Online Annexes: Chapter 2
- International Monetary Fund. Research Dept.
- Published Date:
- October 2018
World Economic and Financial Surveys
World Economic Outlook
Online Annexes: Chapter 2
International Monetary Fund
Annex 2.1. Data Sources and Country Coverage
All indicators and their respective data sources used in the chapter are listed in Annex Table 2.1.1.
|Gross Domestic Product, Constant Prices||IMF, World Economic Outlook database|
|Gross Capital Formation, Constant Prices||IMF, World Economic Outlook database|
|Final Consumption Expenditure, Constant Prices||IMF, World Economic Outlook database|
|Total Domestic Demand, Constant Prices||IMF, World Economic Outlook database|
|Population||IMF, World Economic Outlook database|
|Total Labor Force||IMF, World Economic Outlook database|
|Total Employment||IMF, World Economic Outlook database|
|Share of Population in 15–64||World Bank, World Development Indicators database|
|Capital Stock||Penn World Tables 9.0|
|Sectoral Real Value Added (Manufacturing, Services)||World Bank, World Development Indicators database|
|Banking Crisis||Laeven and Valencia (2013)|
|General Government Debt||IMF, Global Debt database; World Economic Outlook database|
|Share of Labor Compensation in GDP||Penn World Tables 9.0|
|Gini Coefficient||Standardized World Income Inequality Database|
|Research and Development Expenditure||World Bank, World Development Indicators database|
|Robot Stock and Shipment||International Federation of Robotics|
|Domestic Credit to Private Sector||World Bank, World Development Indicators database|
|Export Value of Goods (bilateral)||IMF, Direction of Trade Statistics database|
|Financial Openness||Lane and Milesi-Ferretti (2017)|
|Current Account Balance||IMF, World Economic Outlook database|
|Current Account Gap||Lee and others (2008)|
|Regulation of Employment Dismissal||Cambridge University’s Centre for Business Research|
|De Facto Peg Strength||Ghosh and others (2011)|
|General Government Structural Balance||IMF, World Economic Outlook database|
|Real Short-term Deposit Rate||IMF, World Economic Outlook database|
|Banking Regulation Index||Barth, Caprio, and Levine (2013)|
|Fraction of Bank Application Denied||Barth, Caprio, and Levine (2013)|
|Bank Concentration||Barth, Caprio, and Levine (2013)|
|Supervisory Power||Barth, Caprio, and Levine (2013)|
|Capital Regulation||Barth, Caprio, and Levine (2013)|
|Bilateral Cross-Border Bank Claims||Bank for International Settlements|
|Total Headline Support for Financial and Other Sectors||IMF, Fiscal Affairs Department (2009)|
|Capital Injections||IMF, Fiscal Affairs Department (2009)|
|Purchase of Assets and Lending by Treasury||IMF, Fiscal Affairs Department (2009)|
|Central Bank Support with Treasury Backing||IMF, Fiscal Affairs Department (2009)|
|Central Bank Liquidity Support||IMF, Fiscal Affairs Department (2009)|
|Guarantees (excl. Deposit Insurance)||IMF, Fiscal Affairs Department (2009)|
|Upfront Government Financing||IMF, Fiscal Affairs Department (2009)|
|Crisis-Related Discretionary Fiscal Stimulus||IMF, Fiscal Monitor (2010)|
|Active Labor Market Policy||OECD, Employment database|
|Employment Protection Legislation||OECD, Employment database|
|Inflow/Outflow Rate for Unemployment||OECD, Employment database|
|Labor Skills||World Input-Output Database|
|Labor Compensation||World Input-Output Database|
|Capital Compensation||World Input-Output Database|
|Sectoral Capital Stock||World Input-Output Database|
|Sectoral Price Levels||World Input-Output Database|
|Sectoral Employment Headcount||World Input-Output Database|
|Sectoral Employment Hours worked||World Input-Output Database|
The country coverage for the different sections is presented in Annex Table 2.1.2, there are considerable variations in the sample of countries included in the various analytical exercises due to data constraints.
|Albania*; Algeria*; Angola*; Antigua and Barbados*; Argentina*†; Armenia*; Australia†; Austria†; Azerbaijan*;|
|Bahamas, The*; Bahrain*; Bangladesh*•; Barbados*; Belarus*; Belgium†; Belize*; Benin*•; Bhutan*•; Bolivia*;|
|Bosnia and Herzegovina*; Botswana*; Brazil*†; Brunei Darussalam*; Bulgaria*†; Burkina Faso*•; Burundi*•;|
|Cabo Verde*; Cambodia*•; Cameroon*+; Canada†; Central African Republic*•; Chad*+; Chile*†; China*†;|
|Colombia*†; Comoros*•; Congo, Democratic Republic of the*•; Congo, Republic of*+; Costa Rica*†; Croatia*†;|
|Cyprus†; Czech Republic†; Côte d’Ivoire*+; Denmark†; Djibouti*•; Dominica*; Dominican Republic*; Ecuador*;|
|Egypt*; El Salvador*; Equatorial Guinea*; Eritrea*•; Estonia†; Ethiopia*•; Fiji*; Finland†; France†; Gabon*;|
|Gambia, The*•; Georgia*; Germany†; Ghana*+; Greece†; Grenada*; Guatemala*; Guinea*+; Guinea-Bissau*•;|
|Guyana*; Haiti*•; Honduras*+; Hong Kong SAR; Hungary*†; Iceland†; India*†; Indonesia*†; Iran*; Iraq*; Ireland†;|
|Israel; Italy†; Jamaica*; Japan†; Jordan*; Kazakhstan*; Kenya*•; Kiribati*•; Korea†; Kosovo*; Kuwait*; Kyrgyz|
|Republic*•; Lao P.D.R.*•; Latvia†; Lebanon*; Lesotho*•; Liberia*•; Libya*; Lithuania†; Luxembourg†; Macao|
|SAR; Macedonia, FYR*; Madagascar*•; Malawi*•; Malaysia*; Maldives*; Mali*•; Malta†; Marshall Islands*;|
|Mauritania*+; Mauritius*; Mexico*†; Micronesia*; Moldova*•; Mongolia*; Montenegro, Rep. of*; Morocco*;|
|Mozambique*+; Myanmar*+; Namibia*; Nepal*•; Netherlands†; New Zealand†; Nicaragua*+; Niger*+; Nigeria*+;|
|Norway^ Oman*; Pakistan*; Palau*; Panama*; Papua New Guinea*+; Paraguay*; Peru*†; Philippines*; Poland*†;|
|Portugal†; Puerto Rico; Qatar*; Romania*†; Russia*†; Rwanda*•; Samoa*; San Marino; Saudi Arabia*^|
|Senegal*•; Serbia*; Seychelles*; Sierra Leone*•; Singapore; Slovak Republic†; Slovenia†; Solomon Islands*•;|
|South Africa*†; Spain†; Sri Lanka*; St. Kitts and Nevis*; St. Lucia*; St. Vincent and the Grenadines*; Sudan*+;|
|Suriname*; Swaziland*; Sweden†; Switzerland†; Sao Tome and Principe*•; Taiwan Province; Tajikistan*+;|
|Tanzania*•; Thailand*; Timor-Leste*•; Togo*•; Tonga*; Trinidad and Tobago*; Tunisia*; Turkey*†;|
|Turkmenistan*; Tuvalu*; Uganda*•; Ukraine*; United Arab Emirates*; United Kingdom†; United States†;|
|Uruguay*; Uzbekistan*•; Vanuatu*; Venezuela*; Vietnam*•; Yemen*+; Zambia*+; Zimbabwe*•.|
Annex 2.2. Additional Details on Quantifying Post-Crisis Deviations in Activity from Pre-Crisis Trends
This annex provides additional details on the analysis shown in the section “Quantifying post-crisis deviations in activity from precrisis trends.”
A. Definition of Banking Crises
Annex Table 2.2.1 lists the banking crises used in the analysis. The definition of a banking crisis is from Laeven and Valencia (2013). It is based on two criteria: significant financial distress (including bank runs and liquidations) and significant government intervention in the banking system (including recapitalization, liability guarantees, and nationalization). The sample includes all banking crises that started between 2007–08.
|Country||Start of Crisis|
B. Definitions of Main Data Categories
Deviations from Pre-Crisis Trends
Deviations of GDP and other variables trending from the pre-crisis trend are calculated as follows:
- First, the transitory pre-crisis components are removed by means of low pass filters.1 While no method of removing transitory components can accommodate the specificities of every country in the sample, the filtering approach by Gourinchas and Obstfeld (2012), where the two-sided Hodrick-Prescott (HP) lowpass filter is used to eliminate transitory components, offers a general method of isolating low frequency (log) GDP movements from the data.2 The smoothing parameter is set at a higher value (100) than in standard business cycle detrending (6.25 with annual data). With the higher parameter, the estimated trend is less sensitive to short-run business cycle fluctuations and filters out relatively more medium-term influences, such as those of credit cycles.3Annex Figure 2.2.1 shows how removing transitory components affects estimation of the pre-crisis trend in the cases of the US. The MV filter yields estimates of output deviations that are in agreement with those obtained by applying the HP filter as described above.
- The underlying filtered time series run from 1995 to 2017. While GDP series could have experienced a structural break at the time of the GFC, an analysis shows that the estimated deviations are robust to the presence of a structural break.4Annex Figure 2.2.1 shows the relationships between 2011–13 and 2015–17 GDP deviations estimated with and without allowing for a post-GFC structural break. The closeness of both sets of estimated deviations demonstrates the robustness of estimated GDP deviations to the presence of a structural break.
- Second, the trend of the filtered series is calculated over 2000–08. The 2000–08 period is chosen because it is long enough to minimize the influence of shocks in individual years.5
- Finally, the deviations of post-crisis GDP from its pre-crisis trend are calculated as the average differences for 2011–13 and 2015–17.
Annex Figure 2.2.1.Estimates of Precrisis Trends for the United States
Source: IMF staff calculations.
Note: 2008 log GDP normalized to zero. Trend log GDP denotes extrapolated trend of potential GDP during 2000–08. HP = Hodrick-Prescott; MV = multivariate. Potential GDP estimated with the HP filter, lambda = 100. MV filter regressors are headline consumer price index, housing price index, stock prices, credit growth, and capacity utilization.
Annex Figure 2.2.2.Structural Break
Source: IMF staff calculations.
Note: Distribution of average percent deviations from precrisis trend.
Deviations of GDP per Worker
Annex Figure 2.2.3 presents the distributions of deviations of 2015–17 deviations of GDP per worker (i.e. labor productivity). Most countries in the banking crisis group experienced negative deviations in labor productivity, with few countries situated to the right of vertical axis. The distribution of deviations in the non-crisis group, while still centered below zero, is considerably more symmetric with a higher mean.
Comparing GDP Deviations with Previous Recessions
Annex Figure 2.2.4 compares the aftermaths of the 2008 and 1982 global recessions. While in the shorter run, both recessions induced similar deviations from the pre-crisis trends, the 2008 impact of the 2008 recession has been felt much longer. In addition, the 2008 recession affected a larger share of global output, as seen by comparing the distributions of weighted and unweighted output deviations.
Employment deviations are calculated using the approach by Schanzenbach and others (2017) who track the evolution of the employment ratio and compare it to the “benchmark” value from 2007 as follows:
While Schanzenbach and others (2017) estimate employment deviations only for the US, the Chapter extends their analysis to 102 countries.
Deviations of Total Factor Productivity
Post-crisis deviations of total factor productivity (TFP) from its pre-crisis trend are calculated using the standard Cobb-Douglas production function for output per worker and comparing the observed post-crisis values in labor productivity and output per worker with their pre-crisis trends—starred variables in the following equation:
Annex Figure 2.2.3.Postcrisis Output per Worker Deviation from Precrisis Trend, 2015–17 Annex Figure 2.2.4.Distributions of GDP Deviations after Recessions
Source: IMF staff calculations.
Note: Distribution of GDP percent deviations from precrisis trend. Country weights proportional to purchasing power parity GDP. Short-term = 3–5 years after the recession. Medium-term = 7–9 years after the recession.
Sectoral Capital Stock
As seen in Annex Figure 2.2.5, capital shortfalls are more widespread than just in construction. A broad sample of 38 advanced economies and emerging markets reveals slower average growth rates across many sectors. One exception is the mining and quarrying sector, in part influenced by the continued strength in commodity prices during the early part of the global downturn.
Construction of Explanatory Variables
Explanatory variables, used in the regression exercises described below, are constructed as follows:
- First, all explanatory variables are averaged over the period 2005-08 to attenuate the effect of idiosyncratic shocks.
- Second, all regressors (except for the banking crisis dummy) are standardized to have zero means and standard deviations of unity.
- Finally, the regressors are winsorized to alleviate influence of outliers.6
Tests of Equality of Distributions
Figures 2.3, 2.6 and 2.7 show the distributions of deviations of output, capital stock and total factor productivity respectively. The results of statistical tests of equality of these distributions between countries with and without banking crisis are presented in Annex Table 2.2.2. The table shows the rejection of the null hypothesis of equality of distributions in the cases of output and total factor productivity deviations. However, the distributions of capital stock deviations were not found to be significantly different between the crisis and non-crisis countries.
Annex Figure 2.2.5.Change in Postcrisis and Precrisis Growth Rates in Sectoral Capital Stock
Sources: World Input-Output Database; and IMF staff calculations.
Note: The bars depict the difference in average growth rates between 2011–14 and 2000–07.
|Average Percentile||Expected Percentile||P-Value|
|Total Factor Productivity||41.5||50.5||0.079|
Probability of a Banking Crisis
The probability of a banking crisis occurring in 2007–08 is given by the following qualitative response model:
where regulation is a measure of various aspects of banking regulation and θ is the set of parameters to be estimated. The index of banking regulation is drawn from Barth, Caprio, and Levine (2013). Results in Annex Table 2.2.3 show that the strength of restriction on banking activities (specifically, stronger restrictions on banks’ ability to underwrite, broker, and deal in securities; offer mutual fund products; and engage in insurance underwriting, real estate investment, development, and management) in 2006 is associated with a lower probability of the occurrence of banking crisis in 2007–08 and the coefficient is statistically significant. To test the robustness of the relationship, Annex Table 2.2.4 includes additional influences on the probability of occurrence of banking crisis in 2007–08. The strength of restriction on banking activities remains significant once the additional influences are controlled for.
|Strength of Restrictions on Banking Activities||–0.72 ***||–1.27 ***||–0.18 ***|
|Constant||–1.04 ***||–1.79 ***||0.19 ***|
GDP Deviations: Banking Crisis, Vulnerabilities, Policies and Economic Structure
Annex Table 2.2.5 presents an analysis of factors that determine the deviation of GDP during 2011–13. The analysis considers three sources of variation: vulnerabilities (the first two columns), economic structure and policies. Results show that the occurrence of banking crisis has a significant negative effect on GDP, underscoring the importance of sound banking regulation. Countries whose pre-recession credit growth was relatively more rapid suffered comparatively more damages. The analysis using the pre-crisis CA gap (based on Lee and others 2008) as an explanatory variable shows that excess external imbalances constituted an important vulnerability that was associated with larger post-crisis GDP losses.
Annex Table 2.2.6 presents an analysis of factors driving the deviations in investment during 2011–13. The important finding is that demand exposure to advanced economies weighs on investment even in countries without banking crises, illustrating the importance of the trade channel for investment.
|Strength of Restrictions on Banking Activities||–0.72 ***||–0.71 ***||–0.61 ***||–0.60 ***||–0.60 **||–0.65 **||–0.43||–0.46 *|
|Fraction of Bank Application Denied||–1.50 ***||–1.55 ***||–1.60 *||–1.07 ***||–1.13 ***||–1.32 **||–0.90 **||–1.25 *|
|Share of Interest Borrowing from G5||–0.01||0.27||0.35|
|Financial Openness||1.21 **||2.48 **||1.94 *|
|Demand Exposure to Advanced Economies||3.13 **||4.89 **||3.36|
|Constant||–1.04 ***||–114 ***||–0.74 ***||–0.83 ***||–0.88 ***||–1.01 ***||–0.98 ***||–0.90 ***||–1.15 ***||–1.17 **||–0.81 ***||–1.11 ***||–1.20 ***||–0.87 ***||–1.10 ***|
|Banking Crisis in 2007–08||–4.32 **||–2.01||–6.53 ***||–4.21 **|
|Domestic Credit Growth||–2.70 **||–5.37 ***|
|Demand Exposure to Advanced Economies||–13.35 ***||–6.19|
|Demand Exposure to China||1.07||3.04|
|Financial Openness||–3.35 *||–3.04|
|Precrisis GDP Growth||–0.55||3.31 ***||–0.57||–0.94|
|CA Gap||2.10 ***|
|Share of Manufacturing in GDP||0.15|
|Difficulty of Dismissal||–1.56 **|
|Precrisis GG Debt Change||–8.33 ***|
|De Facto Peg Dummy||–1.79 **|
|Constant||–3.49 ***||–4.04 ***||–2.00 **||–0.95|
|All Countries||Countries without Banking|
Crisis in 2007–08
|Banking Crisis in 2007–08||–11.59 ***||–3.52|
|Domestic Credit Growth||–6.81 **||–12.05 ***||–6.04 *||–8.31|
|Demand Exposure to Advanced Economies||–24.81 *||–14.94||–25.17 *||–19.91|
|Demand Exposure to China||3.87||20.80 ***||4.19||22.70 ***|
|CA Balance||5.46 **||4.43 *|
|Precrisis GDP Growth||–5.17*||6.88 **||–5.36 *||7.48 **|
|CA Gap||10.62 ***||12.79 ***|
|Constant||–9.16 ***||–9.22 ***||–9.37 ***||–9.59 ***|
GDP Deviations: AEs versus EMDEs, Vulnerabilities, Policies and Economic Structure
The differences in the post-recession GDP deviations between advanced and emerging markets (EMs) are presented in Annex Table 2.2.7. As above, rapid credit growth is a robust predictor of more negative short-run GDP deviations for both AEs and EMs. The two groups of countries differ in their responses to the pre-crisis current account (CA) balance, labor market flexibility, demand exposure to AEs and exposure to global financial markets.1
An analysis of the effectiveness of post-recession policies is presented in Annex Table 2.2.8. GDP deviations during 2015–17 are regressed on different policy variables as well as on the GDP deviation during 2011–13. The latter variable controls for the strength of the initial crisis shock. The analysis shows that total post-recession fiscal support was effective in reducing the post-recession decline in GDP. In addition, capital injections and guarantees are also found to be significant and effective policy measures.
Constructing Measures of Labor Market Churn
The chapter follows the methodology of Elsby, Hobijn, and Sahin (2012) to estimate the parameters that characterize labor market dynamism as follows. The law of motion for the rate of unemployment utis represented by
where st is the monthly rate of inflow into unemployment and ft is the monthly rate of outflow from unemployment. For reasons of data availability, the “continuous” time equation (2.4) is mapped into one at annual frequencies as shown in equations (2.5–2.6). The flow-steady rate of unemployment u*t is given by
If flow hazards are constant within a year, the law of motion for the rate of unemployment becomes
where λt is the annual rate of convergence to the steady state
Using the expression for probability that an unemployed worker exits unemployment within d months the expressions (2.6–2.7) are inverted to back out the annual estimates of ft and st.
|Domestic Credit Growth||–4.96 **||–4.79 ***||–5.43 ***||–5.71 **|
|Demand Exposure to Advanced Economies||9.01||–7.20||4.40||–8.27|
|Demand Exposure to China||3.88||7.33 **||6.56 *||4.98|
|CA Balance||4.03 ***||–0.42|
|Precrisis GDP Growth||–1.64||–0.27||2.77 *||3.11||–3.17||–1.55||–2.22||–0.45|
|CA Gap||2.49 ***||1.23|
|Share of Manufacturing in GDP||3.18||0.34|
|Difficulty of Dismissal||–1.72 *||–2.27 **|
|Precrisis GG Debt Change||–11.85 ***||–10.27 ***|
|De Facto Peg Dummy||–2.50 ***||–1.27|
|Constant||–6.99 ***||–4.46 ***||–6.28 ***||–6.91||–8.84 ***||–1.07||–2.58||–0.42|
|Total Headline Support for Financial and Other Sectors||0.20 **|
|Capital Injections||1.90 *|
|Purchase of Assets and Lending by Treasury||0.21|
|Central Bank Support with Treasury Backing||–14.35|
|Central Bank Liquidity Support||–0.25|
|Guarantees (excluding Deposit Insurance)||0.24 *|
|Upfront Government Financing||0.31|
|Crisis-Related Discretionary Fiscal Stimulus||–0.78|
|Banking Crisis in 2007–08||–0.17||–1.74||2.88||3.54*||3.06||–1.35||1.71||2.25|
|GDP Deviation 2011–13||1.12 ***||1.05 ***||1.10 ***||1.08 ***||1.10 ***||1.06 ***||1.09 ***||1.33 ***|
|Constant||–5.95 ***||–5.08 ***||–4.79 **||–4.04 **||–2.04||–5.12 **||–4.72 **||–1.33|
Annex 2.3. Robot Diffusion and its Employment Impact in the Aftermath of the Crisis
This annex provides additional details on the analyses shown in Figure 2.10 in the main text on technology adoption and for Box 2.2 on the impact of robot diffusion on employment. Annex Table 2.3.1 presents the sectors included in the analysis on robots.
|Sector Name||WIOD Sectors Included (ISIC Revision 4)||IFR Sectors Included|
|Agriculture||Agriculture, hunting, forestry and fishing||Agriculture, forestry, fishing|
|Mining||Mining and quarrying||Mining and quarrying|
|Food and Beverages||Food, beverages, and tobacco||Food and beverages|
|Metal||Basic metals and fabricated metal; Machinery, not elsewhere classified||Metal|
|Electronics||Electrical and optical equipment||Electrical/electronics|
|Glass and Ceramics||Other non-metallic mineral||Glass, ceramics, stone, mineral products|
|Paper and Printing||Pulp, paper, paper, printing and publishing||Paper|
|Utilities||Electricity, gas, and water supply||Electricity, gas, water supply|
|Education, Research, and Development||Education, scientific research and development||Education/research/development|
|Plastic and Chemicals||Chemicals and chemical products; rubber and plastics||Plastic and chemical products|
|Textiles and Leather||Textiles, wearing apparel and leather products||Textiles|
|Wood and Furniture||Wood and products of wood and cork||Wood and furniture|
A. Additional Details on Data
The main data on robots come from the International Federation of Robotics (IFR), which compiles information on worldwide shipment and stock of industrial robots from national federations of robot manufacturers, consisting of nearly all industrial robot suppliers worldwide (IFR, 2017). An industrial robot as defined by the International Organization for Standardization (ISO) is an “automatically controlled, reprogrammable, multipurpose manipulator, programmable in three or more axes, which can be either fixed in place or mobile for use in industrial automation application.” This definition limits the set of industrial robots and excludes dedicated industrial robots that serve one purpose.1 Industrial robot data are broken up by destination country, year, industry and technological application. In the 2017 data publication, IFR provides industrial robot shipment and stock data on 75 countries, although the industry level coverage starts at a much later year than the country aggregate data coverage.2 The industry level data is broken down by industrial branches in accordance with the International Standard Industrial Classification of All Economic Activities (ISIC) revision 4 for 2010 and revisions 2 or 3 in earlier years.
Service robots are defined as robots that perform useful tasks for humans or equipment excluding industrial automation application. Services robots can be divided further into personal service robots that are used for non-commercial tasks, usually by lay persons— for instance, domestic service robot, automated wheelchair, and personal mobility assist robot —or professional service robots, used for commercial tasks, usually operated by a properly trained operator, examples of the latter are cleaning robot for public places, delivery robot in offices or hospitals, firefighting robot among others.
The data coverage on service robots is sparse and is limited to only the supplier region. Service robot data is compiled based on market surveys that IFR sends to companies worldwide. As of 2016, the list of service robot suppliers has been expanded to more than 700 companies. Despite improvement in the response rate over the years, IFR urges that “the data reported still underestimate the true sales figures and installed base of service robots. They should therefore be considered a minimum level of the installed base of service robots” (IFR, 2017).3 Moreover, service robot data are only available at the regional level where they are produced (Europe, Americas, and Asia/Australia) and by application. Because of this data limitation in service robots, it is not possible to infer relationship between service robot diffusion and other economic variables in the country and industry where the service robot is deployed. Hence, only descriptive statistics about service robots are provided, whereas the regression analysis is conducted solely based on industrial robots.
Annex Figures 2.3.1 and 2.3.2 provide the breakdown of professional and personal service robots, respectively. Over half of the surveyed service robot manufacturers are in the Americas for both professional and personal use. The most prevalent applications of professional service robots are in logistics and defense, whereas personal service robots are mainly employed to carry out tasks in domestic households. The number of service robots has increased five-fold for professional services robots since 2010, starting from a relatively small base, whereas for personal services robots, the number of units have increased three-fold since 2010, starting from a much bigger base compared to professional service robots.
Annex Figure 2.3.1.Sales of Robots for Professional Services
Sources: International Federation of Robotics; and IMF staff calculations.
Note: Data on service robots is only available by origin regions for Asia/Australia, Americas, and Europe.
Annex Figure 2.3.2.Sales of Robots for Domestic/Personal Services
Sources: International Federation of Robotics; and IMF staff calculations.
Note: Data on service robots is only available by origin region for Asia/Australia, Americas, and Europe.
To assess the employment effects of automation by industrial robots at the sectoral level, the industrial robot data is merged with data from the World Input-Output Database (WIOD) by country and sector. The 2016 WIOD (Timmer and others, 2015) provides information on labor (hours worked, number of employees, and labor compensation) between 2000 and 2014 for 43 countries (representing more than 85 percent of world GDP) and 56 industries at the 2-digit ISIC revision 4 level. Data on labor skills (three different skills based on highest level of education obtained) come from the 2013 version of the WIOD’s Socio Economic Account that provides data on number of hours worked by low, medium and high skilled workers as well as their respective share of overall labor compensation for the period 1995–2009.4 The combined IFR and WIOD data covers 38 countries (Annex Table 2.1.2, countries in italics) and 14 industries (Annex Table 2.3.1).
To provide a meaningful comparison of industrial robot usage across countries and industries, it is important to account for the differences in sizes of industries in various countries. Robot density in industry i and country j in year t is defined as the number of multipurpose industrial robot shipment per thousand hours worked by persons employed in industry i,5 i.e.,
In contrast to IFR and Graetz and Michaels (forthcoming) who define robot density as the stock of robots per worker and stock of robots per million hours worked, respectively, the definition of robot density in this chapter uses robot shipment in the numerator rather than robot stock. Thus, it is a flow variable rather than a stock variable and can be interpreted as the rate of change in robot usage per thousand hours worked in a given year. The definition in this chapter is closer in spirit to that of Acemoglu and Restrepo (2017) who define exposure to industrial robots as the difference in stock of industrial robots for industry i for two given time periods divided by number of workers in the same industry. The main reason to employ industrial robot shipment rather than stock data is because the former is more accurate, especially in later years (IFR, 2017 Introduction, p28).6
B. Assessing the Role of Crisis Exposure
Crisis Exposure and Robot Penetration
The underlying test of medians in Figure 2.10 is based on a simple quantile regression of robot density on a high loss dummy for the sample of AEs. The results are displayed in Annex Table 2.3.2 and show that the industries in AE countries with higher output deviations in post-crisis periods tend to experience lower robot diffusion compared to those in AE countries with lower post-crisis output deviations.
|High Output Loss||–0.016 *|
Moreover, this result is not an artefact of convergence as shown in Annex Figure 2.3.3, regardless of initial levels of robot stock, countries are increasing robot diffusion.
A more rigorous method to estimate the impact of crisis exposure on robot diffusion is to employ a difference-in-differences (diff-in-diff) specification at the industry-country level. The first difference exploits variation at the industry-country (i,j) level in the difference in average changes in robot density for the post-crisis period (2010–14) relative to pre-crisis (2005–08), and the second difference assesses whether automation via robots advanced at a different pace for industries located in countries with high post-crisis activity deviations relative to pre-crisis. The specification is as follows:
Annex Figure 2.3.3.Average Change in Robot Density (2010–14) and Initial Robot Stock in 2010
Sources: International Federation of Robotics; and IMF staff calculations.
Note: Data labels in the figure use International Organization for Standardization (ISO) country codes.
Annex Figure 2.3.4.Effect of Crisis Exposure on Robot Diffusion
Sources: International Federation of Robotics; World Input-Output Database; and IMF staff calculations.
Note: Robot density is defined as robot shipment/1,000 hours worked. Error bars around coefficient estimate are two standard errors. Losses are based on calculations Annex 2.2.B. AEs = advanced economies; EMs = emerging markets; TFP = total factor productivity.
* p < .10, ** p < .05, *** p < .01.
|Output Loss||Employment Loss||Investment Loss||TFP Loss|
|High Loss||–0.010||–0.018||0.002||–0.003||–0.007||0.003||–0.023||–0.032 *||–0.001||–0.020||–0.032 *||0.004|
|Country Fixed Effect||No||No||No||No||No||No||No||No||No||No||No||No|
|Industry Fixed Effect||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes|
C. Sectoral Analysis
Effect of Robot Penetration on Employment
The analysis on the impact of change in robot density on employment follows the main methodology used in Graetz and Michaels (forthcoming) and Acemoglu and Restrepo (2017) and employs an ordinary least squares estimation that relate the average employment growth for the post-crisis period (2010–14) to the average change in robot density over the same period while controlling for country-industry specific characteristics and industry and country fixed effects. The specification is as follows:
where Avg %ΔEMPLij is the average percentage change in employment growth in industry i and country j; Avg ΔRobot Densityij is average year-on-year change in robot density; the country-industry specific controls include the 2010 levels of industry average wage and the capital-labor ratio; δi controls for industry fixed effects, and μj controls for country fixed effects. The regression is further weighted using industries’ 2010 share of workers within each country. The results reported in Box 2 (Figure 2.2.1) are those for high output loss sample as well as for the sample of AEs with high post-crisis losses in Annex Table 2.3.4 columns (2) and (4), respectively. The results based on investment and TFP deviations are similar to those using output deviations, but estimation results obtained from the employment deviations exercise are not significant and much smaller in magnitude.
|Advanced Economies||Emerging Markets|
|Full Sample||High Output|
|Average Δ Robot Density2010–14||–0.010||–0.021 ***||0.006||–0.021 ***||0.008||–0.117||0.457 **|
|Capital to Labor Ratio2010 (log)||0.010 ***||0.007||0.008 **||0.002||0.003||–0.004||0.013|
|Wage2010 (log)||0.013 *||0.047 ***||0.003||0.040 **||–0.011||0.050 *||0.012|
|Constant||–0.049 **||–0.145 ***||–0.029||–0.138 **||0.031||–0.125 ***||0.068 ***|
|Country Fixed Effect||Yes||Yes||Yes||Yes||Yes||Yes||Yes|
|Industry Fixed Effect||Yes||Yes||Yes||Yes||Yes||Yes||Yes|
Hollowing Out of The Employment-Skills Distribution
To determine the hollowing out effects documented by Autor, Levy, and Murnane (2003) and Goos, Manning, and Salomons (2014) across a larger sample of economies in the post-crisis period, the regression analysis in equation (4) is conducted exclusively on high loss countries (above median-loss) and further divided into subsamples based on an industry’s share of medium skilled workers (above- or below-median). The measure of medium skills is based on level of educational attainment of the worker compiled by WIOD. Three types of labor skills are distinguished. To classify as medium skilled labor, WIOD uses the 1997 International Standard Classification of Education (ISCED) classification of the worker having attained (Upper) secondary and/or post-secondary non-tertiary education (WIOD 2013). The main analysis uses the latest available data for the medium skill labor share in 2009. The result reported in Box 2 Figure 2.2.2 is based on the regression results reported in Annex Table 2.3.5, in column (1) and (3). In industries in high output loss countries that have relatively higher share of medium skilled workers, the relationship between robot penetration and employment is negative, this result is driven mainly by the AEs. This negative relationship also holds for industries in countries with above-median employment and investment losses.
|Advanced Economies||Emerging Markets|
|High medium||Low medium||High medium||Low medium||High medium||Low medium|
|Average Δ Robot Density2010–14||–0.019 **||–0.026||–0.016 **||–0.011||0.149||–0.762|
|Capital to Labor Ratio2010 (log)||0.007||0.010||–0.004||0.006||0.014||–0.004|
|Wage2010 (log)||0.039 ***||0.005||0.010||0.005||0.039||0.128|
|Country Fixed Effect||Yes||Yes||Yes||Yes||Yes||Yes|
|Industry Fixed Effect||Yes||Yes||Yes||Yes||Yes||Yes|
Labor Market Policies
To explore whether the impacts of robot penetration on employment differ across countries based on the extent of labor market policies that affect labor market flexibility and resilience, regression analysis is conducted on samples divided by high versus low losses and by different labor market policy measures. There are four specific measures of labor market policy under consideration:
- a) above- (high) and below-median (low) active labor market policy (ALMP) spending as share of GDP (2000–05 average)
- b) ease of dismissal index (CBR dismissal) as measured by the University of Cambridge’s Leximetric datasets (2000–05 average),8 with above-median (high) dismissal index indicating more stringent regulations for worker dismissal
- c) labor churn rate as calculated in Annex 2.2 (2005–08 average) that proxies the degree of initial flexibility in labor markets (measured by pre-crisis exit from and entry into unemployment) with above-median (high) rates indicating a more flexible labor market
- d) employment protection legislation (EPL) index compiled by the OECD (for 2008), with above-median (high) index pointing to more protective labor market policies towards workers
See Chapter for the list of references.
|High Output Loss||Low Output Loss|
|More Flexible Labor Market||Less Flexible Labor Market||More Flexible Labor Market||Less Flexible Labor Market|
|ALMP||Job Churn||Dismissal||EPL||ALMP||Job Churn||Dismissal||EPL||ALMP||Job Churn||Dismissal||EPL||ALMP||Job Churn||Dismissal||EPL|
|Average Δ Robot Density2010–14||–0.021||–0.032||–0.005||–0.012||–0.018 ***||–0.040 **||–0.018 ***||–0.024 ***||0.002||–0.010||0.004||0.011||0.016||0.005||0.008||–0.013|
|Capital to Labor Ratio2010 (log)||–0.003||0.006||–0.009||0.020 ***||0.004||0.013||0.012 **||0.007||–0.002||–0.010 ***||0.009 **||0.007||0.013||* 0.004||0.007||0.011|
|Wage2010 (log)||–0.006||0.050||0.017||0.021 *||0.029||0.070 **||0.047 ***||0.057 ***||0.018||0.031 **||0.004||–0.001||–0.004||–0.008||0.002||0.011|
|Constant||–0.015||–0.075||0.000||–0.055||–0.090||–0.247 **||–0.179 ***||–0.108 ***||–0.054||–0.112 **||–0.032||–0.018||0.000||0.016||0.066 ***||–0.059|
|Country Fixed Effect||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes|
|Industry Fixed Effect||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes||Yes|
An alternative approach is to fit a linear trend to the log-GDP series that has been truncated a few years before the peak of the cycle. This approach produces estimates that are highly sensitive to the length of the truncation period. Furthermore, its cutoff frequency cannot be controlled. Hence it is not used in the chapter.
Estimating the trends with a multivariate filter (as in Berger and others 2015) that accounts for macrofinancial imbalances could in principle provide more accurate estimates of underlying trend. In practice, the HP filter with the smoothing parameter set to 100 works equally well. In addition, the limited availability of data on asset prices precludes a wide application of multivariate filtering.
The “standard” value λ=6.25 has the cutoff frequency of only 8 years.
The structural break is modelled as 5-σ shock to potential GDP in 2009, calibrated to correspond to the 5-σ shock to headline real GDP.
In the case of the US, Fernald (2015) shows that labor productivity accelerated in the 1990s and that it returned to its long-run trend of approximately 1.5 percent per annum around 2003—well before the 2008 recession. For this reason, calculating post-GFC losses based on periods of faster productivity growth before 2000 could overstate post-GFC output losses. In this chapter’s analysis, the trend growth of US labor productivity, calculated as described above, amounts to 1.54 percent per annum. This estimate is in close agreement with the estimate by Fernald.
The analysis omits countries with large output deviations that were caused by war or political strife.
Before 2001, Japan’s data included both multipurpose industrial robots and dedicated industrial robots (e.g. equipment dedicated for loading/unloading machine tools, assembly on printed circuit boards, storage and retrieval systems, etc.), while other countries, in principle, have only reported data on multipurpose industrial robots. As of 2001, dedicated robots are excluded from the flow statistics. The operational stock data, however, continues to include a fairly large share of dedicated robots. Statistics on new installations and flows from 2001 onward are internationally comparable with Europe and North America.
The earliest available data at the industry level starts in 1993 but limited to nine countries, the coverage extends to 38 countries in 2005.
The amount of sales information available also differs significantly between various application areas.
Timmer and others (2015) provide more details about the construction of the database and discuss additional features.
The use of hours instead of number of workers is preferred as workers can differ in the number of hours that they work.
While Graetz and Michaels (forthcoming) construct their robot stock data using the perpetual inventory method, the choice of appropriate depreciation rate is not clear.
The change in industrial robot density can be thought of as the change in the rate of change of industrial robots per million hours worked.
This index is based on nine detailed sub-categories that encompass various dimensions of dismissal law covering 117 countries for 1970–2013: https://www.repository.cam.ac.uk/handle/1810/256566.