How do ICT Investments affect Labour Productivity
Growth of Financial Intermediaries?
Empirical Evidence from a Panel of OECD Countries
Master’s Thesis Economics
By Floris Migchielsen
1August, 2016
Supervisor: dr. E.W.M.T Westerhout dr. W.E. Romp
Abstract: This thesis analyzes the effects of ICT investments on the labour productivity growth of 18 OECD countries over the period 1980-2007 by analyzing both the direct and indirect effect. A growth accounting calculation suggests that the direct effect of ICT related capital deepening accounts for a yearly labour productivity growth of 1,44% (s.e: 0,15%). A regression analysis suggests that ICT investments are associated with a short run elasticity w.r.t. the growth rate of TFP of -0,14 (s.e: 0,05), implying a negative indirect effect. The thesis concludes that the negative indirect effect of ICT investments dominates the positive direct effect.
Keywords: Labour Productivity, Total Factor Productivity, Value Added, Financial Intermediation, Information and Communications Technology JEL Classifications: G2, O3, O4
1MSc Student Economics - Monetary Policy, Banking & Regulation, Amsterdam School of Economics,
University of Amsterdam. Program code: MSc ECO. Mail: floris.migchielsen@student.uva.nl. Student number: 11043245. Special thanks to dr. E.W.M.T. Westerhout for his helpful comments and to dr. R.C. Inklaar for providing useful information on output measurement in the EU KLEMS database.
“..what everyone feels to have been a technological revolution, a drastic change in our productive lives, has been accompanied everywhere, including Japan, by a slowing-down of productivity growth, not by a step up. You can see the computer age everywhere but in the productivity statistics.”
Solow (1987)
-Statement of originality
This document is written by Student Floris Migchielsen who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Contents
1 Introduction . . . 1
2 Literature review . . . 4
2.1 Early work on productivity . . . 4
2.2 ICT investments and productivity . . . 5
2.3 ICT investments and financial sector productivity . . . 6
3 Data . . . 8
3.1 Output of financial intermediaries . . . 8
3.2 Dataset . . . 10
4 Growth accounting . . . 13
4.1 Decomposition of output . . . 13
4.1.1 Production function . . . 13
4.1.2 Factor payments . . . 14
4.1.3 Labour productivity growth & the Solow residual . . . 15
4.2 Results . . . 16 5 Regression analyis . . . 19 5.1 Model specifications . . . 19 5.2 Estimation methods . . . 20 5.3 Results . . . 21 6 Discussion . . . 23 6.1 Conclusion . . . 23 6.2 Limitations . . . 24 Appendices . . . 25 A Elasticity of substitution . . . 25
B The translog production function . . . 27
C Growth accounting: U.S. example . . . 29
D ICT investments and prices changes . . . 31
E Regression output . . . 32
1
Introduction
This thesis examines the effect of ICT investments on labour productivity growth of financial intermediaries for a panel of 18 OECD countries over the period 1980-2007. It does so by analyzing both the direct and indirect effect of ICT investments on labour productivity growth. This section starts by defining these subject matters. After that, it provides the reader with an introduction to the subject. This section concludes with roadmap for the remainder of the thesis.
Throughout the thesis, ICT investments are defined as the growth rate of the ICT capital stock, including software and hardware from one year to another. More specifically, ICT investments refer to net ICT investments since the growth rate of the ICT capital stock is corrected for depreciation. ICT investments is used however as a shorthand notation for net ICT invest-ments. Similarly, labour productivity growth is defined as the growth rate of output per unit of labour, including the change of hours worked and the change of labour composition from one year to another.
Direct effects are defined as labour productivity increases due to an in-crease of the ICT to labour ratio, whereas indirect effects are defined as the effect of ICT investments on total factor productivity (TFP) growth. An example of a direct would be an increase of the available computers from 4 to 5 for a group of more than 4 employees. The investment in the fifth com-puter enables the employees to be more productive during their working day simply because they can spend more time on a computer, creating output.
TFP essentially equals the total quantity of output divided by total quan-tity of input. Section 4.1 defines TFP more explicitly. An example of an indi-rect effect would be that an employee uses his computer to automate his own and his coworkers’ tasks, making himself, his coworkers and their computers superfluous, thereby decreasing total input without affecting total output.
Financial intermediaries comprise all types of banks, insurance companies and pension funds More information on the characteristics of financial inter-mediaries is included in section 3. All growth rates are based on continuous compounding.
In 1971, the U.S. company Intel launched the 4004 processor, the essential component of the first personal computer (PC). The PC was, in contrary to predecessing computers, highly scalable and accordingly became an
essen-tial production factor in developed economies. The diffusion of the PC was followed by a myriad of applications, innovations and ICT related changes in society like the popularization of internet to which academia commonly refers to as the ICT Revolution (Jovanovic & Rousseau, 2005)1.
In the first 20 years of the ICT revolution, it was generally expected that the increasing dominance of ICT in offices and factories would lead to an unambiguous increase of all types of productivity, mainly because ICT allows for the automation of routine tasks. Since the marginal cost of labour of an automated routine task is virtually zero, standard economic theory predicts a significant increase of productivity of ICT intensive services like financial intermediation (van Ark et al. 2002). That is, standard economic theory predicted that less input was needed to create an equal amount of output.
However, simultaneously with the start of the ICT revolution, a slowdown of labour productivity growth in virtually all developed economies started. The productivity downturn was most severe in services industries, once more, like financial intermediation. This slowdown of labour productivity growth was known to be caused by a decreasing of TFP. Figure 1 visualizes the opposite moving trends trends of the ICT capital stock and TFP in the financial intermedation industry for the U.S.2. The underlying data of figure
1 are included in appendix C. Section 2 shows that these opposite moving trends have become known as the productivity paradox (Brynjolfsson, 1993). Since the core business of these institutions is financial intermediation, they are referred to as financial intermediaries in the remaining of the thesis. More information is included in section 3. Figure 1 includes an index of TFP, representing productivity as an ratio of output over all input factors.
Especially for financial intermediaries, the opposite movements are strik-ing because financial intermediation has faced a significant transformation since the start of the ICT Revolution. Examples include on-line banking and ATMs replacing bank branches, and computers that report the sensi-tivity of a financial intermediary to all types of risks3. Therefore, financial
1ICT: Information and Communications Technology. The difference with IT
(Informa-tion Technology) is ambiguously documented in academic literature. This thesis follows the database (EU KLEMS) and the United Nations System of National Accounts, who both solely refer to ICT.
2Figure 1 displays data for only one country for the ease of exposition. The author
chose to display the U.S. because it embodies the largest financial sector in the sample.
Figure 1: Opposite trends of indexes of the ICT capital stock and TFP of U.S. financial intermediaries over the sample period (base year=1995)
intermedation is generally classified as an ICT-intensive industry.
Additionally, focusing this thesis on financial intermediaries is relevant because it contributes to current debate on reforms of the banking sector. A negative relationship between ICT investment and financial intermediaries’ productivity could for example be used as rationale for stimulating alterna-tive ways of financial intermediation like fin-tech.
The outline of this thesis is as follows. Section 2 includes an overview of mayor developments in growth accounting and the relationship between ICT investments and productivity both in general and in the financial sector. Section 3 focuses on providing the reader with a better understanding of the data used and the choices made with respect to the sample selection. The data is used in section 4 as the input for a growth accounting calculation. The growth calculation provides evidence that ICT to labour input ratio in-creased during the sample period, having to a significant impact on labour productivity growth. Subsequently, section 5 examines whether the invest-ments in ICT increased TFP of financial intermediaries, based on the same data. The thesis finishes with a conclusion and some limitations.
2
Literature review
In this section, some related research is discussed. A notable result of previ-ous research is that ICT can only increase total factor productivity if indus-tries adapt their labor force and organizational structure to it. Because it might take several years to adapt to new ICT equipment, accompanying pro-ductivity increases will take several years to develop accordingly. Since this thesis examines the direct effect of ICT investments on labour productivity growth by a growth accounting calculation, the literature review starts with some mayor developments in the growth accounting methodology.
2.1
Early work on productivity
The relationship between technical change and productivity was first for-malized by Hicks (1932). Hicks’ work includes theory on the elasticity of substitution (ES) and the assumption of neutral technical change4. Both
concepts have been key in the development of production functions like the Cobb-Douglas type, the constant elasticity of return type and the translog type. Throughout the 1930’s, Hicks’ contributions did not attract much at-tention as they became dominated the Great Depression and Keynes’ The General Theory of Employment, Interest and Money (1936).
Nevertheless, in the late 1950’s, during the economic recovery following WWII, renewed interest in productivity theory emerged. By using standard neoclassical assumptions, Solow showed in his 1957 paper that output per worker is determined by the amount of capital per worker and technical change. The technical change assumed by Solow was of the same type as introduced by Hicks. Thereby Solow gained Hicks, without mentioning him in his papers, a larger audience (Sato, 2006).
Solow’s theories were the starting point for a myriad of researches on technology and productivity. For instance Arrow et al. who, in a paper (1961), co-authored by Solow, developed the Constant Elasticity of Substi-tution (CES) production function. In an empirical section of their paper Arrow et al. find that the ES of capital and labour an in the US is smaller than 1. This result implies that a decrease of a production factor’s costs will
4Appendix A includes a further elaboration on the elasticity of substitution. A thorough
understanding of this concept is not necessary for understanding the empirical results of the thesis. However, it does provide the reader with a better understanding of the underlying dynamics of the production function used in the growth accounting calculation.
eventually lead to a decrease of the production factor’s share in contribu-tion to total output, et vice versa. The result has been replicated by most empirical evidence on developed economies.
A notable extension to Solow’s theory was developed by Mankiw, Romer & Weil (1992), who show that the two-factor Cobb-Douglas function, ex-tended with a third production factor, human capital, can explain income differences among countries. Moreover, they find a positive relationship be-tween human capital and the Solow residual, indicating that these production factors are complementary in the process of enlarging labour productivity.
A classic theory on the consequence of productivity growth is provided by Baumol & Bowen (1966), whose theory on increasing costs due to technical progress became known as Baumol’s disease. Baumol & Bowen point out that the productivity of some sectors, like performing arts, are less suitable for productivity increases than other sectors, like manufacturing. The increasing productivity in the latter type of sectors allows those sectors to increase wages. However, in order to prevent a disproportional share of the labour force from moving towards the sector with higher wages, the sectors that are less sensitive to productivity increases, need to raise their wage level as well. The non-productivity increasing sectors are faced with increasing labour costs, lacked by increasing productivity and need to increase their prices to prevent losses. As a consequence, the sectors that are less sensitive to productivity growth shrink due to a shortage of demand.
2.2
ICT investments and productivity
Following a widely recognized productivity slowdown that started in the early 1970’s5, researchers like Baily & Gordon (1988) started to study relationships
between ICT and labour productivity by using a growth accounting approach, but found no significant relationship. In the appendix of Baily ad Gordon’s paper, David Romer explains that there is a lack of relationship between ICT investments and productivity because ICT investments embody only a marginal share of total investments.
In his further research (2012), Gordon stresses that productivity growth follows an S-curve, in which the steep part with about 100 years of produc-tivity increasing inventions like electricity and the washing machine lies in
5Roughly synchronized with the inception of the ICT revolution as displayed in table
the past. By the end of the 1980’s, the lack of a relationship between IT and productivity had become known as the productivity paradox.
In 1993, the concept of the productivity paradox was formalized by Erik Brynjolfsson. Based on a meta-analysis, Brynjolfsson points out that, espe-cially in the services sector, there is no evidence for a positive relationship between ICT and productivity. He that this phenomenon is caused by:
i Measurement error : the existing data on input and output of ICT, is inappropriate for measuring productivity
ii Lags: The investments required by TFP might not payoff directly, but after two or three years
iii Redistribution: ICT is mainly used for non-productive, redistributive activities like finance and marketing
iv Mismanagement : (i) and (ii) lead to incomplete management informa-tion, causing misallocation and overconsumption of ICT6
In later research (for example 1996), Brynjolfsson became more optimistic on the relationship between productivity and economic growth stating that productivity paradox has vanished after 1991 because it took organizations multiple decades to exploit the capacities of their ICT investments. Since by 1991 most organizations have integrated efficient use of ICT in their business processes, increases of productivity could become exponential.
2.3
ICT investments and financial sector productivity
More recently, a broader interest in the relationship between ICT and pro-ductivity of financial intermediaries emerged in academia. Examples include, Parsons, Gottlieb & Denny (1993), who show that, when assuming a translog production function, for the period 1974-1989, in the Canadian Banking sec-tor, a small yet positive relationship exits between ICT investments and TFP of banks. As their results become more positive towards the end of their sam-ple period, they suggest that the largest productivity increases caused by ICT investments are still ahead.
6For example, the services of Wikipedia are, despite the fact that these services do
Erber & Madlener (2008) find little or no relationship between ICT capital and TFP in the financial sector. But, in consonance with Mankiw, Romer & Weil (1992), they find a strong relationship between human capital and productivity. They suggest that there has been some over-investment in ICT in the second half of the 1990’s and that this capacity can only be used is the labor force is adapted to the ICT capital.
Edquist & Henrekson (2016) mention that a regression of ICT invest-ments on the growth rate of TFP only measures the indirect effect of ICT investments. They define the direct effect of ICT investments on productivity as the increased availability of capital per unit of labour. Equivalently, they define the indirect effect as the impact of ICT investments on the growth of production in excess of the effect via capital accumulation. The definitions of the direct and indirect effect that are mentioned in section 1 are based on these two definitions. Although Edquist & Madlender do not find a short-run relationship between ICT investments and TFP, they do find a positive effect with a lag of 7 to 8 years.
Another contributor to the empirics of finance and productivity in the financial sector is Philippon (2012), who shows that “Despite its fast com-puters and credit derivatives, the current financial system does not seem better at transferring funds from savers to borrowers than the financial sys-tem of 1910.” He finds that ICT investments did indeed lead to efficiency gains in for U.S. supermarkets, but in the case financial intermediaries, much of new ICT capital has been dedicated to trading activities, not leading to any direct output. Philippon’s research is unique from the perspective that he created his own dataset with a longitude of more than 100 years and his own framework for productivity measurement. This is particular because nowadays, both Solow’s and Hicks’ work on productivity analysis is still at the center of most economic research on productivity.
3
Data
Section 3.1 clarifies the concept of financial intermediaries’ output. After that, section 3.2 describes the data used and explains the sample selection choices.
3.1
Output of financial intermediaries
Financial intermediaries’ output constitutes indirect financial intermediation and financial services. According to Mishkin (2004), the raison d’ˆetre of the financial system is the transfer of funds from savers to spenders. This transfer can occur both directly, via securities, like bonds and shares and indirectly, via financial intermediaries.
Financial intermediaries include all types of banks, insurance companies, pension funds and investment management companies. The distinction be-tween the types of intermediaries is made on basis of the composition of the financial intermediaries’ balance sheets. In the sample period (1980-2007), the majority of financial intermediation took place by financial conglomerates that engaged in numerous types of intermediation.
Problems that the financial intermediaries aim to overcome are risk shar-ing, economies of scale, information asymmetries and matching problems. Taking the example of a single entrepreneur and a single lender. A lending contract between the two, without the existence of a financial system, will probably:
1. Have a high interest rate, because the lender wants to be compensated for the fact that he loses all his money in the case of default of the entrepreneur
2. Be expensive to originate, because originating a loan does typically not belong to an ordinary savers’ and borrowers’ everyday activities
3. Be subject to a lack of confidence of the lender, since the entrepreneur has incentives to:
• Not fully disclose all risks that his investment project is subject to, in order to bargain a low interest rate on his loan
• Engage in risk seeking behaviour since he can default on the loan in the case of a bankruptcy and will better off in the case of a success
4. Not exist at all, because the probability that an entrepreneur meets a lender who is willing to lend exactly the amount he needs for the period he needs is small
By attracting deposits from savers and providing loans to spenders, financial intermediaries match multiple lenders to multiple borrowers and get rid of these problems. These activities generate a net interest income that equals the difference between the interest generated on outstanding loans and the interest paid on deposits.
Besides net interest income, most financial intermediaries’ output includes non-interest income. This income is generated by financial services that are generally, but not exclusively related to direct finance, like advising on and the actual issuance of securities, M&A activities, market making or investment management. Investment banks, a type of financial intermediary focus on these non-core activities.
Since the activities that generate an interest income, create output for both borrowers and lenders, both output sides should be accounted for in output measurement. The standard prescribed by both the System of Na-tional Accounts (SNA) and the European System of Accounts (ESA) is the Financial Intermediation Services Indirectly Measured (FISIM) method. The most recent updates of the SNA and ESA, prescribe a version of FISIM that uses the reference rate approach. In an equation, the reference rate approach takes the following form:
F ISIM = (rL− rR)yL+ (rR− rD)yD (1)
Where rR is the reference rate, rL is the interest rate on loans, yL is the
average balance on loans, rD is the interest rate on deposits and yD is the
deposits outstanding. Equation (1) shows that the reference rate approach implies that earnings on loans above a certain reference rate and payments on deposits below the same reference rate, are part of financial intermedia-tion. A reference rate recommended by SNA (2008) is the interbank lending rate. Throughout this thesis, output encompasses the sum of FISIM and the quantity of other financial services.
3.2
Dataset
This thesis uses data of the EU KLEMS database7. This database includes
data on all 72 industries as classified by the ISIC revision 3 classification system of 32 countries8.
Using data from the EU KLEMS database provides the empirical analyses with 4 conveniences.
Firstly, the EU KLEMS database uses the FISIM method described in section 3.1 to measure financial intermediation services for all countries9.
Secondly, the database is not focused on a single industry. Therefore, the data results of the empirical analysis are comparable to other research on productivity analyses that use the same database.
Thirdly, the database is constructed for productivity analyses like this thesis. This fact minimizes the probability that data used do not adequately reflect variables in growth accounting and regression models. Likewise, it separates capital related data into ICT and other capital.
Finally, the number of missing observations in the sample period (1980-2007) is small, which makes it easier to extract a balanced data sample.
A limitation of the database is that the output of activities that produce a non-interest income are measured by value added. Value added is equivalent to the difference between inflation adjusted revenue and costs of intermedi-ate goods like energy, intermediintermedi-ate services and other intermediintermedi-ate goods. Following this method leads to other measurement issues, for example:
For insurance companies, output is measured by their revenue minus their operating expenses excluding personnel expenses. This method implies that an insurance company can increase its output simply by paying less compen-sations to their clients.
7The EU KLEMS database is constructed by the EU KLEMS Project, funded by the
European commission. Dutch organizations involved in the project include the University of Groningen, the Dutch bureau for economic policy analysis (Centraal Plan Bureau) and the Vrije Universiteit in Amsterdam. The database and a complete list of consortium members can be found on:http://www.euklems.net/index.html
8ISIC is the acronym of: International Standard Industrial Classification of All
Eco-nomic Activities. More info on: http://unstats.un.org/unsd/cr/registry/regcst. asp?Cl=2. In this system the industry financial intermediation is classified by industry code J.
9This fact is, to the author’s best knowledge, not confirmed in any literature on the EU
KLEMS database. Therefore, the author verified this fact by speaking to dr. R.C Inklaar, one of the composers of the EU KLEMS database, on the phone.
Another example is provided by Berger & Humphrey (1992), who point out that the invention of the ATM did not save much of the banking sector’s resources since it merely decreased the size and increased the frequency of transactions. The fact that the ATM increases bank productivity by increas-ing the consumers’ utility is not measured by usincreas-ing the value added method. As an alternative, the EU KLEMS database provides data on gross output, but these data suffer from similar limitations.
In order to overcome these limitations, other databases have been con-sidered, but none of them could match the conveniences of the KLEMS database. For example, the OECD STAN database does, unlike the EU KLEMS database, not use FISIM to measure financial intermedation for all countries.
It is widely recognized that the sample period (1980-2007) was a period of retracting supervision. With Reagan, Thatcher and Greenspan coming into power in 1980, 1979 and 1987 respectively, neo-liberalism dominated super-visory policies in developed economies (Rosner & Kotz, 2015). This trend ended with the start of the credit crisis in 2007. Therefore, it is unlikely that supervision is a productivity limiting factor during the sample period. Com-bined with the fact that the period (1980-2007) lies within ICT revolution, makes it a suitable sample period.
By default, all OECD member countries included in the EU KLEMS database are included in the panel. However, in order to keep the panel as balanced as possible, countries with many missing observations have been excluded. These are countries with small banking sectors such as Estonia and Poland. For other countries, such as South Korea, the database does not include information on ICT investments, and are therefore excluded from the sam-ple. Following this procedure, the panel does include all home countries of Global Systemically Important Banks (GSIB). An exception to this principle is Switzerland, the home country of the GSIBs UBS and Credit Suisse. The EU KLEMS database does not include any data on the Swiss economy, so Switzerland was forced to be excluded from the panel as well. The rationale behind including many GSIBs in panel is to include that these institutions dominate the financial intermediation industry in OECD countries.
The EU KLEMS database measures capital input quantities based on the perpetual inventory model (PIM) in which the capital stock of both ICT
other capital are defined as the weighted sum of past investments, corrected for a geometric depreciation rate. For ICT and other capital, different de-preciation rates per country an per type capital type (ICT and other other capital) are used (O’Mahony & Timmer, 2009).
Data on labour input comprise both hours worked and the labour composi-tion. Labour composition consists of a range of factors having an impact on the wage level such as educational attainment, age and gender. In the EU KLEMS database, it is calculated by:
Li,t = Y j Hl,i,t Hi,t wl,i,t!
Hi,t = LCi,tHi,t. (2)
Where wl,i,tdenotes the period average share of labour costs of the employees
of type l, for example a male of age 45 with a university education, in total labour costs. Hl,i,t denotes the hours worked by employees of type l, Hi,t
denotes the hours worked and LCi,t denotes labour composition. Besides
just educational attainment, LCi,t accounts for all factors that are generally
described as “labour quality” in growth accounting methodology (O’Mahony & Timmer, 2009), making it a more accurate method to measure labour input than solely educational attainment.
4
Growth accounting
Following Edquist & Henrekson (2016), the empirical analysis of this the-sis distinguishes between the direct and indirect effect of ICT investments. Section 4 uses a growth accounting approach to analyse the direct effect of TFP. It shows that ICT investments have significantly contributed to labour productivity during the sample period because they increased the capital available per worker. It also shows that the contribution of TFP to total output is negative in most countries. The methodology explained in section 4.1 is equivalent to the methodology used by the EU KLEMS database10.
4.1
Decomposition of output
4.1.1 Production function
Assume that output of the financial intermediaries Q of a country i in year t is produced by the following production function.
Qi,t = Ai,tICT υICT ,i,t
i,t C
υC,i,t
i,t L
1−υC,i,t−υICT ,i,t
i,t (3)
Where A denotes Hicks-neutral total factor productivity T F P11, ICT
de-notes the amount of ICT Capital, C dede-notes the amount of other capital and L denotes labour input. Labour productivity is implied by Qi,t/Li,t.
Hicks-neutral TFP comprises the technical ability and any other costless factor that affects the quantity produced. It may reflect reflect spillovers thrown off by research projects, or it may simply reflect inspiration and ingenuity (Hulten, 2001). Labour input reflects the combination of hours worked and labour composition as explained in section 3.2.
The share of factor payments υICT,i,t, υC,i,tand 1−υC,i,t−υICT,i,tare larger
than zero and smaller than 1 such that production function (3) exhibits con-stant returns to scale and positive, diminishing, returns to each production factor. More formally, for the marginal product of any production factor φ, M Pφ,i,t, it holds that M Pφ,i,t> 0, whereas M Pφ,i,t0 < 0.
10O’Mahony & Timmer (2009) have provided the author with an elaboration on the
methodology of the EU KLEMS database. Although EU KLEMS does contain TFP data, the author replicated these data in order to verify that his understanding of the methodology is correct .
11Definition of Hicks-neutral total factor productivity: productivity that affects marginal
productivity of production factors equally, leaving the marginal rate of technical substitu-tion constant. Non Hicks-neutral productivity analysis is beyond the scope of this thesis.
Function (3) is a generalization of the Cobb Douglas production function, since its shares of factor payments are variables instead of parameters. That is, υφ,i,tis used rather than υφ,i. Just like the Cobb-Douglas function, (3) is an
application of the translog production function. Appendix B provides more information on the derivation of (3) from the standard translog production function.
Since the assumption of constant returns to scale is essential in the cal-culation of labour productivity, some notes on the rationale behind this as-sumption are required. Assuming decreasing returns to scale would imply that smaller banks outperform larger banks, whereas increasing returns to scale imply that a single monopolist dominates the market. Although some recent studies show evidence for increasing returns to scale in the banking sector, the current status quo in academia is that banks exhibit constant returns to scale (Wheelock & Wilson, 2012). Evidence for constant returns to scale for financial intermediaries is also provided by Philippon (2012).
4.1.2 Factor payments
In the production function (3), the shares of the factor payments υICT,i,t and
υC,i,t are defined by:
υICT,i,t≡
M PICT,i,tICTi,t
Qi,t
, υC,i,t ≡
M PC,i,tCi,t
Qi,t
(4)
Assume that the market of input factors of financial intermediation is char-acterized by profit maximizing producers and imperfect competition with exogenous factor prices. This assumption is inspired by Federici & Saltari (2014). Profit maximization implies that financial intermediaries in every country i in every year t adjust the inputs of every production factor such that the marginal product of any production factor, M Pφ,i,t equals its factor
price βPφ,i,t12. Where β denotes a constant markup on every perfectly
com-petitive factor price Pφ,i,t13. Therefore, the shares of factor payments can be
12 Profit π
i,t is given by the function πi,t = Qi,t− βPICT ,i,tICT − βPC,i,tC − (1 −
βPICT ,i,t− βPC,i,t)Li,t. Given the previous assumptions of this section and the additional
assumption that factor prices are larger than zero, πi,t has only one maximum, at the
point where δπi,t/δφi,t = 0 holds for every production factor φ. M Pφ,i,t = βPφ,i,t is the
first order condition of maximization of profit πi,t. 13β = 1 implies perfect competition
measured by:
υICT,i,t =
βPICT,i,tICTi,t
Qi,t
υC,i,t =
βPC,i,tCi,t
Qi,t
(4’)
The assumptions that the production function (3) exhibits CRS and M Pφ,i,t=
βPφ,i,t are required theoretically to make sure that the shares of factor
pay-ments add up to 1. This has the advantage of making everything come out neatly in terms of intensive magnitudes (Solow, 1957).
In order to fulfill this requirement, the EU KLEMS database measures the shares of factor payments of (4’) as the two-period average of a production factor’s factor payments as a share of total factor payments to make sure that the shares of factor payments add up to 1.
4.1.3 Labour productivity growth & the Solow residual
The quantities of the production factors may change over time and hence affect the output created by the production function. Assessing the effects of growth rates of production factors and technical change on the growth rate of output can be done by taking the natural logarithms and subtracting period t − 1 from period t. For the ease of exposition, define the share of labour input of total factor payments as:
υLi,t ≡ 1 − υICTi,t− υCi,t (5)
Taking the natural logarithm of the production function (3) yields:
ln Qi,t = ln Ai,t + υICTi,tln ICTi,t+ υCi,tln Ci,t+ υLi,tln Li,t (3’)
Subtract a 1 year lag from function (3’) and assume that the shares of factor payments are constant from one period to another such that:
∆ ln Qi,t = ∆ ln Ai,t+ υICTi,t∆ ln ICTi,t+ υCi,t∆ ln Ci,t+ υLi,t∆ ln Li,t (3’’)
Equation (3’’) decomposes output growth into growth rates of the production factors and TFP. It can be used to obtain an equation for labour productivity by subtracting ∆ ln Li,t from both sides:
Where the lower case variables indicate indicate the per unit of labour version of this variable. The left hand side shows the labour productivity variable, which is indicated by: ∆lnqi,t = ∆lnQi,t − ∆lnLi,t (Stiroh, 2001).
Equation (6) shows that the output per worker depends on the growth rates of the amount of ICT available per worker, the amount of other capital available per worker and TFP. It follows from (6) that ICT investments have a positive effect on labour productivity by increasing the availability of ICT per unit of labour. This is the direct effect as defined in section 1.
On the other hand, the indirect effect implies that ICT investments could also have a negative impact on labour productivity growth if ICT investments have a negative impact on TFP growth.
In section 5, the indirect effect is analyzed by regressing the yearly growth rates of production factors on the yearly growth rate of TFP. The approxi-mation used for the growth rate of TFP is obtained by the Solow residual. This Solow residual (1957) can be obtained by rearranging (3’’) and provides standard approximation for the growth rate of TFP:
∆ ln Ai,t = ∆ ln Qi,t− υICT,i,t∆ ln ICTi,t− υC,i,t∆ ln Ci,t− υL,i,t∆ ln Li,t (3’’’)
Equation (3’’’) shows that the growth rate of TFP can be calculated as the difference between the change of output over time and changes of the input quantities of production factors. More specifically, it shows that TFP growth comprises every output change that is not accounted for by a change in the use of a production factor. Throughout this thesis, TFP growth is measured by equation (3’’’).
4.2
Results
Table 1 shows the yearly labour productivity growth for the sample countries based on equation (6). Three conclusions can be derived from the growth accounting calculation.
• The direct effect of effect of ICT-related capital deepening accounts for a cross-country average yearly productivity growth of 1,44% with with a standard error of 0,15%
Table 1: Decomposition of yearly labour productivity growth per country
Country Period ∆ ln(q) υICT∆ ln(ict) υC∆ln(c) ∆ln(A)
Austria 1981-2007 1,29% 0,84% 0,16% 0,29% Belgium 1981-2006 4,26% 1,65% 0,18% 2,43% Canada 1980-2004 1,62% 1,16% 1,64% -1,18% Czech Rep. 1996-2007 4,60% 1,26% 0,13% 3,21% Denmark 1981-2007 4,33% 2,30% -1,30% 3,33% Finland 1980-2007 1,99% 2,58% -1,01% 0,43% France 1981-2007 1,51% 0,83% 0,28% 0,39% Germany 1980-2007 1,16% 0,57% 0,08% 0,51% Hungary 1996-2007 4,50% 1,19% -0,48% 3,79% Ireland 1989-2007 3,47% 0,17% -0,83% 4,13% Italy 1980-2007 0,64% 1,07% -0,16% -0,27% Japan 1980-2006 3,63% 1,19% 0,27% 2,17% Netherlands 1980-2007 1,72% 1,65% 0,01% 0,06% Slovenia 1996-2006 3,74% 2,47% 1,98% -0,71% Spain 1981-2007 2,57% 1,37% 0,06% 1,14% Sweden 1994-2007 1,83% 2,07% 0,67% -0,92% U.K. 1980-2007 1,60% 1,68% 0,28% -0,36% U.S. 1980-2007 1,24% 1,83% 0,84% -1,43% mean 2,54% 1,44% 0,16% 0,94% s.e. 0,32% 0,15% 0,19% 0,42%
Note: For any variable Vi,t, it holds that: ln( Vi,t Vi,t−k) 1 k = 1 kln Vi,t
Vi,t−k. Where k is denotes
the period length. This property implies that the average yearly growth rate of any variable during the sample period can be obtained by dividing the total growth rate during the period by the number of years in the period.
• The average yearly TFP growth is positive, providing evidence that the trend of negative TFP growth as shown in figure 1 does, on average, not hold for this panel
The results are not conclusive on the effect of ICT investments on labour productivity growth. For example, if the average yearly TFP growth would have been 2,55 % points higher without ICT investments, the overall effect of ICT investments would be negative. This provides a rationale for analyzing the effect of ICT investments on TFP growth as is done in section 5.
The labour productivity of financial intermediaries grew on average with 2,54% per year, with a standard error of 0,32%. This result implies that ICT investments account for more than half of the labour productivity growth in the sample. This is an indication of relevance of ICT investments in the process of increasing labour productivity.
Following the assumption of section 4.1, changes in the use of production factor are caused by a change of its factor price. The changes of the factor prices are included in table 3 in appendix D. This table shows, in line with the substantial contribution of ICT investments to yearly output growth, that the yearly decrease of the factor price of ICT is on average 3,62% with a standard error of 1,25%.
The average growth of TFP is 0,94% with a standard error of 0,42%, implying that the values of average TFP within a 95% confidence interval are larger than zero. Canada, Italy, Slovenia, Sweden, the U.K and the U.S. faced a negative average TFP growth during the sample period. From this finding follows that the effect of decreasing TFP in the U.S. as shown in figure 1 and table 2 with 40,20% during the sample period is not present in all countries of the panel.
For the sake of clarification, an example of underlying calculations is included in appendix C. It comprises an overview of the year on year de-composition of output growth for U.S. financial intermediaries. Following, equation (6), subtracting labour input growth from both sides of the period total and dividing by the number of years, yields the the results of table 1.
5
Regression analyis
Section 4.2 suggests that the direct effect of ICT investments on labour pro-ductivity growth is unambiguously positive. The regression analysis in sec-tion 5 examines the indirect effect of ICT investments on TFP. Secsec-tion 5.1 elaborates on the regression models applied, section 5.2 clarifies the estima-tion methods used and secestima-tion 5.3 clarifies the regression results. Overall, section 5 concludes that the impact of ICT investments on TFP growth within a 10 year period is negative and hence opposing the positive direct effect that was found in section 4.
5.1
Model specifications
The regression analysis is largely inspired by Edquist & Henrekson (2016). The baseline regression model is given by:
∆ ln Ai,t = βICT,i,t∆ ln ICTi,t+ βC,i,t∆ ln Ci,t+ βL,i,t∆ ln Li,t
+δt+ ui,t (7)
i = 1...N t = 1...T
Where the symbols that are equivalent to equation (3’’) can be interpreted accordingly, δt denotes a year dummy capturing year specific shocks that
affect TFP, and βφ,i,t represents the TFP elasticity of any production factor
φ. Moreover, N refers to the number of countries (18) in the sample and T refers to the number of years (28) in the sample. Eliminating possible short term shocks and uninformative high-frequency noise data is done by re-running the baseline model with 3, 5 and 10 year moving averages of all input factors. This method is based on Goodridge, Haskel & Wallis (2014). As discussed in section 2, some previous research suggests that the effect of ICT investments on TFP growth does become positive in the long run14.
The first method to analyze the long run effect is a cross sectional re-gression analysis of ICT investments on the growth of TFP in the period 1998-2007. For this purpose, the growth rates of the input factors are split into three periods, being 1998-2007, 1989-1997 and 1980-1987. For every
14For example Edquist en Henrekson (2016), Parsons, Gottlieb & Denny (1993) and
Brynjolfsson (1993). In some of his other research Brynjolfsson suggests that the pro-ductivity increasing capacities of the steam engine and electricity went through a similar process.
period, the yearly average growth rates are regressed on the yearly average growth rate of growth of TFP in the period 1998-2007, such that:
∆ ln A1998−2007i,t = βICT,i∆ ln ICTi1998−2007+ βICT,i∆ ln ICTi1988−1997
+ βICT,i∆ ln ICTi1980−1987+ βC,i∆ ln Ci1998−2007
+ βL,i∆ ln L1999−2007i + ui (8)
i = 1...N
Obvisously, running several variations of this model by excluding one or two periods can provide an indication of the robustness of this regression model. For a deeper understanding of the effects of ICT investments within a 10 year period, the analysis also follows Edquist & Henrekson (2016). All variables are smoothed by taking a 3 year moving average of t, t − 1 and t − 2.
Pretesting indicates that running the baseline model with 3 year moving averages leads to similar results as using 5 and 10 year moving averages. Hence, in order to analyze the relationship within a 10 year period, a 3 year moving average is used. Following Edquist & Henrekson, ICT investments enter the regression model with a 1 to 10 year lag:
∆ ln Asmoothedi,t = βICT,i,t∆ ln ICTi,t−ksmoothed∆ ln C
smoothed i,t
+βL,i,t∆ ln Lsmoothedi,t + ui,t (9)
i = 1...N t = 1...T
Where k denotes the number of lags by which ICT investments enter the regression. Evidently, rerunning this regression model with k = 0 and in-cluding year dummies is equivalent to running the baseline model on 3 year moving averages of input data.
5.2
Estimation methods
As a result of the design of the panel regression models (7) and (8), shown in the section 5.1, a first difference estimator, i.e. OLS with clustered standard errors is most suitable. For the cross sectional regression model (8), an OLS regression, with Huber-White sandwich estimators of standard errors is most suitable in order to correct for possible presence of heteroskedasticity.
Applying the first difference estimator method yields several conveniences15:
• Country specific unobserved effects influencing TFP are eliminated • Possible unit roots are eliminated
• Correlation among cross-sectional units is eliminated
By clustering the standard errors per country for all panel regression models, the standard errors are corrected for serial correlation.
All regressions have been tested for multicollinearity and misspecification. The variance inflation factor, indicating multicollinearity, shows no values larger than 4.0. As a rule of thumb, the author applied an upper limit for the variance inflation factor of 516. The link test for model specification
shows no misspecification issues in any of the regressions17.
The cross-sectional regressions have been tested for homoskedasticity by a Breusch-Pagan test. Although the results did not reject the null hypothesis of homoskedasticity, these regressions have been run with heteroskedasticity robust standard errors since an analysis, taking into account the small sample size, of plotted versus predicted values does indicate that the regressions could be subject to heteroskedasticity.
5.3
Results
Three conclusions can be derived from the regression output.
• The short run elasticity between ICT investments on the growth rate of TFP is approximately -0,14 with a standard error of 0,05
• The elasticity of yearly ICT investments in the period 1980-1987 on yearly TFP increases is negative
The regression results are included in appendix E in the tables 4, 5, 6 and 7. In line with section 5.1, the (natural logarithm of) the growth rate of TFP is the dependent variable in all regressions. Since all input variables are natural logarithms, the coefficients need to be interpreted as elasticities.
Table 4 displays the results of running the baseline model with no trans-formation data and a transtrans-formation of all input variables (except the time
16The results of all tests available from the author upon request.
17More specifically, the coefficient of the variable of prediction is significant in most
regressions, while coefficient of the squared variable of prediction is not significant in any regression
dummies) and their 3, 5, and 10 year moving averages. The regressions unan-imously show a negative short run relationship between ICT investments and the growth rate of TFP. The three year moving average version of the baseline model suggests a short run elasticity of -0,14 with a standard error of 0,05, implying significant negative relationship at the 95% confidence level. The coefficients of the 5 and 10 year moving averages show a similar relationship at the 99% confidence level.
The results of the cross-sectional regression analyses are included in Table 5. Recalling that N=18, the results are subject to larger standard errors. Although table 5 does not provide strong evidence of a significant relationship between ICT investments and the growth rate of TFP in the period 198-2007, it does does show negative coefficient estimates the period 1980-1987. This result implies that ICT investments in the period 1980-1987 have a negative impact on productivity growth in the period 1998-2007.
The tables 6 and 7 show that increasing the length of the lag at which ICT investments enter the model increases its coefficient. Implying that the short run negative effect diminishes over the years. In addition to that, table 6 suggests that the relationship might even become positive if a 10 year lag is used. This result is similar to Edquist & Henrekson’s (2016) result of a positive association with a lag of 7 to 8 years. Possible presence of positive coefficients when k > 10 would be contradictory to the results of cross-sectional analysis and hence inconclusive.
Table 7 shows that the negative coefficients dominate the positive coeffi-cients of ICT investments on TFP growth within a range of 10 years. Hence, this section concludes that the impact of ICT investments on TFP within a 10 year period is negative, while it is about -0,14 in the initial year of the investments.
6
Discussion
Section 6.1 provides an overall conclusion on the effect of ICT investments on labour productivity growth of financial intermediaries. Section 6.2 mentions some of the limitations that the empirical analyses are subject to.
6.1
Conclusion
The thesis has examined the effect of ICT investments on labour produc-tivity growth, by analyzing the direct as well as the indirect effect of ICT investments. The direct effect has been analyzed by a growth accounting calculation, whereas the indirect effect has been examined by a regression analysis.
The growth accounting calculation suggests that the direct effect of ICT related capital deepening accounts for a yearly labour productivity growth of 1,44% (s.e: 0,15%). The regression analysis suggests that ICT investments are associated with a short run elasticity with respect to the growth rate of TFP of -0,14 (s.e: 0,05), implying a negative indirect effect. The opposite signs of the effects, i.e. positive versus negative, imply that effects work in opposite directions.
Considering the cross-country average yearly ICT investments as included in appendix D of 13,48%, with a standard error of 1,22% and applying this data to the elasticity of -0,14 yields a negative effect on the growth rate of TFP of -1,89%. Considering that the absolute value of this result exceeds the estimate of the direct effect (1,44%) suggests that the indirect effect out-weighs the direct effect. Testing the statistical significance of the difference between the estimates of the two analysis types is similar to comparing ap-ples to oranges. Hence, more research is needed to test whether the total effect is significantly smaller than zero.
Thus, the thesis does provide evidence that ICT investments affect the labour productivity growth of financial intermediaries negatively or non-significantly in the short run. This result is in accordance with Solow’s quip, as mentioned on page ii, that computers can be seen anywhere but in the productivity statistics.
6.2
Limitations
Most importantly, the results of both analyses are based on a growth account-ing methodology, includaccount-ing assumptions on imperfect competition, constant returns to scale and a simplified translog production function. Although the regression analysis is less sensitive to these assumptions than the growth ac-counting calculation, its dependent variable, the yearly growth rate of TFP, is calculated by the mentioned growth accounting assumptions.
Secondly, there might be unobserved factors like competition, banking crises and market sentiment that affect TFP besides investments in ICT and other capital, the number of hours worked and labour composition. The author noted that adding data on these variables drastically decreased the number of observations in the sample. In addition to that, year specific effects have been accounted for by running regressions with year dummies and the specification tests did not raise any concerns on misspecification. Therefore, this additional data is kept out of the regressions.
Third, Fayrer & Aiyar (2002) note that a regression of the growth rate of TFP on ICT investments might be subject to endogeneity18. As an
alterna-tive to using lagged ICT growth, they propose the use of a GMM estimator with all lags of the growth rate of TFP as an instrument for the growth rate of TFP and a similar procedure for ICT investments. However, combining the use of a GMM estimator is incompatible with taking moving averages of the data since this would imply that data enters regressions both as an in-strument and as an input component of the moving average. Since academia has provided a rationale for using moving averages in regression analyses on this subject matter, as explained in section 2), this thesis prefers to follow Edquist & Henrekson (2016) by using lags rather than a GMM estimator.
Edquist & Henrekson also disclose that they considered the use of other instrumental variables, but none of them fulfilled the requirements of both exogeneity and relevance. As a consequence, the author considered estima-tion methods using instrumental variables to be beyond the scope of this thesis.
18More specifically, with ICT investments affecting the growth rate of TFP, as shown
in this thesis for example on the one hand. And, on the other hand, standard economic theory, like Solow (1956), implies that the growth rate of TFP does affect investments. This mutually causal relationship rises concerns on simultaneity.
Appendices
A
Elasticity of substitution
The elasticity of substitution (ES) was introduced by Hicks (1932). By as-suming constant returns to scale and perfect competition it can be used to analyze the relative shares of factor payments. Note that the under perfect competition, factor prices are exogenous and producers optimize the quan-tities of each production factor Xφ such that the marginal product of each
production factor M Pφequals its marginal cost. Since these assumptions are
in line with the model developed in section 4.1, the elasticity of substitution can be used to explain the effect of changing factor prices on the share of capital in total output. The definition of the ES ∀φ 6= χ:
σφ,χ ≡
d ln(Xφ/Xχ)
d ln M RT Sχ,φ
(10)
The marginal rate of technical substitution (MRTS) between any two pro-duction factors φ and χ 6= φ equals the ratio of the marginal products and the quantity employed of each production factor is set such that it equals the ratio of the cost of the production factors:
M RT Sχ,φ≡
M Pχ
M Pφ
(11)
For the production function (12’’) derived in appendix B, the M RT S takes the following form:
M RT Sχ,φ=
υχXφ
υφXχ
(11’)
Hence, (10) can be rewritten:
σφ,χ =
d ln(υφ
υχM RT Sχ,φ)
d ln M RT Sχ,φ
(10’)
Using that d ln X = d(X)/X, we obtain:
σφ,χ= d(υφ υχM RT Sχ,φ) υφ υχM RT Sχ,φ M RT Sχ,φ dM RT Sχ,φ (10’’)
Assuming that υi is a constant yields: σφ,χ = υφ υχdM RT Sχ,φ υφ υχM RT Sχ,φ M RT Sχ,φ dM RT Sχ,φ = 1 (10’’’)
The methodology introduced in section 4.1 defines υφ≡ βPφXφ
Q and assumes
that M Pφ= βPφ. From this assumption follows that any change of a factor
price βPφ is accompanied by an equal change of the marginal product of
that production factor M Pφ. In the case of an increase of the factor price,
producers need to decrease the quantity of the production factor Xφ, such
that the marginal product equals its factor price again such that their profits are maximized.
Therefore, equation (10’’’) shows that assuming a Cobb-Douglas produc-tion funcproduc-tion is not suitable for analyzing the effect of ICT investments on labour productivity. That is, assuming a Cobb-Douglas function would imply that any change of ICT prices is accompanied by a proportional ICT invest-ment, keeping the share of ICT in total factor payments constant. A priori, this is an unrealistic assumption and line with that, table 2 of appendix C, shows that ICT’s share of total factor payments fluctuates within range of about 5 percentage points.
B
The translog production function
The translog (a shorthand notation for transcendental logarithmic) produc-tion funcproduc-tion was first proposed by Jorgenson & Christensen (1971, 1973) with the purpose of developing an aggregate production function with disag-gregated factor indexes and deal with problems related to additivity. Other production functions were unable to deal with these issues (Berndt & Chris-tensen, 1973). Moreover, the translog production function does not assume perfect competition and unit or constant elasticity of substitution of produc-tion factors as is the case with the Cobb Douglas and the CES producproduc-tion function, respectively.
The production function (3) in section 4.1 is a simplified form of the translog production function. It can be obtained by assuming constant re-turns to scale, Hicks-neutral technical change and non-negative synergy coef-ficients. Following the theory described in section 2 of (Berndt & Christensen, 1973) yields ∀φ 6= χ:
ln Qi,t = ln υ0,i,t+ υA,i,tln Ai,t+ n
X
φ=1
υφ,i,tln Xφ,i,t+ 12γAA,i,t(ln Ai,t)2
+12 n X φ=1 n X χ=1
γφ,χ,i,tln Xφ,i,tln Xχ,i,t+ n
X
φ=1
γφ,A,i,tln Ai,t (12)
Where equivalently to the model developed in section 4, Q is total output, A represents TFP and Xφ, i, t denotes the input quantities of all n production
factors of financial intermediaries in country i at time t. (The production function used in this thesis limits the number of production factor n to 3 i.e. ICT, other capital and labour) Assuming constant returns to scale yields:
ln Q(λXφ,i,t) = ln Q(Xφ,i,t) + ln λ (13)
Implying the following restrictions on (12):
n X φ=1 υφ,i,t= 1, n X φ=1 γφ,χ,i,t = 0, n X j=1 γφ,χ,i,t = 0 n X φ=1 n X χ=1 γφ,χ,i,t= 0, n X φ=1 γφ,A,i,t= 0 (14)
Assuming Hick-neutral technical change yields:
ln Qi,t(Xφ,i,t, Ai,t) = ln F (Xφ,i,t) + ln Ai,t (15)
Implying the following additional restrictions on (12):
υA,i,t = 1, γA,A,i,t= 0, γφ,A,i,t = 0, (16)
A reduced form of (12) follows from imposing the restrictions specified in (14) and (16): ln Qi,t = ln Ai,t+ n X φ=1 υφln Xφ,i,t+12 n X φ=1 n X χ=1
γφ,χ,i,tln Xφ,i,tln Xχ,i,t (12’)
Recall from (14) that Pn
φ=1
Pn
χ=1γφ,χ = 0. The economic interpretation of
this restriction is that the sum of the coefficients of the synergy effects equals 0. Assuming that a negative synergy γφ,χ < 0, combined with this restriction
yields: (12’) to: ln Qi,t = ln Ai,t+ n X φ=1 υφ,i,tln Xφ,i,t (12’’)
Taking the anti-log of (12’’’) shows that (12’’) is in fact a generalization of the Cobb-Douglas function with time varying shares of factor payments υφ,i,t:
Qi,t = Ai,t n Y i=1 Xυφ,i,t φ,i,t (12’’’)
Equivalently, subtracting period t − 1 fromt period t yields the general form of (3’’’) obtained in section 4.1. (12’) to:
∆ ln Qi,t = ∆ ln Ai,t+ n
X
φ=1
C
Growth accounting: U.S. example
Table 2 exemplifies the construction of table 1. For any variable Vi,t, it holds
that: ∆ ln Vi,t = ln Vi,t − ln Vi,t−1 = ln(Vi,t/Vi,t−1). This property implies
that the growth rate of any variable during the sample period can be ob-tained by summing all per year growth rates, such that: P2007
t=1980∆ ln Vi,t =
ln(Vi,2007/Vi,1980).
Table 2: Growth accounting calculation for U.S. financial intermediaries
Year ∆ ln Q υICT ∆ ln ICT υC ∆ ln C υL ∆ ln L ∆ ln A
1980 6,85% 0,10 30,36% 0,37 10,05% 0,53 2,76% -1,29% 1981 3,58% 0,09 26,67% 0,38 9,20% 0,53 5,62% -5,36% 1982 1,98% 0,09 23,97% 0,37 7,55% 0,54 3,51% -4,90% 1983 0,90% 0,10 28,65% 0,35 7,89% 0,55 1,44% -5,49% 1984 3,51% 0,11 32,46% 0,34 9,55% 0,54 4,95% -6,14% 1985 -0,02% 0,12 27,22% 0,33 10,32% 0,55 3,05% -8,31% 1986 2,99% 0,11 22,59% 0,32 10,43% 0,57 5,73% -6,14% 1987 6,65% 0,11 22,66% 0,30 8,00% 0,59 2,60% 0,15% 1988 3,65% 0,11 21,00% 0,29 7,57% 0,60 2,09% -2,15% 1989 1,92% 0,12 21,54% 0,30 7,22% 0,59 -0,10% -2,73% 1990 3,38% 0,13 14,64% 0,30 5,10% 0,57 2,30% -1,31% 1991 2,94% 0,13 13,48% 0,32 1,66% 0,55 -0,61% 1,03% 1992 0,47% 0,13 11,29% 0,32 1,35% 0,54 -0,30% -1,29% 1993 4,27% 0,13 12,98% 0,31 3,25% 0,56 4,83% -1,13% 1994 0,46% 0,13 13,84% 0,31 3,99% 0,56 2,36% -3,90% 1995 3,34% 0,13 13,90% 0,32 3,35% 0,55 0,37% 0,24% 1996 3,17% 0,14 18,81% 0,32 4,42% 0,55 1,09% -1,41% 1997 6,97% 0,14 19,69% 0,33 4,87% 0,54 3,44% 0,87% 1998 6,28% 0,14 24,95% 0,32 6,84% 0,54 5,63% -2,41% 1999 6,23% 0,13 24,91% 0,31 5,96% 0,56 3,15% -0,70% 2000 7,98% 0,12 17,68% 0,32 3,91% 0,56 3,72% 2,49% 2001 3,56% 0,12 11,26% 0,31 2,12% 0,57 -2,47% 2,98% 2002 1,70% 0,11 6,06% 0,32 -0,4% 0,57 0,15% 0,97% 2003 3,11% 0,11 7,62% 0,33 -1,12% 0,56 2,75% 1,14% 2004 0,34% 0,10 10,40% 0,32 1,20% 0,57 0,89% -1,64% 2005 6,96% 0,10 6,91% 0,32 0,91% 0,58 0,39% 5,75% 2006 6,51% 0,10 10,51% 0,33 1,07% 0,57 3,70% 3,01% 2007 0,14% 0,10 11,60% 0,33 1,09% 0,58 2,10% -2,55% Total 99,82% 0,11 507,65% 0,33 137,35% 0,56 65.14% -40,20%
Remarkably, despite ICT investments of 507,65% during the sample period, the ICT share of factor payments υICT at the end of the sample period equals
the share at the beginning of the sample period. Following the methodology of section 4, the sum of ICT investments and price decrease of ICT of the sample period must equal the growth rate of output since the equation:
υICT,i,t=
βPICT,i,tICTi,t
Qi,t
(4)
Implies that if upsilonICT,i,2007/upsilonICT,i,1979 = 1, equation (17) must
hold. Note that 1979 is included in period growth rate of any variable Vi,t in
1980 is defined as ∆ ln Vi,1980 = ln Vi,1980− ln Vi,1979.
∆ ln Qi,t = ∆ ln(βPICT,i,t) + ∆ ln ICTi,t (17)
Table 3 in appendix D shows that the price level of ICT used by financial in-termediaries in U.S. decreased with -290,38% during the sample period. From table 2 follows that the growth rate of ICT investments and the growth rate of output equals 507,65% and 99,82% respectively. The discrepancy between the theoretical price decrease 407,83%) and the actual price decrease (-290,38%) can be explained by the fact that for any production factor φ, the share of factor payments υφ,i,t is calculated as the two-period average of a
productions factor’s factor payments’ share of total factor payments, in order to make sure that the shares of factor payments add up to 1 (as explained in section 4.1). The discrepancy is an example of the empirical limitations of the grow accounting methodology that are described in section 6.2.
D
ICT investments and prices changes
Although the EU KLEMS database does not directly include price levels of production factors, changes of factor prices can be determined based on the EU KLEMS database, since factor payments and input quantities are given per production factor. Factor price calculations are then based on the following equation:
∆ ln βPICT,i,t= ∆ ln Factor PaymentsICT,i,t− ∆ ln ICTi,t (18)
The results of applying (18) to the dataset are presented in table 3. Table 3 indicates a cross-sectional average yearly price decrease of 3,63% with a standard error of 1,25%. From table 3 also follows that the cross-country average yearly growth rate of the ICT capital stock (i.e. the ICT investments) equals 13,84%, with a standard error of 1,22%
Table 3: ICT Prices and investments of financial intermediaries during the sample period
Full period Per year
Country Period ∆ ln ICT ∆ ln βPICT ∆ ln ICT ∆ ln βPICT
Austria 1981-2007 414,47% -181,33% 15,94% -6,97% Belgium 1981-2006 405,90% -223,34% 16,24% -8,93% Canada 1980-2004 481,39% -221,48% 20,06% -9,23% Czech 1996-2007 106,63% -43,29% 4,10% -1,67% Denmark 1981-2007 461,30% -187,69% 17,74% -7,22% Finland 1980-2007 317,16% 27,77% 11,75% 1,03% France 1981-2007 278,10% -70,15% 10,70% -2,70% Germany 1980-2007 287,20% -123,62% 10,64% -4,58% Hungary 1996-2007 85,93% -6,92% 7,81% -0,63% Ireland 1989-2007 188,18% 91,88% 10,45% 5,10% Italy 1980-2007 450,93% -213,67% 17,34% -8,22% Japan 1980-2006 109,36% 139,81% 4,37% 5,59% Netherlands 1980-2007 491,33% -242,39% 18,20% -8,98% Slovenia 1996-2006 116,04% 46,16% 11,60% 4,62% Spain 1981-2007 329,87% 8,82% 12,69% 0,34% Sweden 1994-2007 166,68% -85,82% 11,91% -6,13% U.K. 1980-2007 600,75% -160,23% 22,25% -5,93% U.S. 1980-2007 507,64% -290,38% 18,80% -10,75% mean 322,16% -96.44% 13,48% -3,63% s.e. 38,47% 29.95% 1,22% 1,25%
E
Regression output
Table 4: Regression results baseline model (7)
(1) (2) (3) (4) ∆lnICT -0,119* -0,144* -0,154** -0,150** (0,052) (0,053) (0,052) (0,048) ∆lnC -0,190** -0,183 -0,177 -0,158 (0,065) (0,092) (0,114) (0,137) ∆lnL -0,223 0,109 0,193 0,120 (0,216) (0,235) (0,296) 0,386
Time dummies Yes Yes Yes Yes
Adjusted R2 0,086 0,199 0,206 0,180
Obs. 410 415 416 416
Cluster robust standard errors in parentheses;∗ρ < 0, 05,∗∗ρ < 0, 01. Regression (1) shows the results of the panel regression without smoothing. Regression (2),(3) and (4) show the results of the baseline regression with smoothing by taking the 3,5 and 10 years movering average, respectely. The table shows evidence of an unambiguous significant negative short run elasticity between ICT investments and TFP.
Table 5: Regression results cross sectional regression model (8)
(5) (6) (7) (8) (9) (10) ∆lnICT1998−2007 0,077 0,033 0,177* (0,070) (0,083) (0,072) ∆lnICT1988−1997 0,225 0,194 0,313 0,248 (0,127) (0,153) (0,158) (0,141) ∆lnICT1980−1987 -0,386* -0,348* -0,528** (0,136) (0,143) (0,128) ∆lnC1998−2007 -0,011 -0,168 0,007 -0,137 -0,156 -0,047 (0,143) (0,125) (0,035) (0,159) (0,101) (0,084) ∆lnL1998−2007 0,241 0,378 1,278* 0,268 0,692 0,160 (0,358) (0,296) (0,554 (0,446) (0,101) (0,550) Time dummies No No No No No No Adjusted R2 0,089 0,159 0,161 0,108 0,211 0,507 Obs. 18 18 13 18 13 13
Table 6: Regression results smoothed lag model (9) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) (21) ∆lnICTsmoothed 0,020 0,031 0,038 0,046* 0,050* 0,049** 0,051** 0,053** 0,055** 0,057** 0,054** (0,029) (0,026) (0,023) (0,021) (0,019) (0,015) (0,012) (0,011) (0,013) (0,014) (0,014) ∆lnCsmoothed -0,296 -0,306** -0,313** -0,321** -0,319** -0,308** -0,306** -0,230** -0,299** -0,277** -0,267** (0,086) (0,091) (0,093) (0,093) (0,089) (0,080) (0,079) (0,074) (0,082) (0,078) (0,076) ∆lnLsmoothed 0,233 0,172 0,194 0,190 0,179 0,201 0,221 0,201 0,253 0,180 0,194 (0,257) (0,246) (0,227) (0,213) (0,194) (0,176) (0,181) (0,176) (0,220) (0,206) (0,215)
ICT lag (years) 0 1 2 3 4 5 6 7 8 9 10
Time dummies No No No No No No No No No No No
Adjusted R2 0,090 0,095 0,113 0,133 0,146 0,151 0,160 0,164 0,188 0,183 0,180
Obs. 415 403 386 369 352 335 318 301 284 267 250
Table 7: Regression results smoothed lag model (9) including time dummies
(22) (23) (24) (25) (26) (27) (28) (29) (30) (31) (32) ∆lnICTsmoothed -0,144* -0,131* -0,081 -0,048 -0,035 -0,046 -0,035 -0,008 0,037 0,068 0,088* (0,053) (0,060) (0,052) (0,043) (0,038) (0,043) (0,049) (0,052) (0,047) (0,045) (0,046) ∆lnCsmoothed -0,183 -0,184 -0,201 -0,223* -0,227* -0,219* -0,219* -0,204* -0,214* -0,187* -0,178** (0,092) (0,101) (0,100) (0,099) (0,095) (0,090) (0,088) (0,078) (0,071) (0,060) (0,054) ∆lnLsmoothed 0,109 0,008 0,023 0,028 0,027 0,043 0,048 0,007 0,052 -0,042 -0,036 (0,234) (0,242) (0,234) (0,236) (0,224) (0,211) (0,217) (0,196) (0,216) (0,021) (0,243)
ICT lag (years) 0 1 2 3 4 5 6 7 8 9 10
Time dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Adjusted R2 0,199 0,191 0,184 0,184 0,189 0,203 0,202 0,201 0,213 0,226 0,229
Obs. 415 403 386 369 352 335 318 301 284 267 250
Note for table 6 and 7: Cluster robust standard errors in parentheses;∗ρ < 0, 05,∗∗ρ < 0, 01. The relationship between table 7 and table 4 is demonstrated by including a zero year lag in the table. Regression (22) is equivalent to regression (2).
References
Arrow, K.J. Chenery, H.B. Minhas, B.S. Solow, R.M. (1961). Capital-Labor Substi-tution and Economic Efficiency, The Review of Economics and Statistics, Vol. 43, No. 3 (Aug. 1961), pp.225-250.
Acemoglu, D. Autor, D. Dorn, Hanson, D. Gordon, H. & Price, B. (2014). ”Return of the Solow Paradox? IT, Productivity, and Employment in U.S. Manu-facturing,” IZA Discussion Papers 7906, Institute for the Study of Labor (IZA).
van Ark, B. Inklaar, R.C. & McGuckin, R. (2002). “Changing gear”: productivity, ICT and services: Europe and the United States, GGDC Research Memo-randum 200260, Groningen Growth and Development Centre, University of Groningen.
Baily, M.N. & Gordon, R.J. (1988). The Productivity Slowdown, Measurement Is-sues, and the Explosion of Computer Power, Brookings Papers on Economic Activity, Economic Studies Program, The Brookings Institution, vol. 19(2), pages 347-432.
Batiz-Lazo, B. & Reid, R.J.K. (2008). Evidence from the Patent Record on the Development of Cash Dispensing Technology, MPRA Paper 9461, University Library of Munich, Germany.
Baumol, W. & Bowen, W. (1966). Performing Arts, The Economic Dilemma: a study of problems common to theater, opera, music, and dance, New York: Twentieth Century Fund.
Berger, A.N. & Humphrey, D.B. (1990). Measurement and efficiency issues in commercial banking, Finance and Economics Discussion Series 151, Board of Governors of the Federal Reserve System (U.S.).
Berndt, E. R. & Christensen, L. R. (1973). The translog function and the substi-tution of equipment, structures, and labor in U.S. manufacturing 1929-68, Journal of Econometrics, Elsevier, vol. 1(1), pages 81-113, March.
Boot, A.W.A. & Dijkstra, M.A. (2011). Financi¨ele Sector in Crisis, De economis-che toekomst van Nederland, Preadviezen van de Koninklijke Vereniging voor de Staathuishoudkunde 2011, SDU Uitgevers, Den Haag, 9 December 2011, pp. 101-141.
