Why is Dutch labour productivity higher in comparison to that of other countries?

Why is Dutch labour productivity higher in

comparison to that of other countries?

University of Groningen

Faculty of Economics and Business

Master Thesis International Economics and Business

Student Name: Charlotte de Wilt Student ID: 2405075

Email: c.j.de.wilt@student.rug.nl

Date: August 24th, 2015 (Revised version) Supervisor: Dr. Bart Los

Abstract

The objective of this paper is to analyse which determinants seem most important in explaining the high labour productivity of the Netherlands and why other high-income countries, on average, are not able to catch up with the Netherlands. The research covers the period of 1971 to 2005 for a sample of 16 high-income countries. In order to capture the most relevant determinants that can explain labour productivity, the high-income countries will be used as a benchmark in the analyses. Following comparison with the Netherlands, the results indicate that the labour productivity of the Netherlands is mainly driven by capital intensity and the level of openness. In addition, the results suggest several explanations for a lack of convergence. Firstly, the factors that explain labour productivity can change over time; this makes it harder for countries to determine which factors should be stimulated. Secondly, evidence shows that unconditional convergence does not hold that all countries move towards the same steady state equilibrium. The results suggest either divergence or conditional convergence because there is no clear pattern to be seen.

3

1. Introduction

In the last decades labour productivity has proved a very popular research topic. The most common relation is that a higher labour productivity will mean, in general, a higher total output in a country; this is why it is very attractive for nations to focus on policies that increase labour productivity. In this paper labour productivity will be investigated and 16 high-income countries (HICs) have been chosen to provide statistics for comparison. These countries have been selected on the basis that they are more or less homogeneous. On average these countries have similar wealth, living standards, and production characteristics. In Figure 1 the labour productivity, the real gross domestic product (GDP) per hours worked in the Netherlands alone, is compared to the annual average for the 16 HICs. It shows that in the period between 1971 and 2005 labour productivity continued to be higher in the Netherlands in comparison to the average of the HICs. In addition, the growth rates of labour productivity are similar because the average growth rate for the Netherlands is 2.12% and for the HICs it is 2.17%.

Figure 1: real GDP per hour worked, US dollar, 2005 constant prices, PPPs.

Note: HICs average includes: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, New Zealand, Sweden, Switzerland, the United Kingdom, and the United States.

Source: Feenstra et al. (2015) and the Conference Board (2014).

4

more effective, in turn increasing labour productivity. Basically, the neoclassical scholars argue that the diminishing returns to capital will eventually lead to steady state equilibrium. When an economy has a capital stock that is equal to this equilibrium, the amount of capital per worker will no longer change over time. It will remain constant. So, the closer the country is to steady state equilibrium, the slower the growth. This theory suggests that, in the long run, when the capital– labour ratio or labour productivity averages for the 16 HICs are lower, the marginal product of capital will be higher in comparison to the Netherlands. That will cause a faster growth for the average of the HICs because the further from steady state equilibrium, the higher the growth (Solow, 1956). Therefore, the averages of the HICs should be catching up with the Netherlands. However, figure 1 contradicts this expectation because it appears there is constant returns to capital instead of diminishing returns.

A possible explanation is posited by Guterriez (2000). He states that there is a possibility that the labour productivity of the different countries does not move towards one point in time, steady state equilibrium, but to different points. This can also referred to as ‘conditional convergence’. This means that there is equilibrium, but each country arrives to its own equilibrium. It should be noted that the statistics for the Netherlands are here compared with the average for the HICs; this means that, in addition to the Netherlands, other benchmark countries included in the HICs sample could also have a high labour productivity. However, in this dataset only the United States seems to have a similar labour productivity level to the Netherlands, although on average the labour productivity of the Netherlands is the one that excels.

This paper will examine the Dutch labour productivity in comparison to the average of the HICs in the period 1971 to 2005. The concrete research questions will be: Which

determinants seem to have had the most impact on the remarkable labour productivity of the Netherlands? And why is there a seeming lack of convergence when comparing the Netherlands with the average of the HICs in the period of 1971 to 2005?

5

intensity were the main contributors for explaining labour productivity in the HICs. Lastly, there is evidence that unconditional convergence does not holds that all countries move towards the same steady state equilibrium. The results suggest either divergence or conditional convergence because there is no clear pattern to be seen. These findings suggest that it is therefore more difficult to converge towards the level of the labour productivity of the Netherlands.

This study is subdivided into the following sections. Section 2 discusses several studies that are relevant to the research. Section 3 explains the methodology and data sources that will be used to respond to the research question. The results of the analysis will be presented and discussed in Section 4. Lastly, in Section 5, the conclusion and prospects for future research will be put forward.

2. Literature

The literature section of this paper will be subdivided into two subsections. (1) What are the different concepts of productivity? (2) What are the important determinants of productivity in the HICs? The last subsection will be further subdivided into paragraphs. Each paragraph will address an indicator that might contribute to the labour productivity level.

2.1 What are the different concepts of productivity?

We should begin with a deeper understanding of what productivity actually is. It has been described as the efficiency with which ‘economic resources can be translated into the production of goods and services’ (Harris, 1999, p.2). In practice there are two approaches commonly adopted in order to assess levels of productivity: the measurement of labour productivity and the measurement of total factor productivity (TFP). The level of labour productivity can be measured as output per hour or per worker. On the other hand, the level of TFP is a broader measurement: it is the fraction of output that cannot be explained by inputs, such as energy, labour, capital and intermediate inputs. It includes the level of technology and the level of efficiency (Sargent and Rodriguez, 2000).

6

more capital per worker is available, the higher labour productivity level will be. It should be emphasised, however (as mentioned earlier), that there are diminishing returns: every injection of capital adds less and less to the total output, and at some point injecting further capital will not lead to greater output level at all. This means that the level of labour productivity will decrease over time. Therefore, level of TFP plays an important role to make sure that labour productivity grows in the long run. In other words, the level of TFP (the level of technology) must offset the effect of diminishing returns to capital. It will make the given capital stock more productive over time, and that will enhance the level of labour productivity once more. This implies that, in the long run, all increases in the level of labour productivity are ultimately determined by a certain level of technology, TFP level.

That is why, for a longer period (such as the 35 years covered by this paper) TFP level is considered a better measure for productivity and for shorter periods it is more appropriate to use labour productivity. However, when using the level of TFP as a measure, bias may creep in through factors such as imperfect estimates of capital stock. When carrying out cross-country analyses, every statistical institution will use different procedures in dealing with (for example) the depreciation of capital, which will in turn influence the total amount of actual capital included in the statistical analysis and, consequently, the level of TFP (Sargent and Rodriguez, 2000). Several economists argue that labour productivity, although a cruder measurement than TFP, is more closely related to living standards and this is what most countries care about (Sargent and Rodriguez, 2000; Krugman, 1994). Therefore, in this case the focus will be on labour productivity level instead of TFP level.

Furthermore, the concept of labour productivity level can provide insights into the potential growth and catch-up of countries, which is very useful because the focus of this research is to examine why the HICs on average are not able to catch-up with the Netherlands (Pilat, 1996).

7 2.2 What are the important determinants of productivity in the HICs?

2.2.1. Capital intensity

As mentioned in the introduction, it might be expected that the more capital per worker is available (also referred to as capital intensity), the higher the productivity. In other words, if more machinery and tools are available, one would expect that the time to complete a task would be reduced. Several studies about the United States and Europe seem to confirm this, in suggesting that it has been mainly capital intensity that has contributed to labour productivity (Oliner and Sichel, 2000; Jorgenson, 2001; van Ark et al., 2008; Stiroh, 2002). This first hypothesis is presented as follows:

H1: The higher the capital intensity, the higher the labour productivity in an HIC. 2.2.2. Openness

Openness indicates the ‘openness’ of the country to trade. In other words, it might be expected that imports will lead to increased input of materials and new technologies, which will in turn encourage more efficient production. At the same time it can be expected that exporting will lead to more importing, due to increased networking. Wagner (2005) confirms this, when he states that exporting businesses are generally more productive than non-exporting businesses. One reason he gives for this is that engaging in trade leads to knowledge spillovers, especially when exporting to advanced economies. For instance, it gives the exporters the chance to get familiar with the latest technologies in the world, thus providing a great opportunity to update their own production processes. Basically, involved in the international market may lead to benefits that will increase labour productivity. Several studies show that there is a significant correlation between productivity growth and openness (Sachs and Warner, 1995; Ades and Glaeser, 1999; Dollar and Kraay, 2004). It might, therefore, be expected that openness is a strong determinant in explaining the labour productivity level in the HICs. The second hypothesis is presented as follows:

H2: The greater the ‘openness’ in trading, the higher the labour productivity level in an HIC. 2.2.3. Urban population

8

urban areas, the proportion of people engaged in agriculture (the low-productivity sector) is smaller. Furthermore, urban areas are known to have a larger share of high-productivity sectors than rural regions, since urban conglomerations offer industrialised companies a wider range of resources, a larger network of suppliers, and more potential consumers. A number of empirical studies have found that a shift from agriculture to non-agricultural sectors would lead to higher labour productivity (Poirson, 2000; Jaumotte and Spatafora, 2007). These studies also found that countries with a higher value-added share in services and industry will also exhibit higher levels of labour productivity. In addition, figure 2 demonstrates through the use of figures for the labour productivity by sector, that agriculture has the lowest value added per hour worked and industry has the highest value added closely followed by services. This indicates that non-farm activities tend to have a higher labour productivity than non-farm activities. Furthermore, the sectorial data is based on the average gross value added per hour worked of 7 randomly selected advanced economies in the period from 1980 to 2000.

Figure 2: The gross value added per hour worked, at current basic prices in euros.

Note: The data on the sectors is based on average values of Germany, Italy, the Netherlands, Finland, France, Austria,, and Belgium in the period from 1980 to 2000. Agriculture includes agriculture; fishing; forestry; hunting. Industry includes mining and quarrying; manufacturing; construction; electricity, gas, and water supply. Services includes wholesale and retail trade; repair of motor vehicles and motorcycles; transportation and storage; accommodation and food service activities; information and communication; financial and insurance activities; real estate activities; professional, scientific, technical, administrative and support service activities ;community social and personal services.

Source: O’Mahony and Timmer (2009).

Thus, figure 2 implies that higher labour productivity is associated with non-farm activities so, might be an important factor. The third hypothesis is presented as follows:

9

2.2.4. Educational attainment

One might expect that the better an individual’s education, the better their career prospects, and the higher their potential level of productivity. Mefford (1986) mentions that education is one of the factors positively related to labour productivity level. Gutierrez (2000) also argues that education is very important for increasing labour productivity. His argument is based on the idea that in order to work with new tools or machinery a certain level of education is required. So education, according to the literature, might be an important indicator for explaining labour productivity. The fourth hypothesis is presented as follows:

H4: The higher the educational level of the labour force, the higher the labour productivity level of an HIC.

2.2.5. Institutions

Several studies show that the existence of strong institutions is correlated positively with productivity (Islam, 2008; Barro, 1991). Acemoglu et al. (2004) also address the importance of good institutions to promote productivity. It is apparent when the quality of institutions is high (meaning, for example, a strong legal structure and clearly defined property rights), this will enhance productivity. Poor quality of institutions (such as those evidencing corruption) often leads to favouritism by the government, in turn restraining productivity due to less competition. In order to capture the quality of institutions it is common to use a variety of indicators that reflect its multi-dimensional nature (Knack and Keefer, 1995; Loko and Diouf, 2009; Daude and Stein, 2007).

However, there are some limitations when including institutions. Firstly, the nations included in the research are known to have the best quality of institutions in the world. It is therefore unlikely that their institutions would restrain labour productivity. It is, therefore, not particularly useful to include institutions in the analyses.

3. Data and Method

10 3.1 Variables

3.1.1. Dependent variable

Labour productivity is the dependent variable. It is determined by a volume measure of

output and a measure of input use. There are several ways to measure labour productivity. First of all the volume measure of output can be measured by GDP. In this case real GDP rather than nominal GDP has been chosen. Nominal GDP is mainly based on the differences in exchange rates. This is not an optimal measurement: it does not account for the price levels in the non-tradable sector, since the exchange rate only mirrors the prices of the non-tradable sectors in a country. In order to capture the trends of the non-tradable sector, purchasing power parity (PPP) should be included in the assessment. This inclusion ensures that, instead of examining only the exchange rate to determine the differences in the value of the currencies, comparison will also be made between the prices of similar goods to correct for any currency differences. For instance, how many currency units of a country are needed to buy a bundle of products worth one dollar in the US. The real GDP at constant national prices converted by purchasing power parity will be used, which is recommended when using it as a dependent variable in a growth regression (Feenstra et al., 2013, Table 5, p. 30). Although the regression in this research is not based on growth rates but on levels, according to Feenstra et al. (2013) this is the only measurement that is suited for a cross-country analysis over time. The measurement basically uses the growth rate of real GDP from each country’s national accounts to extrapolate GDP from 2005 to other years.In addition, it is at constant prices which means that it will correct for inflation (Kim, 1990).

Secondly, the measure of input use has also two options: it can be measured by total employment/head count, or by total hours worked of self-employed and employed persons. Total employment does not allow for the occurrence of overtime, nor does it include any employees who may work part-time. Total hours worked is therefore considered more appropriate (Freeman, 2008). Total hours worked represents the annual aggregate number of hours actually worked as an employee or a self-employed person (Conference Board, 2014).

That said, it should be noted that the hours worked variable is not always reliable, and that it is harder to obtain this kind of data (OECD Publications, 2001). However, since only data from advanced countries is included in the research, it could be argued that the available information should be trustworthy.

11

3.1.2. Independent variables

Capital intensity. As is the case with labour productivity, capital intensity can be

measured in several ways. Most studies calculate the capital stock either per worker (using total employment), or per labour hour (using the total amount of hours worked by employed and self-employed people). For this paper it has been decided to use capital stock per hour instead of capital per worker because of the reasons that have been described in 3.1.1.

Furthermore, it has been decided to use the value of capital stock at the constant prices of 2005 to correct for inflation. The capital stock has been derived from Feenstra et al. (2015), since they provide data for all of the countries included in this study. The total amount of hours worked is again obtained from the Conference Board (2014).

The level of openness. There are several methods for measuring openness. The most

common indicator for openness is the use of nominal imports and exports relative to nominal GDP. However, several studies argue that this measure is imperfect, because it is too simplistic and will not accurately capture the real mechanism behind the productivity/openness relationship (Rodriguez and Rodrik, 2000; Alcalá and Ciccone, 2004). Alcalá and Ciccone specifically observe that if purchasing power parity is not taken into account, this will distort the relative prices of the non-tradable goods and services. As stated earlier, purchasing power parity corrects for the differences in the relative value of currencies. It will determine the real relative prices in a country (Kim, 1990). When a country makes gains in productivity due to openness, this will be shown in the tradable sector, although it also leads to higher prices in the non-tradable sector. This will decrease openness through a demand for non-tradable goods that is inelastic; change in price has no impact on demand, because trade is affecting the tradable sector more than the non-tradable sector. Therefore real openness, where imports and exports are relative to real GDP at constant national prices converted by purchasing power parity, should be included instead of an openness measurement that is based on nominal GDP to circumvent these issues.

12

The share of urban population will be used as a proxy for the share of employment in

non-farm activities. The indicator of the share of employment in non-farming sectors is available, in the EU KLEMS dataset of March 2008 but it contains only data for 14 of the 16 high-income countries. It is better to have more observations that will make the eventual results more reliable. In addition, the data will be split into two sub-periods and it is important that there should be sufficient observations for each of these sub-periods. The share of urban population can be found in the World Bank database of 2015.

The level of education. There are different ways to measure the level of education. The

most common measurements that have been used in several empirical studies and do not suffer from severe data availability, when examining the period from 1971 until 2005, are school enrolment ratios and literacy rates (Romer, 1990; Barro, 1991; Mankiw et al., 1992). However, literacy rates will not be very useful, since 95% or more of the population in the HICs can read. Also, most people did attend primary and secondary education, therefore including data on these topics is not very valuable. This will not lead to large changes that can explain labour productivity since the literacy rate, the enrolment in primary and secondary education, has already been maximized. The only measurement that might make a difference in explaining labour productivity is enrolment in tertiary education because the group of people who do actually undertake tertiary education in the HICs is not large. This means every additional enrolment in tertiary education might lead to an increase in labour productivity in an HIC.

Statistics for enrolment in tertiary education are accessible, but their use has limitations because there are 65 missing values: particularly in the case of Germany and Canada, where only 7 and 13 observations are available. Nonetheless, 88% of the values are available and that is more than sufficient to include the variable in the research. The statistics for enrolment in tertiary education are a gross percentage, which means they include students of all ages onwards from school entry age. The variable can be found in the World Bank’s 2015 database.

3.2 Method

13

follows: the variables in the regression analysis are assigned subscript i to determine the country and subscript t runs from 1971 to 2005. The model is based on the following equation:

lnLAB_𝑖𝑡 = β₀ + 𝛽₁∗ lnCAPI_𝑖𝑡 + 𝛽₂∗ lnOPEN_𝑖𝑡+ 𝛽₃∗ lnURB_𝑖𝑡+ 𝛽₄∗ lnEDU_𝑖𝑡+ ϵ_𝑖𝑡 (1)

 LAB represents the labour productivity level;

 CAPI is the capital intensity;

 OPEN represents the openness of the country;

 URB is the urban population;

 EDU represents education; and

 ε stands for the error term.

lnLAB_𝑖𝑡 = β₀ + 𝛽₁∗ lnCAPI_𝑖𝑡 + 𝛽₂∗ lnOPEN_𝑖𝑡+ 𝛽₃∗ lnURB_𝑖𝑡+ 𝛽₄∗ lnEDU_𝑖𝑡+ ∑𝑛−1_𝑖=1 δ_𝑖𝑑_𝑖+ ϵ_𝑖𝑡 (2)

The second equation is similar to the first equation, except that a country fixed effect, ∑𝑛−1𝑖 =1δ𝑖𝑑𝑖, has been added to control for specific country characteristics.

3.3 Descriptive analysis

In this section the variables that will be used in the research are discussed, with a view to providing insight into the means, standard deviations and minimum and maximum values of the data.

Table 1: Descriptive statistics

Complete dataset 1971-2005 Partial dataset 1971-1988 Partial dataset 1989-2005 Obs Mean Std. Dev.

Min Max Obs Mean Std. Dev. Obs Mean Std. Dev. lnLAB 560 3.401 .279 2.451 3.941 288 3.223 .235 272 3.588 .182 lnCAPI 560 4.448 .389 3.181 5.061 288 4.223 .336 272 4.687 .286 lnOPEN 560 3.873 .611 2.276 5.133 288 3.661 .574 272 4.099 .568 lnURB 560 4.340 .111 4.045 4.579 288 4.317 .119 272 4.364 .097 lnEDU 495 3.593 .474 2.306 4.576 248 3.250 .336 247 3.938 .317 Note: Obs are the number of observations. The mean is the average value of the data. The variances in the data are represented by the standard deviation (Std Dev.), and the min and max are the lowest and highest value.

Source: Feenstra et al. (2015), the Conference Board (2014), and the World Bank (2015).

14

homogeneity of variances can result in unreliable panel data analysis. For instance, before the transformation, capital intensity and the level of openness to trade had a standard deviation of approximately 32, and the other two variables had a standard deviation of approximately 8, which indicated that the data had large differences in variances. After the transformation into logarithm, the standard deviations of all the variables became below 1, as can be seen in table 1. Thus, this transformation made the variances more equal between the variables.

With regard to the variables, table 1 shows that the mean of lnLAB is 3.401 and the Netherlands has the highest value of 3.941 in 2005 and Japan has the lowest value of 2.451 in 1971. The variable that has the largest mean is capital intensity with 4.448. Belgium has the largest value with 5.061 in 2005 and it is Japan that has the lowest value of 3.181 in 1971.

Furthermore, what can be seen is that the United States accounts for the lowest value of the level of openness with 2.276 in 1971 and Belgium has the highest value with 5.133 in 2005. The share of urban population seems somewhat decreased over time because Switzerland has the lowest value of 4.045 in 1980. So, in the period from 1971 to 1979 Switzerland had a higher share of urban population than in 1980. The highest value has Belgium with 4.566 in 2005. The enrolment in tertiary education seem to decrease as well: the highest value has Canada with 4.576 in 1992 and the lowest value has Switzerland in 1971 with 2.306. This means that the enrolment in tertiary education did not increased in Canada after 1992. When viewing the descriptive statistics of the partial datasets, all the variables seem to have no extreme values and are very similar to the values of the complete dataset. What can be noted when comparing the partial datasets, the variances within the variables all became smaller in the second period. For instance, the standard deviation of lnLAB was .235 and decreased to .182 after 1988. This indicate that the differences between the observations has diminished over time.

15

Table 2: averages values of each country

lnLAB mean lnCAPI mean lnOPEN mean lnURB mean lnEDU mean Australia 3.448 3.146 4.637 4.456 3.630 Austria 3.325 4.211 4.467 4.184 3.378 Belgium 3.577 4.731 4.630 4.564 3.553 Canada 3.453 3.869 4.170 4.345 4.336 Denmark 3.446 4.351 4.616 4.432 3.603 Finland 3.201 4.056 4.555 4.330 3.729 France 3.479 3.650 4.656 4.307 3.552 Germany 3.463 3.711 4.505 4.290 3.691 Italy 3.389 3.588 4.650 4.199 3.486 Japan 3.083 2.987 4.274 4.351 3.480 Netherlands 3.662 4.493 4.704 4.239 3.598 New Zealand 3.075 3.808 3.854 4.434 3.692 Sweden 3.368 4.272 4.087 4.422 3.638 Switzerland 3.524 4.563 4.466 4.194 3.134 United Kingdom 3.285 3.800 4.117 4.361 3.374 United States 3.630 2.739 4.789 4.328 4.146

Source: Feenstra et al. (2015), the Conference Board (2014, and the World Bank (2015).

However, this high average of Canada needs to be handled with caution because, as stated earlier, it is one of the countries that has a small number of observations regarding enrolment in tertiary education. Hence, these results of table 2 do not indicate that one country on average deviates much of the rest of the HICs.

Thus, the descriptive statistics suggest that the variables did not have any extreme or odd values over time or across countries. In the next sub-section the data will be more thoroughly examined because certain assumptions need to be met before doing a panel data analysis.

.

3.4 Panel data assumptions

In this section the focus will be on panel data assumptions. The most important assumptions that might be an issue will be examined are: (1) stationarity, (2) normally distributed residuals, (3) multicollinearity, and (4) serial correlation and heteroskedasticity. In order to check these issues a number of tests will be performed.

3.2.1 Stationarity

16

the stationarity, because it is the most appropriate test when there are relatively few panels and a large number of time periods (Statacorp., 2015). The test will be executed to reveal if the null hypothesis that the residuals at levels contain unit root, are non-stationary, will hold. In order to perform the unit root test the dataset must have no missing values. Therefore, the gaps in independent variable lnEDU have been filled with previous values. Furthermore, to capture if the current value of labour productivity depends not just on the current values of the independent variables, but also on the values in previous time period, a time lag has been included. The adjusted t-values of the residuals of the OLS model, fixed effects model and random effects model are displayed in appendix Ia. This shows that all the adjusted t-values are near -2 and are significant at a 5% level. This means that the null hypothesis that the residuals are non-stationary can be rejected. So, the residuals of all the three models are stationary at a 5% significance level. Thus, there is no stationarity issue.

3.2.2 Normality

The second step is to check whether the residuals are normally distributed. If the residuals are not normally distributed, then certain observations influence the panel data analyses more than others and this leads to biased results (Hill et al., 2012). The residuals of all three models will be compared. From the different figures in appendix II it can be noted that the residuals are most normally distributed when using the fixed effects and random effects models, because they produce histograms with a better bell shape. Thus, the fixed and random effects models have a better normal distribution in comparison to the OLS model therefore only these models will be further examined.

3.2.3 Multicollinearity

The third step is to check if there is multicollinearity. If multicollinearity is present it is more difficult to have significant results in the panel data analyses. When two or more of the independent variables are highly correlated, the assumption of multicollinearity is violated (Hill et al., 2012). This is very important for this particular research, because if determinants are highly correlated among each other; it would no longer be possible to observe the way in which each determinant explains labour productivity in the panel data results, since the estimates are not based on a single determinant anymore.

17

show a slightly higher correlation of below .6, and lnURB is highly correlated to the intercept with .9. The latter might indicate that there is a collinearity problem. When viewing the initial dataset, it can be noted that the average growth rate of the share of urban population of the HICs varies from 0.18% to 0.36% between 1971 and 2005. This explains the strong correlation, because the share of the urban population remains more or less constant over the period corresponding to the intercept. A reason for this can be that the share of urban population is at such a high level in the HICs that it is almost impossible to increase the share of urban population further. The urban areas have already reached their maximum share; this is can also referred to as the ceiling effect (Vogt and Johnson, 2011). Secondly, the intercept of the random effects and fixed effects models is the average value of the fixed effects. However, the random effects model recognises that the countries in the sample are randomly selected, and thus the country differences in the intercept are treated as random rather than fixed (Hill et al., 2012). It might well be that the average fixed and random effects, the country characteristics, in the intercept play a role in determining the share of urban population in a country and that has led to the high correlation between the intercept and lnURB. For instance, the population of a country with a centralized government has the tendency to live near the city where the government is seated. In any case, the high correlation with the intercept is not a serious issue for the present study because the intercept is in this case not the foremost variable that must be interpreted. It will only mean that there will be a high variability of the estimate of the intercept.

In addition, the collinearity diagnostics in appendix IIIb show that the VIF is below 2.50 and the tolerance of VIF is above .40. According to Williams (2014) this indicates that the data has no large multicollinearity issues. Thus, multicollinearity is not a huge problem in this dataset.

3.2.4 Heteroskedasticity and serial correlation

The fourth step is checking for heteroskedasticity and serial correlation. It can lead to overestimation. This indicates that some estimates could be seen as significant, but in reality are insignificant (Hill et al.,2012). Groupwise heteroscedasticity test is preferred when dealing with panel data because it examines whether the variance differ across countries(Baum, 2001). Two tests are performed one concerning the fixed effects model and the other the random effects model.

18

within each country of a random-effect regression model (Baum, 2006). Appendix IVa shows that heteroskedasticity is present in both models because the null hypothesis that the data is homoscedastic are in both cases rejected at a 5% significance level.

For indicating if the data has serial correlation, the Wooldridge test is used because it can reveal autocorrelation in both fixed and random effects models and it requires only a few assumptions. For instance, it does not matter if the data is homoskedastic or heteroskedastic, balanced or unbalanced, with or without gaps (Drukker, 2003). In appendix IVb the results of the test suggest that there is serial correlation. The null hypotheses of no first-order autocorrelation has been rejected at a 5% significance level. Hence, this means that the data suffers from serial correlation and heteroscedasticity therefore a robust option will be included to correct for both issues (Hill et al.,2012).

Thus, it can be argued that after these tests there is a much higher chance that the results of the regression analysis will be trustworthy. In the next section the results will be discussed.

4. Results

In this section the panel data analyses of the fixed and random effects models will be interpreted. In the first section the complete dataset will be interpreted to reveal which determinants seem to matter most in explaining the labour productivity in HICs. In addition to the complete dataset, it has been decided to use partial datasets that present two sub-periods to get a better understanding of how the determinants explain labour productivity over time. The complete dataset has been split in half to get the two sub-periods. The reason for this is that when the dataset is split in half the trustworthiness of the panel data results will not be jeopardized because in each sub-period there are sufficient observations remaining.

The second section will be used to answer the research questions. It will provide insight into the way in which the results of the panel data explain the high labour productivity in the Netherlands and the lack of convergence.

4.1 Panel data analyses

4.1.1 Complete dataset

19

(5) and (6) fits the data the best because 96.33% of the variation is explained within the models. It has, however, to be acknowledged that the fixed and random effects are partially responsible for the high proportion. Also, it has to be noted that because of missing values in the independent variable education models (5) and (6) have less observations than the other models.

Table 3: Empirical results of the complete dataset

Random and Fixed Effects models

Independent variable: Labour productivity (lnLAB)

(1)RE (2)FE (3)RE (4)FE (5)RE (6)FE

(Constant) -.003 [.162] -.007 [.149] .703 [1.511] .771 [1.540] .925 [1.460] .949 [1.485] Capital intensity (lnCAPI) H1 .582 [.029]*** .577 [.030]*** .588 [.029]*** .584 [.031]*** .556 [.024]*** .556 [.025]*** Openness (lnOPEN) H2 .210 [.031]*** .217 [.034]*** .222 [.033]*** .230 [.037]*** .204 [.044]*** .217 [.054]*** Urban population (lnURB) H3 -.180 [.360] -.198 [.369] -.220 [.353] -.234 [.363] Education (lnEDU) H4 .046 [.030] .039 [.034] R2 0.9599 0.9599 0.9608 0.9608 0.9633 0.9633 Observations 560 560 560 560 495 495

Standard dev. in brackets, *significant at 10%, **significant at 5%, ***significant at 1%;

20

Either case, it indicates that hypothesis 3 will be rejected because having a higher share of urban population will not lead to a higher labour productivity. In case of education it has a positive sign but it is insignificant in the models. This means that hypothesis 4 will be rejected that an increase in education increases the labour productivity.

Thus, it can be stated that the most important determinants for explaining labour productivity in 1971 to 2005 are capital intensity and openness. In the next subsections the results of the partial datasets will be interpreted. The first partial dataset is from the period 1971 to 1988 and the second partial dataset is from the period 1989 to 2005.

4.1.2 Partial dataset of the first period (1971 to 1988)

Table 4 presents the estimated coefficients of the independent variables and their corresponding significance levels during the period from 1971 to 1988.

Table 4: Empirical results of sub-period 1

Random and Fixed Effects models

Independent variable: Labour productivity (lnLAB)

(1)RE (2)FE (3)RE (4)FE (5)RE (6)FE

(Constant) .298 [.236] .311 [.263] .518 [1.314] .578 [1.315] .343 [1.295] .341 [1.314] Capital intensity (lnCAPI) H1 .633 [.039]*** .637 [.046]*** .636 [.042]*** .640 [.048]*** .650 [.053]*** .661 [.062]*** Openness (lnOPEN) H2 .069 [.079] .060 [.102] .023 [.073] .066 [.097] .059 [.077] .048 [.105] Urban population (lnURB) H3 -.057 [.308] -.069 [.315] -.006 [.309] -.007 [.322] Education (lnEDU) H4 -.017 [.027] -.023 [.028] R2 _{0.9444 0.9445 0.9446 0.9446 0.9404 0.9405} Observations 288 288 288 288 248 248

Standard dev. are in brackets, *significant at 10%, **significant at 5%, ***significant at 1%;

21

fact that there more observations in these models 288 instead of 248. The results suggest, as also in case of the complete dataset, that capital intensity is an important determinant for explaining labour productivity because it significantly and positively relates to labour productivity in all the models. That is why hypothesis 1 will not be rejected. In addition, the impact of capital intensity on labour productivity is larger in comparison to the complete dataset. For instance, from model (6) it can be seen that a 10% increase in capital intensity will mean a 6.61% increase in labour productivity instead of 5.56% in the complete dataset. In addition, the share of urban population and education have not the expected sign both are negative and are insignificant. However, what is interesting to see is that openness is also insignificant, which was not the case when using the complete dataset. So, it indicates that in this period openness is not really an important factor. One reason could be is that from 1971 to 1988 international trade was not as large than after 1989. For instance, the more remote countries were harder to reach partially because the World Wide Web did not yet exist. This means that only hypothesis 1 holds and the other hypotheses will be rejected.

Thus, the main determinant for explaining labour productivity in this period is capital intensity.

4.1.3 Partial dataset of the second period (1989 to 2005)

22

Table 5: Empirical results of sub-period 2

Random and Fixed Effects models

Independent variable: Labour productivity (lnLAB)

(1)RE (2)FE (3)RE (4)FE (5)RE (6)FE

(Constant) .362 [.246] .431 [.249] .171 [.773] .497 [.811 ] .615 [.661] .998 [.581] Capital intensity (lnCAPI) H1 .433 [.080]*** .389 [.085]*** .432 [.083]*** .388 [.086]*** .368 [.087]*** .331 [.085]*** Openness (lnOPEN) H2 .292 [.051]*** .325 [.056]*** .290 [.059]*** .327 [.064]*** .295 [.046]*** .340 [.051]*** Urban population (lnURB) H3 .048 [.175] -.016 [.195] -.035 [.142] -.120 [.139] Education (lnEDU) H4 .050 [.018]*** .041 [.021]* R2 _{0.9243 0.9250 0.9243 0.9250 0.9322 0.9332} Observations 272 272 272 272 247 247

Standard dev. in brackets, *significant at 10%, **significant at 5%, ***significant at 1%;

As stated earlier, the reason might be that international trade is becoming larger because the HICs were able to reach much more remote countries than had been the case before 1989. Also, education seems to have a small impact on labour productivity: in model (6) it has a significance of 10%. A 10% increase in education will lead to 0.43% increase in labour productivity. However, the relation of labour productivity to the share of urban population seems to be ambiguous because it is positively related to labour productivity but is still insignificant. In this scenario it would mean that the hypothesis 1, 2, and 4 will not be rejected and hypothesis 3 will be rejected.

Thus, the main determinants after 1988 are capital intensity, openness and lastly education.

4.2 Answering the research questions

23

sub-periods, providing more insight into those determinants which are most responsible for the high labour productivity in the Netherlands.

Table 6: The average percentage change of the non-log data

Note:HICs average includes: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, New Zealand, Sweden, Switzerland, the United Kingdom and the United States.

Source: Feenstra et al. (2015), the Conference Board (2014, and the World Bank (2015).

Combining the average percentage change of the non-logarithm data with the results of the panel data analyses can provide a clear picture of the factors determining the labour productivity of the Netherlands and the average of HICs. It has to be acknowledged, however, that the estimates derived from the panel data analysis are approximations based on the averages of the HICs. Therefore the following results should be handled with caution. Model (6) will be used as the general model because it is the best fitted model when including all the variables. In the first partial dataset, table 3, it is only capital intensity that was significant and positively related to labour productivity. The estimate of capital intensity was .661. Table 6 shows that during the first period, from 1971 to 1988, the Netherlands had a 3.79% increase in capital intensity and a 2.79% increase in labour productivity. After linking the non-log data with the panel analyses, the growth in capital intensity has led to a 2.51% (.661*3.79) increase in labour productivity, which means that 2.51% of 2.79% increase in labour productivity is explained by capital intensity. In case of the average of the HICs, it can be seen that it had a 3.44% increase in capital intensity and 2.45% increase in labour productivity. This led to a 2.27% (.661*3.44) increase in labour productivity, which implies that 2.27% of the 2.45% increase in labour productivity is explained by capital intensity. In the second period, from 1989 to 2005, openness and capital intensity were positively and significantly at a 5% level associated to labour productivity . The estimate of openness was .340 and for capital intensity .331. In this period the Netherlands had a 1.07% increase in capital intensity and a 3.15% increase in openness, leading to a 1.43%

24

(.331*1.07+.340*3.15) increase in labour productivity. This indicates that 1.43% of the 1.45% increase in labour productivity is explained by openness and capital intensity. The average of the HICs had a 2.22% increase in capital intensity and a 3.01% increase in openness. This led to a 1.76% (.331*2.22+.340*3.01) increase in labour productivity. It implies that 1.76% of the 1.88% increase in labour productivity is explained by openness and capital intensity. Hence, 90% or more of the increase in labour productivity is explained by the determinants for the average of the HICs and the Netherlands. Furthermore, it shows that after 1988 openness overtook capital intensity as the most important determinant in explaining labour productivity. Also, it can be noted that the growth of capital intensity slowed down after 1988. This confirms what the neoclassical theory argues, namely that the returns to capital does eventually diminish over time. However what the neoclassical theory did not anticipate, was that labour productivity did not decrease, due to the fact that the level of openness compensated for the diminishing effect of returns to capital.

Now that the factors which are largely responsible for explaining the labour productivity of the Netherlands have been determined, it is time to provide some further insights into why the average of the HICs has not caught up. As stated earlier, labour productivity is something unique there seem to be no diminishing returns to capital but constant returns to capital. According to Guterriez, (2000) there are several reasons that explain why labour productivity behaves the way it does. Firstly, he states that labour productivity is not only determined by returns to capital. It can be noticed that the determinants of labour productivity are not all fixed over time. The empirical results suggest that in the first period capital intensity seems to be the only important determinant for explaining labour productivity but that after 1988 also openness became an important determinant for labour productivity. This means there are several determinants that will impact labour productivity differently over time. It is therefore more difficult for a country to enhance labour productivity.

25

Table 7: Does unconditional convergence hold?

average lab_prod average growth rates (%) 1971-05 1972-88 1989-05 New Zealand 21.86 1.20 1.31 Japan 22.94 4.08 2.40 Finland 25.71 3.50 2.83

the United Kingdom 27.52 2.74 2.38

Austria 28.60 2.80 2.38 Sweden 29.48 1.82 2.26 Italy 30.23 2.93 1.21 Canada 31.92 1.49 1.45 Australia 31.92 1.56 1.81 Denmark 32.34 2.78 2.20 Germany 33.14 3.29 2.29 France 33.46 3.48 1.89 Switzerland 34.15 1.52 1.06 Belgium 36.76 3.30 1.70

the United States 38.22 1.37 1.93 the Netherlands 39.58 2.79 1.45

Source: Feenstra et al. (2015) and the Conference Board (2014).

26

direct effect of capital intensity and ,after 1988, the level of openness on labour productivity, as it can be seen that the determinants explain more than 90% of the labour productivity increase. In either case, it will be more difficult for countries to catch-up with the Netherlands.

5. Conclusions

The purpose of the paper is to investigate which determinants are the causes of the high labour productivity level of the Netherlands and to arrive at an understanding of the lack of convergence when comparing the Netherlands with the average of the HICs. It emerged from the empirical research discussion that approximately 90% of the increase in labour productivity of the Netherlands can be explained in the first sub-period by capital intensity and in the second by capital intensity and the level of openness.

There are several explanations for lack of convergence. Firstly, the empirical evidence shows that the determinants impact labour productivity differently over time. This makes it more difficult for countries to determine which factors should be stimulated. Secondly, there is evidence that unconditional convergence does not hold that all countries move towards the same steady state equilibrium. The results suggest either divergence or conditional convergence because there is no clear pattern to be seen. As a result, it will be more difficult for the average of the HICs to converge towards the labour productivity of the Netherlands.

27

might be to examine a more recent period to examine if the determinants that explain labour productivity have changed.

Acknowledgement

I would like to express my gratitude to my supervisor dr. Bart Los who gave me valuable advice and insightful comments on the thesis. Also, I would like to thank Cathy Thompson for proofreading and guidance concerning the linguistic aspects of this thesis.Finally, I thank all the instructors who guided me during the past year for their teaching.

28

References

Acemoglu D., Johnson, S. and Robinson J. (2004). Institutions as the Fundamental Cause of Long-Run Growth. NBER Working Paper No. 10481, Cambridge, Massachusetts: MIT Press.

Ades, A.F. and Glaeser, E.L. (1999). Evidence on growth, increasing returns and the extent of the market. Quarterly Journal of Economics, 114(3): 1025-45.

Alcala, F. and Ciccone, A. (2004). Trade and Productivity, Quarterly Journal of Economics, 119(2): 613-46.

Ark, B. van, Haan, H., and Jong, H.J., de (1996). Characteristics of economic growth in the Netherlands during the postwar period, in Crafts N. and Toniolo, G. (Eds) Economic

growth in Europe since 1945. Cambridge: CEPR/Cambridge University Press,

290-328.

Ark, B. van, O’Mahony, M., Timmer, M.P. (2008). The Productivity Gap between Europe and the United States: Trends and Causes, Journal of Economic Perspectives, 22(1): 25–44.

Barro, R.J. (1991). Economic Growth in a Cross Section of Countries, Quarterly Journal of

Economics, 106(2):407-443.

Barro, R. and Lee, J. (2010). A New Data Set of Educational Attainment in the World, 1950-2010. Journal of Development Economics, 104:184-198.

Baum, C.F. (2001). Residual diagnostics for cross-section time series regression models. The

Stata Journal, 1(1): 101–104.

Baum, C.F. (2006). Stata tip 38: Testing for groupwise heteroscedasticity. The Stata Journal, 6(4):590–592.

Choudhry, M.T. (2009). Determinants of Labor Productivity: An Empirical Investigation of Productivity Divergence, University of Groningen. Retrieved from:

Daude, C., and Stein, E. (2007). The quality of institutions and foreign direct investment.

Economics and Politics, 19(3), 317–344.

Dollar D., and Kraay, A. (2004). Trade, Growth, and Poverty. Economic Journal, 114(493): F22–49.

Drukker, D.M. (2003). Testing for serial correlation in linear panel-data models. The Stata

Journal, 3( 2): 168–177

29

Feenstra, R.C., Inklaar, R. and Timmer, M.P. (2015). The Next Generation of the Penn World Table, American Economic Review, forthcoming, available for download at www.ggdc.net/pwt.

Freeman, R. (2008). OECD Labour productivity indicators, comparison to two OECD databases productivity differentials and the Balassa-Samuelson effect. Statistics Directorate Division of Structural Economic Statistics. Retrieved from: http://www.oecd.org/employment/labour-stats/41354425.pdf.

Fuenta, de la A. (2011). Human capital and productivity, Nordic Economic Policy Review, 2: 103-132

Glaeser, E.L., La Porta, R., Lopez-deSilanes, F. and Shleifer, A. (2004). Do Institutions Cause Growth?. Journal of Economic Growth, 9(3):271-303.

Guterriez, L. (2000). Why is agricultural labour productivity higher in some countries than others? Annual meeting American Agricultural Economics Association in 2000, Tampa, Florida.

Gwartney, J., Lawson, R., and Hall. J. (2014). 2014 Economic Freedom Dataset, Economic

Freedom of the World: 2014 Annual Report, Fraser Institute, retrieved from:

http://www.freetheworld.com/datasets_efw.html

Harris, R.G. (1999). Determinants of Canadian Productivity Growth: Issues and Prospects. Discussion. Paper #8., Ottawa: Industry Canada.

Hill, R.C., Griffiths, W.E. and Lim, G.C. (2012). Principles of Econometrics, International

student version, 4th (ed) . New York: John Wiley and Sons.

Islam, N. (2008) Determinants of Productivity across Countries: An Exploratory Analysis,

The Journal of Developing Areas, 42(1):201-242.

Jaumotte F., and Spatafora, N. (2007). Asia Rising: A Sectoral Perspective. IMF Working Paper No. 07/130, Washington: International Monetary Fund.

Jorgenson, D.W. (2001). Information Technology and the U.S. Economy. American

Economic Review, 90(1): 1–32.

Kennedy, P. (2008). A Guide to Econometrics, 6th ed. Malden, MA: Blackwell Publishing. Kim, Y. (1990). Purchasing Power Parity in the Long Run: A Cointegration Approach.

Journal of Money, Credit and Banking, 22(4): 491-503.

30

Loko, B. and Diouf M.A. (2009). Revisiting the Determinants of Productivity Growth: What’s New? IMF Working Paper No.09/225.

Mankiw, N.G., Romer, D., and Weil, D. (1992). A Contribution to the Empirics of Economic Growth, Quarterly Journal of Economics, 107(2):407-37.

Mefford, R.N. (1986). Determinants of Productivity Differences in International Industry, Journal of International Business Studies, 17(1):63-82.

OECD Publications (2001). Measuring productivity, OECD Manual: measurement of aggregate and industry-level productivity growth. Retrieved from: http://www.oecd.org/std/productivity-stats/2352458.pdf.

Oliner, S. and Sichel, D.E. (2000). The Resurgence of Growth in the late 1990s: Is Information Technology the Story? Journal of Economic Perspectives

14(4): 3–22.

O’Mahony, M. and Timmer, M.P. (2009). Output, Input and Productivity Measures at the Industry Level: the EU KLEMS Database, Economic Journal, 119(538):F374-F403. Pilat, D. (1996). Labour Productivity Levels in OECD Countries: Estimates for

Manufacturing and Selected Service Sectors. OECD Economics Department Working

Papers, No. 169, OECD Publishing.

Poirson H. (2000). Factor Reallocation and Growth in Developing Countries. IMF Working Paper No. 00/94, Washington: International Monetary Fund.

Rodriguez, F. and Rodrik, D. (2000). Trade Policy and Economic Growth: A Skeptic's Guide to the Cross-National Evidence, in NBER Macroeconomics Annual, Cambridge, MA: MIT Press.

Romer, P. (1990). Endogenous Technological Change. Journal of Political Economy, 98(5):S71-S102.

Sachs, J.D. and Warner, A.M.(1995). Economic Reform and the Process of Global Integration. Brookings Papers on Economic Activity, 1:1-118.

Sargent, T.C., and Rodriguez, E.R. (2000). Labour or Total Factor Productivity: Do We Need to Choose? International Productivity Monitor, 1:41-44.

Solow, R.M. (1956). A Contribution to the Theory of Economic Growth. The Quarterly

Journal of Economics, 70(1):65-94.

Soubbotina, P.T., and Sheram, K.A. (2000). Beyond economic growth : meeting the

challenges of global development, chapter 9. The World Bank, Washington D.C.

31

Stiroh, K.J. (2002). Information Technology and the U.S. Productivity Revival: What Do the Industry Data Say? American Economic Review, 92(5): 1559-1576.

The Conference Board (2014). Total Economy Database™, January 2014, retrieved from: http://www.conference-board.org/data/economydatabase/.

The World Bank (2015). World Development Indicators. Washington, D.C.: The World Bank. Vogt, W.P., Johnson, R.B. (2011). Dictionary of Statistics & Methodology: A Nontechnical

Guide for the Social Sciences, 4th (ed). Los Angeles, California: Sage.

Wagner, J. (2005). Exports and Productivity: A survey of the evidence from firm level data. HWWA Discussion Paper 319.

32

APPENDIX

I. Unit root test

a. Levin-Lin-Chu unit-root test with residuals

Number of panels = 16 Number of periods = 35

Ho: Residuals contain unit roots (non-stationary) Ha: Residuals are stationary

logarithm data Residuals of: Adj. t-value lags

OLS model -2.1749** 1

Random effects model -1.9268** 1

Fixed effects model -1.8998** 1

Note: *significant at 10%, **significant at 5%, ***significant at 1%,

a time lag is included to capture if the current value of labour productivity depends not just on the current values of the independent variables, but also on the values in previous time period.

II. Tests of normal distribution of the residuals

a. OLS model

33

b. Random effects model

Note: lnre are the residuals of the random effects model.

c. Fixed effects model

Note: lnfe are the residuals of the fixed effects model.

III. Multicollinearity

a. correlations

Correlation matrix of coefficients of xtreg model (fixed effects model)

lnCAPI lnOPEN lnURB lnEDU _cons

lnCAPI 1.0000

lnOPEN -0.3349 1.0000 lnURB -0.1108 -0.1563 1.0000

lnEDU -0.4292 -0.5653 -0.0956 1.0000

34

Correlation matrix of coefficients of xtreg model (random effects model)

lnCAPI lnOPEN lnURB lnEDU _cons

lnCAPI 1.0000

lnOPEN -0.3260 1.0000 lnURB -0.1060 -0.1498 1.0000

lnEDU -0.4561 -0.5409 -0.1115 1.0000

_cons 0.0053 0.0544 -0.9724 0.2283 1.0000

b. Collinearity Diagnostics (obs=495)

Variable VIF SQRT VIF Tolerance R- Squared lnCAPI 1.72 1.31 .582 0.418 lnEDU 1.89 1.37 .530 0.470 lnURB 1.23 1.11 .810 0.190 lnOPEN 1.06 1.03 .944 0.056 Mean VIF 1.47

IV. Serial correlation and Heteroskedasticity

a. Heteroskedasticity test Fixed effects model

Modified Wald test for groupwise heteroskedasticity in fixed effect regression model

H0: sigma(i)^2 = sigma^2 for all i (homoscedastic) Result:

chi2 (16) = 845.020 Prob>chi2 = .000

Random effects model

Levene's test for groupwise heteroscedasticity in a random effect regression model

H0: constant variance (homoscedastic) Result:

W0(mean) = 9.829 df(15, 479) Pr > F = .000

b. Serial correlation test for both models

Wooldridge test for autocorrelation in panel data Result:

H0: no first-order autocorrelation F(1,15) = 72.237