during the industrial revolution in Great Britain was endogenous or exogenous.

2.3.1 I

G

B

The industrial revolution started in Great Britain in the second half of the eighteenth century.

Great Britain can be seen as the first industrialised country of the world. Different authors, among other Crafts and Harley (1992) and Greasley and Oxley (1994), agree that the growth

9 De Vries and Van der Woude (1997), “The first modern economy”, pp. 693.

experience of Great Britain in the second half of the eighteenth and the first half of the nineteenth century has been “historically unique and internationally remarkable”.

Several authors have investigated whether the British industrial revolution experienced exogenous or endogenous growth. The conclusion is ambiguous because different authors have derived at different results.

According to Hartwell was the industrial revolution one of the great discontinuities of history, the end of a world of slow economic growth and the beginning of must faster economic growth, with enormously fast increases in population size, sustained increases in output and in per capita real income (Hartwell, 1971, pp. 42).

Komlos (1989) suggests that during the eighteenth century industrial revolution there was no discontinuity with the past. It was an integral part of the epoch that preceded it. The industrial revolution can be seen as a break out of the Malthusian demographic regime. Slow but persistent capital accumulation was important to this process. Without the earlier food constraints, population could grow unconstrained and did not hinder economic growth anymore. Growth of population in the industrial revolution was the beginning of a subsistence crisis, but the accumulated stock of capital was large enough to break out of the Malthusian trap. Without any economic structural breaks, output could grow faster and longer than ever before.

Mokyr focuses on different kinds of (technological) inventions. He defines inventions as an increment in the set of the total technological knowledge of a given society. Mokyr distinguishes between minor inventions, whose cumulative impact is important for productivity growth, and major technological breakthroughs. Microinventions are defined as small, incremental steps that adapt and improve already existing techniques thereby reducing costs, increasing durability, or reducing input requirements. Macroinventions are radical new ideas with no precedent, that emerge ab nihilo. Mokyr argues that both concepts are important for technological development, although the microinventions occur more frequent and account for most of the productivity increases (Mokyr, 1990, pp. 13). The industrial revolution can be characterised by accelerating and unprecedented technological change. “To return to the terminology introduced earlier, a clustering of macroinventions occurred, leading

10 O’Brien (1991), “The industrial revolution”, pp. 28.

to intensified work in improvement and adjustment, and thus creating a complementary flow of microinventions. The result was a sharp increase in patenting activity.”

Crafts and Harley (1992) view British growth performance during the industrial revolution as a discontinuity. They estimated growth between 1750 and 1850 and argue that growth of the British economy was very unique.

Greasley and Oxley (1994) think that the conclusion of Crafts and Harley are not persuasive because there are no criteria against which a discontinuity can be tested. They argue that the growth estimates are no evidence of a historical discontinuity. Greasley and Oxley therefore perform a unit root test on the growth estimates of Crafts and Harley. The null hypothesis of difference stationarity is rejected for the time periods 1700 until 1780 and 1815 until 1913. The time period 1780 until 1851, however, can be seen as a fundamental historical discontinuity.

Crafts (1995) argues that factor accumulation can not explain the speeding up of economic growth in the late eighteenth and nineteenth century. He states that trying to explain British growth during the eighteenth and nineteenth century without explicit acknowledgement of technological change is not a very promising approach (Crafts, 1995, pp. 756). Crafts claims that endogenous growth models explain only part of the output growth. Crafts adopts the view of Mokyr on technological change by assuming that the industrial revolution can be regarded as a cluster of exogenous macroinventions which accelerated microinventions.

Greasley and Oxley (1997) assess the value of an endogenous approach to modelling growth in the British industrial revolution. They state that the appeal of the endogenous growth models resides in the fact that it replaces the unexplained residuals with historical explanations. The endogenous growth model is thus able to predict continuing economic growth without exogenous technological progress (Greasley and Oxley, 1997, pp. 935).

Concluding, Greasley and Oxley do not favour an exogenous model. They argue that the British industrial revolution is characterised by highly persistent output fluctuations. In their view, growth is endogenously determined, caused by an accumulation of historical events.

11 Mokyr (1990), “The lever”, pp. 82.

2.3.2 D

N

There are many different views on the development of the Netherlands. De Vries and Van der Woude (1997) have argued that the Netherlands experienced modern economic growth already in the seventeenth century, so before the industrial revolution took place. Mokyr states that the Netherlands was a latecomer in the industrial revolution, despite its commercialized, sophisticated, and urban economy (Mokyr, 1999, pp. 1). According to Griffiths (1980), the Dutch economy in the nineteenth century did not stagnate, but developed in another way compared to other European economies. In particular, the agricultural and services sector contributed to this different growth path.

This section discusses the socio-economic situation in the Netherlands before and during the nineteenth century.

During the seventeenth century, the Netherlands was one of the richest countries of the world.

This time period is known as the ‘golden age’. Continuous economic growth during the sixteenth and seventeenth century featured the Netherlands, while other European economies were stagnating. At that time, Amsterdam was one of the most important international markets. However, in the eighteenth century, Great Britain took over technology leadership.

In Britain there was a huge amount of structural change, while in the Netherlands radical structural change was absent. Population grew very slowly in the Netherlands as compared to Great Britain. Dutch growth in the nineteenth century was based mostly on agriculture and commercial services. In contrast to British entrepreneurs, Dutch entrepreneurs suffered from scale constraints during the first half of the nineteenth century because of less export possibilities due to protectionist policies in international trade. Furthermore, the Dutch domestic market was rather small and fragmented. The Netherlands was unable to imitate the economic and institutional structures that caused Britain to industrialise. These unfavourable institutional conditions caused that the Netherlands did not industrialise and remained highly specialized in agriculture.

In the second half of the nineteenth century some important changes occurred. In 1865 international trade barriers disappeared. Between 1850 and 1890 the share of steam engines rose from 5% to 60% of total machinery. However, the use of steam engines was unevenly distributed among several industries, resulting in low overall growth rates compared to other countries (Smits, De Jong, and Van Ark, 1999). After the 1860s, railways were built.

Together with other developments from the 1850s onwards, for instance tax reforms, the

introduction of the steam engine in agriculture and later the introduction of ‘modern banking’

in the 1890s, the Dutch economic conditions improved, resulting in a gradual increase in economic growth. The Netherlands remained highly specialized in agriculture until the end of the nineteenth century and in the beginning of the twentieth century. Van Zanden (2000) argues that the process of industrialisation started only after the first world war. According to Smits et al. (1999), the period 1865-1895 can be regarded as the first phase of ‘modern economic growth’ in the Netherlands.

The second Industrial Revolution started after 1870. There is no clear break between the first and second Industrial Revolution, resulting in disagreement among historians about the exact timing. Electricity contributed to the invention of powered machines, which improved productivity. Furthermore, communication devices were discovered and improved as a result of electricity. This increased the spread of knowledge, which enhanced research and development. Technology became more complex and diverse. After the second world war, investments in research and development were enlarged, resulting in an increase in innovations which stimulated economic growth.

During the second industrial revolution, the conditions in Great Britain and the Netherlands are more similar than in the previous centuries. Large differences in economic growth because of differences in the degree of industrialisation are less pronounced. The end of the second industrial revolution is not properly defined.

A useful classification of the industrial revolution in the Netherlands is provided by Smits et al. (1999). Smits et al. have analysed breaks in Dutch growth data from 1816 to 1997. They focus on four variables (output, labour productivity, total factor productivity, and capital intensity) and test for structural breaks using a Chow test. Combining the results of the econometric tests and earlier literature, Smits et al. distinguish three growth phases during the nineteenth and twentieth century. These are the first industrial revolution based on steam technology (1800 – 1913), the second industrial revolution based on electricity (1890 – 1990) and the IT revolution (1973 – present). The diffusion of technology during the first industrial revolution was rather slow, only after 1850 the share of steam engines increased strongly.

During the twentieth century, economic performance of the Netherlands was rather

successful.

Van Eijkel and Romp (2005) identify two weaknesses in the investigation of Smits et al.

“First, they use the Chow test to find significant structural breaks. The Chow test is a standard F-test, while the asymptotic distributions of the test statistics for structural change models are nonstandard. This is because the structural change parameter is not present when the null hypothesis is true, but only appears under the alternative hypothesis (see also Andrews, 1993).

Using the Chow test then results in finding too many significant breaks. In their analysis, Smits et al. found eleven significant breaks at a significance level of 10 per cent. Second, they use a sequential procedure to find the breaks, which does not yield a global minimum of the total sum of squared residuals (SSR).”

Bai and Perron (1998) propose a method which can be used to find multiple structural breaks by minimizing the sum of squared residuals. In a later paper (Bai and Perron, 2003) they further investigate the practical issues for empirical application of these procedures. Van Eijkel and Romp (2005) have used this method to investigate GDP per capita growth data of the Netherlands between 1816 and 2003. They find, after controlling for the war years, only one significant break which is in 1974. During the nineteenth century, there were no significant breaks.

2.3.3 I

N

It must be stressed that the Netherlands developed different then Great Britain. Griffiths (1980) argues that the industrialisation pattern that Great Britain followed, does not apply to every country. He argues that it is not the Netherlands that developed different from the rest of the world, but perhaps Great Britain should be viewed as having developed differently.

Also Cameron states that the term ‘industrial revolution’ is misleading, and that the notion that Great Britain served as a model that other countries followed is too simple and therefore also misleading (Cameron, 1985, pp. 2).

Several researchers have investigated when the Dutch economy changed to a modern economy. One way to analyse this is to compare international business cycles. Because the degree of regional integration is rather weak in the pre-modern era, different countries do not follow the same business cycle. The existence of business cycles at all can be doubted, because of the lack of sector integration, the absence of large-scale production, a lack of well- developed capital goods and a modern banking system (Jacobs and Smits, 2006, pp. 15).

12 Van Eijkel and Romp (2005), “Multiple structural breaks”, pp. 3.

Brugmans (1969) assumed that when there is a lack of cyclical behaviour, there is no modern economic growth. From his analysis, it appeared that 1870 is a breakpoint for the Netherlands, because from the 1870s onwards the Dutch economy started to follow international business cycles.

Jacobs and Smits (2004) compared Dutch business cycles with international business cycles. They observed that the Dutch economic growth path is influenced by international business cycles already in the 1850s and 1860s. In contrast, between 1870 and 1890 the Dutch economy developed independently of the international business cycle. This development can be explained by high levels of domestic demand and decreasing scale constraints to industrial production. From the 1890s onwards, Dutch business cycles moved rather similar with foreign business cycles.

According to Smits (2000), Dutch entrepreneurs suffered from scale constraints in the first half of the nineteenth century. Because of protectionist policies with regard to international trade, export possibilities were restricted. The domestic market was rather small. Because of these diseconomies of scale, there were few possibilities for technological improvements in industry. After 1850 taxes on primary foodstuffs were abolished and nominal wages declined.

Both events resulted in increased consumer demand. Smits shows that the Dutch industrial sector started to catch up in terms of labour productivity as compared to Great Britain after the 1870s. Until 1865 the levels of TFP were declining, but after 1865 TFP makes a positive contribution to industrial growth. Capital intensity was rather low in the first half of the nineteenth century, but started to increase after the 1850s. After 1870, growth rates of capital intensity accelerated. Smits argues that the period 1865 – 1895 can be regarded as the first phase of modern economic growth. Important technological changes such as the diffusion of steam technology were realised in this time period. During the 1870s and 1880s the Dutch economy developed very fast, or as Jacobs and Smits (2006) say, at light speed. Steam technology was diffused rapidly and labour productivity in the industrial sector was rising. In the beginning of the 1890s, the diffusion of steam technology reached its limit. After 1900 the share of electric power technology increased.

Bonenkamp et al. (2005) investigated to what extent changes in consumer demand may have

affected patterns of industrial development. By applying a unit root test and a VAR analysis,

it is tested whether demand shocks play a role in the explanation of the industrial revolution in

the Netherlands around the 1860s. On the basis of the results of the analysis, the authors

conclude that until 1865 growth is exogenous, whereas for the period 1865-1913 growth is endogenous.

Combining these results, economic historians see a breakpoint in Dutch economic

development around 1870. At this moment, the Dutch economy experienced a transition

towards a modern economy. In the light of the distinction between exogenous and

endogenous growth patterns, it can be expected to see a shift from an exogenous growth

pattern in the pre-1870 period to an endogenous growth pattern in the post-1870 period.

3. METHODOLOGY

This section will discuss (historical) time series analysis, tests for persistence of data series such as the unit root test, and tests for a change in persistence.

3.1 T

A time series is a sequence of data points usually with constant intervals. Analysis of time series is necessary to understand the process underlying the data observations or to be able to make predictions about future values. Historical time series analysis is the analysis of historical time series by means of sophisticated statistical and econometric techniques.

In the analysis of time series, one must separate the cyclical movements of the data series, output fluctuations, from the long-term trend. A time series y

, most often the logarithm of the series under consideration, can be decomposed into a trend component μ

and a cyclical component ε

, which are assumed to be statistically independent.

(1) y

= μ

+ ε

with t = 1, 2, …, T.

In the simple linear trend model, the following regression equation is estimated.

(2) y

= α + β t + ε

Now assume that the residuals ε

are stationary, but not serially uncorrelated. Alternatively, y

can be modelled in such a way that it is an accumulation of historical events which are stationary. If changes are accumulated from an initial value, this gives

(3) ∑

=

+ +

=

i i

t

y t e

y

1

0

β

There are two important differences between equation (2) and (3). First, the intercept is no longer a fixed parameter, but is now given by the initial value. Second, the error term is no longer stationary because its variance and covariance depend on time. Equation (2) represents a trend stationary model and equation (3) represents a difference stationary series.

Making a distinction between trend stationary series and difference stationary series is very

important. For example, macroeconomists are very interested in knowing whether an

economic recession has a permanent or temporary effect on the future level of GNP

(Hamilton, 1994, pp. 444). When time series are trend stationary, all variations in output are

cyclical. If time series are characterised by a difference stationary process, however, the trend

component is a non-stationary stochastic process instead of a function of time. A shock then has a permanent effect on the future.

Nelson and Plosser (1982) argue that many economic time series, like for example GNP, are random walks instead of a stationary process with a trend, as is often assumed in traditional time series analysis. Durlauf and Phillips analyse the effects of treating a non-stationary time series as being deterministic (stationary). When the time series is a random walk, regressions generate non-normal coefficient estimates. The F-statistic examining the significance of the trend coefficient in a regression will diverge. Furthermore, conventional strong laws and central limit theory do not apply to standardized sums of the realizations of an integrated process (Durlauf and Phillips, 1988, pp. 1334). The Durbin-Watson statistic can be a measure of stationarity. Granger and Newbold (1974) analyse the consequences of auto-correlated errors in regression analysis. They show evidence that a regression based on a non-stationary time series on another integrated time series leads to coefficient tests that are biased towards rejecting the null hypothesis. Furthermore, they develop a rule of thumb for detecting a spurious regression. They argue that when the R-squared value is greater than the Durbin- Watson statistic, one should suspect a spurious regression.

3.2 U

To distinguish whether observed time series has a trend or difference stationary process, one must test for the presence of a unit root. This has been modelled by Dickey and Fuller (1979).

They consider an autoregressive model which takes the following form (4) y

= ρy

+ e

with t = 1, 2, …, T.

The error term is a sequence of independent normal random variables with zero mean and variance σ

. The time series converges (as t → ∞) to a stationary time series if | ρ | < 1. If ρ equals one, the series has a unit root and is non-stationary. The first differences of the time series then are stationary. Non-stationary series can also be called a random walk. If | ρ | > 1, the time series is not stationary and the variance of the time series grows exponentially with time.

Stationarity of time series can be tested directly with a unit root test, such as the augmented

Dickey-Fuller (ADF) test. The ADF test investigates the null hypothesis of a unit root (ρ = 1)

against a trend-stationary alternative (ρ < 1). The ADF test regresses the first difference of a

series on its lagged level, supplemented with a trend and/or intercept and lagged differences (Jacobs and Smits, 2006, pp. 7).

The calculated ADF τ statistic is compared with specially constructed critical values.

The τ-statistic must take larger, negative values than the critical value in order to reject the null hypothesis. If the null hypothesis can not be rejected, the process has a unit root. A stationary (deterministic) process is also called integrated of order zero, I (0). If the first differences of a time series are stationary, so that the time series can be made stationary, the series are called integrated of order one, I (1).

3.3 S

A problem with the regular augmented Dickey-Fuller test is that it does not identify breaks.

The existence of breaks has an influence on the outcome of an ADF test. With breaks the outcomes are biased towards rejected the null hypothesis. When the examined time series include one or multiple structural breaks, the unit root test can not distinguish stationary from non-stationary series. Perron (1989) demonstrates that the Dickey-Fuller test fails to reject the null hypothesis if a break occurred under the trend-stationary alternative. He suggests an alternative model in which log output is stationary around a deterministic trend that changes slope at a certain moment in time. Perron states that most macroeconomic time series are stationary with transient fluctuations. Only two shocks in the twentieth century have had a permanent effect on output. These two events are the great crash in 1929 and the oil price shock in 1973. The shocks itself are assumed to have an exogenous cause. The null hypothesis cannot be rejected if the data series have a deterministic nature but include a break either in the intercept or the slope. As long as the magnitude of a shock or shift is significant, the null hypothesis will generally not be rejected, although the data series include a trend. The unit root test will therefore indicate that the shocks have a permanent effect, but in reality only the one-time shift is permanent. Furthermore, Perron develops a number of extensions to the Dickey-Fuller test which ensure that the testing procedure is consistent.

Banerjee et al. give four reasons why it is important to investigate stationary/trend-shift model

as suggested by Perron. First, if a break occurred, traditional unit root tests will incorrectly

fail to reject the null hypothesis. Second, incorrectly classifying a time series has important

complications for empirical research, because a lot of econometric techniques are based upon

the assumption of stationary time series. Third, the parsimonious model can be interpreted as

there being a few large events that determine the growth path of output for some decades.

Fourth, if this stationary model with breaking trends turns out to fit real world data series better, the empirical relevance of a lot of research in theoretical econometric decreases (Banerjee, Lumsdaine, Stock, 1992, pp. 271).

Following Perron, Leybourne and Newbold (2000) further analyse the behaviour of Dickey- Fuller tests in a model that is stationary around a broken trend. Leybourne and Newbold analyse a break in level and a break in the slope separately and show that the outcomes are very different. “Overall, the conclusion is that when a break occurs no less than halfway through a series the Perron phenomenon is apparent on the obvious basis – that is, relatively few rejections of the null and, the larger the break, all else equal, the fewer the rejections. This picture is rather different for breaks in the first half of the series, where it is entirely possible that the Perron phenomenon will not be observed. The precise picture here depends on the nature and size of the break.”

Leybourne et al. furthermore argue that the augmented Dickey-Fuller test will not diverge to minus infinity when there is a change in persistence in the series, because the I (1) part of the series will dominate the outcome. Therefore, the augmented Dickey-Fuller test does not provide a consistent testing procedure for discriminating between a constant I (1) process and a change in persistence (Leybourne, Kim, Taylor, 2006, pp. 596).

3.4 R

Banerjee et al. (1992) have developed measures to investigate the persistence in output in the presence of breaks in the data series. Three classes of statistics, and their related distributions, are considered, which are recursive, rolling, and sequential statistics. Both the rolling and recursive unit root test use a changing sub-sample of the data series.

Rolling statistics are computed by using a sub-sample of fixed size, rolling through the sample. For every interval, the ADF statistic is calculated, which is presented in a graph.

Because each sub-sample is a constant fraction of the full sample, the marginal weight of each observation can be kept constant.

In a recursive analysis, the ADF statistic is computed using sub-samples that expand, where the first interval consists of a specified number of years. Every next interval is one year

13 Leybourne and Newbold, “Behavior of Dickey-Fuller t-tests”, pp. 786

longer. Recursive statistics are computed for sub samples t = 1, …, k, for k = k0, …, T, where k0 is a start-up value and T is the total sample. Another possibility is to compute these statistics backwards – that is, using data over t = k + 1, … , T for k = 0, …, T – k

.

The sequential statistics are computed by sequentially increasing the date of the possible break, using all observations in the sample.

Banerjee et al. calculate critical values using a Monte Carlo simulation. Critical values are given for the total sample Dickey-Fuller statistic, the maximal, and minimal Dickey-Fuller statistic. For the recursive statistics, the start-up value of the sample size is taken to be one fourth of the total number of observations and for the rolling statistics the sample size is one third of the total number of observations.

3.5 T

When analysing long-term data, there is a possibility that the data series under investigation change from a difference-stationary process to a trend-stationary process, or vice versa. It is necessary to detect such changes and decompose data series into their stationary and non- stationary components, because this has implications for model building, forecasting, and policy implementation. Several authors have developed test statistics which can test for a change in persistence. Two general types exist. The first type is based on the augmented Dickey-Fuller test as suggested by Banerjee et al. (1992) and further developed by Leybourne et al. (2003). The second type is the ratio-based test suggested by Kim (2000). Both types of tests will be discussed below.

3.5.1 D

-F

Leybourne et al. (2003), henceforth LKSN, have developed a method to test for a change in

persistence based on the Dickey-Fuller unit root statistic. The null hypothesis is that the data

series are difference-stationary. The alternative hypothesis considers a possible change in

structure from I (0) to I (1) or from I (1) to I (0). The timing and direction of this possible

change are not specified. An advantage of this test procedure, is that the test statistic itself

provides an estimate of the break fraction. LKSN use the fact that a change from I (0) to I (1)

at time fraction τ is the same as a change from I (1) to I (0) at time fraction 1 – τ. Therefore,

the unit root hypothesis is tested using the original series and its reversed realisation. The null

hypothesis is denoted by H

and the respective alternatives by H

and H

. The method

developed by LKSN can only be applied when there is only one change in persistence. It can not be applied when the investigated data series has multiple breaks.

Suppose that the direction of the change is specified and is a switch from I (0) to I (1).

Because the timing of the change is not known, Dickey-Fuller τ-ratios are computed for all possible values of τ in the interval Λ = [0.2, 0.8]. The authors give an explanation for the use of an upper and lower bound for the break fraction τ*. “Since the last (1 − τ)T observations are taken under both the null and alternative hypotheses to be generated by an I (1) process, one possibility is to apply DF-type tests to the first τT observations, allowing τ to vary. (Tests of this sort are termed ‘recursive’ by BLS.) However, the asymptotic analysis requires that the number of observations on which such tests are based increases with T . Accordingly, as is sensible as a practical matter, a lower bound greater than 0 must be imposed on τ.

Correspondingly, when (…) we allow a switch from I (1) to I (0), an upper bound for the switch fraction is required.”

LKSN argue that the test statistic that is least favourable to the null hypothesis should be taken, that is, the infimum over τ of the Dickey-Fuller τ-ratios.

Which, according to LKSN is a natural choice, because under the alternative H

the value of τ at the infimum provides a consistent estimator of the true break fraction.

Smith and Tambakis (2003) have applied the test procedure of LKSN to treasury bond on/off spreads in the United States. They found a change in persistence from I (0) to I (1) that occurred in the late 1990s.

Cook further examines the test for a change in persistence developed by LKSN. He evaluates the test under the presence of a structural break under the null hypothesis. Cook concludes that the size properties of the tests differ depending on whether the break occurs in the level of the unit root process or in its drift parameter. Undersizing can occur in the presence of breaks in drift, while oversizing can occur with breaks in the level of the series. These effects are most likely to occur when the sequential LKSN test statistics are considered (Cook, 2004, pp. 6). He therefore argues that the results of the test should be interpreted with caution because the change in persistence can be spurious. Cook also suggests that an alternative approach should be considered when constructing a test for a change in persistence given the sensitivity of the Dickey-Fuller test to structural change.

14 Leybourne et al. (2003), “Tests for a change in persistence”, pp. 293.

Leybourne et al. (2006) modify the statistics of the recursive unit root test developed by BLS (1992). They argue that the tests proposed by BLS diverge when they are applied to data series which are constant I (0). Therefore, these tests can not discern between a I (0) process throughout or a change in persistence. Because both the infimum of the forward series as the infimum of the backward series diverge at the same rate, the ratio of these two statistics is constant. Furthermore, they give the lower and upper tail critical values for their ratio statistic.

3.5.2 R

-

Kim (2000), see also Kim et al. (2002), has derived another statistic that detects a shift in persistence of a time series. The null hypothesis is that the time series is stationary. The alternative hypothesis is that the time series is stationary until some period, after which it becomes a process of higher persistence (unit root). Or the alternative hypothesis is that the time series has high persistence until some time after which it becomes a process of lower persistence. Kim describes the null hypothesis as follows.

(5) H

: y

= r

+ z

, for t = 1,…, T

where r

is a constant, and z

is a stationary variable satisfying some regularity conditions.

There are two possible alternatives. One alternative hypothesis is that y

is stationary until t = [τT] for τ∈(0, 1) where [τT] is the integer part of τT, but after t = [τT] it becomes a process of higher persistence such as a unit root.

(6) H

: y

= r

+ z

for t = 1,…,[τT]

y

= r

+ z

for t = [τT] + 1,…,T

where r

and r

are constants and z

is a stationary process and z

is a non-stationary process.

On the other hand, y

can be process of relatively high persistence until t = [τT], but after t = [τT] it becomes a process of lower persistence.

(7) H’

: y

= r

+ z

for t = 1,…,[τT]

y

= r

+ z

for t = [τT] + 1,…,T

Likewise, the null hypothesis can characterize a trend stationary process. The hypotheses can then be described as follows.

(8) Hγ

: y

= γ

+ r

+ z

,

for t = 1,…, T (9) Hγ

: y

= γt + r

+ z

for t = 1,…,[τT]

y

= γ

+ r

+ z

for t = [τT] + 1,…,T

(10) Hγ’

: y

= γ

+ r

+ z

for t = 1,…,[τT]

y

= γ

+ r

+ z

for t = [τT] + 1,…,T

Following Harvey et al. (2006), Kim’s ratio statistic can be written as follows.

(11) K[τT] =

∑ ∑

∑ ∑

= =

−

+

= = +

−

] [

1

2 1 , 2

1 ]

[ [ ]1

2 , 2

) (

] [

) (

]) [ (

T t

t i it T

T t

t T

i it

v T

v T

T

τ

τ τ

τ τ

) (

where v) in the denominator is the residual from OLS regression of y

on x

for observations up to [τT]. Similarly, v( in the numerator is the OLS residual from the regression of y

on x

for t = [τT] + 1, …, T. The date of the change in persistence is unknown. The test statistic is applied in three different ways, which are Hansen’s (1991) mean score statistic (MS), Andrews and Ploberger’s (1994) mean-exponential statistic (ME), and the maximum Chow- type test (MX) considered by Andrews (1993).

(12) ∑

=

=

* T

T

t t

u

l

K T

MS

τ τ

(13) ln{

exp( 0 . 5 ) }

] [ 1

*

∑

=

=

T

t t

u

l

K T

ME

τ

(14)

T T

t K

MX

u l ],...,[ ]}

{[ max

τ τ

= ∈

where

≡ [ τ

T ] − [ τ

T ] + 1 .

Busetti and Taylor (2004) investigate ratio-based tests against changes from I (0) to I (1).

They show that these are inconsistent with constant I (1) processes and changes from I (1) to I (0). Therefore, they propose new ratio-based tests and breakpoint estimators which are consistent under I (1) to I (0) changes.

Leybourne and Taylor (2004) develop new persistence change tests, based on among others Kim (2000) and Busetti and Taylor (2004). These earlier tests were based on the maximum of a sequence of ratios, while the new test proposed by Leybourne and Taylor are based on the maximum of the sequence of the numerators of these ratios divided by the minimum of the sequence of denominators of the ratios. This new test statistic is somewhat more powerful than the maximal ratio test statistic and thus is a very useful complement to the maximal ratio statistic.

Harvey et al., henceforth HLT, investigate the behaviour of Kim’s ratio statistic in the

absence of a change in persistence. They wonder what happens if the null hypothesis assumes

that the data series are I (0) throughout, when the series are actually integrated of order one.

The authors demonstrate that in this case, the null hypothesis will be spuriously rejected in favour of the alternative hypothesis that assumes a change in persistence. Therefore, they conclude that existing tests are not able to adequately discern between a change in persistence and constant persistence of the form not covered by the null hypothesis (Harvey, Leybourne, and Taylor, 2006, pp. 443). In order to overcome this spurious rejection problem, the null hypothesis should assume constant persistence, which can be either I (1) or I (0). A rejection of the null hypothesis indicates unambiguously that a change in persistence has occurred.

HLT adopt the following data generating process.

(15) y

= x’

β + v

(16) v

= ρ

v

+ ε

t = 1, …, T Four hypotheses are defined.

(17) H

: y

is I (0) throughout (18) H

: y

is I (1) throughout

(19) H

: y

is I (0) changing to I (1) at time [τ*T]

ρ

= ρ, | ρ | < 1 for t ≤ [τ*T] and ρ

= 1− α /T, for t > [τ*T]

(20) H

: y

is I (1) changing to I (0) at time [τ*T]

ρ

= 1− α /T for t ≤ [τ*T ] and ρ

= ρ, | ρ |< 1 for t > [τ*T]

where τ* is the (unknown) proportion of the sample size where the change in persistence occurs, and α

allows for a local unit root. The vector x’

contains either a constant (x

= 1) or a constant and a trend (x

= (1,t)’). It is assumed that there are lower and upper values for τ*, i.e. τ

≤ * τ ≤ τ

. It is often assumed that τ lies in the interval Λ = [0.2, 0.8]. To test H

against H

, HLT use the statistics MS, ME and MX developed by Kim (2000). In order to test H

against H

, the tests proposed by Busetti and Taylor (2004) are used which are based on the sequence of reciprocals of K[τT] and are referred to as MS

, ME

and MX

as the respective analogues of MS, ME and MX, with K

replaced by K

throughout.

In order to test against an unknown direction of change, Busetti and Taylor (2004) propose MS

= max[MS, MS

], ME

= max[ME, ME

] and MS

= max[MX, MX

].

Another modification is made because “the limiting distributions of the test statistics are

pivotal under both H

and H

(…). Consequently, we can employ the approach of Vogelsang

(1998) to produce tests based on modified versions of these statistics which, for a given test

and significance level, have the same critical value in the limit as the corresponding

unmodified test under H

, but where the same limiting critical value is also appropriate under H

.”

The modified mean score statistics can be written as follows. The approach for the other statistics, mean exponential and maximum, is the same.

(21) MS

= exp( − bJ

) MS (22) MS

= exp( − bJ

) MS

HLT also report critical values for the developed tests for a change in persistence, i.e. the variables MS, ME, and MX and their reciprocal and pairwise maximum counterparts. The modified test statistics are constructed in such a way that the asymptotic critical values for a given significance level under either H

or H

are the same as the critical values of Kim’s ratio-based test. The critical values are given for both the de-meaned (x

= 1) data and the de- meaned and de-trended (x

= 1, t)’ data.

3.5.3 C

An important difference between the ratio-based test and the augmented Dickey-Fuller test, is the null hypothesis that is assumed. The ratio-based test for a change in persistence assumes a null hypothesis of I (0). The augmented Dickey-Fuller type of test, in contrast, assumes a null hypothesis of I (1). The modified versions of the ratio-based statistic (Harvey et al., 2006) assume a null hypothesis of constant persistence which can either be I (0) or I (1).

LKSN (2003) note that both test procedures can be regarded as complementary to each other. They refer to the work of Kim (2000) and Busetti and Taylor (2004) and state that these authors have investigated similar testing problems, although with a different null hypothesis.

These procedures, however, are based on LBI-type (locally best invariant) stationarity tests instead of the unit root test. A similarity of both test procedures is that none of them assumes that the timing of the change in persistence is known. According to LKSN, a disadvantage of the ratio-based test procedures is that they do not provide an estimate of the break fraction.

Instead, those test procedures require that auxiliary estimators are constructed that involve either maximizing or minimizing various ratio statistics that depend on the direction of the change. LKSN state that this is an unavoidable consequence of using stationarity tests, rather than unit root tests (Leybourne, Kim, Smith, Newbold, 2003, pp. 292).

15 Harvey et al. (2006), “Modified tests”, pp. 447.